Commit | Line | Data |
---|---|---|
77241056 MM |
1 | /* |
2 | * | |
3 | * This file is provided under a dual BSD/GPLv2 license. When using or | |
4 | * redistributing this file, you may do so under either license. | |
5 | * | |
6 | * GPL LICENSE SUMMARY | |
7 | * | |
8 | * Copyright(c) 2015 Intel Corporation. | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or modify | |
11 | * it under the terms of version 2 of the GNU General Public License as | |
12 | * published by the Free Software Foundation. | |
13 | * | |
14 | * This program is distributed in the hope that it will be useful, but | |
15 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * General Public License for more details. | |
18 | * | |
19 | * BSD LICENSE | |
20 | * | |
21 | * Copyright(c) 2015 Intel Corporation. | |
22 | * | |
23 | * Redistribution and use in source and binary forms, with or without | |
24 | * modification, are permitted provided that the following conditions | |
25 | * are met: | |
26 | * | |
27 | * - Redistributions of source code must retain the above copyright | |
28 | * notice, this list of conditions and the following disclaimer. | |
29 | * - Redistributions in binary form must reproduce the above copyright | |
30 | * notice, this list of conditions and the following disclaimer in | |
31 | * the documentation and/or other materials provided with the | |
32 | * distribution. | |
33 | * - Neither the name of Intel Corporation nor the names of its | |
34 | * contributors may be used to endorse or promote products derived | |
35 | * from this software without specific prior written permission. | |
36 | * | |
37 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
38 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
39 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
40 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
41 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
42 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
43 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
44 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
45 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
46 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
47 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
48 | * | |
49 | */ | |
50 | ||
51 | #include <linux/err.h> | |
52 | #include <linux/vmalloc.h> | |
53 | #include <linux/hash.h> | |
54 | #include <linux/module.h> | |
55 | #include <linux/random.h> | |
56 | #include <linux/seq_file.h> | |
57 | ||
58 | #include "hfi.h" | |
59 | #include "qp.h" | |
60 | #include "trace.h" | |
61 | #include "sdma.h" | |
62 | ||
63 | #define BITS_PER_PAGE (PAGE_SIZE*BITS_PER_BYTE) | |
64 | #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1) | |
65 | ||
66 | static unsigned int hfi1_qp_table_size = 256; | |
67 | module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO); | |
68 | MODULE_PARM_DESC(qp_table_size, "QP table size"); | |
69 | ||
895420dd | 70 | static void flush_tx_list(struct rvt_qp *qp); |
77241056 MM |
71 | static int iowait_sleep( |
72 | struct sdma_engine *sde, | |
73 | struct iowait *wait, | |
74 | struct sdma_txreq *stx, | |
75 | unsigned seq); | |
76 | static void iowait_wakeup(struct iowait *wait, int reason); | |
77 | ||
78 | static inline unsigned mk_qpn(struct hfi1_qpn_table *qpt, | |
79 | struct qpn_map *map, unsigned off) | |
80 | { | |
81 | return (map - qpt->map) * BITS_PER_PAGE + off; | |
82 | } | |
83 | ||
84 | /* | |
85 | * Convert the AETH credit code into the number of credits. | |
86 | */ | |
87 | static const u16 credit_table[31] = { | |
88 | 0, /* 0 */ | |
89 | 1, /* 1 */ | |
90 | 2, /* 2 */ | |
91 | 3, /* 3 */ | |
92 | 4, /* 4 */ | |
93 | 6, /* 5 */ | |
94 | 8, /* 6 */ | |
95 | 12, /* 7 */ | |
96 | 16, /* 8 */ | |
97 | 24, /* 9 */ | |
98 | 32, /* A */ | |
99 | 48, /* B */ | |
100 | 64, /* C */ | |
101 | 96, /* D */ | |
102 | 128, /* E */ | |
103 | 192, /* F */ | |
104 | 256, /* 10 */ | |
105 | 384, /* 11 */ | |
106 | 512, /* 12 */ | |
107 | 768, /* 13 */ | |
108 | 1024, /* 14 */ | |
109 | 1536, /* 15 */ | |
110 | 2048, /* 16 */ | |
111 | 3072, /* 17 */ | |
112 | 4096, /* 18 */ | |
113 | 6144, /* 19 */ | |
114 | 8192, /* 1A */ | |
115 | 12288, /* 1B */ | |
116 | 16384, /* 1C */ | |
117 | 24576, /* 1D */ | |
118 | 32768 /* 1E */ | |
119 | }; | |
120 | ||
121 | static void get_map_page(struct hfi1_qpn_table *qpt, struct qpn_map *map) | |
122 | { | |
123 | unsigned long page = get_zeroed_page(GFP_KERNEL); | |
124 | ||
125 | /* | |
126 | * Free the page if someone raced with us installing it. | |
127 | */ | |
128 | ||
129 | spin_lock(&qpt->lock); | |
130 | if (map->page) | |
131 | free_page(page); | |
132 | else | |
133 | map->page = (void *)page; | |
134 | spin_unlock(&qpt->lock); | |
135 | } | |
136 | ||
137 | /* | |
138 | * Allocate the next available QPN or | |
139 | * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI. | |
140 | */ | |
141 | static int alloc_qpn(struct hfi1_devdata *dd, struct hfi1_qpn_table *qpt, | |
142 | enum ib_qp_type type, u8 port) | |
143 | { | |
144 | u32 i, offset, max_scan, qpn; | |
145 | struct qpn_map *map; | |
146 | u32 ret; | |
147 | ||
148 | if (type == IB_QPT_SMI || type == IB_QPT_GSI) { | |
149 | unsigned n; | |
150 | ||
151 | ret = type == IB_QPT_GSI; | |
152 | n = 1 << (ret + 2 * (port - 1)); | |
153 | spin_lock(&qpt->lock); | |
154 | if (qpt->flags & n) | |
155 | ret = -EINVAL; | |
156 | else | |
157 | qpt->flags |= n; | |
158 | spin_unlock(&qpt->lock); | |
159 | goto bail; | |
160 | } | |
161 | ||
162 | qpn = qpt->last + qpt->incr; | |
163 | if (qpn >= QPN_MAX) | |
164 | qpn = qpt->incr | ((qpt->last & 1) ^ 1); | |
165 | /* offset carries bit 0 */ | |
166 | offset = qpn & BITS_PER_PAGE_MASK; | |
167 | map = &qpt->map[qpn / BITS_PER_PAGE]; | |
168 | max_scan = qpt->nmaps - !offset; | |
169 | for (i = 0;;) { | |
170 | if (unlikely(!map->page)) { | |
171 | get_map_page(qpt, map); | |
172 | if (unlikely(!map->page)) | |
173 | break; | |
174 | } | |
175 | do { | |
176 | if (!test_and_set_bit(offset, map->page)) { | |
177 | qpt->last = qpn; | |
178 | ret = qpn; | |
179 | goto bail; | |
180 | } | |
181 | offset += qpt->incr; | |
182 | /* | |
183 | * This qpn might be bogus if offset >= BITS_PER_PAGE. | |
184 | * That is OK. It gets re-assigned below | |
185 | */ | |
186 | qpn = mk_qpn(qpt, map, offset); | |
187 | } while (offset < BITS_PER_PAGE && qpn < QPN_MAX); | |
188 | /* | |
189 | * In order to keep the number of pages allocated to a | |
190 | * minimum, we scan the all existing pages before increasing | |
191 | * the size of the bitmap table. | |
192 | */ | |
193 | if (++i > max_scan) { | |
194 | if (qpt->nmaps == QPNMAP_ENTRIES) | |
195 | break; | |
196 | map = &qpt->map[qpt->nmaps++]; | |
197 | /* start at incr with current bit 0 */ | |
198 | offset = qpt->incr | (offset & 1); | |
199 | } else if (map < &qpt->map[qpt->nmaps]) { | |
200 | ++map; | |
201 | /* start at incr with current bit 0 */ | |
202 | offset = qpt->incr | (offset & 1); | |
203 | } else { | |
204 | map = &qpt->map[0]; | |
205 | /* wrap to first map page, invert bit 0 */ | |
206 | offset = qpt->incr | ((offset & 1) ^ 1); | |
207 | } | |
208 | /* there can be no bits at shift and below */ | |
209 | WARN_ON(offset & (dd->qos_shift - 1)); | |
210 | qpn = mk_qpn(qpt, map, offset); | |
211 | } | |
212 | ||
213 | ret = -ENOMEM; | |
214 | ||
215 | bail: | |
216 | return ret; | |
217 | } | |
218 | ||
219 | static void free_qpn(struct hfi1_qpn_table *qpt, u32 qpn) | |
220 | { | |
221 | struct qpn_map *map; | |
222 | ||
223 | map = qpt->map + qpn / BITS_PER_PAGE; | |
224 | if (map->page) | |
225 | clear_bit(qpn & BITS_PER_PAGE_MASK, map->page); | |
226 | } | |
227 | ||
228 | /* | |
229 | * Put the QP into the hash table. | |
230 | * The hash table holds a reference to the QP. | |
231 | */ | |
895420dd | 232 | static void insert_qp(struct hfi1_ibdev *dev, struct rvt_qp *qp) |
77241056 MM |
233 | { |
234 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
235 | unsigned long flags; | |
236 | ||
237 | atomic_inc(&qp->refcount); | |
238 | spin_lock_irqsave(&dev->qp_dev->qpt_lock, flags); | |
239 | ||
240 | if (qp->ibqp.qp_num <= 1) { | |
241 | rcu_assign_pointer(ibp->qp[qp->ibqp.qp_num], qp); | |
242 | } else { | |
243 | u32 n = qpn_hash(dev->qp_dev, qp->ibqp.qp_num); | |
244 | ||
245 | qp->next = dev->qp_dev->qp_table[n]; | |
246 | rcu_assign_pointer(dev->qp_dev->qp_table[n], qp); | |
247 | trace_hfi1_qpinsert(qp, n); | |
248 | } | |
249 | ||
250 | spin_unlock_irqrestore(&dev->qp_dev->qpt_lock, flags); | |
251 | } | |
252 | ||
253 | /* | |
254 | * Remove the QP from the table so it can't be found asynchronously by | |
255 | * the receive interrupt routine. | |
256 | */ | |
895420dd | 257 | static void remove_qp(struct hfi1_ibdev *dev, struct rvt_qp *qp) |
77241056 MM |
258 | { |
259 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
260 | u32 n = qpn_hash(dev->qp_dev, qp->ibqp.qp_num); | |
261 | unsigned long flags; | |
262 | int removed = 1; | |
263 | ||
264 | spin_lock_irqsave(&dev->qp_dev->qpt_lock, flags); | |
265 | ||
266 | if (rcu_dereference_protected(ibp->qp[0], | |
267 | lockdep_is_held(&dev->qp_dev->qpt_lock)) == qp) { | |
268 | RCU_INIT_POINTER(ibp->qp[0], NULL); | |
269 | } else if (rcu_dereference_protected(ibp->qp[1], | |
270 | lockdep_is_held(&dev->qp_dev->qpt_lock)) == qp) { | |
271 | RCU_INIT_POINTER(ibp->qp[1], NULL); | |
272 | } else { | |
895420dd DD |
273 | struct rvt_qp *q; |
274 | struct rvt_qp __rcu **qpp; | |
77241056 MM |
275 | |
276 | removed = 0; | |
277 | qpp = &dev->qp_dev->qp_table[n]; | |
278 | for (; (q = rcu_dereference_protected(*qpp, | |
279 | lockdep_is_held(&dev->qp_dev->qpt_lock))) | |
280 | != NULL; | |
281 | qpp = &q->next) | |
282 | if (q == qp) { | |
283 | RCU_INIT_POINTER(*qpp, | |
284 | rcu_dereference_protected(qp->next, | |
285 | lockdep_is_held(&dev->qp_dev->qpt_lock))); | |
286 | removed = 1; | |
287 | trace_hfi1_qpremove(qp, n); | |
288 | break; | |
289 | } | |
290 | } | |
291 | ||
292 | spin_unlock_irqrestore(&dev->qp_dev->qpt_lock, flags); | |
293 | if (removed) { | |
294 | synchronize_rcu(); | |
295 | if (atomic_dec_and_test(&qp->refcount)) | |
296 | wake_up(&qp->wait); | |
297 | } | |
298 | } | |
299 | ||
300 | /** | |
301 | * free_all_qps - check for QPs still in use | |
302 | * @qpt: the QP table to empty | |
303 | * | |
304 | * There should not be any QPs still in use. | |
305 | * Free memory for table. | |
306 | */ | |
307 | static unsigned free_all_qps(struct hfi1_devdata *dd) | |
308 | { | |
309 | struct hfi1_ibdev *dev = &dd->verbs_dev; | |
310 | unsigned long flags; | |
895420dd | 311 | struct rvt_qp *qp; |
77241056 MM |
312 | unsigned n, qp_inuse = 0; |
313 | ||
314 | for (n = 0; n < dd->num_pports; n++) { | |
315 | struct hfi1_ibport *ibp = &dd->pport[n].ibport_data; | |
316 | ||
317 | if (!hfi1_mcast_tree_empty(ibp)) | |
318 | qp_inuse++; | |
319 | rcu_read_lock(); | |
320 | if (rcu_dereference(ibp->qp[0])) | |
321 | qp_inuse++; | |
322 | if (rcu_dereference(ibp->qp[1])) | |
323 | qp_inuse++; | |
324 | rcu_read_unlock(); | |
325 | } | |
326 | ||
327 | if (!dev->qp_dev) | |
328 | goto bail; | |
329 | spin_lock_irqsave(&dev->qp_dev->qpt_lock, flags); | |
330 | for (n = 0; n < dev->qp_dev->qp_table_size; n++) { | |
331 | qp = rcu_dereference_protected(dev->qp_dev->qp_table[n], | |
332 | lockdep_is_held(&dev->qp_dev->qpt_lock)); | |
333 | RCU_INIT_POINTER(dev->qp_dev->qp_table[n], NULL); | |
334 | ||
335 | for (; qp; qp = rcu_dereference_protected(qp->next, | |
336 | lockdep_is_held(&dev->qp_dev->qpt_lock))) | |
337 | qp_inuse++; | |
338 | } | |
339 | spin_unlock_irqrestore(&dev->qp_dev->qpt_lock, flags); | |
340 | synchronize_rcu(); | |
341 | bail: | |
342 | return qp_inuse; | |
343 | } | |
344 | ||
345 | /** | |
346 | * reset_qp - initialize the QP state to the reset state | |
347 | * @qp: the QP to reset | |
348 | * @type: the QP type | |
349 | */ | |
895420dd | 350 | static void reset_qp(struct rvt_qp *qp, enum ib_qp_type type) |
77241056 | 351 | { |
4c6829c5 | 352 | struct hfi1_qp_priv *priv = qp->priv; |
77241056 MM |
353 | qp->remote_qpn = 0; |
354 | qp->qkey = 0; | |
355 | qp->qp_access_flags = 0; | |
356 | iowait_init( | |
4c6829c5 | 357 | &priv->s_iowait, |
77241056 MM |
358 | 1, |
359 | hfi1_do_send, | |
360 | iowait_sleep, | |
361 | iowait_wakeup); | |
362 | qp->s_flags &= HFI1_S_SIGNAL_REQ_WR; | |
363 | qp->s_hdrwords = 0; | |
364 | qp->s_wqe = NULL; | |
365 | qp->s_draining = 0; | |
366 | qp->s_next_psn = 0; | |
367 | qp->s_last_psn = 0; | |
368 | qp->s_sending_psn = 0; | |
369 | qp->s_sending_hpsn = 0; | |
370 | qp->s_psn = 0; | |
371 | qp->r_psn = 0; | |
372 | qp->r_msn = 0; | |
373 | if (type == IB_QPT_RC) { | |
374 | qp->s_state = IB_OPCODE_RC_SEND_LAST; | |
375 | qp->r_state = IB_OPCODE_RC_SEND_LAST; | |
376 | } else { | |
377 | qp->s_state = IB_OPCODE_UC_SEND_LAST; | |
378 | qp->r_state = IB_OPCODE_UC_SEND_LAST; | |
379 | } | |
380 | qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; | |
381 | qp->r_nak_state = 0; | |
4c6829c5 | 382 | priv->r_adefered = 0; |
77241056 MM |
383 | qp->r_aflags = 0; |
384 | qp->r_flags = 0; | |
385 | qp->s_head = 0; | |
386 | qp->s_tail = 0; | |
387 | qp->s_cur = 0; | |
388 | qp->s_acked = 0; | |
389 | qp->s_last = 0; | |
390 | qp->s_ssn = 1; | |
391 | qp->s_lsn = 0; | |
392 | clear_ahg(qp); | |
393 | qp->s_mig_state = IB_MIG_MIGRATED; | |
394 | memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); | |
395 | qp->r_head_ack_queue = 0; | |
396 | qp->s_tail_ack_queue = 0; | |
397 | qp->s_num_rd_atomic = 0; | |
398 | if (qp->r_rq.wq) { | |
399 | qp->r_rq.wq->head = 0; | |
400 | qp->r_rq.wq->tail = 0; | |
401 | } | |
402 | qp->r_sge.num_sge = 0; | |
403 | } | |
404 | ||
895420dd | 405 | static void clear_mr_refs(struct rvt_qp *qp, int clr_sends) |
77241056 MM |
406 | { |
407 | unsigned n; | |
408 | ||
409 | if (test_and_clear_bit(HFI1_R_REWIND_SGE, &qp->r_aflags)) | |
410 | hfi1_put_ss(&qp->s_rdma_read_sge); | |
411 | ||
412 | hfi1_put_ss(&qp->r_sge); | |
413 | ||
414 | if (clr_sends) { | |
415 | while (qp->s_last != qp->s_head) { | |
895420dd | 416 | struct rvt_swqe *wqe = get_swqe_ptr(qp, qp->s_last); |
77241056 MM |
417 | unsigned i; |
418 | ||
419 | for (i = 0; i < wqe->wr.num_sge; i++) { | |
895420dd | 420 | struct rvt_sge *sge = &wqe->sg_list[i]; |
77241056 | 421 | |
895420dd | 422 | rvt_put_mr(sge->mr); |
77241056 MM |
423 | } |
424 | if (qp->ibqp.qp_type == IB_QPT_UD || | |
425 | qp->ibqp.qp_type == IB_QPT_SMI || | |
426 | qp->ibqp.qp_type == IB_QPT_GSI) | |
15723f06 | 427 | atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount); |
77241056 MM |
428 | if (++qp->s_last >= qp->s_size) |
429 | qp->s_last = 0; | |
430 | } | |
431 | if (qp->s_rdma_mr) { | |
895420dd | 432 | rvt_put_mr(qp->s_rdma_mr); |
77241056 MM |
433 | qp->s_rdma_mr = NULL; |
434 | } | |
435 | } | |
436 | ||
437 | if (qp->ibqp.qp_type != IB_QPT_RC) | |
438 | return; | |
439 | ||
440 | for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) { | |
895420dd | 441 | struct rvt_ack_entry *e = &qp->s_ack_queue[n]; |
77241056 MM |
442 | |
443 | if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST && | |
444 | e->rdma_sge.mr) { | |
895420dd | 445 | rvt_put_mr(e->rdma_sge.mr); |
77241056 MM |
446 | e->rdma_sge.mr = NULL; |
447 | } | |
448 | } | |
449 | } | |
450 | ||
451 | /** | |
452 | * hfi1_error_qp - put a QP into the error state | |
453 | * @qp: the QP to put into the error state | |
454 | * @err: the receive completion error to signal if a RWQE is active | |
455 | * | |
456 | * Flushes both send and receive work queues. | |
457 | * Returns true if last WQE event should be generated. | |
458 | * The QP r_lock and s_lock should be held and interrupts disabled. | |
459 | * If we are already in error state, just return. | |
460 | */ | |
895420dd | 461 | int hfi1_error_qp(struct rvt_qp *qp, enum ib_wc_status err) |
77241056 MM |
462 | { |
463 | struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); | |
4c6829c5 | 464 | struct hfi1_qp_priv *priv = qp->priv; |
77241056 MM |
465 | struct ib_wc wc; |
466 | int ret = 0; | |
467 | ||
468 | if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) | |
469 | goto bail; | |
470 | ||
471 | qp->state = IB_QPS_ERR; | |
472 | ||
473 | if (qp->s_flags & (HFI1_S_TIMER | HFI1_S_WAIT_RNR)) { | |
474 | qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_WAIT_RNR); | |
475 | del_timer(&qp->s_timer); | |
476 | } | |
477 | ||
478 | if (qp->s_flags & HFI1_S_ANY_WAIT_SEND) | |
479 | qp->s_flags &= ~HFI1_S_ANY_WAIT_SEND; | |
480 | ||
481 | write_seqlock(&dev->iowait_lock); | |
4c6829c5 | 482 | if (!list_empty(&priv->s_iowait.list) && !(qp->s_flags & HFI1_S_BUSY)) { |
77241056 | 483 | qp->s_flags &= ~HFI1_S_ANY_WAIT_IO; |
4c6829c5 | 484 | list_del_init(&priv->s_iowait.list); |
77241056 MM |
485 | if (atomic_dec_and_test(&qp->refcount)) |
486 | wake_up(&qp->wait); | |
487 | } | |
488 | write_sequnlock(&dev->iowait_lock); | |
489 | ||
490 | if (!(qp->s_flags & HFI1_S_BUSY)) { | |
491 | qp->s_hdrwords = 0; | |
492 | if (qp->s_rdma_mr) { | |
895420dd | 493 | rvt_put_mr(qp->s_rdma_mr); |
77241056 MM |
494 | qp->s_rdma_mr = NULL; |
495 | } | |
496 | flush_tx_list(qp); | |
497 | } | |
498 | ||
499 | /* Schedule the sending tasklet to drain the send work queue. */ | |
500 | if (qp->s_last != qp->s_head) | |
501 | hfi1_schedule_send(qp); | |
502 | ||
503 | clear_mr_refs(qp, 0); | |
504 | ||
505 | memset(&wc, 0, sizeof(wc)); | |
506 | wc.qp = &qp->ibqp; | |
507 | wc.opcode = IB_WC_RECV; | |
508 | ||
509 | if (test_and_clear_bit(HFI1_R_WRID_VALID, &qp->r_aflags)) { | |
510 | wc.wr_id = qp->r_wr_id; | |
511 | wc.status = err; | |
512 | hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); | |
513 | } | |
514 | wc.status = IB_WC_WR_FLUSH_ERR; | |
515 | ||
516 | if (qp->r_rq.wq) { | |
895420dd | 517 | struct rvt_rwq *wq; |
77241056 MM |
518 | u32 head; |
519 | u32 tail; | |
520 | ||
521 | spin_lock(&qp->r_rq.lock); | |
522 | ||
523 | /* sanity check pointers before trusting them */ | |
524 | wq = qp->r_rq.wq; | |
525 | head = wq->head; | |
526 | if (head >= qp->r_rq.size) | |
527 | head = 0; | |
528 | tail = wq->tail; | |
529 | if (tail >= qp->r_rq.size) | |
530 | tail = 0; | |
531 | while (tail != head) { | |
532 | wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id; | |
533 | if (++tail >= qp->r_rq.size) | |
534 | tail = 0; | |
535 | hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1); | |
536 | } | |
537 | wq->tail = tail; | |
538 | ||
539 | spin_unlock(&qp->r_rq.lock); | |
540 | } else if (qp->ibqp.event_handler) | |
541 | ret = 1; | |
542 | ||
543 | bail: | |
544 | return ret; | |
545 | } | |
546 | ||
895420dd | 547 | static void flush_tx_list(struct rvt_qp *qp) |
77241056 | 548 | { |
4c6829c5 DD |
549 | struct hfi1_qp_priv *priv = qp->priv; |
550 | ||
551 | while (!list_empty(&priv->s_iowait.tx_head)) { | |
77241056 MM |
552 | struct sdma_txreq *tx; |
553 | ||
554 | tx = list_first_entry( | |
4c6829c5 | 555 | &priv->s_iowait.tx_head, |
77241056 MM |
556 | struct sdma_txreq, |
557 | list); | |
558 | list_del_init(&tx->list); | |
559 | hfi1_put_txreq( | |
560 | container_of(tx, struct verbs_txreq, txreq)); | |
561 | } | |
562 | } | |
563 | ||
895420dd | 564 | static void flush_iowait(struct rvt_qp *qp) |
77241056 | 565 | { |
4c6829c5 | 566 | struct hfi1_qp_priv *priv = qp->priv; |
77241056 MM |
567 | struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); |
568 | unsigned long flags; | |
569 | ||
570 | write_seqlock_irqsave(&dev->iowait_lock, flags); | |
4c6829c5 DD |
571 | if (!list_empty(&priv->s_iowait.list)) { |
572 | list_del_init(&priv->s_iowait.list); | |
77241056 MM |
573 | if (atomic_dec_and_test(&qp->refcount)) |
574 | wake_up(&qp->wait); | |
575 | } | |
576 | write_sequnlock_irqrestore(&dev->iowait_lock, flags); | |
577 | } | |
578 | ||
579 | static inline int opa_mtu_enum_to_int(int mtu) | |
580 | { | |
581 | switch (mtu) { | |
582 | case OPA_MTU_8192: return 8192; | |
583 | case OPA_MTU_10240: return 10240; | |
584 | default: return -1; | |
585 | } | |
586 | } | |
587 | ||
588 | /** | |
589 | * This function is what we would push to the core layer if we wanted to be a | |
590 | * "first class citizen". Instead we hide this here and rely on Verbs ULPs | |
591 | * to blindly pass the MTU enum value from the PathRecord to us. | |
592 | * | |
593 | * The actual flag used to determine "8k MTU" will change and is currently | |
594 | * unknown. | |
595 | */ | |
596 | static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu) | |
597 | { | |
598 | int val = opa_mtu_enum_to_int((int)mtu); | |
599 | ||
600 | if (val > 0) | |
601 | return val; | |
602 | return ib_mtu_enum_to_int(mtu); | |
603 | } | |
604 | ||
605 | ||
606 | /** | |
607 | * hfi1_modify_qp - modify the attributes of a queue pair | |
608 | * @ibqp: the queue pair who's attributes we're modifying | |
609 | * @attr: the new attributes | |
610 | * @attr_mask: the mask of attributes to modify | |
611 | * @udata: user data for libibverbs.so | |
612 | * | |
613 | * Returns 0 on success, otherwise returns an errno. | |
614 | */ | |
615 | int hfi1_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, | |
616 | int attr_mask, struct ib_udata *udata) | |
617 | { | |
618 | struct hfi1_ibdev *dev = to_idev(ibqp->device); | |
895420dd | 619 | struct rvt_qp *qp = to_iqp(ibqp); |
4c6829c5 | 620 | struct hfi1_qp_priv *priv = qp->priv; |
77241056 MM |
621 | enum ib_qp_state cur_state, new_state; |
622 | struct ib_event ev; | |
623 | int lastwqe = 0; | |
624 | int mig = 0; | |
625 | int ret; | |
626 | u32 pmtu = 0; /* for gcc warning only */ | |
d7b8ba51 | 627 | struct hfi1_devdata *dd = dd_from_dev(dev); |
77241056 MM |
628 | |
629 | spin_lock_irq(&qp->r_lock); | |
630 | spin_lock(&qp->s_lock); | |
631 | ||
632 | cur_state = attr_mask & IB_QP_CUR_STATE ? | |
633 | attr->cur_qp_state : qp->state; | |
634 | new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; | |
635 | ||
636 | if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, | |
637 | attr_mask, IB_LINK_LAYER_UNSPECIFIED)) | |
638 | goto inval; | |
639 | ||
640 | if (attr_mask & IB_QP_AV) { | |
d7b8ba51 MM |
641 | u8 sc; |
642 | ||
8859b4a6 | 643 | if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) |
77241056 | 644 | goto inval; |
15723f06 | 645 | if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr)) |
77241056 | 646 | goto inval; |
d7b8ba51 MM |
647 | sc = ah_to_sc(ibqp->device, &attr->ah_attr); |
648 | if (!qp_to_sdma_engine(qp, sc) && | |
649 | dd->flags & HFI1_HAS_SEND_DMA) | |
650 | goto inval; | |
77241056 MM |
651 | } |
652 | ||
653 | if (attr_mask & IB_QP_ALT_PATH) { | |
d7b8ba51 MM |
654 | u8 sc; |
655 | ||
8859b4a6 DD |
656 | if (attr->alt_ah_attr.dlid >= |
657 | be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
77241056 | 658 | goto inval; |
15723f06 | 659 | if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) |
77241056 | 660 | goto inval; |
d7b8ba51 MM |
661 | if (attr->alt_pkey_index >= hfi1_get_npkeys(dd)) |
662 | goto inval; | |
663 | sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr); | |
664 | if (!qp_to_sdma_engine(qp, sc) && | |
665 | dd->flags & HFI1_HAS_SEND_DMA) | |
77241056 MM |
666 | goto inval; |
667 | } | |
668 | ||
669 | if (attr_mask & IB_QP_PKEY_INDEX) | |
d7b8ba51 | 670 | if (attr->pkey_index >= hfi1_get_npkeys(dd)) |
77241056 MM |
671 | goto inval; |
672 | ||
673 | if (attr_mask & IB_QP_MIN_RNR_TIMER) | |
674 | if (attr->min_rnr_timer > 31) | |
675 | goto inval; | |
676 | ||
677 | if (attr_mask & IB_QP_PORT) | |
678 | if (qp->ibqp.qp_type == IB_QPT_SMI || | |
679 | qp->ibqp.qp_type == IB_QPT_GSI || | |
680 | attr->port_num == 0 || | |
681 | attr->port_num > ibqp->device->phys_port_cnt) | |
682 | goto inval; | |
683 | ||
684 | if (attr_mask & IB_QP_DEST_QPN) | |
685 | if (attr->dest_qp_num > HFI1_QPN_MASK) | |
686 | goto inval; | |
687 | ||
688 | if (attr_mask & IB_QP_RETRY_CNT) | |
689 | if (attr->retry_cnt > 7) | |
690 | goto inval; | |
691 | ||
692 | if (attr_mask & IB_QP_RNR_RETRY) | |
693 | if (attr->rnr_retry > 7) | |
694 | goto inval; | |
695 | ||
696 | /* | |
697 | * Don't allow invalid path_mtu values. OK to set greater | |
698 | * than the active mtu (or even the max_cap, if we have tuned | |
699 | * that to a small mtu. We'll set qp->path_mtu | |
700 | * to the lesser of requested attribute mtu and active, | |
701 | * for packetizing messages. | |
702 | * Note that the QP port has to be set in INIT and MTU in RTR. | |
703 | */ | |
704 | if (attr_mask & IB_QP_PATH_MTU) { | |
705 | int mtu, pidx = qp->port_num - 1; | |
706 | ||
707 | dd = dd_from_dev(dev); | |
708 | mtu = verbs_mtu_enum_to_int(ibqp->device, attr->path_mtu); | |
709 | if (mtu == -1) | |
710 | goto inval; | |
711 | ||
712 | if (mtu > dd->pport[pidx].ibmtu) | |
713 | pmtu = mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048); | |
714 | else | |
715 | pmtu = attr->path_mtu; | |
716 | } | |
717 | ||
718 | if (attr_mask & IB_QP_PATH_MIG_STATE) { | |
719 | if (attr->path_mig_state == IB_MIG_REARM) { | |
720 | if (qp->s_mig_state == IB_MIG_ARMED) | |
721 | goto inval; | |
722 | if (new_state != IB_QPS_RTS) | |
723 | goto inval; | |
724 | } else if (attr->path_mig_state == IB_MIG_MIGRATED) { | |
725 | if (qp->s_mig_state == IB_MIG_REARM) | |
726 | goto inval; | |
727 | if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) | |
728 | goto inval; | |
729 | if (qp->s_mig_state == IB_MIG_ARMED) | |
730 | mig = 1; | |
731 | } else | |
732 | goto inval; | |
733 | } | |
734 | ||
735 | if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) | |
736 | if (attr->max_dest_rd_atomic > HFI1_MAX_RDMA_ATOMIC) | |
737 | goto inval; | |
738 | ||
739 | switch (new_state) { | |
740 | case IB_QPS_RESET: | |
741 | if (qp->state != IB_QPS_RESET) { | |
742 | qp->state = IB_QPS_RESET; | |
743 | flush_iowait(qp); | |
744 | qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_ANY_WAIT); | |
745 | spin_unlock(&qp->s_lock); | |
746 | spin_unlock_irq(&qp->r_lock); | |
747 | /* Stop the sending work queue and retry timer */ | |
4c6829c5 | 748 | cancel_work_sync(&priv->s_iowait.iowork); |
77241056 | 749 | del_timer_sync(&qp->s_timer); |
4c6829c5 | 750 | iowait_sdma_drain(&priv->s_iowait); |
77241056 MM |
751 | flush_tx_list(qp); |
752 | remove_qp(dev, qp); | |
753 | wait_event(qp->wait, !atomic_read(&qp->refcount)); | |
754 | spin_lock_irq(&qp->r_lock); | |
755 | spin_lock(&qp->s_lock); | |
756 | clear_mr_refs(qp, 1); | |
757 | clear_ahg(qp); | |
758 | reset_qp(qp, ibqp->qp_type); | |
759 | } | |
760 | break; | |
761 | ||
762 | case IB_QPS_RTR: | |
763 | /* Allow event to re-trigger if QP set to RTR more than once */ | |
764 | qp->r_flags &= ~HFI1_R_COMM_EST; | |
765 | qp->state = new_state; | |
766 | break; | |
767 | ||
768 | case IB_QPS_SQD: | |
769 | qp->s_draining = qp->s_last != qp->s_cur; | |
770 | qp->state = new_state; | |
771 | break; | |
772 | ||
773 | case IB_QPS_SQE: | |
774 | if (qp->ibqp.qp_type == IB_QPT_RC) | |
775 | goto inval; | |
776 | qp->state = new_state; | |
777 | break; | |
778 | ||
779 | case IB_QPS_ERR: | |
780 | lastwqe = hfi1_error_qp(qp, IB_WC_WR_FLUSH_ERR); | |
781 | break; | |
782 | ||
783 | default: | |
784 | qp->state = new_state; | |
785 | break; | |
786 | } | |
787 | ||
788 | if (attr_mask & IB_QP_PKEY_INDEX) | |
789 | qp->s_pkey_index = attr->pkey_index; | |
790 | ||
791 | if (attr_mask & IB_QP_PORT) | |
792 | qp->port_num = attr->port_num; | |
793 | ||
794 | if (attr_mask & IB_QP_DEST_QPN) | |
795 | qp->remote_qpn = attr->dest_qp_num; | |
796 | ||
797 | if (attr_mask & IB_QP_SQ_PSN) { | |
798 | qp->s_next_psn = attr->sq_psn & PSN_MODIFY_MASK; | |
799 | qp->s_psn = qp->s_next_psn; | |
800 | qp->s_sending_psn = qp->s_next_psn; | |
801 | qp->s_last_psn = qp->s_next_psn - 1; | |
802 | qp->s_sending_hpsn = qp->s_last_psn; | |
803 | } | |
804 | ||
805 | if (attr_mask & IB_QP_RQ_PSN) | |
806 | qp->r_psn = attr->rq_psn & PSN_MODIFY_MASK; | |
807 | ||
808 | if (attr_mask & IB_QP_ACCESS_FLAGS) | |
809 | qp->qp_access_flags = attr->qp_access_flags; | |
810 | ||
811 | if (attr_mask & IB_QP_AV) { | |
812 | qp->remote_ah_attr = attr->ah_attr; | |
813 | qp->s_srate = attr->ah_attr.static_rate; | |
814 | qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); | |
4c6829c5 DD |
815 | priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr); |
816 | priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc); | |
77241056 MM |
817 | } |
818 | ||
819 | if (attr_mask & IB_QP_ALT_PATH) { | |
820 | qp->alt_ah_attr = attr->alt_ah_attr; | |
821 | qp->s_alt_pkey_index = attr->alt_pkey_index; | |
822 | } | |
823 | ||
824 | if (attr_mask & IB_QP_PATH_MIG_STATE) { | |
825 | qp->s_mig_state = attr->path_mig_state; | |
826 | if (mig) { | |
827 | qp->remote_ah_attr = qp->alt_ah_attr; | |
828 | qp->port_num = qp->alt_ah_attr.port_num; | |
829 | qp->s_pkey_index = qp->s_alt_pkey_index; | |
830 | qp->s_flags |= HFI1_S_AHG_CLEAR; | |
4c6829c5 DD |
831 | priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr); |
832 | priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc); | |
77241056 MM |
833 | } |
834 | } | |
835 | ||
836 | if (attr_mask & IB_QP_PATH_MTU) { | |
837 | struct hfi1_ibport *ibp; | |
838 | u8 sc, vl; | |
839 | u32 mtu; | |
840 | ||
841 | dd = dd_from_dev(dev); | |
842 | ibp = &dd->pport[qp->port_num - 1].ibport_data; | |
843 | ||
844 | sc = ibp->sl_to_sc[qp->remote_ah_attr.sl]; | |
845 | vl = sc_to_vlt(dd, sc); | |
846 | ||
847 | mtu = verbs_mtu_enum_to_int(ibqp->device, pmtu); | |
848 | if (vl < PER_VL_SEND_CONTEXTS) | |
849 | mtu = min_t(u32, mtu, dd->vld[vl].mtu); | |
850 | pmtu = mtu_to_enum(mtu, OPA_MTU_8192); | |
851 | ||
852 | qp->path_mtu = pmtu; | |
853 | qp->pmtu = mtu; | |
854 | } | |
855 | ||
856 | if (attr_mask & IB_QP_RETRY_CNT) { | |
857 | qp->s_retry_cnt = attr->retry_cnt; | |
858 | qp->s_retry = attr->retry_cnt; | |
859 | } | |
860 | ||
861 | if (attr_mask & IB_QP_RNR_RETRY) { | |
862 | qp->s_rnr_retry_cnt = attr->rnr_retry; | |
863 | qp->s_rnr_retry = attr->rnr_retry; | |
864 | } | |
865 | ||
866 | if (attr_mask & IB_QP_MIN_RNR_TIMER) | |
867 | qp->r_min_rnr_timer = attr->min_rnr_timer; | |
868 | ||
869 | if (attr_mask & IB_QP_TIMEOUT) { | |
870 | qp->timeout = attr->timeout; | |
871 | qp->timeout_jiffies = | |
872 | usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / | |
873 | 1000UL); | |
874 | } | |
875 | ||
876 | if (attr_mask & IB_QP_QKEY) | |
877 | qp->qkey = attr->qkey; | |
878 | ||
879 | if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) | |
880 | qp->r_max_rd_atomic = attr->max_dest_rd_atomic; | |
881 | ||
882 | if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) | |
883 | qp->s_max_rd_atomic = attr->max_rd_atomic; | |
884 | ||
885 | spin_unlock(&qp->s_lock); | |
886 | spin_unlock_irq(&qp->r_lock); | |
887 | ||
888 | if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) | |
889 | insert_qp(dev, qp); | |
890 | ||
891 | if (lastwqe) { | |
892 | ev.device = qp->ibqp.device; | |
893 | ev.element.qp = &qp->ibqp; | |
894 | ev.event = IB_EVENT_QP_LAST_WQE_REACHED; | |
895 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
896 | } | |
897 | if (mig) { | |
898 | ev.device = qp->ibqp.device; | |
899 | ev.element.qp = &qp->ibqp; | |
900 | ev.event = IB_EVENT_PATH_MIG; | |
901 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
902 | } | |
903 | ret = 0; | |
904 | goto bail; | |
905 | ||
906 | inval: | |
907 | spin_unlock(&qp->s_lock); | |
908 | spin_unlock_irq(&qp->r_lock); | |
909 | ret = -EINVAL; | |
910 | ||
911 | bail: | |
912 | return ret; | |
913 | } | |
914 | ||
915 | int hfi1_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, | |
916 | int attr_mask, struct ib_qp_init_attr *init_attr) | |
917 | { | |
895420dd | 918 | struct rvt_qp *qp = to_iqp(ibqp); |
77241056 MM |
919 | |
920 | attr->qp_state = qp->state; | |
921 | attr->cur_qp_state = attr->qp_state; | |
922 | attr->path_mtu = qp->path_mtu; | |
923 | attr->path_mig_state = qp->s_mig_state; | |
924 | attr->qkey = qp->qkey; | |
925 | attr->rq_psn = mask_psn(qp->r_psn); | |
926 | attr->sq_psn = mask_psn(qp->s_next_psn); | |
927 | attr->dest_qp_num = qp->remote_qpn; | |
928 | attr->qp_access_flags = qp->qp_access_flags; | |
929 | attr->cap.max_send_wr = qp->s_size - 1; | |
930 | attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1; | |
931 | attr->cap.max_send_sge = qp->s_max_sge; | |
932 | attr->cap.max_recv_sge = qp->r_rq.max_sge; | |
933 | attr->cap.max_inline_data = 0; | |
934 | attr->ah_attr = qp->remote_ah_attr; | |
935 | attr->alt_ah_attr = qp->alt_ah_attr; | |
936 | attr->pkey_index = qp->s_pkey_index; | |
937 | attr->alt_pkey_index = qp->s_alt_pkey_index; | |
938 | attr->en_sqd_async_notify = 0; | |
939 | attr->sq_draining = qp->s_draining; | |
940 | attr->max_rd_atomic = qp->s_max_rd_atomic; | |
941 | attr->max_dest_rd_atomic = qp->r_max_rd_atomic; | |
942 | attr->min_rnr_timer = qp->r_min_rnr_timer; | |
943 | attr->port_num = qp->port_num; | |
944 | attr->timeout = qp->timeout; | |
945 | attr->retry_cnt = qp->s_retry_cnt; | |
946 | attr->rnr_retry = qp->s_rnr_retry_cnt; | |
947 | attr->alt_port_num = qp->alt_ah_attr.port_num; | |
948 | attr->alt_timeout = qp->alt_timeout; | |
949 | ||
950 | init_attr->event_handler = qp->ibqp.event_handler; | |
951 | init_attr->qp_context = qp->ibqp.qp_context; | |
952 | init_attr->send_cq = qp->ibqp.send_cq; | |
953 | init_attr->recv_cq = qp->ibqp.recv_cq; | |
954 | init_attr->srq = qp->ibqp.srq; | |
955 | init_attr->cap = attr->cap; | |
956 | if (qp->s_flags & HFI1_S_SIGNAL_REQ_WR) | |
957 | init_attr->sq_sig_type = IB_SIGNAL_REQ_WR; | |
958 | else | |
959 | init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; | |
960 | init_attr->qp_type = qp->ibqp.qp_type; | |
961 | init_attr->port_num = qp->port_num; | |
962 | return 0; | |
963 | } | |
964 | ||
965 | /** | |
966 | * hfi1_compute_aeth - compute the AETH (syndrome + MSN) | |
967 | * @qp: the queue pair to compute the AETH for | |
968 | * | |
969 | * Returns the AETH. | |
970 | */ | |
895420dd | 971 | __be32 hfi1_compute_aeth(struct rvt_qp *qp) |
77241056 MM |
972 | { |
973 | u32 aeth = qp->r_msn & HFI1_MSN_MASK; | |
974 | ||
975 | if (qp->ibqp.srq) { | |
976 | /* | |
977 | * Shared receive queues don't generate credits. | |
978 | * Set the credit field to the invalid value. | |
979 | */ | |
980 | aeth |= HFI1_AETH_CREDIT_INVAL << HFI1_AETH_CREDIT_SHIFT; | |
981 | } else { | |
982 | u32 min, max, x; | |
983 | u32 credits; | |
895420dd | 984 | struct rvt_rwq *wq = qp->r_rq.wq; |
77241056 MM |
985 | u32 head; |
986 | u32 tail; | |
987 | ||
988 | /* sanity check pointers before trusting them */ | |
989 | head = wq->head; | |
990 | if (head >= qp->r_rq.size) | |
991 | head = 0; | |
992 | tail = wq->tail; | |
993 | if (tail >= qp->r_rq.size) | |
994 | tail = 0; | |
995 | /* | |
996 | * Compute the number of credits available (RWQEs). | |
997 | * There is a small chance that the pair of reads are | |
998 | * not atomic, which is OK, since the fuzziness is | |
999 | * resolved as further ACKs go out. | |
1000 | */ | |
1001 | credits = head - tail; | |
1002 | if ((int)credits < 0) | |
1003 | credits += qp->r_rq.size; | |
1004 | /* | |
1005 | * Binary search the credit table to find the code to | |
1006 | * use. | |
1007 | */ | |
1008 | min = 0; | |
1009 | max = 31; | |
1010 | for (;;) { | |
1011 | x = (min + max) / 2; | |
1012 | if (credit_table[x] == credits) | |
1013 | break; | |
1014 | if (credit_table[x] > credits) | |
1015 | max = x; | |
1016 | else if (min == x) | |
1017 | break; | |
1018 | else | |
1019 | min = x; | |
1020 | } | |
1021 | aeth |= x << HFI1_AETH_CREDIT_SHIFT; | |
1022 | } | |
1023 | return cpu_to_be32(aeth); | |
1024 | } | |
1025 | ||
1026 | /** | |
1027 | * hfi1_create_qp - create a queue pair for a device | |
1028 | * @ibpd: the protection domain who's device we create the queue pair for | |
1029 | * @init_attr: the attributes of the queue pair | |
1030 | * @udata: user data for libibverbs.so | |
1031 | * | |
1032 | * Returns the queue pair on success, otherwise returns an errno. | |
1033 | * | |
1034 | * Called by the ib_create_qp() core verbs function. | |
1035 | */ | |
1036 | struct ib_qp *hfi1_create_qp(struct ib_pd *ibpd, | |
1037 | struct ib_qp_init_attr *init_attr, | |
1038 | struct ib_udata *udata) | |
1039 | { | |
895420dd | 1040 | struct rvt_qp *qp; |
4c6829c5 | 1041 | struct hfi1_qp_priv *priv; |
77241056 | 1042 | int err; |
895420dd | 1043 | struct rvt_swqe *swq = NULL; |
77241056 MM |
1044 | struct hfi1_ibdev *dev; |
1045 | struct hfi1_devdata *dd; | |
1046 | size_t sz; | |
1047 | size_t sg_list_sz; | |
1048 | struct ib_qp *ret; | |
1049 | ||
1050 | if (init_attr->cap.max_send_sge > hfi1_max_sges || | |
1051 | init_attr->cap.max_send_wr > hfi1_max_qp_wrs || | |
1052 | init_attr->create_flags) { | |
1053 | ret = ERR_PTR(-EINVAL); | |
1054 | goto bail; | |
1055 | } | |
1056 | ||
1057 | /* Check receive queue parameters if no SRQ is specified. */ | |
1058 | if (!init_attr->srq) { | |
1059 | if (init_attr->cap.max_recv_sge > hfi1_max_sges || | |
1060 | init_attr->cap.max_recv_wr > hfi1_max_qp_wrs) { | |
1061 | ret = ERR_PTR(-EINVAL); | |
1062 | goto bail; | |
1063 | } | |
1064 | if (init_attr->cap.max_send_sge + | |
1065 | init_attr->cap.max_send_wr + | |
1066 | init_attr->cap.max_recv_sge + | |
1067 | init_attr->cap.max_recv_wr == 0) { | |
1068 | ret = ERR_PTR(-EINVAL); | |
1069 | goto bail; | |
1070 | } | |
1071 | } | |
1072 | ||
1073 | switch (init_attr->qp_type) { | |
1074 | case IB_QPT_SMI: | |
1075 | case IB_QPT_GSI: | |
1076 | if (init_attr->port_num == 0 || | |
1077 | init_attr->port_num > ibpd->device->phys_port_cnt) { | |
1078 | ret = ERR_PTR(-EINVAL); | |
1079 | goto bail; | |
1080 | } | |
1081 | case IB_QPT_UC: | |
1082 | case IB_QPT_RC: | |
1083 | case IB_QPT_UD: | |
895420dd | 1084 | sz = sizeof(struct rvt_sge) * |
77241056 | 1085 | init_attr->cap.max_send_sge + |
895420dd | 1086 | sizeof(struct rvt_swqe); |
77241056 MM |
1087 | swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz); |
1088 | if (swq == NULL) { | |
1089 | ret = ERR_PTR(-ENOMEM); | |
1090 | goto bail; | |
1091 | } | |
1092 | sz = sizeof(*qp); | |
1093 | sg_list_sz = 0; | |
1094 | if (init_attr->srq) { | |
39db3e66 | 1095 | struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); |
77241056 MM |
1096 | |
1097 | if (srq->rq.max_sge > 1) | |
1098 | sg_list_sz = sizeof(*qp->r_sg_list) * | |
1099 | (srq->rq.max_sge - 1); | |
1100 | } else if (init_attr->cap.max_recv_sge > 1) | |
1101 | sg_list_sz = sizeof(*qp->r_sg_list) * | |
1102 | (init_attr->cap.max_recv_sge - 1); | |
1103 | qp = kzalloc(sz + sg_list_sz, GFP_KERNEL); | |
1104 | if (!qp) { | |
1105 | ret = ERR_PTR(-ENOMEM); | |
1106 | goto bail_swq; | |
1107 | } | |
1108 | RCU_INIT_POINTER(qp->next, NULL); | |
4c6829c5 DD |
1109 | priv = kzalloc(sizeof(*priv), GFP_KERNEL); |
1110 | if (!priv) { | |
1111 | ret = ERR_PTR(-ENOMEM); | |
1112 | goto bail_qp_priv; | |
1113 | } | |
1114 | priv->owner = qp; | |
1115 | priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), GFP_KERNEL); | |
1116 | if (!priv->s_hdr) { | |
77241056 MM |
1117 | ret = ERR_PTR(-ENOMEM); |
1118 | goto bail_qp; | |
1119 | } | |
4c6829c5 | 1120 | qp->priv = priv; |
77241056 MM |
1121 | qp->timeout_jiffies = |
1122 | usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / | |
1123 | 1000UL); | |
1124 | if (init_attr->srq) | |
1125 | sz = 0; | |
1126 | else { | |
1127 | qp->r_rq.size = init_attr->cap.max_recv_wr + 1; | |
1128 | qp->r_rq.max_sge = init_attr->cap.max_recv_sge; | |
1129 | sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + | |
895420dd DD |
1130 | sizeof(struct rvt_rwqe); |
1131 | qp->r_rq.wq = vmalloc_user(sizeof(struct rvt_rwq) + | |
77241056 MM |
1132 | qp->r_rq.size * sz); |
1133 | if (!qp->r_rq.wq) { | |
1134 | ret = ERR_PTR(-ENOMEM); | |
1135 | goto bail_qp; | |
1136 | } | |
1137 | } | |
1138 | ||
1139 | /* | |
1140 | * ib_create_qp() will initialize qp->ibqp | |
1141 | * except for qp->ibqp.qp_num. | |
1142 | */ | |
1143 | spin_lock_init(&qp->r_lock); | |
1144 | spin_lock_init(&qp->s_lock); | |
1145 | spin_lock_init(&qp->r_rq.lock); | |
1146 | atomic_set(&qp->refcount, 0); | |
1147 | init_waitqueue_head(&qp->wait); | |
1148 | init_timer(&qp->s_timer); | |
1149 | qp->s_timer.data = (unsigned long)qp; | |
1150 | INIT_LIST_HEAD(&qp->rspwait); | |
1151 | qp->state = IB_QPS_RESET; | |
1152 | qp->s_wq = swq; | |
1153 | qp->s_size = init_attr->cap.max_send_wr + 1; | |
1154 | qp->s_max_sge = init_attr->cap.max_send_sge; | |
1155 | if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) | |
1156 | qp->s_flags = HFI1_S_SIGNAL_REQ_WR; | |
1157 | dev = to_idev(ibpd->device); | |
1158 | dd = dd_from_dev(dev); | |
1159 | err = alloc_qpn(dd, &dev->qp_dev->qpn_table, init_attr->qp_type, | |
1160 | init_attr->port_num); | |
1161 | if (err < 0) { | |
1162 | ret = ERR_PTR(err); | |
1163 | vfree(qp->r_rq.wq); | |
1164 | goto bail_qp; | |
1165 | } | |
1166 | qp->ibqp.qp_num = err; | |
1167 | qp->port_num = init_attr->port_num; | |
1168 | reset_qp(qp, init_attr->qp_type); | |
1169 | ||
1170 | break; | |
1171 | ||
1172 | default: | |
1173 | /* Don't support raw QPs */ | |
1174 | ret = ERR_PTR(-ENOSYS); | |
1175 | goto bail; | |
1176 | } | |
1177 | ||
1178 | init_attr->cap.max_inline_data = 0; | |
1179 | ||
1180 | /* | |
1181 | * Return the address of the RWQ as the offset to mmap. | |
1182 | * See hfi1_mmap() for details. | |
1183 | */ | |
1184 | if (udata && udata->outlen >= sizeof(__u64)) { | |
1185 | if (!qp->r_rq.wq) { | |
1186 | __u64 offset = 0; | |
1187 | ||
1188 | err = ib_copy_to_udata(udata, &offset, | |
1189 | sizeof(offset)); | |
1190 | if (err) { | |
1191 | ret = ERR_PTR(err); | |
1192 | goto bail_ip; | |
1193 | } | |
1194 | } else { | |
895420dd | 1195 | u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; |
77241056 MM |
1196 | |
1197 | qp->ip = hfi1_create_mmap_info(dev, s, | |
1198 | ibpd->uobject->context, | |
1199 | qp->r_rq.wq); | |
1200 | if (!qp->ip) { | |
1201 | ret = ERR_PTR(-ENOMEM); | |
1202 | goto bail_ip; | |
1203 | } | |
1204 | ||
1205 | err = ib_copy_to_udata(udata, &(qp->ip->offset), | |
1206 | sizeof(qp->ip->offset)); | |
1207 | if (err) { | |
1208 | ret = ERR_PTR(err); | |
1209 | goto bail_ip; | |
1210 | } | |
1211 | } | |
1212 | } | |
1213 | ||
1214 | spin_lock(&dev->n_qps_lock); | |
1215 | if (dev->n_qps_allocated == hfi1_max_qps) { | |
1216 | spin_unlock(&dev->n_qps_lock); | |
1217 | ret = ERR_PTR(-ENOMEM); | |
1218 | goto bail_ip; | |
1219 | } | |
1220 | ||
1221 | dev->n_qps_allocated++; | |
1222 | spin_unlock(&dev->n_qps_lock); | |
1223 | ||
1224 | if (qp->ip) { | |
1225 | spin_lock_irq(&dev->pending_lock); | |
1226 | list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps); | |
1227 | spin_unlock_irq(&dev->pending_lock); | |
1228 | } | |
1229 | ||
1230 | ret = &qp->ibqp; | |
1231 | ||
1232 | /* | |
1233 | * We have our QP and its good, now keep track of what types of opcodes | |
1234 | * can be processed on this QP. We do this by keeping track of what the | |
1235 | * 3 high order bits of the opcode are. | |
1236 | */ | |
1237 | switch (init_attr->qp_type) { | |
1238 | case IB_QPT_SMI: | |
1239 | case IB_QPT_GSI: | |
1240 | case IB_QPT_UD: | |
1241 | qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & OPCODE_QP_MASK; | |
1242 | break; | |
1243 | case IB_QPT_RC: | |
1244 | qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & OPCODE_QP_MASK; | |
1245 | break; | |
1246 | case IB_QPT_UC: | |
1247 | qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & OPCODE_QP_MASK; | |
1248 | break; | |
1249 | default: | |
1250 | ret = ERR_PTR(-EINVAL); | |
1251 | goto bail_ip; | |
1252 | } | |
1253 | ||
1254 | goto bail; | |
1255 | ||
1256 | bail_ip: | |
1257 | if (qp->ip) | |
1258 | kref_put(&qp->ip->ref, hfi1_release_mmap_info); | |
1259 | else | |
1260 | vfree(qp->r_rq.wq); | |
1261 | free_qpn(&dev->qp_dev->qpn_table, qp->ibqp.qp_num); | |
1262 | bail_qp: | |
4c6829c5 DD |
1263 | kfree(priv->s_hdr); |
1264 | kfree(priv); | |
1265 | bail_qp_priv: | |
77241056 MM |
1266 | kfree(qp); |
1267 | bail_swq: | |
1268 | vfree(swq); | |
1269 | bail: | |
1270 | return ret; | |
1271 | } | |
1272 | ||
1273 | /** | |
1274 | * hfi1_destroy_qp - destroy a queue pair | |
1275 | * @ibqp: the queue pair to destroy | |
1276 | * | |
1277 | * Returns 0 on success. | |
1278 | * | |
1279 | * Note that this can be called while the QP is actively sending or | |
1280 | * receiving! | |
1281 | */ | |
1282 | int hfi1_destroy_qp(struct ib_qp *ibqp) | |
1283 | { | |
895420dd | 1284 | struct rvt_qp *qp = to_iqp(ibqp); |
77241056 | 1285 | struct hfi1_ibdev *dev = to_idev(ibqp->device); |
4c6829c5 | 1286 | struct hfi1_qp_priv *priv = qp->priv; |
77241056 MM |
1287 | |
1288 | /* Make sure HW and driver activity is stopped. */ | |
1289 | spin_lock_irq(&qp->r_lock); | |
1290 | spin_lock(&qp->s_lock); | |
1291 | if (qp->state != IB_QPS_RESET) { | |
1292 | qp->state = IB_QPS_RESET; | |
1293 | flush_iowait(qp); | |
1294 | qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_ANY_WAIT); | |
1295 | spin_unlock(&qp->s_lock); | |
1296 | spin_unlock_irq(&qp->r_lock); | |
4c6829c5 | 1297 | cancel_work_sync(&priv->s_iowait.iowork); |
77241056 | 1298 | del_timer_sync(&qp->s_timer); |
4c6829c5 | 1299 | iowait_sdma_drain(&priv->s_iowait); |
77241056 MM |
1300 | flush_tx_list(qp); |
1301 | remove_qp(dev, qp); | |
1302 | wait_event(qp->wait, !atomic_read(&qp->refcount)); | |
1303 | spin_lock_irq(&qp->r_lock); | |
1304 | spin_lock(&qp->s_lock); | |
1305 | clear_mr_refs(qp, 1); | |
1306 | clear_ahg(qp); | |
1307 | } | |
1308 | spin_unlock(&qp->s_lock); | |
1309 | spin_unlock_irq(&qp->r_lock); | |
1310 | ||
1311 | /* all user's cleaned up, mark it available */ | |
1312 | free_qpn(&dev->qp_dev->qpn_table, qp->ibqp.qp_num); | |
1313 | spin_lock(&dev->n_qps_lock); | |
1314 | dev->n_qps_allocated--; | |
1315 | spin_unlock(&dev->n_qps_lock); | |
1316 | ||
1317 | if (qp->ip) | |
1318 | kref_put(&qp->ip->ref, hfi1_release_mmap_info); | |
1319 | else | |
1320 | vfree(qp->r_rq.wq); | |
1321 | vfree(qp->s_wq); | |
4c6829c5 DD |
1322 | kfree(priv->s_hdr); |
1323 | kfree(priv); | |
77241056 MM |
1324 | kfree(qp); |
1325 | return 0; | |
1326 | } | |
1327 | ||
1328 | /** | |
1329 | * init_qpn_table - initialize the QP number table for a device | |
1330 | * @qpt: the QPN table | |
1331 | */ | |
1332 | static int init_qpn_table(struct hfi1_devdata *dd, struct hfi1_qpn_table *qpt) | |
1333 | { | |
1334 | u32 offset, qpn, i; | |
1335 | struct qpn_map *map; | |
1336 | int ret = 0; | |
1337 | ||
1338 | spin_lock_init(&qpt->lock); | |
1339 | ||
1340 | qpt->last = 0; | |
1341 | qpt->incr = 1 << dd->qos_shift; | |
1342 | ||
1343 | /* insure we don't assign QPs from KDETH 64K window */ | |
1344 | qpn = kdeth_qp << 16; | |
1345 | qpt->nmaps = qpn / BITS_PER_PAGE; | |
1346 | /* This should always be zero */ | |
1347 | offset = qpn & BITS_PER_PAGE_MASK; | |
1348 | map = &qpt->map[qpt->nmaps]; | |
1349 | dd_dev_info(dd, "Reserving QPNs for KDETH window from 0x%x to 0x%x\n", | |
1350 | qpn, qpn + 65535); | |
1351 | for (i = 0; i < 65536; i++) { | |
1352 | if (!map->page) { | |
1353 | get_map_page(qpt, map); | |
1354 | if (!map->page) { | |
1355 | ret = -ENOMEM; | |
1356 | break; | |
1357 | } | |
1358 | } | |
1359 | set_bit(offset, map->page); | |
1360 | offset++; | |
1361 | if (offset == BITS_PER_PAGE) { | |
1362 | /* next page */ | |
1363 | qpt->nmaps++; | |
1364 | map++; | |
1365 | offset = 0; | |
1366 | } | |
1367 | } | |
1368 | return ret; | |
1369 | } | |
1370 | ||
1371 | /** | |
1372 | * free_qpn_table - free the QP number table for a device | |
1373 | * @qpt: the QPN table | |
1374 | */ | |
1375 | static void free_qpn_table(struct hfi1_qpn_table *qpt) | |
1376 | { | |
1377 | int i; | |
1378 | ||
1379 | for (i = 0; i < ARRAY_SIZE(qpt->map); i++) | |
1380 | free_page((unsigned long) qpt->map[i].page); | |
1381 | } | |
1382 | ||
1383 | /** | |
1384 | * hfi1_get_credit - flush the send work queue of a QP | |
1385 | * @qp: the qp who's send work queue to flush | |
1386 | * @aeth: the Acknowledge Extended Transport Header | |
1387 | * | |
1388 | * The QP s_lock should be held. | |
1389 | */ | |
895420dd | 1390 | void hfi1_get_credit(struct rvt_qp *qp, u32 aeth) |
77241056 MM |
1391 | { |
1392 | u32 credit = (aeth >> HFI1_AETH_CREDIT_SHIFT) & HFI1_AETH_CREDIT_MASK; | |
1393 | ||
1394 | /* | |
1395 | * If the credit is invalid, we can send | |
1396 | * as many packets as we like. Otherwise, we have to | |
1397 | * honor the credit field. | |
1398 | */ | |
1399 | if (credit == HFI1_AETH_CREDIT_INVAL) { | |
1400 | if (!(qp->s_flags & HFI1_S_UNLIMITED_CREDIT)) { | |
1401 | qp->s_flags |= HFI1_S_UNLIMITED_CREDIT; | |
1402 | if (qp->s_flags & HFI1_S_WAIT_SSN_CREDIT) { | |
1403 | qp->s_flags &= ~HFI1_S_WAIT_SSN_CREDIT; | |
1404 | hfi1_schedule_send(qp); | |
1405 | } | |
1406 | } | |
1407 | } else if (!(qp->s_flags & HFI1_S_UNLIMITED_CREDIT)) { | |
1408 | /* Compute new LSN (i.e., MSN + credit) */ | |
1409 | credit = (aeth + credit_table[credit]) & HFI1_MSN_MASK; | |
1410 | if (cmp_msn(credit, qp->s_lsn) > 0) { | |
1411 | qp->s_lsn = credit; | |
1412 | if (qp->s_flags & HFI1_S_WAIT_SSN_CREDIT) { | |
1413 | qp->s_flags &= ~HFI1_S_WAIT_SSN_CREDIT; | |
1414 | hfi1_schedule_send(qp); | |
1415 | } | |
1416 | } | |
1417 | } | |
1418 | } | |
1419 | ||
895420dd | 1420 | void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag) |
77241056 MM |
1421 | { |
1422 | unsigned long flags; | |
1423 | ||
1424 | spin_lock_irqsave(&qp->s_lock, flags); | |
1425 | if (qp->s_flags & flag) { | |
1426 | qp->s_flags &= ~flag; | |
1427 | trace_hfi1_qpwakeup(qp, flag); | |
1428 | hfi1_schedule_send(qp); | |
1429 | } | |
1430 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1431 | /* Notify hfi1_destroy_qp() if it is waiting. */ | |
1432 | if (atomic_dec_and_test(&qp->refcount)) | |
1433 | wake_up(&qp->wait); | |
1434 | } | |
1435 | ||
1436 | static int iowait_sleep( | |
1437 | struct sdma_engine *sde, | |
1438 | struct iowait *wait, | |
1439 | struct sdma_txreq *stx, | |
1440 | unsigned seq) | |
1441 | { | |
1442 | struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq); | |
895420dd | 1443 | struct rvt_qp *qp; |
4c6829c5 | 1444 | struct hfi1_qp_priv *priv; |
77241056 MM |
1445 | unsigned long flags; |
1446 | int ret = 0; | |
1447 | struct hfi1_ibdev *dev; | |
1448 | ||
1449 | qp = tx->qp; | |
4c6829c5 | 1450 | priv = qp->priv; |
77241056 MM |
1451 | |
1452 | spin_lock_irqsave(&qp->s_lock, flags); | |
1453 | if (ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK) { | |
1454 | ||
1455 | /* | |
1456 | * If we couldn't queue the DMA request, save the info | |
1457 | * and try again later rather than destroying the | |
1458 | * buffer and undoing the side effects of the copy. | |
1459 | */ | |
1460 | /* Make a common routine? */ | |
1461 | dev = &sde->dd->verbs_dev; | |
1462 | list_add_tail(&stx->list, &wait->tx_head); | |
1463 | write_seqlock(&dev->iowait_lock); | |
1464 | if (sdma_progress(sde, seq, stx)) | |
1465 | goto eagain; | |
4c6829c5 | 1466 | if (list_empty(&priv->s_iowait.list)) { |
77241056 MM |
1467 | struct hfi1_ibport *ibp = |
1468 | to_iport(qp->ibqp.device, qp->port_num); | |
1469 | ||
1470 | ibp->n_dmawait++; | |
1471 | qp->s_flags |= HFI1_S_WAIT_DMA_DESC; | |
4c6829c5 | 1472 | list_add_tail(&priv->s_iowait.list, &sde->dmawait); |
77241056 MM |
1473 | trace_hfi1_qpsleep(qp, HFI1_S_WAIT_DMA_DESC); |
1474 | atomic_inc(&qp->refcount); | |
1475 | } | |
1476 | write_sequnlock(&dev->iowait_lock); | |
1477 | qp->s_flags &= ~HFI1_S_BUSY; | |
1478 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1479 | ret = -EBUSY; | |
1480 | } else { | |
1481 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1482 | hfi1_put_txreq(tx); | |
1483 | } | |
1484 | return ret; | |
1485 | eagain: | |
1486 | write_sequnlock(&dev->iowait_lock); | |
1487 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1488 | list_del_init(&stx->list); | |
1489 | return -EAGAIN; | |
1490 | } | |
1491 | ||
1492 | static void iowait_wakeup(struct iowait *wait, int reason) | |
1493 | { | |
895420dd | 1494 | struct rvt_qp *qp = iowait_to_qp(wait); |
77241056 MM |
1495 | |
1496 | WARN_ON(reason != SDMA_AVAIL_REASON); | |
1497 | hfi1_qp_wakeup(qp, HFI1_S_WAIT_DMA_DESC); | |
1498 | } | |
1499 | ||
1500 | int hfi1_qp_init(struct hfi1_ibdev *dev) | |
1501 | { | |
1502 | struct hfi1_devdata *dd = dd_from_dev(dev); | |
1503 | int i; | |
1504 | int ret = -ENOMEM; | |
1505 | ||
1506 | /* allocate parent object */ | |
1507 | dev->qp_dev = kzalloc(sizeof(*dev->qp_dev), GFP_KERNEL); | |
1508 | if (!dev->qp_dev) | |
1509 | goto nomem; | |
1510 | /* allocate hash table */ | |
1511 | dev->qp_dev->qp_table_size = hfi1_qp_table_size; | |
1512 | dev->qp_dev->qp_table_bits = ilog2(hfi1_qp_table_size); | |
1513 | dev->qp_dev->qp_table = | |
1514 | kmalloc(dev->qp_dev->qp_table_size * | |
1515 | sizeof(*dev->qp_dev->qp_table), | |
1516 | GFP_KERNEL); | |
1517 | if (!dev->qp_dev->qp_table) | |
1518 | goto nomem; | |
1519 | for (i = 0; i < dev->qp_dev->qp_table_size; i++) | |
1520 | RCU_INIT_POINTER(dev->qp_dev->qp_table[i], NULL); | |
1521 | spin_lock_init(&dev->qp_dev->qpt_lock); | |
1522 | /* initialize qpn map */ | |
1523 | ret = init_qpn_table(dd, &dev->qp_dev->qpn_table); | |
1524 | if (ret) | |
1525 | goto nomem; | |
1526 | return ret; | |
1527 | nomem: | |
1528 | if (dev->qp_dev) { | |
1529 | kfree(dev->qp_dev->qp_table); | |
1530 | free_qpn_table(&dev->qp_dev->qpn_table); | |
1531 | kfree(dev->qp_dev); | |
1532 | } | |
1533 | return ret; | |
1534 | } | |
1535 | ||
1536 | void hfi1_qp_exit(struct hfi1_ibdev *dev) | |
1537 | { | |
1538 | struct hfi1_devdata *dd = dd_from_dev(dev); | |
1539 | u32 qps_inuse; | |
1540 | ||
1541 | qps_inuse = free_all_qps(dd); | |
1542 | if (qps_inuse) | |
1543 | dd_dev_err(dd, "QP memory leak! %u still in use\n", | |
1544 | qps_inuse); | |
1545 | if (dev->qp_dev) { | |
1546 | kfree(dev->qp_dev->qp_table); | |
1547 | free_qpn_table(&dev->qp_dev->qpn_table); | |
1548 | kfree(dev->qp_dev); | |
1549 | } | |
1550 | } | |
1551 | ||
1552 | /** | |
1553 | * | |
1554 | * qp_to_sdma_engine - map a qp to a send engine | |
1555 | * @qp: the QP | |
1556 | * @sc5: the 5 bit sc | |
1557 | * | |
1558 | * Return: | |
1559 | * A send engine for the qp or NULL for SMI type qp. | |
1560 | */ | |
895420dd | 1561 | struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5) |
77241056 MM |
1562 | { |
1563 | struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device); | |
1564 | struct sdma_engine *sde; | |
1565 | ||
1566 | if (!(dd->flags & HFI1_HAS_SEND_DMA)) | |
1567 | return NULL; | |
1568 | switch (qp->ibqp.qp_type) { | |
77241056 MM |
1569 | case IB_QPT_SMI: |
1570 | return NULL; | |
1571 | default: | |
1572 | break; | |
1573 | } | |
1574 | sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5); | |
1575 | return sde; | |
1576 | } | |
1577 | ||
1578 | struct qp_iter { | |
1579 | struct hfi1_ibdev *dev; | |
895420dd | 1580 | struct rvt_qp *qp; |
77241056 MM |
1581 | int specials; |
1582 | int n; | |
1583 | }; | |
1584 | ||
1585 | struct qp_iter *qp_iter_init(struct hfi1_ibdev *dev) | |
1586 | { | |
1587 | struct qp_iter *iter; | |
1588 | ||
1589 | iter = kzalloc(sizeof(*iter), GFP_KERNEL); | |
1590 | if (!iter) | |
1591 | return NULL; | |
1592 | ||
1593 | iter->dev = dev; | |
ec3f2c12 | 1594 | iter->specials = dev->rdi.ibdev.phys_port_cnt * 2; |
77241056 MM |
1595 | if (qp_iter_next(iter)) { |
1596 | kfree(iter); | |
1597 | return NULL; | |
1598 | } | |
1599 | ||
1600 | return iter; | |
1601 | } | |
1602 | ||
1603 | int qp_iter_next(struct qp_iter *iter) | |
1604 | { | |
1605 | struct hfi1_ibdev *dev = iter->dev; | |
1606 | int n = iter->n; | |
1607 | int ret = 1; | |
895420dd DD |
1608 | struct rvt_qp *pqp = iter->qp; |
1609 | struct rvt_qp *qp; | |
77241056 MM |
1610 | |
1611 | /* | |
1612 | * The approach is to consider the special qps | |
1613 | * as an additional table entries before the | |
1614 | * real hash table. Since the qp code sets | |
1615 | * the qp->next hash link to NULL, this works just fine. | |
1616 | * | |
1617 | * iter->specials is 2 * # ports | |
1618 | * | |
1619 | * n = 0..iter->specials is the special qp indices | |
1620 | * | |
1621 | * n = iter->specials..dev->qp_dev->qp_table_size+iter->specials are | |
1622 | * the potential hash bucket entries | |
1623 | * | |
1624 | */ | |
1625 | for (; n < dev->qp_dev->qp_table_size + iter->specials; n++) { | |
1626 | if (pqp) { | |
1627 | qp = rcu_dereference(pqp->next); | |
1628 | } else { | |
1629 | if (n < iter->specials) { | |
1630 | struct hfi1_pportdata *ppd; | |
1631 | struct hfi1_ibport *ibp; | |
1632 | int pidx; | |
1633 | ||
ec3f2c12 | 1634 | pidx = n % dev->rdi.ibdev.phys_port_cnt; |
77241056 MM |
1635 | ppd = &dd_from_dev(dev)->pport[pidx]; |
1636 | ibp = &ppd->ibport_data; | |
1637 | ||
1638 | if (!(n & 1)) | |
1639 | qp = rcu_dereference(ibp->qp[0]); | |
1640 | else | |
1641 | qp = rcu_dereference(ibp->qp[1]); | |
1642 | } else { | |
1643 | qp = rcu_dereference( | |
1644 | dev->qp_dev->qp_table[ | |
1645 | (n - iter->specials)]); | |
1646 | } | |
1647 | } | |
1648 | pqp = qp; | |
1649 | if (qp) { | |
1650 | iter->qp = qp; | |
1651 | iter->n = n; | |
1652 | return 0; | |
1653 | } | |
1654 | } | |
1655 | return ret; | |
1656 | } | |
1657 | ||
1658 | static const char * const qp_type_str[] = { | |
1659 | "SMI", "GSI", "RC", "UC", "UD", | |
1660 | }; | |
1661 | ||
895420dd | 1662 | static int qp_idle(struct rvt_qp *qp) |
77241056 MM |
1663 | { |
1664 | return | |
1665 | qp->s_last == qp->s_acked && | |
1666 | qp->s_acked == qp->s_cur && | |
1667 | qp->s_cur == qp->s_tail && | |
1668 | qp->s_tail == qp->s_head; | |
1669 | } | |
1670 | ||
1671 | void qp_iter_print(struct seq_file *s, struct qp_iter *iter) | |
1672 | { | |
895420dd DD |
1673 | struct rvt_swqe *wqe; |
1674 | struct rvt_qp *qp = iter->qp; | |
4c6829c5 | 1675 | struct hfi1_qp_priv *priv = qp->priv; |
77241056 MM |
1676 | struct sdma_engine *sde; |
1677 | ||
4c6829c5 | 1678 | sde = qp_to_sdma_engine(qp, priv->s_sc); |
77241056 MM |
1679 | wqe = get_swqe_ptr(qp, qp->s_last); |
1680 | seq_printf(s, | |
1681 | "N %d %s QP%u R %u %s %u %u %u f=%x %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u) QP%u LID %x SL %u MTU %d %u %u %u SDE %p,%u\n", | |
1682 | iter->n, | |
1683 | qp_idle(qp) ? "I" : "B", | |
1684 | qp->ibqp.qp_num, | |
1685 | atomic_read(&qp->refcount), | |
1686 | qp_type_str[qp->ibqp.qp_type], | |
1687 | qp->state, | |
1688 | wqe ? wqe->wr.opcode : 0, | |
1689 | qp->s_hdrwords, | |
1690 | qp->s_flags, | |
4c6829c5 DD |
1691 | atomic_read(&priv->s_iowait.sdma_busy), |
1692 | !list_empty(&priv->s_iowait.list), | |
77241056 MM |
1693 | qp->timeout, |
1694 | wqe ? wqe->ssn : 0, | |
1695 | qp->s_lsn, | |
1696 | qp->s_last_psn, | |
1697 | qp->s_psn, qp->s_next_psn, | |
1698 | qp->s_sending_psn, qp->s_sending_hpsn, | |
1699 | qp->s_last, qp->s_acked, qp->s_cur, | |
1700 | qp->s_tail, qp->s_head, qp->s_size, | |
1701 | qp->remote_qpn, | |
1702 | qp->remote_ah_attr.dlid, | |
1703 | qp->remote_ah_attr.sl, | |
1704 | qp->pmtu, | |
1705 | qp->s_retry_cnt, | |
1706 | qp->timeout, | |
1707 | qp->s_rnr_retry_cnt, | |
1708 | sde, | |
1709 | sde ? sde->this_idx : 0); | |
1710 | } | |
1711 | ||
895420dd | 1712 | void qp_comm_est(struct rvt_qp *qp) |
77241056 MM |
1713 | { |
1714 | qp->r_flags |= HFI1_R_COMM_EST; | |
1715 | if (qp->ibqp.event_handler) { | |
1716 | struct ib_event ev; | |
1717 | ||
1718 | ev.device = qp->ibqp.device; | |
1719 | ev.element.qp = &qp->ibqp; | |
1720 | ev.event = IB_EVENT_COMM_EST; | |
1721 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1722 | } | |
1723 | } | |
c2f3ffb0 MM |
1724 | |
1725 | /* | |
1726 | * Switch to alternate path. | |
1727 | * The QP s_lock should be held and interrupts disabled. | |
1728 | */ | |
895420dd | 1729 | void hfi1_migrate_qp(struct rvt_qp *qp) |
c2f3ffb0 | 1730 | { |
4c6829c5 | 1731 | struct hfi1_qp_priv *priv = qp->priv; |
c2f3ffb0 MM |
1732 | struct ib_event ev; |
1733 | ||
1734 | qp->s_mig_state = IB_MIG_MIGRATED; | |
1735 | qp->remote_ah_attr = qp->alt_ah_attr; | |
1736 | qp->port_num = qp->alt_ah_attr.port_num; | |
1737 | qp->s_pkey_index = qp->s_alt_pkey_index; | |
1738 | qp->s_flags |= HFI1_S_AHG_CLEAR; | |
4c6829c5 DD |
1739 | priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr); |
1740 | priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc); | |
c2f3ffb0 MM |
1741 | |
1742 | ev.device = qp->ibqp.device; | |
1743 | ev.element.qp = &qp->ibqp; | |
1744 | ev.event = IB_EVENT_PATH_MIG; | |
1745 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1746 | } |