]>
Commit | Line | Data |
---|---|---|
b518d3e6 | 1 | /* |
fe314195 | 2 | * Copyright(c) 2016 Intel Corporation. |
b518d3e6 DD |
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 | |
11a10d4b VSD |
58 | static void rvt_rc_timeout(unsigned long arg); |
59 | ||
60 | /* | |
61 | * Convert the AETH RNR timeout code into the number of microseconds. | |
62 | */ | |
63 | static const u32 ib_rvt_rnr_table[32] = { | |
64 | 655360, /* 00: 655.36 */ | |
65 | 10, /* 01: .01 */ | |
66 | 20, /* 02 .02 */ | |
67 | 30, /* 03: .03 */ | |
68 | 40, /* 04: .04 */ | |
69 | 60, /* 05: .06 */ | |
70 | 80, /* 06: .08 */ | |
71 | 120, /* 07: .12 */ | |
72 | 160, /* 08: .16 */ | |
73 | 240, /* 09: .24 */ | |
74 | 320, /* 0A: .32 */ | |
75 | 480, /* 0B: .48 */ | |
76 | 640, /* 0C: .64 */ | |
77 | 960, /* 0D: .96 */ | |
78 | 1280, /* 0E: 1.28 */ | |
79 | 1920, /* 0F: 1.92 */ | |
80 | 2560, /* 10: 2.56 */ | |
81 | 3840, /* 11: 3.84 */ | |
82 | 5120, /* 12: 5.12 */ | |
83 | 7680, /* 13: 7.68 */ | |
84 | 10240, /* 14: 10.24 */ | |
85 | 15360, /* 15: 15.36 */ | |
86 | 20480, /* 16: 20.48 */ | |
87 | 30720, /* 17: 30.72 */ | |
88 | 40960, /* 18: 40.96 */ | |
89 | 61440, /* 19: 61.44 */ | |
90 | 81920, /* 1A: 81.92 */ | |
91 | 122880, /* 1B: 122.88 */ | |
92 | 163840, /* 1C: 163.84 */ | |
93 | 245760, /* 1D: 245.76 */ | |
94 | 327680, /* 1E: 327.68 */ | |
95 | 491520 /* 1F: 491.52 */ | |
96 | }; | |
97 | ||
bfbac097 DD |
98 | /* |
99 | * Note that it is OK to post send work requests in the SQE and ERR | |
100 | * states; rvt_do_send() will process them and generate error | |
101 | * completions as per IB 1.2 C10-96. | |
102 | */ | |
103 | const int ib_rvt_state_ops[IB_QPS_ERR + 1] = { | |
104 | [IB_QPS_RESET] = 0, | |
105 | [IB_QPS_INIT] = RVT_POST_RECV_OK, | |
106 | [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK, | |
107 | [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | | |
108 | RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK | | |
109 | RVT_PROCESS_NEXT_SEND_OK, | |
110 | [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | | |
111 | RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK, | |
112 | [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | | |
113 | RVT_POST_SEND_OK | RVT_FLUSH_SEND, | |
114 | [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV | | |
115 | RVT_POST_SEND_OK | RVT_FLUSH_SEND, | |
116 | }; | |
117 | EXPORT_SYMBOL(ib_rvt_state_ops); | |
118 | ||
f2dc9cdc MM |
119 | /* |
120 | * Translate ib_wr_opcode into ib_wc_opcode. | |
121 | */ | |
122 | const enum ib_wc_opcode ib_rvt_wc_opcode[] = { | |
123 | [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE, | |
124 | [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE, | |
125 | [IB_WR_SEND] = IB_WC_SEND, | |
126 | [IB_WR_SEND_WITH_IMM] = IB_WC_SEND, | |
127 | [IB_WR_RDMA_READ] = IB_WC_RDMA_READ, | |
128 | [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP, | |
129 | [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD, | |
130 | [IB_WR_SEND_WITH_INV] = IB_WC_SEND, | |
131 | [IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV, | |
132 | [IB_WR_REG_MR] = IB_WC_REG_MR | |
133 | }; | |
134 | EXPORT_SYMBOL(ib_rvt_wc_opcode); | |
135 | ||
d2b8d4da MM |
136 | static void get_map_page(struct rvt_qpn_table *qpt, |
137 | struct rvt_qpn_map *map, | |
138 | gfp_t gfp) | |
0acb0cc7 | 139 | { |
d2b8d4da | 140 | unsigned long page = get_zeroed_page(gfp); |
0acb0cc7 DD |
141 | |
142 | /* | |
143 | * Free the page if someone raced with us installing it. | |
144 | */ | |
145 | ||
146 | spin_lock(&qpt->lock); | |
147 | if (map->page) | |
148 | free_page(page); | |
149 | else | |
150 | map->page = (void *)page; | |
151 | spin_unlock(&qpt->lock); | |
152 | } | |
153 | ||
154 | /** | |
155 | * init_qpn_table - initialize the QP number table for a device | |
156 | * @qpt: the QPN table | |
157 | */ | |
158 | static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt) | |
159 | { | |
160 | u32 offset, i; | |
161 | struct rvt_qpn_map *map; | |
162 | int ret = 0; | |
163 | ||
fef2efd6 | 164 | if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start)) |
0acb0cc7 DD |
165 | return -EINVAL; |
166 | ||
167 | spin_lock_init(&qpt->lock); | |
168 | ||
169 | qpt->last = rdi->dparms.qpn_start; | |
170 | qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift; | |
171 | ||
172 | /* | |
173 | * Drivers may want some QPs beyond what we need for verbs let them use | |
174 | * our qpn table. No need for two. Lets go ahead and mark the bitmaps | |
175 | * for those. The reserved range must be *after* the range which verbs | |
176 | * will pick from. | |
177 | */ | |
178 | ||
179 | /* Figure out number of bit maps needed before reserved range */ | |
180 | qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE; | |
181 | ||
182 | /* This should always be zero */ | |
183 | offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK; | |
184 | ||
185 | /* Starting with the first reserved bit map */ | |
186 | map = &qpt->map[qpt->nmaps]; | |
187 | ||
188 | rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n", | |
189 | rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end); | |
fef2efd6 | 190 | for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) { |
0acb0cc7 | 191 | if (!map->page) { |
d2b8d4da | 192 | get_map_page(qpt, map, GFP_KERNEL); |
0acb0cc7 DD |
193 | if (!map->page) { |
194 | ret = -ENOMEM; | |
195 | break; | |
196 | } | |
197 | } | |
198 | set_bit(offset, map->page); | |
199 | offset++; | |
200 | if (offset == RVT_BITS_PER_PAGE) { | |
201 | /* next page */ | |
202 | qpt->nmaps++; | |
203 | map++; | |
204 | offset = 0; | |
205 | } | |
206 | } | |
207 | return ret; | |
208 | } | |
209 | ||
210 | /** | |
211 | * free_qpn_table - free the QP number table for a device | |
212 | * @qpt: the QPN table | |
213 | */ | |
214 | static void free_qpn_table(struct rvt_qpn_table *qpt) | |
215 | { | |
216 | int i; | |
217 | ||
218 | for (i = 0; i < ARRAY_SIZE(qpt->map); i++) | |
219 | free_page((unsigned long)qpt->map[i].page); | |
220 | } | |
221 | ||
90793f71 DD |
222 | /** |
223 | * rvt_driver_qp_init - Init driver qp resources | |
224 | * @rdi: rvt dev strucutre | |
225 | * | |
226 | * Return: 0 on success | |
227 | */ | |
0acb0cc7 DD |
228 | int rvt_driver_qp_init(struct rvt_dev_info *rdi) |
229 | { | |
230 | int i; | |
231 | int ret = -ENOMEM; | |
232 | ||
0acb0cc7 DD |
233 | if (!rdi->dparms.qp_table_size) |
234 | return -EINVAL; | |
235 | ||
236 | /* | |
237 | * If driver is not doing any QP allocation then make sure it is | |
238 | * providing the necessary QP functions. | |
239 | */ | |
515667f8 DD |
240 | if (!rdi->driver_f.free_all_qps || |
241 | !rdi->driver_f.qp_priv_alloc || | |
242 | !rdi->driver_f.qp_priv_free || | |
11a10d4b VSD |
243 | !rdi->driver_f.notify_qp_reset || |
244 | !rdi->driver_f.notify_restart_rc) | |
0acb0cc7 DD |
245 | return -EINVAL; |
246 | ||
247 | /* allocate parent object */ | |
d1b697b6 MH |
248 | rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL, |
249 | rdi->dparms.node); | |
0acb0cc7 DD |
250 | if (!rdi->qp_dev) |
251 | return -ENOMEM; | |
252 | ||
253 | /* allocate hash table */ | |
254 | rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size; | |
255 | rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size); | |
256 | rdi->qp_dev->qp_table = | |
d1b697b6 MH |
257 | kmalloc_node(rdi->qp_dev->qp_table_size * |
258 | sizeof(*rdi->qp_dev->qp_table), | |
259 | GFP_KERNEL, rdi->dparms.node); | |
0acb0cc7 DD |
260 | if (!rdi->qp_dev->qp_table) |
261 | goto no_qp_table; | |
262 | ||
263 | for (i = 0; i < rdi->qp_dev->qp_table_size; i++) | |
264 | RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL); | |
265 | ||
266 | spin_lock_init(&rdi->qp_dev->qpt_lock); | |
267 | ||
268 | /* initialize qpn map */ | |
269 | if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table)) | |
270 | goto fail_table; | |
271 | ||
515667f8 DD |
272 | spin_lock_init(&rdi->n_qps_lock); |
273 | ||
274 | return 0; | |
0acb0cc7 DD |
275 | |
276 | fail_table: | |
277 | kfree(rdi->qp_dev->qp_table); | |
278 | free_qpn_table(&rdi->qp_dev->qpn_table); | |
279 | ||
280 | no_qp_table: | |
281 | kfree(rdi->qp_dev); | |
282 | ||
283 | return ret; | |
284 | } | |
285 | ||
286 | /** | |
287 | * free_all_qps - check for QPs still in use | |
288 | * @qpt: the QP table to empty | |
289 | * | |
290 | * There should not be any QPs still in use. | |
291 | * Free memory for table. | |
292 | */ | |
515667f8 | 293 | static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi) |
0acb0cc7 DD |
294 | { |
295 | unsigned long flags; | |
296 | struct rvt_qp *qp; | |
297 | unsigned n, qp_inuse = 0; | |
298 | spinlock_t *ql; /* work around too long line below */ | |
299 | ||
515667f8 DD |
300 | if (rdi->driver_f.free_all_qps) |
301 | qp_inuse = rdi->driver_f.free_all_qps(rdi); | |
0acb0cc7 | 302 | |
4e74080b DD |
303 | qp_inuse += rvt_mcast_tree_empty(rdi); |
304 | ||
0acb0cc7 | 305 | if (!rdi->qp_dev) |
515667f8 | 306 | return qp_inuse; |
0acb0cc7 DD |
307 | |
308 | ql = &rdi->qp_dev->qpt_lock; | |
515667f8 | 309 | spin_lock_irqsave(ql, flags); |
0acb0cc7 DD |
310 | for (n = 0; n < rdi->qp_dev->qp_table_size; n++) { |
311 | qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n], | |
312 | lockdep_is_held(ql)); | |
313 | RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL); | |
515667f8 DD |
314 | |
315 | for (; qp; qp = rcu_dereference_protected(qp->next, | |
316 | lockdep_is_held(ql))) | |
0acb0cc7 | 317 | qp_inuse++; |
0acb0cc7 DD |
318 | } |
319 | spin_unlock_irqrestore(ql, flags); | |
320 | synchronize_rcu(); | |
321 | return qp_inuse; | |
322 | } | |
323 | ||
90793f71 DD |
324 | /** |
325 | * rvt_qp_exit - clean up qps on device exit | |
326 | * @rdi: rvt dev structure | |
327 | * | |
328 | * Check for qp leaks and free resources. | |
329 | */ | |
0acb0cc7 DD |
330 | void rvt_qp_exit(struct rvt_dev_info *rdi) |
331 | { | |
515667f8 | 332 | u32 qps_inuse = rvt_free_all_qps(rdi); |
0acb0cc7 | 333 | |
0acb0cc7 DD |
334 | if (qps_inuse) |
335 | rvt_pr_err(rdi, "QP memory leak! %u still in use\n", | |
336 | qps_inuse); | |
337 | if (!rdi->qp_dev) | |
338 | return; | |
339 | ||
340 | kfree(rdi->qp_dev->qp_table); | |
341 | free_qpn_table(&rdi->qp_dev->qpn_table); | |
342 | kfree(rdi->qp_dev); | |
343 | } | |
344 | ||
515667f8 DD |
345 | static inline unsigned mk_qpn(struct rvt_qpn_table *qpt, |
346 | struct rvt_qpn_map *map, unsigned off) | |
347 | { | |
348 | return (map - qpt->map) * RVT_BITS_PER_PAGE + off; | |
349 | } | |
350 | ||
f1badc71 DD |
351 | /** |
352 | * alloc_qpn - Allocate the next available qpn or zero/one for QP type | |
353 | * IB_QPT_SMI/IB_QPT_GSI | |
354 | *@rdi: rvt device info structure | |
355 | *@qpt: queue pair number table pointer | |
356 | *@port_num: IB port number, 1 based, comes from core | |
357 | * | |
358 | * Return: The queue pair number | |
515667f8 DD |
359 | */ |
360 | static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, | |
f1badc71 | 361 | enum ib_qp_type type, u8 port_num, gfp_t gfp) |
515667f8 DD |
362 | { |
363 | u32 i, offset, max_scan, qpn; | |
364 | struct rvt_qpn_map *map; | |
365 | u32 ret; | |
366 | ||
367 | if (rdi->driver_f.alloc_qpn) | |
b7b3cf44 | 368 | return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num, gfp); |
515667f8 DD |
369 | |
370 | if (type == IB_QPT_SMI || type == IB_QPT_GSI) { | |
371 | unsigned n; | |
372 | ||
373 | ret = type == IB_QPT_GSI; | |
f1badc71 | 374 | n = 1 << (ret + 2 * (port_num - 1)); |
515667f8 DD |
375 | spin_lock(&qpt->lock); |
376 | if (qpt->flags & n) | |
377 | ret = -EINVAL; | |
378 | else | |
379 | qpt->flags |= n; | |
380 | spin_unlock(&qpt->lock); | |
381 | goto bail; | |
382 | } | |
383 | ||
384 | qpn = qpt->last + qpt->incr; | |
385 | if (qpn >= RVT_QPN_MAX) | |
386 | qpn = qpt->incr | ((qpt->last & 1) ^ 1); | |
387 | /* offset carries bit 0 */ | |
388 | offset = qpn & RVT_BITS_PER_PAGE_MASK; | |
389 | map = &qpt->map[qpn / RVT_BITS_PER_PAGE]; | |
390 | max_scan = qpt->nmaps - !offset; | |
391 | for (i = 0;;) { | |
392 | if (unlikely(!map->page)) { | |
d2b8d4da | 393 | get_map_page(qpt, map, gfp); |
515667f8 DD |
394 | if (unlikely(!map->page)) |
395 | break; | |
396 | } | |
397 | do { | |
398 | if (!test_and_set_bit(offset, map->page)) { | |
399 | qpt->last = qpn; | |
400 | ret = qpn; | |
401 | goto bail; | |
402 | } | |
403 | offset += qpt->incr; | |
404 | /* | |
405 | * This qpn might be bogus if offset >= BITS_PER_PAGE. | |
406 | * That is OK. It gets re-assigned below | |
407 | */ | |
408 | qpn = mk_qpn(qpt, map, offset); | |
409 | } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX); | |
410 | /* | |
411 | * In order to keep the number of pages allocated to a | |
412 | * minimum, we scan the all existing pages before increasing | |
413 | * the size of the bitmap table. | |
414 | */ | |
415 | if (++i > max_scan) { | |
416 | if (qpt->nmaps == RVT_QPNMAP_ENTRIES) | |
417 | break; | |
418 | map = &qpt->map[qpt->nmaps++]; | |
419 | /* start at incr with current bit 0 */ | |
420 | offset = qpt->incr | (offset & 1); | |
421 | } else if (map < &qpt->map[qpt->nmaps]) { | |
422 | ++map; | |
423 | /* start at incr with current bit 0 */ | |
424 | offset = qpt->incr | (offset & 1); | |
425 | } else { | |
426 | map = &qpt->map[0]; | |
427 | /* wrap to first map page, invert bit 0 */ | |
428 | offset = qpt->incr | ((offset & 1) ^ 1); | |
429 | } | |
501edc42 BW |
430 | /* there can be no set bits in low-order QoS bits */ |
431 | WARN_ON(offset & (BIT(rdi->dparms.qos_shift) - 1)); | |
515667f8 DD |
432 | qpn = mk_qpn(qpt, map, offset); |
433 | } | |
434 | ||
435 | ret = -ENOMEM; | |
436 | ||
437 | bail: | |
438 | return ret; | |
439 | } | |
440 | ||
441 | static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn) | |
442 | { | |
443 | struct rvt_qpn_map *map; | |
444 | ||
445 | map = qpt->map + qpn / RVT_BITS_PER_PAGE; | |
446 | if (map->page) | |
447 | clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); | |
448 | } | |
449 | ||
79a225be DD |
450 | /** |
451 | * rvt_clear_mr_refs - Drop help mr refs | |
452 | * @qp: rvt qp data structure | |
453 | * @clr_sends: If shoudl clear send side or not | |
454 | */ | |
455 | static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) | |
456 | { | |
457 | unsigned n; | |
8b103e9c | 458 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); |
79a225be DD |
459 | |
460 | if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) | |
461 | rvt_put_ss(&qp->s_rdma_read_sge); | |
462 | ||
463 | rvt_put_ss(&qp->r_sge); | |
464 | ||
465 | if (clr_sends) { | |
466 | while (qp->s_last != qp->s_head) { | |
467 | struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last); | |
468 | unsigned i; | |
469 | ||
470 | for (i = 0; i < wqe->wr.num_sge; i++) { | |
471 | struct rvt_sge *sge = &wqe->sg_list[i]; | |
472 | ||
473 | rvt_put_mr(sge->mr); | |
474 | } | |
475 | if (qp->ibqp.qp_type == IB_QPT_UD || | |
476 | qp->ibqp.qp_type == IB_QPT_SMI || | |
477 | qp->ibqp.qp_type == IB_QPT_GSI) | |
478 | atomic_dec(&ibah_to_rvtah( | |
479 | wqe->ud_wr.ah)->refcount); | |
480 | if (++qp->s_last >= qp->s_size) | |
481 | qp->s_last = 0; | |
482 | smp_wmb(); /* see qp_set_savail */ | |
483 | } | |
484 | if (qp->s_rdma_mr) { | |
485 | rvt_put_mr(qp->s_rdma_mr); | |
486 | qp->s_rdma_mr = NULL; | |
487 | } | |
488 | } | |
489 | ||
490 | if (qp->ibqp.qp_type != IB_QPT_RC) | |
491 | return; | |
492 | ||
8b103e9c | 493 | for (n = 0; n < rvt_max_atomic(rdi); n++) { |
79a225be DD |
494 | struct rvt_ack_entry *e = &qp->s_ack_queue[n]; |
495 | ||
fe508272 | 496 | if (e->rdma_sge.mr) { |
79a225be DD |
497 | rvt_put_mr(e->rdma_sge.mr); |
498 | e->rdma_sge.mr = NULL; | |
499 | } | |
500 | } | |
501 | } | |
502 | ||
503 | /** | |
504 | * rvt_remove_qp - remove qp form table | |
505 | * @rdi: rvt dev struct | |
506 | * @qp: qp to remove | |
507 | * | |
508 | * Remove the QP from the table so it can't be found asynchronously by | |
509 | * the receive routine. | |
510 | */ | |
511 | static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) | |
512 | { | |
513 | struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; | |
514 | u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); | |
515 | unsigned long flags; | |
516 | int removed = 1; | |
517 | ||
518 | spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); | |
519 | ||
520 | if (rcu_dereference_protected(rvp->qp[0], | |
521 | lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { | |
522 | RCU_INIT_POINTER(rvp->qp[0], NULL); | |
523 | } else if (rcu_dereference_protected(rvp->qp[1], | |
524 | lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { | |
525 | RCU_INIT_POINTER(rvp->qp[1], NULL); | |
526 | } else { | |
527 | struct rvt_qp *q; | |
528 | struct rvt_qp __rcu **qpp; | |
529 | ||
530 | removed = 0; | |
531 | qpp = &rdi->qp_dev->qp_table[n]; | |
532 | for (; (q = rcu_dereference_protected(*qpp, | |
533 | lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL; | |
534 | qpp = &q->next) { | |
535 | if (q == qp) { | |
536 | RCU_INIT_POINTER(*qpp, | |
537 | rcu_dereference_protected(qp->next, | |
538 | lockdep_is_held(&rdi->qp_dev->qpt_lock))); | |
539 | removed = 1; | |
540 | trace_rvt_qpremove(qp, n); | |
541 | break; | |
542 | } | |
543 | } | |
544 | } | |
545 | ||
546 | spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); | |
547 | if (removed) { | |
548 | synchronize_rcu(); | |
4d6f85c3 | 549 | rvt_put_qp(qp); |
79a225be DD |
550 | } |
551 | } | |
552 | ||
515667f8 | 553 | /** |
222f7a9a MM |
554 | * rvt_init_qp - initialize the QP state to the reset state |
555 | * @qp: the QP to init or reinit | |
515667f8 | 556 | * @type: the QP type |
222f7a9a MM |
557 | * |
558 | * This function is called from both rvt_create_qp() and | |
559 | * rvt_reset_qp(). The difference is that the reset | |
560 | * patch the necessary locks to protect against concurent | |
561 | * access. | |
515667f8 | 562 | */ |
222f7a9a MM |
563 | static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, |
564 | enum ib_qp_type type) | |
515667f8 | 565 | { |
3b0b3fb3 DD |
566 | qp->remote_qpn = 0; |
567 | qp->qkey = 0; | |
568 | qp->qp_access_flags = 0; | |
515667f8 DD |
569 | qp->s_flags &= RVT_S_SIGNAL_REQ_WR; |
570 | qp->s_hdrwords = 0; | |
571 | qp->s_wqe = NULL; | |
572 | qp->s_draining = 0; | |
573 | qp->s_next_psn = 0; | |
574 | qp->s_last_psn = 0; | |
575 | qp->s_sending_psn = 0; | |
576 | qp->s_sending_hpsn = 0; | |
577 | qp->s_psn = 0; | |
578 | qp->r_psn = 0; | |
579 | qp->r_msn = 0; | |
580 | if (type == IB_QPT_RC) { | |
581 | qp->s_state = IB_OPCODE_RC_SEND_LAST; | |
582 | qp->r_state = IB_OPCODE_RC_SEND_LAST; | |
583 | } else { | |
584 | qp->s_state = IB_OPCODE_UC_SEND_LAST; | |
585 | qp->r_state = IB_OPCODE_UC_SEND_LAST; | |
586 | } | |
587 | qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; | |
588 | qp->r_nak_state = 0; | |
589 | qp->r_aflags = 0; | |
590 | qp->r_flags = 0; | |
591 | qp->s_head = 0; | |
592 | qp->s_tail = 0; | |
593 | qp->s_cur = 0; | |
594 | qp->s_acked = 0; | |
595 | qp->s_last = 0; | |
596 | qp->s_ssn = 1; | |
597 | qp->s_lsn = 0; | |
598 | qp->s_mig_state = IB_MIG_MIGRATED; | |
515667f8 DD |
599 | qp->r_head_ack_queue = 0; |
600 | qp->s_tail_ack_queue = 0; | |
601 | qp->s_num_rd_atomic = 0; | |
602 | if (qp->r_rq.wq) { | |
603 | qp->r_rq.wq->head = 0; | |
604 | qp->r_rq.wq->tail = 0; | |
605 | } | |
606 | qp->r_sge.num_sge = 0; | |
856cc4c2 | 607 | atomic_set(&qp->s_reserved_used, 0); |
515667f8 DD |
608 | } |
609 | ||
222f7a9a MM |
610 | /** |
611 | * rvt_reset_qp - initialize the QP state to the reset state | |
612 | * @qp: the QP to reset | |
613 | * @type: the QP type | |
614 | * | |
615 | * r_lock, s_hlock, and s_lock are required to be held by the caller | |
616 | */ | |
617 | static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, | |
618 | enum ib_qp_type type) | |
619 | __must_hold(&qp->s_lock) | |
620 | __must_hold(&qp->s_hlock) | |
621 | __must_hold(&qp->r_lock) | |
622 | { | |
68e78b3d MM |
623 | lockdep_assert_held(&qp->r_lock); |
624 | lockdep_assert_held(&qp->s_hlock); | |
625 | lockdep_assert_held(&qp->s_lock); | |
222f7a9a MM |
626 | if (qp->state != IB_QPS_RESET) { |
627 | qp->state = IB_QPS_RESET; | |
628 | ||
629 | /* Let drivers flush their waitlist */ | |
630 | rdi->driver_f.flush_qp_waiters(qp); | |
11a10d4b | 631 | rvt_stop_rc_timers(qp); |
222f7a9a MM |
632 | qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT); |
633 | spin_unlock(&qp->s_lock); | |
634 | spin_unlock(&qp->s_hlock); | |
635 | spin_unlock_irq(&qp->r_lock); | |
636 | ||
637 | /* Stop the send queue and the retry timer */ | |
638 | rdi->driver_f.stop_send_queue(qp); | |
11a10d4b | 639 | rvt_del_timers_sync(qp); |
222f7a9a MM |
640 | /* Wait for things to stop */ |
641 | rdi->driver_f.quiesce_qp(qp); | |
642 | ||
643 | /* take qp out the hash and wait for it to be unused */ | |
644 | rvt_remove_qp(rdi, qp); | |
645 | wait_event(qp->wait, !atomic_read(&qp->refcount)); | |
646 | ||
647 | /* grab the lock b/c it was locked at call time */ | |
648 | spin_lock_irq(&qp->r_lock); | |
649 | spin_lock(&qp->s_hlock); | |
650 | spin_lock(&qp->s_lock); | |
651 | ||
652 | rvt_clear_mr_refs(qp, 1); | |
653 | /* | |
654 | * Let the driver do any tear down or re-init it needs to for | |
655 | * a qp that has been reset | |
656 | */ | |
657 | rdi->driver_f.notify_qp_reset(qp); | |
658 | } | |
659 | rvt_init_qp(rdi, qp, type); | |
68e78b3d MM |
660 | lockdep_assert_held(&qp->r_lock); |
661 | lockdep_assert_held(&qp->s_hlock); | |
662 | lockdep_assert_held(&qp->s_lock); | |
222f7a9a MM |
663 | } |
664 | ||
b518d3e6 DD |
665 | /** |
666 | * rvt_create_qp - create a queue pair for a device | |
667 | * @ibpd: the protection domain who's device we create the queue pair for | |
668 | * @init_attr: the attributes of the queue pair | |
669 | * @udata: user data for libibverbs.so | |
670 | * | |
515667f8 DD |
671 | * Queue pair creation is mostly an rvt issue. However, drivers have their own |
672 | * unique idea of what queue pair numbers mean. For instance there is a reserved | |
673 | * range for PSM. | |
674 | * | |
90793f71 | 675 | * Return: the queue pair on success, otherwise returns an errno. |
b518d3e6 DD |
676 | * |
677 | * Called by the ib_create_qp() core verbs function. | |
678 | */ | |
679 | struct ib_qp *rvt_create_qp(struct ib_pd *ibpd, | |
680 | struct ib_qp_init_attr *init_attr, | |
681 | struct ib_udata *udata) | |
682 | { | |
515667f8 DD |
683 | struct rvt_qp *qp; |
684 | int err; | |
685 | struct rvt_swqe *swq = NULL; | |
686 | size_t sz; | |
687 | size_t sg_list_sz; | |
688 | struct ib_qp *ret = ERR_PTR(-ENOMEM); | |
689 | struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device); | |
690 | void *priv = NULL; | |
d2b8d4da | 691 | gfp_t gfp; |
afcf8f76 | 692 | size_t sqsize; |
515667f8 DD |
693 | |
694 | if (!rdi) | |
695 | return ERR_PTR(-EINVAL); | |
696 | ||
697 | if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge || | |
698 | init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr || | |
d2b8d4da | 699 | init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO)) |
515667f8 DD |
700 | return ERR_PTR(-EINVAL); |
701 | ||
d2b8d4da MM |
702 | /* GFP_NOIO is applicable to RC QP's only */ |
703 | ||
704 | if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO && | |
705 | init_attr->qp_type != IB_QPT_RC) | |
706 | return ERR_PTR(-EINVAL); | |
707 | ||
708 | gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ? | |
709 | GFP_NOIO : GFP_KERNEL; | |
710 | ||
515667f8 DD |
711 | /* Check receive queue parameters if no SRQ is specified. */ |
712 | if (!init_attr->srq) { | |
713 | if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge || | |
714 | init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr) | |
715 | return ERR_PTR(-EINVAL); | |
716 | ||
717 | if (init_attr->cap.max_send_sge + | |
718 | init_attr->cap.max_send_wr + | |
719 | init_attr->cap.max_recv_sge + | |
720 | init_attr->cap.max_recv_wr == 0) | |
721 | return ERR_PTR(-EINVAL); | |
722 | } | |
afcf8f76 | 723 | sqsize = |
856cc4c2 MM |
724 | init_attr->cap.max_send_wr + 1 + |
725 | rdi->dparms.reserved_operations; | |
515667f8 DD |
726 | switch (init_attr->qp_type) { |
727 | case IB_QPT_SMI: | |
728 | case IB_QPT_GSI: | |
729 | if (init_attr->port_num == 0 || | |
730 | init_attr->port_num > ibpd->device->phys_port_cnt) | |
731 | return ERR_PTR(-EINVAL); | |
732 | case IB_QPT_UC: | |
733 | case IB_QPT_RC: | |
734 | case IB_QPT_UD: | |
735 | sz = sizeof(struct rvt_sge) * | |
736 | init_attr->cap.max_send_sge + | |
737 | sizeof(struct rvt_swqe); | |
d2b8d4da MM |
738 | if (gfp == GFP_NOIO) |
739 | swq = __vmalloc( | |
afcf8f76 | 740 | sqsize * sz, |
654b6436 | 741 | gfp | __GFP_ZERO, PAGE_KERNEL); |
d2b8d4da | 742 | else |
654b6436 | 743 | swq = vzalloc_node( |
afcf8f76 | 744 | sqsize * sz, |
d1b697b6 | 745 | rdi->dparms.node); |
515667f8 DD |
746 | if (!swq) |
747 | return ERR_PTR(-ENOMEM); | |
748 | ||
749 | sz = sizeof(*qp); | |
750 | sg_list_sz = 0; | |
751 | if (init_attr->srq) { | |
752 | struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); | |
753 | ||
754 | if (srq->rq.max_sge > 1) | |
755 | sg_list_sz = sizeof(*qp->r_sg_list) * | |
756 | (srq->rq.max_sge - 1); | |
757 | } else if (init_attr->cap.max_recv_sge > 1) | |
758 | sg_list_sz = sizeof(*qp->r_sg_list) * | |
759 | (init_attr->cap.max_recv_sge - 1); | |
d1b697b6 | 760 | qp = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node); |
515667f8 DD |
761 | if (!qp) |
762 | goto bail_swq; | |
763 | ||
764 | RCU_INIT_POINTER(qp->next, NULL); | |
8b103e9c MM |
765 | if (init_attr->qp_type == IB_QPT_RC) { |
766 | qp->s_ack_queue = | |
767 | kzalloc_node( | |
768 | sizeof(*qp->s_ack_queue) * | |
769 | rvt_max_atomic(rdi), | |
770 | gfp, | |
771 | rdi->dparms.node); | |
772 | if (!qp->s_ack_queue) | |
773 | goto bail_qp; | |
774 | } | |
11a10d4b VSD |
775 | /* initialize timers needed for rc qp */ |
776 | setup_timer(&qp->s_timer, rvt_rc_timeout, (unsigned long)qp); | |
777 | hrtimer_init(&qp->s_rnr_timer, CLOCK_MONOTONIC, | |
778 | HRTIMER_MODE_REL); | |
779 | qp->s_rnr_timer.function = rvt_rc_rnr_retry; | |
515667f8 DD |
780 | |
781 | /* | |
782 | * Driver needs to set up it's private QP structure and do any | |
783 | * initialization that is needed. | |
784 | */ | |
d2b8d4da | 785 | priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp); |
c755f4af MM |
786 | if (IS_ERR(priv)) { |
787 | ret = priv; | |
515667f8 | 788 | goto bail_qp; |
c755f4af | 789 | } |
515667f8 DD |
790 | qp->priv = priv; |
791 | qp->timeout_jiffies = | |
792 | usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / | |
793 | 1000UL); | |
794 | if (init_attr->srq) { | |
795 | sz = 0; | |
796 | } else { | |
797 | qp->r_rq.size = init_attr->cap.max_recv_wr + 1; | |
798 | qp->r_rq.max_sge = init_attr->cap.max_recv_sge; | |
799 | sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + | |
800 | sizeof(struct rvt_rwqe); | |
d2b8d4da MM |
801 | if (udata) |
802 | qp->r_rq.wq = vmalloc_user( | |
803 | sizeof(struct rvt_rwq) + | |
804 | qp->r_rq.size * sz); | |
805 | else if (gfp == GFP_NOIO) | |
806 | qp->r_rq.wq = __vmalloc( | |
807 | sizeof(struct rvt_rwq) + | |
808 | qp->r_rq.size * sz, | |
654b6436 | 809 | gfp | __GFP_ZERO, PAGE_KERNEL); |
d2b8d4da | 810 | else |
654b6436 | 811 | qp->r_rq.wq = vzalloc_node( |
d2b8d4da | 812 | sizeof(struct rvt_rwq) + |
d1b697b6 MH |
813 | qp->r_rq.size * sz, |
814 | rdi->dparms.node); | |
515667f8 DD |
815 | if (!qp->r_rq.wq) |
816 | goto bail_driver_priv; | |
817 | } | |
818 | ||
819 | /* | |
820 | * ib_create_qp() will initialize qp->ibqp | |
821 | * except for qp->ibqp.qp_num. | |
822 | */ | |
823 | spin_lock_init(&qp->r_lock); | |
46a80d62 | 824 | spin_lock_init(&qp->s_hlock); |
515667f8 DD |
825 | spin_lock_init(&qp->s_lock); |
826 | spin_lock_init(&qp->r_rq.lock); | |
827 | atomic_set(&qp->refcount, 0); | |
d9f87239 | 828 | atomic_set(&qp->local_ops_pending, 0); |
515667f8 DD |
829 | init_waitqueue_head(&qp->wait); |
830 | init_timer(&qp->s_timer); | |
831 | qp->s_timer.data = (unsigned long)qp; | |
832 | INIT_LIST_HEAD(&qp->rspwait); | |
833 | qp->state = IB_QPS_RESET; | |
834 | qp->s_wq = swq; | |
afcf8f76 | 835 | qp->s_size = sqsize; |
46a80d62 | 836 | qp->s_avail = init_attr->cap.max_send_wr; |
515667f8 DD |
837 | qp->s_max_sge = init_attr->cap.max_send_sge; |
838 | if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) | |
839 | qp->s_flags = RVT_S_SIGNAL_REQ_WR; | |
840 | ||
841 | err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table, | |
842 | init_attr->qp_type, | |
d2b8d4da | 843 | init_attr->port_num, gfp); |
515667f8 DD |
844 | if (err < 0) { |
845 | ret = ERR_PTR(err); | |
846 | goto bail_rq_wq; | |
847 | } | |
848 | qp->ibqp.qp_num = err; | |
849 | qp->port_num = init_attr->port_num; | |
222f7a9a | 850 | rvt_init_qp(rdi, qp, init_attr->qp_type); |
515667f8 DD |
851 | break; |
852 | ||
853 | default: | |
854 | /* Don't support raw QPs */ | |
855 | return ERR_PTR(-EINVAL); | |
856 | } | |
857 | ||
858 | init_attr->cap.max_inline_data = 0; | |
859 | ||
860 | /* | |
861 | * Return the address of the RWQ as the offset to mmap. | |
bfbac097 | 862 | * See rvt_mmap() for details. |
515667f8 DD |
863 | */ |
864 | if (udata && udata->outlen >= sizeof(__u64)) { | |
865 | if (!qp->r_rq.wq) { | |
866 | __u64 offset = 0; | |
867 | ||
868 | err = ib_copy_to_udata(udata, &offset, | |
869 | sizeof(offset)); | |
870 | if (err) { | |
871 | ret = ERR_PTR(err); | |
872 | goto bail_qpn; | |
873 | } | |
874 | } else { | |
875 | u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; | |
876 | ||
877 | qp->ip = rvt_create_mmap_info(rdi, s, | |
878 | ibpd->uobject->context, | |
879 | qp->r_rq.wq); | |
880 | if (!qp->ip) { | |
881 | ret = ERR_PTR(-ENOMEM); | |
882 | goto bail_qpn; | |
883 | } | |
884 | ||
885 | err = ib_copy_to_udata(udata, &qp->ip->offset, | |
886 | sizeof(qp->ip->offset)); | |
887 | if (err) { | |
888 | ret = ERR_PTR(err); | |
889 | goto bail_ip; | |
890 | } | |
891 | } | |
ef086c0d | 892 | qp->pid = current->pid; |
515667f8 DD |
893 | } |
894 | ||
895 | spin_lock(&rdi->n_qps_lock); | |
896 | if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) { | |
897 | spin_unlock(&rdi->n_qps_lock); | |
898 | ret = ERR_PTR(-ENOMEM); | |
899 | goto bail_ip; | |
900 | } | |
901 | ||
902 | rdi->n_qps_allocated++; | |
bfee5e32 VM |
903 | /* |
904 | * Maintain a busy_jiffies variable that will be added to the timeout | |
905 | * period in mod_retry_timer and add_retry_timer. This busy jiffies | |
906 | * is scaled by the number of rc qps created for the device to reduce | |
907 | * the number of timeouts occurring when there is a large number of | |
908 | * qps. busy_jiffies is incremented every rc qp scaling interval. | |
909 | * The scaling interval is selected based on extensive performance | |
910 | * evaluation of targeted workloads. | |
911 | */ | |
912 | if (init_attr->qp_type == IB_QPT_RC) { | |
913 | rdi->n_rc_qps++; | |
914 | rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL; | |
915 | } | |
515667f8 DD |
916 | spin_unlock(&rdi->n_qps_lock); |
917 | ||
918 | if (qp->ip) { | |
919 | spin_lock_irq(&rdi->pending_lock); | |
920 | list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps); | |
921 | spin_unlock_irq(&rdi->pending_lock); | |
922 | } | |
923 | ||
924 | ret = &qp->ibqp; | |
925 | ||
b518d3e6 | 926 | /* |
515667f8 DD |
927 | * We have our QP and its good, now keep track of what types of opcodes |
928 | * can be processed on this QP. We do this by keeping track of what the | |
929 | * 3 high order bits of the opcode are. | |
b518d3e6 | 930 | */ |
515667f8 DD |
931 | switch (init_attr->qp_type) { |
932 | case IB_QPT_SMI: | |
933 | case IB_QPT_GSI: | |
934 | case IB_QPT_UD: | |
b218f786 | 935 | qp->allowed_ops = IB_OPCODE_UD; |
515667f8 DD |
936 | break; |
937 | case IB_QPT_RC: | |
b218f786 | 938 | qp->allowed_ops = IB_OPCODE_RC; |
515667f8 DD |
939 | break; |
940 | case IB_QPT_UC: | |
b218f786 | 941 | qp->allowed_ops = IB_OPCODE_UC; |
515667f8 DD |
942 | break; |
943 | default: | |
944 | ret = ERR_PTR(-EINVAL); | |
945 | goto bail_ip; | |
946 | } | |
947 | ||
948 | return ret; | |
949 | ||
950 | bail_ip: | |
22dccc54 JF |
951 | if (qp->ip) |
952 | kref_put(&qp->ip->ref, rvt_release_mmap_info); | |
515667f8 DD |
953 | |
954 | bail_qpn: | |
955 | free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); | |
956 | ||
957 | bail_rq_wq: | |
56c8ca51 MM |
958 | if (!qp->ip) |
959 | vfree(qp->r_rq.wq); | |
515667f8 DD |
960 | |
961 | bail_driver_priv: | |
962 | rdi->driver_f.qp_priv_free(rdi, qp); | |
963 | ||
964 | bail_qp: | |
8b103e9c | 965 | kfree(qp->s_ack_queue); |
515667f8 DD |
966 | kfree(qp); |
967 | ||
968 | bail_swq: | |
969 | vfree(swq); | |
970 | ||
971 | return ret; | |
b518d3e6 DD |
972 | } |
973 | ||
3b0b3fb3 DD |
974 | /** |
975 | * rvt_error_qp - put a QP into the error state | |
976 | * @qp: the QP to put into the error state | |
977 | * @err: the receive completion error to signal if a RWQE is active | |
978 | * | |
979 | * Flushes both send and receive work queues. | |
90793f71 DD |
980 | * |
981 | * Return: true if last WQE event should be generated. | |
3b0b3fb3 DD |
982 | * The QP r_lock and s_lock should be held and interrupts disabled. |
983 | * If we are already in error state, just return. | |
984 | */ | |
985 | int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err) | |
986 | { | |
987 | struct ib_wc wc; | |
988 | int ret = 0; | |
989 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
990 | ||
68e78b3d MM |
991 | lockdep_assert_held(&qp->r_lock); |
992 | lockdep_assert_held(&qp->s_lock); | |
3b0b3fb3 DD |
993 | if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) |
994 | goto bail; | |
995 | ||
996 | qp->state = IB_QPS_ERR; | |
997 | ||
998 | if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { | |
999 | qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); | |
1000 | del_timer(&qp->s_timer); | |
1001 | } | |
1002 | ||
1003 | if (qp->s_flags & RVT_S_ANY_WAIT_SEND) | |
1004 | qp->s_flags &= ~RVT_S_ANY_WAIT_SEND; | |
1005 | ||
1006 | rdi->driver_f.notify_error_qp(qp); | |
1007 | ||
1008 | /* Schedule the sending tasklet to drain the send work queue. */ | |
46a80d62 | 1009 | if (ACCESS_ONCE(qp->s_last) != qp->s_head) |
3b0b3fb3 DD |
1010 | rdi->driver_f.schedule_send(qp); |
1011 | ||
1012 | rvt_clear_mr_refs(qp, 0); | |
1013 | ||
1014 | memset(&wc, 0, sizeof(wc)); | |
1015 | wc.qp = &qp->ibqp; | |
1016 | wc.opcode = IB_WC_RECV; | |
1017 | ||
1018 | if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) { | |
1019 | wc.wr_id = qp->r_wr_id; | |
1020 | wc.status = err; | |
1021 | rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); | |
1022 | } | |
1023 | wc.status = IB_WC_WR_FLUSH_ERR; | |
1024 | ||
1025 | if (qp->r_rq.wq) { | |
1026 | struct rvt_rwq *wq; | |
1027 | u32 head; | |
1028 | u32 tail; | |
1029 | ||
1030 | spin_lock(&qp->r_rq.lock); | |
1031 | ||
1032 | /* sanity check pointers before trusting them */ | |
1033 | wq = qp->r_rq.wq; | |
1034 | head = wq->head; | |
1035 | if (head >= qp->r_rq.size) | |
1036 | head = 0; | |
1037 | tail = wq->tail; | |
1038 | if (tail >= qp->r_rq.size) | |
1039 | tail = 0; | |
1040 | while (tail != head) { | |
1041 | wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id; | |
1042 | if (++tail >= qp->r_rq.size) | |
1043 | tail = 0; | |
1044 | rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); | |
1045 | } | |
1046 | wq->tail = tail; | |
1047 | ||
1048 | spin_unlock(&qp->r_rq.lock); | |
1049 | } else if (qp->ibqp.event_handler) { | |
1050 | ret = 1; | |
1051 | } | |
1052 | ||
1053 | bail: | |
1054 | return ret; | |
1055 | } | |
1056 | EXPORT_SYMBOL(rvt_error_qp); | |
1057 | ||
1058 | /* | |
1059 | * Put the QP into the hash table. | |
1060 | * The hash table holds a reference to the QP. | |
1061 | */ | |
1062 | static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) | |
1063 | { | |
1064 | struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; | |
1065 | unsigned long flags; | |
1066 | ||
4d6f85c3 | 1067 | rvt_get_qp(qp); |
3b0b3fb3 DD |
1068 | spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); |
1069 | ||
1070 | if (qp->ibqp.qp_num <= 1) { | |
1071 | rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp); | |
1072 | } else { | |
1073 | u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); | |
1074 | ||
1075 | qp->next = rdi->qp_dev->qp_table[n]; | |
1076 | rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp); | |
1077 | trace_rvt_qpinsert(qp, n); | |
1078 | } | |
1079 | ||
1080 | spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); | |
1081 | } | |
1082 | ||
b518d3e6 | 1083 | /** |
61347fa6 | 1084 | * rvt_modify_qp - modify the attributes of a queue pair |
b518d3e6 DD |
1085 | * @ibqp: the queue pair who's attributes we're modifying |
1086 | * @attr: the new attributes | |
1087 | * @attr_mask: the mask of attributes to modify | |
1088 | * @udata: user data for libibverbs.so | |
1089 | * | |
90793f71 | 1090 | * Return: 0 on success, otherwise returns an errno. |
b518d3e6 DD |
1091 | */ |
1092 | int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, | |
1093 | int attr_mask, struct ib_udata *udata) | |
1094 | { | |
3b0b3fb3 DD |
1095 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); |
1096 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); | |
1097 | enum ib_qp_state cur_state, new_state; | |
1098 | struct ib_event ev; | |
1099 | int lastwqe = 0; | |
1100 | int mig = 0; | |
1101 | int pmtu = 0; /* for gcc warning only */ | |
1102 | enum rdma_link_layer link; | |
1103 | ||
1104 | link = rdma_port_get_link_layer(ibqp->device, qp->port_num); | |
1105 | ||
1106 | spin_lock_irq(&qp->r_lock); | |
46a80d62 | 1107 | spin_lock(&qp->s_hlock); |
3b0b3fb3 DD |
1108 | spin_lock(&qp->s_lock); |
1109 | ||
1110 | cur_state = attr_mask & IB_QP_CUR_STATE ? | |
1111 | attr->cur_qp_state : qp->state; | |
1112 | new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; | |
1113 | ||
1114 | if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, | |
1115 | attr_mask, link)) | |
1116 | goto inval; | |
1117 | ||
e85ec33d IW |
1118 | if (rdi->driver_f.check_modify_qp && |
1119 | rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata)) | |
1120 | goto inval; | |
1121 | ||
3b0b3fb3 DD |
1122 | if (attr_mask & IB_QP_AV) { |
1123 | if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
1124 | goto inval; | |
1125 | if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr)) | |
1126 | goto inval; | |
1127 | } | |
1128 | ||
1129 | if (attr_mask & IB_QP_ALT_PATH) { | |
1130 | if (attr->alt_ah_attr.dlid >= | |
1131 | be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
1132 | goto inval; | |
1133 | if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) | |
1134 | goto inval; | |
1135 | if (attr->alt_pkey_index >= rvt_get_npkeys(rdi)) | |
1136 | goto inval; | |
1137 | } | |
1138 | ||
1139 | if (attr_mask & IB_QP_PKEY_INDEX) | |
1140 | if (attr->pkey_index >= rvt_get_npkeys(rdi)) | |
1141 | goto inval; | |
1142 | ||
1143 | if (attr_mask & IB_QP_MIN_RNR_TIMER) | |
1144 | if (attr->min_rnr_timer > 31) | |
1145 | goto inval; | |
1146 | ||
1147 | if (attr_mask & IB_QP_PORT) | |
1148 | if (qp->ibqp.qp_type == IB_QPT_SMI || | |
1149 | qp->ibqp.qp_type == IB_QPT_GSI || | |
1150 | attr->port_num == 0 || | |
1151 | attr->port_num > ibqp->device->phys_port_cnt) | |
1152 | goto inval; | |
1153 | ||
1154 | if (attr_mask & IB_QP_DEST_QPN) | |
1155 | if (attr->dest_qp_num > RVT_QPN_MASK) | |
1156 | goto inval; | |
1157 | ||
1158 | if (attr_mask & IB_QP_RETRY_CNT) | |
1159 | if (attr->retry_cnt > 7) | |
1160 | goto inval; | |
1161 | ||
1162 | if (attr_mask & IB_QP_RNR_RETRY) | |
1163 | if (attr->rnr_retry > 7) | |
1164 | goto inval; | |
1165 | ||
b518d3e6 | 1166 | /* |
3b0b3fb3 DD |
1167 | * Don't allow invalid path_mtu values. OK to set greater |
1168 | * than the active mtu (or even the max_cap, if we have tuned | |
1169 | * that to a small mtu. We'll set qp->path_mtu | |
1170 | * to the lesser of requested attribute mtu and active, | |
1171 | * for packetizing messages. | |
1172 | * Note that the QP port has to be set in INIT and MTU in RTR. | |
b518d3e6 | 1173 | */ |
3b0b3fb3 DD |
1174 | if (attr_mask & IB_QP_PATH_MTU) { |
1175 | pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr); | |
1176 | if (pmtu < 0) | |
1177 | goto inval; | |
1178 | } | |
1179 | ||
1180 | if (attr_mask & IB_QP_PATH_MIG_STATE) { | |
1181 | if (attr->path_mig_state == IB_MIG_REARM) { | |
1182 | if (qp->s_mig_state == IB_MIG_ARMED) | |
1183 | goto inval; | |
1184 | if (new_state != IB_QPS_RTS) | |
1185 | goto inval; | |
1186 | } else if (attr->path_mig_state == IB_MIG_MIGRATED) { | |
1187 | if (qp->s_mig_state == IB_MIG_REARM) | |
1188 | goto inval; | |
1189 | if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) | |
1190 | goto inval; | |
1191 | if (qp->s_mig_state == IB_MIG_ARMED) | |
1192 | mig = 1; | |
1193 | } else { | |
1194 | goto inval; | |
1195 | } | |
1196 | } | |
1197 | ||
1198 | if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) | |
1199 | if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic) | |
1200 | goto inval; | |
1201 | ||
1202 | switch (new_state) { | |
1203 | case IB_QPS_RESET: | |
1204 | if (qp->state != IB_QPS_RESET) | |
1205 | rvt_reset_qp(rdi, qp, ibqp->qp_type); | |
1206 | break; | |
1207 | ||
1208 | case IB_QPS_RTR: | |
1209 | /* Allow event to re-trigger if QP set to RTR more than once */ | |
1210 | qp->r_flags &= ~RVT_R_COMM_EST; | |
1211 | qp->state = new_state; | |
1212 | break; | |
1213 | ||
1214 | case IB_QPS_SQD: | |
1215 | qp->s_draining = qp->s_last != qp->s_cur; | |
1216 | qp->state = new_state; | |
1217 | break; | |
1218 | ||
1219 | case IB_QPS_SQE: | |
1220 | if (qp->ibqp.qp_type == IB_QPT_RC) | |
1221 | goto inval; | |
1222 | qp->state = new_state; | |
1223 | break; | |
1224 | ||
1225 | case IB_QPS_ERR: | |
1226 | lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); | |
1227 | break; | |
1228 | ||
1229 | default: | |
1230 | qp->state = new_state; | |
1231 | break; | |
1232 | } | |
1233 | ||
1234 | if (attr_mask & IB_QP_PKEY_INDEX) | |
1235 | qp->s_pkey_index = attr->pkey_index; | |
1236 | ||
1237 | if (attr_mask & IB_QP_PORT) | |
1238 | qp->port_num = attr->port_num; | |
1239 | ||
1240 | if (attr_mask & IB_QP_DEST_QPN) | |
1241 | qp->remote_qpn = attr->dest_qp_num; | |
1242 | ||
1243 | if (attr_mask & IB_QP_SQ_PSN) { | |
1244 | qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask; | |
1245 | qp->s_psn = qp->s_next_psn; | |
1246 | qp->s_sending_psn = qp->s_next_psn; | |
1247 | qp->s_last_psn = qp->s_next_psn - 1; | |
1248 | qp->s_sending_hpsn = qp->s_last_psn; | |
1249 | } | |
1250 | ||
1251 | if (attr_mask & IB_QP_RQ_PSN) | |
1252 | qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask; | |
1253 | ||
1254 | if (attr_mask & IB_QP_ACCESS_FLAGS) | |
1255 | qp->qp_access_flags = attr->qp_access_flags; | |
1256 | ||
1257 | if (attr_mask & IB_QP_AV) { | |
1258 | qp->remote_ah_attr = attr->ah_attr; | |
1259 | qp->s_srate = attr->ah_attr.static_rate; | |
1260 | qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); | |
1261 | } | |
1262 | ||
1263 | if (attr_mask & IB_QP_ALT_PATH) { | |
1264 | qp->alt_ah_attr = attr->alt_ah_attr; | |
1265 | qp->s_alt_pkey_index = attr->alt_pkey_index; | |
1266 | } | |
1267 | ||
1268 | if (attr_mask & IB_QP_PATH_MIG_STATE) { | |
1269 | qp->s_mig_state = attr->path_mig_state; | |
1270 | if (mig) { | |
1271 | qp->remote_ah_attr = qp->alt_ah_attr; | |
1272 | qp->port_num = qp->alt_ah_attr.port_num; | |
1273 | qp->s_pkey_index = qp->s_alt_pkey_index; | |
3b0b3fb3 DD |
1274 | } |
1275 | } | |
1276 | ||
1277 | if (attr_mask & IB_QP_PATH_MTU) { | |
1278 | qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu); | |
1279 | qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu); | |
46a80d62 | 1280 | qp->log_pmtu = ilog2(qp->pmtu); |
3b0b3fb3 DD |
1281 | } |
1282 | ||
1283 | if (attr_mask & IB_QP_RETRY_CNT) { | |
1284 | qp->s_retry_cnt = attr->retry_cnt; | |
1285 | qp->s_retry = attr->retry_cnt; | |
1286 | } | |
1287 | ||
1288 | if (attr_mask & IB_QP_RNR_RETRY) { | |
1289 | qp->s_rnr_retry_cnt = attr->rnr_retry; | |
1290 | qp->s_rnr_retry = attr->rnr_retry; | |
1291 | } | |
1292 | ||
1293 | if (attr_mask & IB_QP_MIN_RNR_TIMER) | |
1294 | qp->r_min_rnr_timer = attr->min_rnr_timer; | |
1295 | ||
1296 | if (attr_mask & IB_QP_TIMEOUT) { | |
1297 | qp->timeout = attr->timeout; | |
1298 | qp->timeout_jiffies = | |
1299 | usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / | |
1300 | 1000UL); | |
1301 | } | |
1302 | ||
1303 | if (attr_mask & IB_QP_QKEY) | |
1304 | qp->qkey = attr->qkey; | |
1305 | ||
1306 | if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) | |
1307 | qp->r_max_rd_atomic = attr->max_dest_rd_atomic; | |
1308 | ||
1309 | if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) | |
1310 | qp->s_max_rd_atomic = attr->max_rd_atomic; | |
1311 | ||
e85ec33d IW |
1312 | if (rdi->driver_f.modify_qp) |
1313 | rdi->driver_f.modify_qp(qp, attr, attr_mask, udata); | |
1314 | ||
3b0b3fb3 | 1315 | spin_unlock(&qp->s_lock); |
46a80d62 | 1316 | spin_unlock(&qp->s_hlock); |
3b0b3fb3 DD |
1317 | spin_unlock_irq(&qp->r_lock); |
1318 | ||
1319 | if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) | |
1320 | rvt_insert_qp(rdi, qp); | |
1321 | ||
1322 | if (lastwqe) { | |
1323 | ev.device = qp->ibqp.device; | |
1324 | ev.element.qp = &qp->ibqp; | |
1325 | ev.event = IB_EVENT_QP_LAST_WQE_REACHED; | |
1326 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1327 | } | |
1328 | if (mig) { | |
1329 | ev.device = qp->ibqp.device; | |
1330 | ev.element.qp = &qp->ibqp; | |
1331 | ev.event = IB_EVENT_PATH_MIG; | |
1332 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1333 | } | |
1334 | return 0; | |
1335 | ||
1336 | inval: | |
1337 | spin_unlock(&qp->s_lock); | |
46a80d62 | 1338 | spin_unlock(&qp->s_hlock); |
3b0b3fb3 DD |
1339 | spin_unlock_irq(&qp->r_lock); |
1340 | return -EINVAL; | |
b518d3e6 DD |
1341 | } |
1342 | ||
79a225be DD |
1343 | /** rvt_free_qpn - Free a qpn from the bit map |
1344 | * @qpt: QP table | |
1345 | * @qpn: queue pair number to free | |
1346 | */ | |
1347 | static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn) | |
1348 | { | |
1349 | struct rvt_qpn_map *map; | |
1350 | ||
1351 | map = qpt->map + qpn / RVT_BITS_PER_PAGE; | |
1352 | if (map->page) | |
1353 | clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); | |
1354 | } | |
1355 | ||
b518d3e6 DD |
1356 | /** |
1357 | * rvt_destroy_qp - destroy a queue pair | |
1358 | * @ibqp: the queue pair to destroy | |
1359 | * | |
b518d3e6 DD |
1360 | * Note that this can be called while the QP is actively sending or |
1361 | * receiving! | |
90793f71 DD |
1362 | * |
1363 | * Return: 0 on success. | |
b518d3e6 DD |
1364 | */ |
1365 | int rvt_destroy_qp(struct ib_qp *ibqp) | |
1366 | { | |
5a17ad11 DD |
1367 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1368 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
b518d3e6 | 1369 | |
5a17ad11 | 1370 | spin_lock_irq(&qp->r_lock); |
46a80d62 | 1371 | spin_lock(&qp->s_hlock); |
5a17ad11 DD |
1372 | spin_lock(&qp->s_lock); |
1373 | rvt_reset_qp(rdi, qp, ibqp->qp_type); | |
1374 | spin_unlock(&qp->s_lock); | |
46a80d62 | 1375 | spin_unlock(&qp->s_hlock); |
5a17ad11 DD |
1376 | spin_unlock_irq(&qp->r_lock); |
1377 | ||
1378 | /* qpn is now available for use again */ | |
1379 | rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); | |
1380 | ||
1381 | spin_lock(&rdi->n_qps_lock); | |
1382 | rdi->n_qps_allocated--; | |
bfee5e32 VM |
1383 | if (qp->ibqp.qp_type == IB_QPT_RC) { |
1384 | rdi->n_rc_qps--; | |
1385 | rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL; | |
1386 | } | |
5a17ad11 DD |
1387 | spin_unlock(&rdi->n_qps_lock); |
1388 | ||
1389 | if (qp->ip) | |
1390 | kref_put(&qp->ip->ref, rvt_release_mmap_info); | |
1391 | else | |
1392 | vfree(qp->r_rq.wq); | |
1393 | vfree(qp->s_wq); | |
1394 | rdi->driver_f.qp_priv_free(rdi, qp); | |
8b103e9c | 1395 | kfree(qp->s_ack_queue); |
5a17ad11 DD |
1396 | kfree(qp); |
1397 | return 0; | |
b518d3e6 DD |
1398 | } |
1399 | ||
90793f71 DD |
1400 | /** |
1401 | * rvt_query_qp - query an ipbq | |
1402 | * @ibqp: IB qp to query | |
1403 | * @attr: attr struct to fill in | |
1404 | * @attr_mask: attr mask ignored | |
1405 | * @init_attr: struct to fill in | |
1406 | * | |
1407 | * Return: always 0 | |
1408 | */ | |
b518d3e6 DD |
1409 | int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, |
1410 | int attr_mask, struct ib_qp_init_attr *init_attr) | |
1411 | { | |
74d2d500 HC |
1412 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1413 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
1414 | ||
1415 | attr->qp_state = qp->state; | |
1416 | attr->cur_qp_state = attr->qp_state; | |
1417 | attr->path_mtu = qp->path_mtu; | |
1418 | attr->path_mig_state = qp->s_mig_state; | |
1419 | attr->qkey = qp->qkey; | |
1420 | attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask; | |
1421 | attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask; | |
1422 | attr->dest_qp_num = qp->remote_qpn; | |
1423 | attr->qp_access_flags = qp->qp_access_flags; | |
856cc4c2 MM |
1424 | attr->cap.max_send_wr = qp->s_size - 1 - |
1425 | rdi->dparms.reserved_operations; | |
74d2d500 HC |
1426 | attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1; |
1427 | attr->cap.max_send_sge = qp->s_max_sge; | |
1428 | attr->cap.max_recv_sge = qp->r_rq.max_sge; | |
1429 | attr->cap.max_inline_data = 0; | |
1430 | attr->ah_attr = qp->remote_ah_attr; | |
1431 | attr->alt_ah_attr = qp->alt_ah_attr; | |
1432 | attr->pkey_index = qp->s_pkey_index; | |
1433 | attr->alt_pkey_index = qp->s_alt_pkey_index; | |
1434 | attr->en_sqd_async_notify = 0; | |
1435 | attr->sq_draining = qp->s_draining; | |
1436 | attr->max_rd_atomic = qp->s_max_rd_atomic; | |
1437 | attr->max_dest_rd_atomic = qp->r_max_rd_atomic; | |
1438 | attr->min_rnr_timer = qp->r_min_rnr_timer; | |
1439 | attr->port_num = qp->port_num; | |
1440 | attr->timeout = qp->timeout; | |
1441 | attr->retry_cnt = qp->s_retry_cnt; | |
1442 | attr->rnr_retry = qp->s_rnr_retry_cnt; | |
1443 | attr->alt_port_num = qp->alt_ah_attr.port_num; | |
1444 | attr->alt_timeout = qp->alt_timeout; | |
1445 | ||
1446 | init_attr->event_handler = qp->ibqp.event_handler; | |
1447 | init_attr->qp_context = qp->ibqp.qp_context; | |
1448 | init_attr->send_cq = qp->ibqp.send_cq; | |
1449 | init_attr->recv_cq = qp->ibqp.recv_cq; | |
1450 | init_attr->srq = qp->ibqp.srq; | |
1451 | init_attr->cap = attr->cap; | |
1452 | if (qp->s_flags & RVT_S_SIGNAL_REQ_WR) | |
1453 | init_attr->sq_sig_type = IB_SIGNAL_REQ_WR; | |
1454 | else | |
1455 | init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; | |
1456 | init_attr->qp_type = qp->ibqp.qp_type; | |
1457 | init_attr->port_num = qp->port_num; | |
1458 | return 0; | |
b518d3e6 | 1459 | } |
8cf4020b DD |
1460 | |
1461 | /** | |
1462 | * rvt_post_receive - post a receive on a QP | |
1463 | * @ibqp: the QP to post the receive on | |
1464 | * @wr: the WR to post | |
1465 | * @bad_wr: the first bad WR is put here | |
1466 | * | |
1467 | * This may be called from interrupt context. | |
90793f71 DD |
1468 | * |
1469 | * Return: 0 on success otherwise errno | |
8cf4020b DD |
1470 | */ |
1471 | int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, | |
1472 | struct ib_recv_wr **bad_wr) | |
1473 | { | |
120bdafa DD |
1474 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1475 | struct rvt_rwq *wq = qp->r_rq.wq; | |
1476 | unsigned long flags; | |
000a830e AE |
1477 | int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) && |
1478 | !qp->ibqp.srq; | |
8cf4020b | 1479 | |
120bdafa DD |
1480 | /* Check that state is OK to post receive. */ |
1481 | if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) { | |
1482 | *bad_wr = wr; | |
1483 | return -EINVAL; | |
1484 | } | |
1485 | ||
1486 | for (; wr; wr = wr->next) { | |
1487 | struct rvt_rwqe *wqe; | |
1488 | u32 next; | |
1489 | int i; | |
1490 | ||
1491 | if ((unsigned)wr->num_sge > qp->r_rq.max_sge) { | |
1492 | *bad_wr = wr; | |
1493 | return -EINVAL; | |
1494 | } | |
1495 | ||
1496 | spin_lock_irqsave(&qp->r_rq.lock, flags); | |
1497 | next = wq->head + 1; | |
1498 | if (next >= qp->r_rq.size) | |
1499 | next = 0; | |
1500 | if (next == wq->tail) { | |
1501 | spin_unlock_irqrestore(&qp->r_rq.lock, flags); | |
1502 | *bad_wr = wr; | |
1503 | return -ENOMEM; | |
1504 | } | |
000a830e AE |
1505 | if (unlikely(qp_err_flush)) { |
1506 | struct ib_wc wc; | |
1507 | ||
1508 | memset(&wc, 0, sizeof(wc)); | |
1509 | wc.qp = &qp->ibqp; | |
1510 | wc.opcode = IB_WC_RECV; | |
1511 | wc.wr_id = wr->wr_id; | |
1512 | wc.status = IB_WC_WR_FLUSH_ERR; | |
1513 | rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); | |
1514 | } else { | |
1515 | wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head); | |
1516 | wqe->wr_id = wr->wr_id; | |
1517 | wqe->num_sge = wr->num_sge; | |
1518 | for (i = 0; i < wr->num_sge; i++) | |
1519 | wqe->sg_list[i] = wr->sg_list[i]; | |
1520 | /* | |
1521 | * Make sure queue entry is written | |
1522 | * before the head index. | |
1523 | */ | |
1524 | smp_wmb(); | |
1525 | wq->head = next; | |
1526 | } | |
120bdafa DD |
1527 | spin_unlock_irqrestore(&qp->r_rq.lock, flags); |
1528 | } | |
1529 | return 0; | |
8cf4020b DD |
1530 | } |
1531 | ||
46a80d62 | 1532 | /** |
afcf8f76 MM |
1533 | * rvt_qp_valid_operation - validate post send wr request |
1534 | * @qp - the qp | |
1535 | * @post-parms - the post send table for the driver | |
1536 | * @wr - the work request | |
46a80d62 | 1537 | * |
afcf8f76 MM |
1538 | * The routine validates the operation based on the |
1539 | * validation table an returns the length of the operation | |
1540 | * which can extend beyond the ib_send_bw. Operation | |
1541 | * dependent flags key atomic operation validation. | |
1542 | * | |
1543 | * There is an exception for UD qps that validates the pd and | |
1544 | * overrides the length to include the additional UD specific | |
1545 | * length. | |
1546 | * | |
1547 | * Returns a negative error or the length of the work request | |
1548 | * for building the swqe. | |
1549 | */ | |
1550 | static inline int rvt_qp_valid_operation( | |
1551 | struct rvt_qp *qp, | |
1552 | const struct rvt_operation_params *post_parms, | |
1553 | struct ib_send_wr *wr) | |
1554 | { | |
1555 | int len; | |
1556 | ||
1557 | if (wr->opcode >= RVT_OPERATION_MAX || !post_parms[wr->opcode].length) | |
1558 | return -EINVAL; | |
1559 | if (!(post_parms[wr->opcode].qpt_support & BIT(qp->ibqp.qp_type))) | |
1560 | return -EINVAL; | |
1561 | if ((post_parms[wr->opcode].flags & RVT_OPERATION_PRIV) && | |
1562 | ibpd_to_rvtpd(qp->ibqp.pd)->user) | |
1563 | return -EINVAL; | |
1564 | if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC_SGE && | |
1565 | (wr->num_sge == 0 || | |
1566 | wr->sg_list[0].length < sizeof(u64) || | |
1567 | wr->sg_list[0].addr & (sizeof(u64) - 1))) | |
1568 | return -EINVAL; | |
1569 | if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC && | |
1570 | !qp->s_max_rd_atomic) | |
1571 | return -EINVAL; | |
1572 | len = post_parms[wr->opcode].length; | |
1573 | /* UD specific */ | |
1574 | if (qp->ibqp.qp_type != IB_QPT_UC && | |
1575 | qp->ibqp.qp_type != IB_QPT_RC) { | |
1576 | if (qp->ibqp.pd != ud_wr(wr)->ah->pd) | |
1577 | return -EINVAL; | |
1578 | len = sizeof(struct ib_ud_wr); | |
1579 | } | |
1580 | return len; | |
1581 | } | |
1582 | ||
1583 | /** | |
856cc4c2 | 1584 | * rvt_qp_is_avail - determine queue capacity |
46a80d62 | 1585 | * @qp - the qp |
856cc4c2 MM |
1586 | * @rdi - the rdmavt device |
1587 | * @reserved_op - is reserved operation | |
46a80d62 MM |
1588 | * |
1589 | * This assumes the s_hlock is held but the s_last | |
1590 | * qp variable is uncontrolled. | |
afcf8f76 | 1591 | * |
856cc4c2 MM |
1592 | * For non reserved operations, the qp->s_avail |
1593 | * may be changed. | |
1594 | * | |
1595 | * The return value is zero or a -ENOMEM. | |
46a80d62 | 1596 | */ |
856cc4c2 MM |
1597 | static inline int rvt_qp_is_avail( |
1598 | struct rvt_qp *qp, | |
1599 | struct rvt_dev_info *rdi, | |
1600 | bool reserved_op) | |
46a80d62 MM |
1601 | { |
1602 | u32 slast; | |
856cc4c2 MM |
1603 | u32 avail; |
1604 | u32 reserved_used; | |
1605 | ||
1606 | /* see rvt_qp_wqe_unreserve() */ | |
1607 | smp_mb__before_atomic(); | |
1608 | reserved_used = atomic_read(&qp->s_reserved_used); | |
1609 | if (unlikely(reserved_op)) { | |
1610 | /* see rvt_qp_wqe_unreserve() */ | |
1611 | smp_mb__before_atomic(); | |
1612 | if (reserved_used >= rdi->dparms.reserved_operations) | |
1613 | return -ENOMEM; | |
1614 | return 0; | |
1615 | } | |
1616 | /* non-reserved operations */ | |
1617 | if (likely(qp->s_avail)) | |
1618 | return 0; | |
46a80d62 MM |
1619 | smp_read_barrier_depends(); /* see rc.c */ |
1620 | slast = ACCESS_ONCE(qp->s_last); | |
1621 | if (qp->s_head >= slast) | |
856cc4c2 | 1622 | avail = qp->s_size - (qp->s_head - slast); |
46a80d62 | 1623 | else |
856cc4c2 MM |
1624 | avail = slast - qp->s_head; |
1625 | ||
1626 | /* see rvt_qp_wqe_unreserve() */ | |
1627 | smp_mb__before_atomic(); | |
1628 | reserved_used = atomic_read(&qp->s_reserved_used); | |
1629 | avail = avail - 1 - | |
1630 | (rdi->dparms.reserved_operations - reserved_used); | |
1631 | /* insure we don't assign a negative s_avail */ | |
1632 | if ((s32)avail <= 0) | |
1633 | return -ENOMEM; | |
1634 | qp->s_avail = avail; | |
1635 | if (WARN_ON(qp->s_avail > | |
1636 | (qp->s_size - 1 - rdi->dparms.reserved_operations))) | |
1637 | rvt_pr_err(rdi, | |
1638 | "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u", | |
1639 | qp->ibqp.qp_num, qp->s_size, qp->s_avail, | |
1640 | qp->s_head, qp->s_tail, qp->s_cur, | |
1641 | qp->s_acked, qp->s_last); | |
1642 | return 0; | |
46a80d62 MM |
1643 | } |
1644 | ||
bfbac097 DD |
1645 | /** |
1646 | * rvt_post_one_wr - post one RC, UC, or UD send work request | |
1647 | * @qp: the QP to post on | |
1648 | * @wr: the work request to send | |
1649 | */ | |
91702b4a MM |
1650 | static int rvt_post_one_wr(struct rvt_qp *qp, |
1651 | struct ib_send_wr *wr, | |
1652 | int *call_send) | |
bfbac097 DD |
1653 | { |
1654 | struct rvt_swqe *wqe; | |
1655 | u32 next; | |
1656 | int i; | |
1657 | int j; | |
1658 | int acc; | |
1659 | struct rvt_lkey_table *rkt; | |
1660 | struct rvt_pd *pd; | |
1661 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
46a80d62 MM |
1662 | u8 log_pmtu; |
1663 | int ret; | |
2821c509 | 1664 | size_t cplen; |
856cc4c2 | 1665 | bool reserved_op; |
d9b13c20 | 1666 | int local_ops_delayed = 0; |
bfbac097 | 1667 | |
afcf8f76 MM |
1668 | BUILD_BUG_ON(IB_QPT_MAX >= (sizeof(u32) * BITS_PER_BYTE)); |
1669 | ||
bfbac097 DD |
1670 | /* IB spec says that num_sge == 0 is OK. */ |
1671 | if (unlikely(wr->num_sge > qp->s_max_sge)) | |
1672 | return -EINVAL; | |
1673 | ||
2821c509 MM |
1674 | ret = rvt_qp_valid_operation(qp, rdi->post_parms, wr); |
1675 | if (ret < 0) | |
1676 | return ret; | |
1677 | cplen = ret; | |
1678 | ||
d9f87239 | 1679 | /* |
d9b13c20 JX |
1680 | * Local operations include fast register and local invalidate. |
1681 | * Fast register needs to be processed immediately because the | |
1682 | * registered lkey may be used by following work requests and the | |
1683 | * lkey needs to be valid at the time those requests are posted. | |
1684 | * Local invalidate can be processed immediately if fencing is | |
1685 | * not required and no previous local invalidate ops are pending. | |
1686 | * Signaled local operations that have been processed immediately | |
1687 | * need to have requests with "completion only" flags set posted | |
1688 | * to the send queue in order to generate completions. | |
d9f87239 | 1689 | */ |
d9b13c20 | 1690 | if ((rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) { |
d9f87239 JX |
1691 | switch (wr->opcode) { |
1692 | case IB_WR_REG_MR: | |
d9b13c20 JX |
1693 | ret = rvt_fast_reg_mr(qp, |
1694 | reg_wr(wr)->mr, | |
1695 | reg_wr(wr)->key, | |
1696 | reg_wr(wr)->access); | |
1697 | if (ret || !(wr->send_flags & IB_SEND_SIGNALED)) | |
1698 | return ret; | |
1699 | break; | |
d9f87239 | 1700 | case IB_WR_LOCAL_INV: |
d9b13c20 JX |
1701 | if ((wr->send_flags & IB_SEND_FENCE) || |
1702 | atomic_read(&qp->local_ops_pending)) { | |
1703 | local_ops_delayed = 1; | |
1704 | } else { | |
1705 | ret = rvt_invalidate_rkey( | |
1706 | qp, wr->ex.invalidate_rkey); | |
1707 | if (ret || !(wr->send_flags & IB_SEND_SIGNALED)) | |
1708 | return ret; | |
1709 | } | |
1710 | break; | |
d9f87239 JX |
1711 | default: |
1712 | return -EINVAL; | |
1713 | } | |
1714 | } | |
1715 | ||
856cc4c2 MM |
1716 | reserved_op = rdi->post_parms[wr->opcode].flags & |
1717 | RVT_OPERATION_USE_RESERVE; | |
46a80d62 | 1718 | /* check for avail */ |
856cc4c2 MM |
1719 | ret = rvt_qp_is_avail(qp, rdi, reserved_op); |
1720 | if (ret) | |
1721 | return ret; | |
bfbac097 DD |
1722 | next = qp->s_head + 1; |
1723 | if (next >= qp->s_size) | |
1724 | next = 0; | |
60c30f57 | 1725 | |
bfbac097 DD |
1726 | rkt = &rdi->lkey_table; |
1727 | pd = ibpd_to_rvtpd(qp->ibqp.pd); | |
1728 | wqe = rvt_get_swqe_ptr(qp, qp->s_head); | |
1729 | ||
2821c509 MM |
1730 | /* cplen has length from above */ |
1731 | memcpy(&wqe->wr, wr, cplen); | |
bfbac097 DD |
1732 | |
1733 | wqe->length = 0; | |
1734 | j = 0; | |
1735 | if (wr->num_sge) { | |
1736 | acc = wr->opcode >= IB_WR_RDMA_READ ? | |
1737 | IB_ACCESS_LOCAL_WRITE : 0; | |
1738 | for (i = 0; i < wr->num_sge; i++) { | |
1739 | u32 length = wr->sg_list[i].length; | |
1740 | int ok; | |
1741 | ||
1742 | if (length == 0) | |
1743 | continue; | |
1744 | ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], | |
1745 | &wr->sg_list[i], acc); | |
46a80d62 MM |
1746 | if (!ok) { |
1747 | ret = -EINVAL; | |
bfbac097 | 1748 | goto bail_inval_free; |
46a80d62 | 1749 | } |
bfbac097 DD |
1750 | wqe->length += length; |
1751 | j++; | |
1752 | } | |
1753 | wqe->wr.num_sge = j; | |
1754 | } | |
46a80d62 MM |
1755 | |
1756 | /* general part of wqe valid - allow for driver checks */ | |
1757 | if (rdi->driver_f.check_send_wqe) { | |
1758 | ret = rdi->driver_f.check_send_wqe(qp, wqe); | |
91702b4a | 1759 | if (ret < 0) |
bfbac097 | 1760 | goto bail_inval_free; |
91702b4a MM |
1761 | if (ret) |
1762 | *call_send = ret; | |
46a80d62 MM |
1763 | } |
1764 | ||
1765 | log_pmtu = qp->log_pmtu; | |
1766 | if (qp->ibqp.qp_type != IB_QPT_UC && | |
1767 | qp->ibqp.qp_type != IB_QPT_RC) { | |
1768 | struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah); | |
1769 | ||
1770 | log_pmtu = ah->log_pmtu; | |
bfbac097 DD |
1771 | atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount); |
1772 | } | |
46a80d62 | 1773 | |
d9f87239 | 1774 | if (rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL) { |
d9b13c20 JX |
1775 | if (local_ops_delayed) |
1776 | atomic_inc(&qp->local_ops_pending); | |
1777 | else | |
1778 | wqe->wr.send_flags |= RVT_SEND_COMPLETION_ONLY; | |
d9f87239 JX |
1779 | wqe->ssn = 0; |
1780 | wqe->psn = 0; | |
1781 | wqe->lpsn = 0; | |
1782 | } else { | |
1783 | wqe->ssn = qp->s_ssn++; | |
1784 | wqe->psn = qp->s_next_psn; | |
1785 | wqe->lpsn = wqe->psn + | |
1786 | (wqe->length ? | |
1787 | ((wqe->length - 1) >> log_pmtu) : | |
1788 | 0); | |
1789 | qp->s_next_psn = wqe->lpsn + 1; | |
1790 | } | |
e16689e4 | 1791 | trace_rvt_post_one_wr(qp, wqe); |
856cc4c2 MM |
1792 | if (unlikely(reserved_op)) |
1793 | rvt_qp_wqe_reserve(qp, wqe); | |
1794 | else | |
1795 | qp->s_avail--; | |
46a80d62 | 1796 | smp_wmb(); /* see request builders */ |
bfbac097 DD |
1797 | qp->s_head = next; |
1798 | ||
1799 | return 0; | |
1800 | ||
1801 | bail_inval_free: | |
1802 | /* release mr holds */ | |
1803 | while (j) { | |
1804 | struct rvt_sge *sge = &wqe->sg_list[--j]; | |
1805 | ||
1806 | rvt_put_mr(sge->mr); | |
1807 | } | |
46a80d62 | 1808 | return ret; |
bfbac097 DD |
1809 | } |
1810 | ||
8cf4020b DD |
1811 | /** |
1812 | * rvt_post_send - post a send on a QP | |
1813 | * @ibqp: the QP to post the send on | |
1814 | * @wr: the list of work requests to post | |
1815 | * @bad_wr: the first bad WR is put here | |
1816 | * | |
1817 | * This may be called from interrupt context. | |
90793f71 DD |
1818 | * |
1819 | * Return: 0 on success else errno | |
8cf4020b DD |
1820 | */ |
1821 | int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, | |
1822 | struct ib_send_wr **bad_wr) | |
1823 | { | |
bfbac097 DD |
1824 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1825 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
1826 | unsigned long flags = 0; | |
1827 | int call_send; | |
1828 | unsigned nreq = 0; | |
1829 | int err = 0; | |
1830 | ||
46a80d62 | 1831 | spin_lock_irqsave(&qp->s_hlock, flags); |
bfbac097 | 1832 | |
8cf4020b | 1833 | /* |
bfbac097 DD |
1834 | * Ensure QP state is such that we can send. If not bail out early, |
1835 | * there is no need to do this every time we post a send. | |
8cf4020b | 1836 | */ |
bfbac097 | 1837 | if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) { |
46a80d62 | 1838 | spin_unlock_irqrestore(&qp->s_hlock, flags); |
bfbac097 DD |
1839 | return -EINVAL; |
1840 | } | |
8cf4020b | 1841 | |
bfbac097 DD |
1842 | /* |
1843 | * If the send queue is empty, and we only have a single WR then just go | |
1844 | * ahead and kick the send engine into gear. Otherwise we will always | |
1845 | * just schedule the send to happen later. | |
1846 | */ | |
1847 | call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next; | |
1848 | ||
1849 | for (; wr; wr = wr->next) { | |
91702b4a | 1850 | err = rvt_post_one_wr(qp, wr, &call_send); |
bfbac097 DD |
1851 | if (unlikely(err)) { |
1852 | *bad_wr = wr; | |
1853 | goto bail; | |
1854 | } | |
1855 | nreq++; | |
1856 | } | |
1857 | bail: | |
46a80d62 MM |
1858 | spin_unlock_irqrestore(&qp->s_hlock, flags); |
1859 | if (nreq) { | |
1860 | if (call_send) | |
46a80d62 | 1861 | rdi->driver_f.do_send(qp); |
e6d2e017 JJ |
1862 | else |
1863 | rdi->driver_f.schedule_send_no_lock(qp); | |
46a80d62 | 1864 | } |
bfbac097 | 1865 | return err; |
8cf4020b DD |
1866 | } |
1867 | ||
1868 | /** | |
1869 | * rvt_post_srq_receive - post a receive on a shared receive queue | |
1870 | * @ibsrq: the SRQ to post the receive on | |
1871 | * @wr: the list of work requests to post | |
1872 | * @bad_wr: A pointer to the first WR to cause a problem is put here | |
1873 | * | |
1874 | * This may be called from interrupt context. | |
90793f71 DD |
1875 | * |
1876 | * Return: 0 on success else errno | |
8cf4020b DD |
1877 | */ |
1878 | int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, | |
1879 | struct ib_recv_wr **bad_wr) | |
1880 | { | |
b8f881b9 JJ |
1881 | struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); |
1882 | struct rvt_rwq *wq; | |
1883 | unsigned long flags; | |
1884 | ||
1885 | for (; wr; wr = wr->next) { | |
1886 | struct rvt_rwqe *wqe; | |
1887 | u32 next; | |
1888 | int i; | |
1889 | ||
1890 | if ((unsigned)wr->num_sge > srq->rq.max_sge) { | |
1891 | *bad_wr = wr; | |
1892 | return -EINVAL; | |
1893 | } | |
1894 | ||
1895 | spin_lock_irqsave(&srq->rq.lock, flags); | |
1896 | wq = srq->rq.wq; | |
1897 | next = wq->head + 1; | |
1898 | if (next >= srq->rq.size) | |
1899 | next = 0; | |
1900 | if (next == wq->tail) { | |
1901 | spin_unlock_irqrestore(&srq->rq.lock, flags); | |
1902 | *bad_wr = wr; | |
1903 | return -ENOMEM; | |
1904 | } | |
1905 | ||
1906 | wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head); | |
1907 | wqe->wr_id = wr->wr_id; | |
1908 | wqe->num_sge = wr->num_sge; | |
1909 | for (i = 0; i < wr->num_sge; i++) | |
1910 | wqe->sg_list[i] = wr->sg_list[i]; | |
1911 | /* Make sure queue entry is written before the head index. */ | |
1912 | smp_wmb(); | |
1913 | wq->head = next; | |
1914 | spin_unlock_irqrestore(&srq->rq.lock, flags); | |
1915 | } | |
1916 | return 0; | |
8cf4020b | 1917 | } |
beb5a042 BW |
1918 | |
1919 | /** | |
1920 | * qp_comm_est - handle trap with QP established | |
1921 | * @qp: the QP | |
1922 | */ | |
1923 | void rvt_comm_est(struct rvt_qp *qp) | |
1924 | { | |
1925 | qp->r_flags |= RVT_R_COMM_EST; | |
1926 | if (qp->ibqp.event_handler) { | |
1927 | struct ib_event ev; | |
1928 | ||
1929 | ev.device = qp->ibqp.device; | |
1930 | ev.element.qp = &qp->ibqp; | |
1931 | ev.event = IB_EVENT_COMM_EST; | |
1932 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1933 | } | |
1934 | } | |
1935 | EXPORT_SYMBOL(rvt_comm_est); | |
1936 | ||
1937 | void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err) | |
1938 | { | |
1939 | unsigned long flags; | |
1940 | int lastwqe; | |
1941 | ||
1942 | spin_lock_irqsave(&qp->s_lock, flags); | |
1943 | lastwqe = rvt_error_qp(qp, err); | |
1944 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1945 | ||
1946 | if (lastwqe) { | |
1947 | struct ib_event ev; | |
1948 | ||
1949 | ev.device = qp->ibqp.device; | |
1950 | ev.element.qp = &qp->ibqp; | |
1951 | ev.event = IB_EVENT_QP_LAST_WQE_REACHED; | |
1952 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1953 | } | |
1954 | } | |
1955 | EXPORT_SYMBOL(rvt_rc_error); | |
11a10d4b VSD |
1956 | |
1957 | static inline unsigned long rvt_aeth_to_usec(u32 aeth) | |
1958 | { | |
1959 | return ib_rvt_rnr_table[(aeth >> RVT_AETH_CREDIT_SHIFT) & | |
1960 | RVT_AETH_CREDIT_MASK]; | |
1961 | } | |
1962 | ||
1963 | /* | |
1964 | * rvt_add_retry_timer - add/start a retry timer | |
1965 | * @qp - the QP | |
1966 | * add a retry timer on the QP | |
1967 | */ | |
1968 | void rvt_add_retry_timer(struct rvt_qp *qp) | |
1969 | { | |
1970 | struct ib_qp *ibqp = &qp->ibqp; | |
1971 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
1972 | ||
1973 | lockdep_assert_held(&qp->s_lock); | |
1974 | qp->s_flags |= RVT_S_TIMER; | |
1975 | /* 4.096 usec. * (1 << qp->timeout) */ | |
1976 | qp->s_timer.expires = jiffies + qp->timeout_jiffies + | |
1977 | rdi->busy_jiffies; | |
1978 | add_timer(&qp->s_timer); | |
1979 | } | |
1980 | EXPORT_SYMBOL(rvt_add_retry_timer); | |
1981 | ||
1982 | /** | |
1983 | * rvt_add_rnr_timer - add/start an rnr timer | |
1984 | * @qp - the QP | |
1985 | * @aeth - aeth of RNR timeout, simulated aeth for loopback | |
1986 | * add an rnr timer on the QP | |
1987 | */ | |
1988 | void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth) | |
1989 | { | |
1990 | u32 to; | |
1991 | ||
1992 | lockdep_assert_held(&qp->s_lock); | |
1993 | qp->s_flags |= RVT_S_WAIT_RNR; | |
1994 | to = rvt_aeth_to_usec(aeth); | |
1995 | hrtimer_start(&qp->s_rnr_timer, | |
1996 | ns_to_ktime(1000 * to), HRTIMER_MODE_REL); | |
1997 | } | |
1998 | EXPORT_SYMBOL(rvt_add_rnr_timer); | |
1999 | ||
2000 | /** | |
2001 | * rvt_stop_rc_timers - stop all timers | |
2002 | * @qp - the QP | |
2003 | * stop any pending timers | |
2004 | */ | |
2005 | void rvt_stop_rc_timers(struct rvt_qp *qp) | |
2006 | { | |
2007 | lockdep_assert_held(&qp->s_lock); | |
2008 | /* Remove QP from all timers */ | |
2009 | if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { | |
2010 | qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); | |
2011 | del_timer(&qp->s_timer); | |
2012 | hrtimer_try_to_cancel(&qp->s_rnr_timer); | |
2013 | } | |
2014 | } | |
2015 | EXPORT_SYMBOL(rvt_stop_rc_timers); | |
2016 | ||
2017 | /** | |
2018 | * rvt_stop_rnr_timer - stop an rnr timer | |
2019 | * @qp - the QP | |
2020 | * | |
2021 | * stop an rnr timer and return if the timer | |
2022 | * had been pending. | |
2023 | */ | |
2024 | static int rvt_stop_rnr_timer(struct rvt_qp *qp) | |
2025 | { | |
2026 | int rval = 0; | |
2027 | ||
2028 | lockdep_assert_held(&qp->s_lock); | |
2029 | /* Remove QP from rnr timer */ | |
2030 | if (qp->s_flags & RVT_S_WAIT_RNR) { | |
2031 | qp->s_flags &= ~RVT_S_WAIT_RNR; | |
2032 | rval = hrtimer_try_to_cancel(&qp->s_rnr_timer); | |
2033 | } | |
2034 | return rval; | |
2035 | } | |
2036 | ||
2037 | /** | |
2038 | * rvt_del_timers_sync - wait for any timeout routines to exit | |
2039 | * @qp - the QP | |
2040 | */ | |
2041 | void rvt_del_timers_sync(struct rvt_qp *qp) | |
2042 | { | |
2043 | del_timer_sync(&qp->s_timer); | |
2044 | hrtimer_cancel(&qp->s_rnr_timer); | |
2045 | } | |
2046 | EXPORT_SYMBOL(rvt_del_timers_sync); | |
2047 | ||
2048 | /** | |
2049 | * This is called from s_timer for missing responses. | |
2050 | */ | |
2051 | static void rvt_rc_timeout(unsigned long arg) | |
2052 | { | |
2053 | struct rvt_qp *qp = (struct rvt_qp *)arg; | |
2054 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
2055 | unsigned long flags; | |
2056 | ||
2057 | spin_lock_irqsave(&qp->r_lock, flags); | |
2058 | spin_lock(&qp->s_lock); | |
2059 | if (qp->s_flags & RVT_S_TIMER) { | |
2060 | qp->s_flags &= ~RVT_S_TIMER; | |
2061 | del_timer(&qp->s_timer); | |
2062 | if (rdi->driver_f.notify_restart_rc) | |
2063 | rdi->driver_f.notify_restart_rc(qp, | |
2064 | qp->s_last_psn + 1, | |
2065 | 1); | |
2066 | rdi->driver_f.schedule_send(qp); | |
2067 | } | |
2068 | spin_unlock(&qp->s_lock); | |
2069 | spin_unlock_irqrestore(&qp->r_lock, flags); | |
2070 | } | |
2071 | ||
2072 | /* | |
2073 | * This is called from s_timer for RNR timeouts. | |
2074 | */ | |
2075 | enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t) | |
2076 | { | |
2077 | struct rvt_qp *qp = container_of(t, struct rvt_qp, s_rnr_timer); | |
2078 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
2079 | unsigned long flags; | |
2080 | ||
2081 | spin_lock_irqsave(&qp->s_lock, flags); | |
2082 | rvt_stop_rnr_timer(qp); | |
2083 | rdi->driver_f.schedule_send(qp); | |
2084 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2085 | return HRTIMER_NORESTART; | |
2086 | } | |
2087 | EXPORT_SYMBOL(rvt_rc_rnr_retry); |