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.
8 * Copyright(c) 2015 Intel Corporation.
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.
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.
21 * Copyright(c) 2015 Intel Corporation.
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
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
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.
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.
51 #include <rdma/ib_mad.h>
52 #include <rdma/ib_user_verbs.h>
54 #include <linux/module.h>
55 #include <linux/utsname.h>
56 #include <linux/rculist.h>
58 #include <linux/random.h>
59 #include <linux/vmalloc.h>
68 unsigned int hfi1_lkey_table_size = 16;
69 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
71 MODULE_PARM_DESC(lkey_table_size,
72 "LKEY table size in bits (2^n, 1 <= n <= 23)");
74 static unsigned int hfi1_max_pds = 0xFFFF;
75 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
76 MODULE_PARM_DESC(max_pds,
77 "Maximum number of protection domains to support");
79 static unsigned int hfi1_max_ahs = 0xFFFF;
80 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
81 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
83 unsigned int hfi1_max_cqes = 0x2FFFF;
84 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
85 MODULE_PARM_DESC(max_cqes,
86 "Maximum number of completion queue entries to support");
88 unsigned int hfi1_max_cqs = 0x1FFFF;
89 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
90 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
92 unsigned int hfi1_max_qp_wrs = 0x3FFF;
93 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
94 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
96 unsigned int hfi1_max_qps = 16384;
97 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
98 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
100 unsigned int hfi1_max_sges = 0x60;
101 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
102 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
104 unsigned int hfi1_max_mcast_grps = 16384;
105 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
106 MODULE_PARM_DESC(max_mcast_grps,
107 "Maximum number of multicast groups to support");
109 unsigned int hfi1_max_mcast_qp_attached = 16;
110 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
112 MODULE_PARM_DESC(max_mcast_qp_attached,
113 "Maximum number of attached QPs to support");
115 unsigned int hfi1_max_srqs = 1024;
116 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
117 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
119 unsigned int hfi1_max_srq_sges = 128;
120 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
121 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
123 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
124 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
125 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
127 static void verbs_sdma_complete(
128 struct sdma_txreq *cookie,
132 /* Length of buffer to create verbs txreq cache name */
133 #define TXREQ_NAME_LEN 24
136 * Note that it is OK to post send work requests in the SQE and ERR
137 * states; hfi1_do_send() will process them and generate error
138 * completions as per IB 1.2 C10-96.
140 const int ib_hfi1_state_ops[IB_QPS_ERR + 1] = {
142 [IB_QPS_INIT] = HFI1_POST_RECV_OK,
143 [IB_QPS_RTR] = HFI1_POST_RECV_OK | HFI1_PROCESS_RECV_OK,
144 [IB_QPS_RTS] = HFI1_POST_RECV_OK | HFI1_PROCESS_RECV_OK |
145 HFI1_POST_SEND_OK | HFI1_PROCESS_SEND_OK |
146 HFI1_PROCESS_NEXT_SEND_OK,
147 [IB_QPS_SQD] = HFI1_POST_RECV_OK | HFI1_PROCESS_RECV_OK |
148 HFI1_POST_SEND_OK | HFI1_PROCESS_SEND_OK,
149 [IB_QPS_SQE] = HFI1_POST_RECV_OK | HFI1_PROCESS_RECV_OK |
150 HFI1_POST_SEND_OK | HFI1_FLUSH_SEND,
151 [IB_QPS_ERR] = HFI1_POST_RECV_OK | HFI1_FLUSH_RECV |
152 HFI1_POST_SEND_OK | HFI1_FLUSH_SEND,
155 struct hfi1_ucontext {
156 struct ib_ucontext ibucontext;
159 static inline struct hfi1_ucontext *to_iucontext(struct ib_ucontext
162 return container_of(ibucontext, struct hfi1_ucontext, ibucontext);
166 * Translate ib_wr_opcode into ib_wc_opcode.
168 const enum ib_wc_opcode ib_hfi1_wc_opcode[] = {
169 [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
170 [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
171 [IB_WR_SEND] = IB_WC_SEND,
172 [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
173 [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
174 [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
175 [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
179 * Length of header by opcode, 0 --> not supported
181 const u8 hdr_len_by_opcode[256] = {
183 [IB_OPCODE_RC_SEND_FIRST] = 12 + 8,
184 [IB_OPCODE_RC_SEND_MIDDLE] = 12 + 8,
185 [IB_OPCODE_RC_SEND_LAST] = 12 + 8,
186 [IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
187 [IB_OPCODE_RC_SEND_ONLY] = 12 + 8,
188 [IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE] = 12 + 8 + 4,
189 [IB_OPCODE_RC_RDMA_WRITE_FIRST] = 12 + 8 + 16,
190 [IB_OPCODE_RC_RDMA_WRITE_MIDDLE] = 12 + 8,
191 [IB_OPCODE_RC_RDMA_WRITE_LAST] = 12 + 8,
192 [IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
193 [IB_OPCODE_RC_RDMA_WRITE_ONLY] = 12 + 8 + 16,
194 [IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
195 [IB_OPCODE_RC_RDMA_READ_REQUEST] = 12 + 8 + 16,
196 [IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST] = 12 + 8 + 4,
197 [IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE] = 12 + 8,
198 [IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST] = 12 + 8 + 4,
199 [IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY] = 12 + 8 + 4,
200 [IB_OPCODE_RC_ACKNOWLEDGE] = 12 + 8 + 4,
201 [IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE] = 12 + 8 + 4,
202 [IB_OPCODE_RC_COMPARE_SWAP] = 12 + 8 + 28,
203 [IB_OPCODE_RC_FETCH_ADD] = 12 + 8 + 28,
205 [IB_OPCODE_UC_SEND_FIRST] = 12 + 8,
206 [IB_OPCODE_UC_SEND_MIDDLE] = 12 + 8,
207 [IB_OPCODE_UC_SEND_LAST] = 12 + 8,
208 [IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
209 [IB_OPCODE_UC_SEND_ONLY] = 12 + 8,
210 [IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE] = 12 + 8 + 4,
211 [IB_OPCODE_UC_RDMA_WRITE_FIRST] = 12 + 8 + 16,
212 [IB_OPCODE_UC_RDMA_WRITE_MIDDLE] = 12 + 8,
213 [IB_OPCODE_UC_RDMA_WRITE_LAST] = 12 + 8,
214 [IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
215 [IB_OPCODE_UC_RDMA_WRITE_ONLY] = 12 + 8 + 16,
216 [IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
218 [IB_OPCODE_UD_SEND_ONLY] = 12 + 8 + 8,
219 [IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE] = 12 + 8 + 12
222 static const opcode_handler opcode_handler_tbl[256] = {
224 [IB_OPCODE_RC_SEND_FIRST] = &hfi1_rc_rcv,
225 [IB_OPCODE_RC_SEND_MIDDLE] = &hfi1_rc_rcv,
226 [IB_OPCODE_RC_SEND_LAST] = &hfi1_rc_rcv,
227 [IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
228 [IB_OPCODE_RC_SEND_ONLY] = &hfi1_rc_rcv,
229 [IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
230 [IB_OPCODE_RC_RDMA_WRITE_FIRST] = &hfi1_rc_rcv,
231 [IB_OPCODE_RC_RDMA_WRITE_MIDDLE] = &hfi1_rc_rcv,
232 [IB_OPCODE_RC_RDMA_WRITE_LAST] = &hfi1_rc_rcv,
233 [IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
234 [IB_OPCODE_RC_RDMA_WRITE_ONLY] = &hfi1_rc_rcv,
235 [IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
236 [IB_OPCODE_RC_RDMA_READ_REQUEST] = &hfi1_rc_rcv,
237 [IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST] = &hfi1_rc_rcv,
238 [IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE] = &hfi1_rc_rcv,
239 [IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST] = &hfi1_rc_rcv,
240 [IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY] = &hfi1_rc_rcv,
241 [IB_OPCODE_RC_ACKNOWLEDGE] = &hfi1_rc_rcv,
242 [IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE] = &hfi1_rc_rcv,
243 [IB_OPCODE_RC_COMPARE_SWAP] = &hfi1_rc_rcv,
244 [IB_OPCODE_RC_FETCH_ADD] = &hfi1_rc_rcv,
246 [IB_OPCODE_UC_SEND_FIRST] = &hfi1_uc_rcv,
247 [IB_OPCODE_UC_SEND_MIDDLE] = &hfi1_uc_rcv,
248 [IB_OPCODE_UC_SEND_LAST] = &hfi1_uc_rcv,
249 [IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
250 [IB_OPCODE_UC_SEND_ONLY] = &hfi1_uc_rcv,
251 [IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
252 [IB_OPCODE_UC_RDMA_WRITE_FIRST] = &hfi1_uc_rcv,
253 [IB_OPCODE_UC_RDMA_WRITE_MIDDLE] = &hfi1_uc_rcv,
254 [IB_OPCODE_UC_RDMA_WRITE_LAST] = &hfi1_uc_rcv,
255 [IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
256 [IB_OPCODE_UC_RDMA_WRITE_ONLY] = &hfi1_uc_rcv,
257 [IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
259 [IB_OPCODE_UD_SEND_ONLY] = &hfi1_ud_rcv,
260 [IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE] = &hfi1_ud_rcv,
262 [IB_OPCODE_CNP] = &hfi1_cnp_rcv
268 __be64 ib_hfi1_sys_image_guid;
271 * hfi1_copy_sge - copy data to SGE memory
273 * @data: the data to copy
274 * @length: the length of the data
277 struct hfi1_sge_state *ss,
278 void *data, u32 length,
281 struct hfi1_sge *sge = &ss->sge;
284 u32 len = sge->length;
288 if (len > sge->sge_length)
289 len = sge->sge_length;
290 WARN_ON_ONCE(len == 0);
291 memcpy(sge->vaddr, data, len);
294 sge->sge_length -= len;
295 if (sge->sge_length == 0) {
297 hfi1_put_mr(sge->mr);
299 *sge = *ss->sg_list++;
300 } else if (sge->length == 0 && sge->mr->lkey) {
301 if (++sge->n >= HFI1_SEGSZ) {
302 if (++sge->m >= sge->mr->mapsz)
307 sge->mr->map[sge->m]->segs[sge->n].vaddr;
309 sge->mr->map[sge->m]->segs[sge->n].length;
317 * hfi1_skip_sge - skip over SGE memory
319 * @length: the number of bytes to skip
321 void hfi1_skip_sge(struct hfi1_sge_state *ss, u32 length, int release)
323 struct hfi1_sge *sge = &ss->sge;
326 u32 len = sge->length;
330 if (len > sge->sge_length)
331 len = sge->sge_length;
332 WARN_ON_ONCE(len == 0);
335 sge->sge_length -= len;
336 if (sge->sge_length == 0) {
338 hfi1_put_mr(sge->mr);
340 *sge = *ss->sg_list++;
341 } else if (sge->length == 0 && sge->mr->lkey) {
342 if (++sge->n >= HFI1_SEGSZ) {
343 if (++sge->m >= sge->mr->mapsz)
348 sge->mr->map[sge->m]->segs[sge->n].vaddr;
350 sge->mr->map[sge->m]->segs[sge->n].length;
357 * post_one_send - post one RC, UC, or UD send work request
358 * @qp: the QP to post on
359 * @wr: the work request to send
361 static int post_one_send(struct hfi1_qp *qp, struct ib_send_wr *wr)
363 struct hfi1_swqe *wqe;
368 struct hfi1_lkey_table *rkt;
370 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
371 struct hfi1_pportdata *ppd;
372 struct hfi1_ibport *ibp;
374 /* IB spec says that num_sge == 0 is OK. */
375 if (unlikely(wr->num_sge > qp->s_max_sge))
378 ppd = &dd->pport[qp->port_num - 1];
379 ibp = &ppd->ibport_data;
382 * Don't allow RDMA reads or atomic operations on UC or
383 * undefined operations.
384 * Make sure buffer is large enough to hold the result for atomics.
386 if (qp->ibqp.qp_type == IB_QPT_UC) {
387 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
389 } else if (qp->ibqp.qp_type != IB_QPT_RC) {
390 /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
391 if (wr->opcode != IB_WR_SEND &&
392 wr->opcode != IB_WR_SEND_WITH_IMM)
394 /* Check UD destination address PD */
395 if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
397 } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
399 else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
401 wr->sg_list[0].length < sizeof(u64) ||
402 wr->sg_list[0].addr & (sizeof(u64) - 1)))
404 else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
407 next = qp->s_head + 1;
408 if (next >= qp->s_size)
410 if (next == qp->s_last)
413 rkt = &to_idev(qp->ibqp.device)->lk_table;
414 pd = to_ipd(qp->ibqp.pd);
415 wqe = get_swqe_ptr(qp, qp->s_head);
418 if (qp->ibqp.qp_type != IB_QPT_UC &&
419 qp->ibqp.qp_type != IB_QPT_RC)
420 memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
421 else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
422 wr->opcode == IB_WR_RDMA_WRITE ||
423 wr->opcode == IB_WR_RDMA_READ)
424 memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
425 else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
426 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
427 memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
429 memcpy(&wqe->wr, wr, sizeof(wqe->wr));
434 acc = wr->opcode >= IB_WR_RDMA_READ ?
435 IB_ACCESS_LOCAL_WRITE : 0;
436 for (i = 0; i < wr->num_sge; i++) {
437 u32 length = wr->sg_list[i].length;
442 ok = hfi1_lkey_ok(rkt, pd, &wqe->sg_list[j],
443 &wr->sg_list[i], acc);
445 goto bail_inval_free;
446 wqe->length += length;
451 if (qp->ibqp.qp_type == IB_QPT_UC ||
452 qp->ibqp.qp_type == IB_QPT_RC) {
453 if (wqe->length > 0x80000000U)
454 goto bail_inval_free;
456 struct hfi1_ah *ah = to_iah(ud_wr(wr)->ah);
458 atomic_inc(&ah->refcount);
460 wqe->ssn = qp->s_ssn++;
466 /* release mr holds */
468 struct hfi1_sge *sge = &wqe->sg_list[--j];
470 hfi1_put_mr(sge->mr);
476 * post_send - post a send on a QP
477 * @ibqp: the QP to post the send on
478 * @wr: the list of work requests to post
479 * @bad_wr: the first bad WR is put here
481 * This may be called from interrupt context.
483 static int post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
484 struct ib_send_wr **bad_wr)
486 struct hfi1_qp *qp = to_iqp(ibqp);
492 spin_lock_irqsave(&qp->s_lock, flags);
494 /* Check that state is OK to post send. */
495 if (unlikely(!(ib_hfi1_state_ops[qp->state] & HFI1_POST_SEND_OK))) {
496 spin_unlock_irqrestore(&qp->s_lock, flags);
500 /* sq empty and not list -> call send */
501 call_send = qp->s_head == qp->s_last && !wr->next;
503 for (; wr; wr = wr->next) {
504 err = post_one_send(qp, wr);
512 if (nreq && !call_send)
513 hfi1_schedule_send(qp);
514 spin_unlock_irqrestore(&qp->s_lock, flags);
515 if (nreq && call_send)
516 hfi1_do_send(&qp->s_iowait.iowork);
521 * post_receive - post a receive on a QP
522 * @ibqp: the QP to post the receive on
523 * @wr: the WR to post
524 * @bad_wr: the first bad WR is put here
526 * This may be called from interrupt context.
528 static int post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
529 struct ib_recv_wr **bad_wr)
531 struct hfi1_qp *qp = to_iqp(ibqp);
532 struct hfi1_rwq *wq = qp->r_rq.wq;
536 /* Check that state is OK to post receive. */
537 if (!(ib_hfi1_state_ops[qp->state] & HFI1_POST_RECV_OK) || !wq) {
543 for (; wr; wr = wr->next) {
544 struct hfi1_rwqe *wqe;
548 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
554 spin_lock_irqsave(&qp->r_rq.lock, flags);
556 if (next >= qp->r_rq.size)
558 if (next == wq->tail) {
559 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
565 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
566 wqe->wr_id = wr->wr_id;
567 wqe->num_sge = wr->num_sge;
568 for (i = 0; i < wr->num_sge; i++)
569 wqe->sg_list[i] = wr->sg_list[i];
570 /* Make sure queue entry is written before the head index. */
573 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
582 * Make sure the QP is ready and able to accept the given opcode.
584 static inline int qp_ok(int opcode, struct hfi1_packet *packet)
586 struct hfi1_ibport *ibp;
588 if (!(ib_hfi1_state_ops[packet->qp->state] & HFI1_PROCESS_RECV_OK))
590 if (((opcode & OPCODE_QP_MASK) == packet->qp->allowed_ops) ||
591 (opcode == IB_OPCODE_CNP))
594 ibp = &packet->rcd->ppd->ibport_data;
601 * hfi1_ib_rcv - process an incoming packet
602 * @packet: data packet information
604 * This is called to process an incoming packet at interrupt level.
606 * Tlen is the length of the header + data + CRC in bytes.
608 void hfi1_ib_rcv(struct hfi1_packet *packet)
610 struct hfi1_ctxtdata *rcd = packet->rcd;
611 struct hfi1_ib_header *hdr = packet->hdr;
612 u32 tlen = packet->tlen;
613 struct hfi1_pportdata *ppd = rcd->ppd;
614 struct hfi1_ibport *ibp = &ppd->ibport_data;
622 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
623 if (lnh == HFI1_LRH_BTH)
624 packet->ohdr = &hdr->u.oth;
625 else if (lnh == HFI1_LRH_GRH) {
628 packet->ohdr = &hdr->u.l.oth;
629 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
631 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
632 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
634 packet->rcv_flags |= HFI1_HAS_GRH;
638 trace_input_ibhdr(rcd->dd, hdr);
640 opcode = (be32_to_cpu(packet->ohdr->bth[0]) >> 24);
641 inc_opstats(tlen, &rcd->opstats->stats[opcode]);
643 /* Get the destination QP number. */
644 qp_num = be32_to_cpu(packet->ohdr->bth[1]) & HFI1_QPN_MASK;
645 lid = be16_to_cpu(hdr->lrh[1]);
646 if (unlikely((lid >= HFI1_MULTICAST_LID_BASE) &&
647 (lid != HFI1_PERMISSIVE_LID))) {
648 struct hfi1_mcast *mcast;
649 struct hfi1_mcast_qp *p;
651 if (lnh != HFI1_LRH_GRH)
653 mcast = hfi1_mcast_find(ibp, &hdr->u.l.grh.dgid);
656 list_for_each_entry_rcu(p, &mcast->qp_list, list) {
658 spin_lock_irqsave(&packet->qp->r_lock, flags);
659 if (likely((qp_ok(opcode, packet))))
660 opcode_handler_tbl[opcode](packet);
661 spin_unlock_irqrestore(&packet->qp->r_lock, flags);
664 * Notify hfi1_multicast_detach() if it is waiting for us
667 if (atomic_dec_return(&mcast->refcount) <= 1)
668 wake_up(&mcast->wait);
671 packet->qp = hfi1_lookup_qpn(ibp, qp_num);
676 spin_lock_irqsave(&packet->qp->r_lock, flags);
677 if (likely((qp_ok(opcode, packet))))
678 opcode_handler_tbl[opcode](packet);
679 spin_unlock_irqrestore(&packet->qp->r_lock, flags);
689 * This is called from a timer to check for QPs
690 * which need kernel memory in order to send a packet.
692 static void mem_timer(unsigned long data)
694 struct hfi1_ibdev *dev = (struct hfi1_ibdev *)data;
695 struct list_head *list = &dev->memwait;
696 struct hfi1_qp *qp = NULL;
700 write_seqlock_irqsave(&dev->iowait_lock, flags);
701 if (!list_empty(list)) {
702 wait = list_first_entry(list, struct iowait, list);
703 qp = container_of(wait, struct hfi1_qp, s_iowait);
704 list_del_init(&qp->s_iowait.list);
705 /* refcount held until actual wake up */
706 if (!list_empty(list))
707 mod_timer(&dev->mem_timer, jiffies + 1);
709 write_sequnlock_irqrestore(&dev->iowait_lock, flags);
712 hfi1_qp_wakeup(qp, HFI1_S_WAIT_KMEM);
715 void update_sge(struct hfi1_sge_state *ss, u32 length)
717 struct hfi1_sge *sge = &ss->sge;
719 sge->vaddr += length;
720 sge->length -= length;
721 sge->sge_length -= length;
722 if (sge->sge_length == 0) {
724 *sge = *ss->sg_list++;
725 } else if (sge->length == 0 && sge->mr->lkey) {
726 if (++sge->n >= HFI1_SEGSZ) {
727 if (++sge->m >= sge->mr->mapsz)
731 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
732 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
736 static noinline struct verbs_txreq *__get_txreq(struct hfi1_ibdev *dev,
739 struct verbs_txreq *tx;
742 tx = kmem_cache_alloc(dev->verbs_txreq_cache, GFP_ATOMIC);
744 spin_lock_irqsave(&qp->s_lock, flags);
745 write_seqlock(&dev->iowait_lock);
746 if (ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK &&
747 list_empty(&qp->s_iowait.list)) {
749 qp->s_flags |= HFI1_S_WAIT_TX;
750 list_add_tail(&qp->s_iowait.list, &dev->txwait);
751 trace_hfi1_qpsleep(qp, HFI1_S_WAIT_TX);
752 atomic_inc(&qp->refcount);
754 qp->s_flags &= ~HFI1_S_BUSY;
755 write_sequnlock(&dev->iowait_lock);
756 spin_unlock_irqrestore(&qp->s_lock, flags);
757 tx = ERR_PTR(-EBUSY);
762 static inline struct verbs_txreq *get_txreq(struct hfi1_ibdev *dev,
765 struct verbs_txreq *tx;
767 tx = kmem_cache_alloc(dev->verbs_txreq_cache, GFP_ATOMIC);
769 /* call slow path to get the lock */
770 tx = __get_txreq(dev, qp);
778 void hfi1_put_txreq(struct verbs_txreq *tx)
780 struct hfi1_ibdev *dev;
786 dev = to_idev(qp->ibqp.device);
792 sdma_txclean(dd_from_dev(dev), &tx->txreq);
794 /* Free verbs_txreq and return to slab cache */
795 kmem_cache_free(dev->verbs_txreq_cache, tx);
798 seq = read_seqbegin(&dev->iowait_lock);
799 if (!list_empty(&dev->txwait)) {
802 write_seqlock_irqsave(&dev->iowait_lock, flags);
803 /* Wake up first QP wanting a free struct */
804 wait = list_first_entry(&dev->txwait, struct iowait,
806 qp = container_of(wait, struct hfi1_qp, s_iowait);
807 list_del_init(&qp->s_iowait.list);
808 /* refcount held until actual wake up */
809 write_sequnlock_irqrestore(&dev->iowait_lock, flags);
810 hfi1_qp_wakeup(qp, HFI1_S_WAIT_TX);
813 } while (read_seqretry(&dev->iowait_lock, seq));
817 * This is called with progress side lock held.
820 static void verbs_sdma_complete(
821 struct sdma_txreq *cookie,
825 struct verbs_txreq *tx =
826 container_of(cookie, struct verbs_txreq, txreq);
827 struct hfi1_qp *qp = tx->qp;
829 spin_lock(&qp->s_lock);
831 hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
832 else if (qp->ibqp.qp_type == IB_QPT_RC) {
833 struct hfi1_ib_header *hdr;
836 hfi1_rc_send_complete(qp, hdr);
840 * This happens when the send engine notes
841 * a QP in the error state and cannot
842 * do the flush work until that QP's
843 * sdma work has finished.
845 if (qp->s_flags & HFI1_S_WAIT_DMA) {
846 qp->s_flags &= ~HFI1_S_WAIT_DMA;
847 hfi1_schedule_send(qp);
850 spin_unlock(&qp->s_lock);
855 static int wait_kmem(struct hfi1_ibdev *dev, struct hfi1_qp *qp)
860 spin_lock_irqsave(&qp->s_lock, flags);
861 if (ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK) {
862 write_seqlock(&dev->iowait_lock);
863 if (list_empty(&qp->s_iowait.list)) {
864 if (list_empty(&dev->memwait))
865 mod_timer(&dev->mem_timer, jiffies + 1);
866 qp->s_flags |= HFI1_S_WAIT_KMEM;
867 list_add_tail(&qp->s_iowait.list, &dev->memwait);
868 trace_hfi1_qpsleep(qp, HFI1_S_WAIT_KMEM);
869 atomic_inc(&qp->refcount);
871 write_sequnlock(&dev->iowait_lock);
872 qp->s_flags &= ~HFI1_S_BUSY;
875 spin_unlock_irqrestore(&qp->s_lock, flags);
881 * This routine calls txadds for each sg entry.
883 * Add failures will revert the sge cursor
885 static int build_verbs_ulp_payload(
886 struct sdma_engine *sde,
887 struct hfi1_sge_state *ss,
889 struct verbs_txreq *tx)
891 struct hfi1_sge *sg_list = ss->sg_list;
892 struct hfi1_sge sge = ss->sge;
893 u8 num_sge = ss->num_sge;
898 len = ss->sge.length;
901 if (len > ss->sge.sge_length)
902 len = ss->sge.sge_length;
903 WARN_ON_ONCE(len == 0);
904 ret = sdma_txadd_kvaddr(
918 ss->num_sge = num_sge;
919 ss->sg_list = sg_list;
924 * Build the number of DMA descriptors needed to send length bytes of data.
926 * NOTE: DMA mapping is held in the tx until completed in the ring or
927 * the tx desc is freed without having been submitted to the ring
929 * This routine insures the following all the helper routine
933 static int build_verbs_tx_desc(
934 struct sdma_engine *sde,
935 struct hfi1_sge_state *ss,
937 struct verbs_txreq *tx,
938 struct ahg_ib_header *ahdr,
942 struct hfi1_pio_header *phdr;
943 u16 hdrbytes = tx->hdr_dwords << 2;
946 if (!ahdr->ahgcount) {
947 ret = sdma_txinit_ahg(
955 verbs_sdma_complete);
958 phdr->pbc = cpu_to_le64(pbc);
959 memcpy(&phdr->hdr, &ahdr->ibh, hdrbytes - sizeof(phdr->pbc));
961 ret = sdma_txadd_kvaddr(
965 tx->hdr_dwords << 2);
969 struct hfi1_other_headers *sohdr = &ahdr->ibh.u.oth;
970 struct hfi1_other_headers *dohdr = &phdr->hdr.u.oth;
972 /* needed in rc_send_complete() */
973 phdr->hdr.lrh[0] = ahdr->ibh.lrh[0];
974 if ((be16_to_cpu(phdr->hdr.lrh[0]) & 3) == HFI1_LRH_GRH) {
975 sohdr = &ahdr->ibh.u.l.oth;
976 dohdr = &phdr->hdr.u.l.oth;
979 dohdr->bth[0] = sohdr->bth[0];
981 dohdr->bth[2] = sohdr->bth[2];
982 ret = sdma_txinit_ahg(
990 verbs_sdma_complete);
995 /* add the ulp payload - if any. ss can be NULL for acks */
997 ret = build_verbs_ulp_payload(sde, ss, length, tx);
1002 int hfi1_verbs_send_dma(struct hfi1_qp *qp, struct ahg_ib_header *ahdr,
1003 u32 hdrwords, struct hfi1_sge_state *ss, u32 len,
1004 u32 plen, u32 dwords, u64 pbc)
1006 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
1007 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1008 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1009 struct verbs_txreq *tx;
1010 struct sdma_txreq *stx;
1012 struct sdma_engine *sde;
1016 if (!list_empty(&qp->s_iowait.tx_head)) {
1017 stx = list_first_entry(
1018 &qp->s_iowait.tx_head,
1021 list_del_init(&stx->list);
1022 tx = container_of(stx, struct verbs_txreq, txreq);
1023 ret = sdma_send_txreq(tx->sde, &qp->s_iowait, stx);
1024 if (unlikely(ret == -ECOMM))
1029 tx = get_txreq(dev, qp);
1033 if (!qp->s_hdr->sde) {
1034 tx->sde = sde = qp_to_sdma_engine(qp, sc5);
1038 tx->sde = sde = qp->s_hdr->sde;
1040 if (likely(pbc == 0)) {
1041 u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
1043 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1044 pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
1046 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
1048 tx->wqe = qp->s_wqe;
1049 tx->mr = qp->s_rdma_mr;
1051 qp->s_rdma_mr = NULL;
1052 tx->hdr_dwords = hdrwords + 2;
1053 ret = build_verbs_tx_desc(sde, ss, len, tx, ahdr, pbc);
1056 trace_output_ibhdr(dd_from_ibdev(qp->ibqp.device), &ahdr->ibh);
1057 ret = sdma_send_txreq(sde, &qp->s_iowait, &tx->txreq);
1058 if (unlikely(ret == -ECOMM))
1065 /* The current one got "sent" */
1068 /* kmalloc or mapping fail */
1070 return wait_kmem(dev, qp);
1076 * If we are now in the error state, return zero to flush the
1077 * send work request.
1079 static int no_bufs_available(struct hfi1_qp *qp, struct send_context *sc)
1081 struct hfi1_devdata *dd = sc->dd;
1082 struct hfi1_ibdev *dev = &dd->verbs_dev;
1083 unsigned long flags;
1087 * Note that as soon as want_buffer() is called and
1088 * possibly before it returns, sc_piobufavail()
1089 * could be called. Therefore, put QP on the I/O wait list before
1090 * enabling the PIO avail interrupt.
1092 spin_lock_irqsave(&qp->s_lock, flags);
1093 if (ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK) {
1094 write_seqlock(&dev->iowait_lock);
1095 if (list_empty(&qp->s_iowait.list)) {
1096 struct hfi1_ibdev *dev = &dd->verbs_dev;
1100 qp->s_flags |= HFI1_S_WAIT_PIO;
1101 was_empty = list_empty(&sc->piowait);
1102 list_add_tail(&qp->s_iowait.list, &sc->piowait);
1103 trace_hfi1_qpsleep(qp, HFI1_S_WAIT_PIO);
1104 atomic_inc(&qp->refcount);
1105 /* counting: only call wantpiobuf_intr if first user */
1107 hfi1_sc_wantpiobuf_intr(sc, 1);
1109 write_sequnlock(&dev->iowait_lock);
1110 qp->s_flags &= ~HFI1_S_BUSY;
1113 spin_unlock_irqrestore(&qp->s_lock, flags);
1117 struct send_context *qp_to_send_context(struct hfi1_qp *qp, u8 sc5)
1119 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1120 struct hfi1_pportdata *ppd = dd->pport + (qp->port_num - 1);
1123 vl = sc_to_vlt(dd, sc5);
1124 if (vl >= ppd->vls_supported && vl != 15)
1126 return dd->vld[vl].sc;
1129 int hfi1_verbs_send_pio(struct hfi1_qp *qp, struct ahg_ib_header *ahdr,
1130 u32 hdrwords, struct hfi1_sge_state *ss, u32 len,
1131 u32 plen, u32 dwords, u64 pbc)
1133 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1134 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1135 u32 *hdr = (u32 *)&ahdr->ibh;
1138 unsigned long flags = 0;
1139 struct send_context *sc;
1140 struct pio_buf *pbuf;
1141 int wc_status = IB_WC_SUCCESS;
1143 /* vl15 special case taken care of in ud.c */
1145 sc = qp_to_send_context(qp, sc5);
1149 if (likely(pbc == 0)) {
1150 u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
1151 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1152 pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
1153 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
1155 pbuf = sc_buffer_alloc(sc, plen, NULL, NULL);
1156 if (unlikely(pbuf == NULL)) {
1157 if (ppd->host_link_state != HLS_UP_ACTIVE) {
1159 * If we have filled the PIO buffers to capacity and are
1160 * not in an active state this request is not going to
1161 * go out to so just complete it with an error or else a
1162 * ULP or the core may be stuck waiting.
1166 "alloc failed. state not active, completing");
1167 wc_status = IB_WC_GENERAL_ERR;
1171 * This is a normal occurrence. The PIO buffs are full
1172 * up but we are still happily sending, well we could be
1173 * so lets continue to queue the request.
1175 hfi1_cdbg(PIO, "alloc failed. state active, queuing");
1176 return no_bufs_available(qp, sc);
1181 pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
1184 seg_pio_copy_start(pbuf, pbc, hdr, hdrwords*4);
1186 void *addr = ss->sge.vaddr;
1187 u32 slen = ss->sge.length;
1191 update_sge(ss, slen);
1192 seg_pio_copy_mid(pbuf, addr, slen);
1195 seg_pio_copy_end(pbuf);
1199 trace_output_ibhdr(dd_from_ibdev(qp->ibqp.device), &ahdr->ibh);
1201 if (qp->s_rdma_mr) {
1202 hfi1_put_mr(qp->s_rdma_mr);
1203 qp->s_rdma_mr = NULL;
1208 spin_lock_irqsave(&qp->s_lock, flags);
1209 hfi1_send_complete(qp, qp->s_wqe, wc_status);
1210 spin_unlock_irqrestore(&qp->s_lock, flags);
1211 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1212 spin_lock_irqsave(&qp->s_lock, flags);
1213 hfi1_rc_send_complete(qp, &ahdr->ibh);
1214 spin_unlock_irqrestore(&qp->s_lock, flags);
1220 * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1221 * being an entry from the ingress partition key table), return 0
1222 * otherwise. Use the matching criteria for egress partition keys
1223 * specified in the OPAv1 spec., section 9.1l.7.
1225 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
1227 u16 mkey = pkey & PKEY_LOW_15_MASK;
1228 u16 ment = ent & PKEY_LOW_15_MASK;
1232 * If pkey[15] is set (full partition member),
1233 * is bit 15 in the corresponding table element
1234 * clear (limited member)?
1236 if (pkey & PKEY_MEMBER_MASK)
1237 return !!(ent & PKEY_MEMBER_MASK);
1244 * egress_pkey_check - return 0 if hdr's pkey matches according to the
1245 * criteria in the OPAv1 spec., section 9.11.7.
1247 static inline int egress_pkey_check(struct hfi1_pportdata *ppd,
1248 struct hfi1_ib_header *hdr,
1251 struct hfi1_other_headers *ohdr;
1252 struct hfi1_devdata *dd;
1255 u8 lnh, sc5 = qp->s_sc;
1257 if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1260 /* locate the pkey within the headers */
1261 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
1262 if (lnh == HFI1_LRH_GRH)
1263 ohdr = &hdr->u.l.oth;
1267 pkey = (u16)be32_to_cpu(ohdr->bth[0]);
1269 /* If SC15, pkey[0:14] must be 0x7fff */
1270 if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1274 /* Is the pkey = 0x0, or 0x8000? */
1275 if ((pkey & PKEY_LOW_15_MASK) == 0)
1278 /* The most likely matching pkey has index qp->s_pkey_index */
1279 if (unlikely(!egress_pkey_matches_entry(pkey,
1280 ppd->pkeys[qp->s_pkey_index]))) {
1281 /* no match - try the entire table */
1282 for (; i < MAX_PKEY_VALUES; i++) {
1283 if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1288 if (i < MAX_PKEY_VALUES)
1291 incr_cntr64(&ppd->port_xmit_constraint_errors);
1293 if (!(dd->err_info_xmit_constraint.status & OPA_EI_STATUS_SMASK)) {
1294 u16 slid = be16_to_cpu(hdr->lrh[3]);
1296 dd->err_info_xmit_constraint.status |= OPA_EI_STATUS_SMASK;
1297 dd->err_info_xmit_constraint.slid = slid;
1298 dd->err_info_xmit_constraint.pkey = pkey;
1304 * hfi1_verbs_send - send a packet
1305 * @qp: the QP to send on
1306 * @ahdr: the packet header
1307 * @hdrwords: the number of 32-bit words in the header
1308 * @ss: the SGE to send
1309 * @len: the length of the packet in bytes
1311 * Return zero if packet is sent or queued OK.
1312 * Return non-zero and clear qp->s_flags HFI1_S_BUSY otherwise.
1314 int hfi1_verbs_send(struct hfi1_qp *qp, struct ahg_ib_header *ahdr,
1315 u32 hdrwords, struct hfi1_sge_state *ss, u32 len)
1317 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1321 unsigned long flags = 0;
1322 u32 dwords = (len + 3) >> 2;
1325 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1326 * can defer SDMA restart until link goes ACTIVE without
1327 * worrying about just how we got there.
1329 if ((qp->ibqp.qp_type == IB_QPT_SMI) ||
1330 !(dd->flags & HFI1_HAS_SEND_DMA))
1333 ret = egress_pkey_check(dd->pport, &ahdr->ibh, qp);
1334 if (unlikely(ret)) {
1336 * The value we are returning here does not get propagated to
1337 * the verbs caller. Thus we need to complete the request with
1338 * error otherwise the caller could be sitting waiting on the
1339 * completion event. Only do this for PIO. SDMA has its own
1340 * mechanism for handling the errors. So for SDMA we can just
1344 hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1346 spin_lock_irqsave(&qp->s_lock, flags);
1347 hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1348 spin_unlock_irqrestore(&qp->s_lock, flags);
1354 * Calculate the send buffer trigger address.
1355 * The +2 counts for the pbc control qword
1357 plen = hdrwords + dwords + 2;
1360 ret = dd->process_pio_send(
1361 qp, ahdr, hdrwords, ss, len, plen, dwords, 0);
1363 #ifdef CONFIG_SDMA_VERBOSITY
1364 dd_dev_err(dd, "CONFIG SDMA %s:%d %s()\n",
1365 slashstrip(__FILE__), __LINE__, __func__);
1366 dd_dev_err(dd, "SDMA hdrwords = %u, len = %u\n", hdrwords, len);
1368 ret = dd->process_dma_send(
1369 qp, ahdr, hdrwords, ss, len, plen, dwords, 0);
1375 static int query_device(struct ib_device *ibdev,
1376 struct ib_device_attr *props,
1377 struct ib_udata *uhw)
1379 struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1380 struct hfi1_ibdev *dev = to_idev(ibdev);
1382 if (uhw->inlen || uhw->outlen)
1384 memset(props, 0, sizeof(*props));
1386 props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1387 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1388 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1389 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1391 props->page_size_cap = PAGE_SIZE;
1393 dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1394 props->vendor_part_id = dd->pcidev->device;
1395 props->hw_ver = dd->minrev;
1396 props->sys_image_guid = ib_hfi1_sys_image_guid;
1397 props->max_mr_size = ~0ULL;
1398 props->max_qp = hfi1_max_qps;
1399 props->max_qp_wr = hfi1_max_qp_wrs;
1400 props->max_sge = hfi1_max_sges;
1401 props->max_sge_rd = hfi1_max_sges;
1402 props->max_cq = hfi1_max_cqs;
1403 props->max_ah = hfi1_max_ahs;
1404 props->max_cqe = hfi1_max_cqes;
1405 props->max_mr = dev->lk_table.max;
1406 props->max_fmr = dev->lk_table.max;
1407 props->max_map_per_fmr = 32767;
1408 props->max_pd = hfi1_max_pds;
1409 props->max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1410 props->max_qp_init_rd_atom = 255;
1411 /* props->max_res_rd_atom */
1412 props->max_srq = hfi1_max_srqs;
1413 props->max_srq_wr = hfi1_max_srq_wrs;
1414 props->max_srq_sge = hfi1_max_srq_sges;
1415 /* props->local_ca_ack_delay */
1416 props->atomic_cap = IB_ATOMIC_GLOB;
1417 props->max_pkeys = hfi1_get_npkeys(dd);
1418 props->max_mcast_grp = hfi1_max_mcast_grps;
1419 props->max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1420 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1421 props->max_mcast_grp;
1426 static inline u16 opa_speed_to_ib(u16 in)
1430 if (in & OPA_LINK_SPEED_25G)
1431 out |= IB_SPEED_EDR;
1432 if (in & OPA_LINK_SPEED_12_5G)
1433 out |= IB_SPEED_FDR;
1439 * Convert a single OPA link width (no multiple flags) to an IB value.
1440 * A zero OPA link width means link down, which means the IB width value
1443 static inline u16 opa_width_to_ib(u16 in)
1446 case OPA_LINK_WIDTH_1X:
1447 /* map 2x and 3x to 1x as they don't exist in IB */
1448 case OPA_LINK_WIDTH_2X:
1449 case OPA_LINK_WIDTH_3X:
1451 default: /* link down or unknown, return our largest width */
1452 case OPA_LINK_WIDTH_4X:
1457 static int query_port(struct ib_device *ibdev, u8 port,
1458 struct ib_port_attr *props)
1460 struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1461 struct hfi1_ibport *ibp = to_iport(ibdev, port);
1462 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1465 memset(props, 0, sizeof(*props));
1466 props->lid = lid ? lid : 0;
1467 props->lmc = ppd->lmc;
1468 props->sm_lid = ibp->sm_lid;
1469 props->sm_sl = ibp->sm_sl;
1470 /* OPA logical states match IB logical states */
1471 props->state = driver_lstate(ppd);
1472 props->phys_state = hfi1_ibphys_portstate(ppd);
1473 props->port_cap_flags = ibp->port_cap_flags;
1474 props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1475 props->max_msg_sz = 0x80000000;
1476 props->pkey_tbl_len = hfi1_get_npkeys(dd);
1477 props->bad_pkey_cntr = ibp->pkey_violations;
1478 props->qkey_viol_cntr = ibp->qkey_violations;
1479 props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1480 /* see rate_show() in ib core/sysfs.c */
1481 props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
1482 props->max_vl_num = ppd->vls_supported;
1483 props->init_type_reply = 0;
1485 /* Once we are a "first class" citizen and have added the OPA MTUs to
1486 * the core we can advertise the larger MTU enum to the ULPs, for now
1487 * advertise only 4K.
1489 * Those applications which are either OPA aware or pass the MTU enum
1490 * from the Path Records to us will get the new 8k MTU. Those that
1491 * attempt to process the MTU enum may fail in various ways.
1493 props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1494 4096 : hfi1_max_mtu), IB_MTU_4096);
1495 props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1496 mtu_to_enum(ppd->ibmtu, IB_MTU_2048);
1497 props->subnet_timeout = ibp->subnet_timeout;
1502 static int port_immutable(struct ib_device *ibdev, u8 port_num,
1503 struct ib_port_immutable *immutable)
1505 struct ib_port_attr attr;
1508 err = query_port(ibdev, port_num, &attr);
1512 memset(immutable, 0, sizeof(*immutable));
1514 immutable->pkey_tbl_len = attr.pkey_tbl_len;
1515 immutable->gid_tbl_len = attr.gid_tbl_len;
1516 immutable->core_cap_flags = RDMA_CORE_PORT_INTEL_OPA;
1517 immutable->max_mad_size = OPA_MGMT_MAD_SIZE;
1522 static int modify_device(struct ib_device *device,
1523 int device_modify_mask,
1524 struct ib_device_modify *device_modify)
1526 struct hfi1_devdata *dd = dd_from_ibdev(device);
1530 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1531 IB_DEVICE_MODIFY_NODE_DESC)) {
1536 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1537 memcpy(device->node_desc, device_modify->node_desc, 64);
1538 for (i = 0; i < dd->num_pports; i++) {
1539 struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1541 hfi1_node_desc_chg(ibp);
1545 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1546 ib_hfi1_sys_image_guid =
1547 cpu_to_be64(device_modify->sys_image_guid);
1548 for (i = 0; i < dd->num_pports; i++) {
1549 struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1551 hfi1_sys_guid_chg(ibp);
1561 static int modify_port(struct ib_device *ibdev, u8 port,
1562 int port_modify_mask, struct ib_port_modify *props)
1564 struct hfi1_ibport *ibp = to_iport(ibdev, port);
1565 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1568 ibp->port_cap_flags |= props->set_port_cap_mask;
1569 ibp->port_cap_flags &= ~props->clr_port_cap_mask;
1570 if (props->set_port_cap_mask || props->clr_port_cap_mask)
1571 hfi1_cap_mask_chg(ibp);
1572 if (port_modify_mask & IB_PORT_SHUTDOWN) {
1573 set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1574 OPA_LINKDOWN_REASON_UNKNOWN);
1575 ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1577 if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1578 ibp->qkey_violations = 0;
1582 static int query_gid(struct ib_device *ibdev, u8 port,
1583 int index, union ib_gid *gid)
1585 struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1588 if (!port || port > dd->num_pports)
1591 struct hfi1_ibport *ibp = to_iport(ibdev, port);
1592 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1594 gid->global.subnet_prefix = ibp->gid_prefix;
1596 gid->global.interface_id = cpu_to_be64(ppd->guid);
1597 else if (index < HFI1_GUIDS_PER_PORT)
1598 gid->global.interface_id = ibp->guids[index - 1];
1606 static struct ib_pd *alloc_pd(struct ib_device *ibdev,
1607 struct ib_ucontext *context,
1608 struct ib_udata *udata)
1610 struct hfi1_ibdev *dev = to_idev(ibdev);
1615 * This is actually totally arbitrary. Some correctness tests
1616 * assume there's a maximum number of PDs that can be allocated.
1617 * We don't actually have this limit, but we fail the test if
1618 * we allow allocations of more than we report for this value.
1621 pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1623 ret = ERR_PTR(-ENOMEM);
1627 spin_lock(&dev->n_pds_lock);
1628 if (dev->n_pds_allocated == hfi1_max_pds) {
1629 spin_unlock(&dev->n_pds_lock);
1631 ret = ERR_PTR(-ENOMEM);
1635 dev->n_pds_allocated++;
1636 spin_unlock(&dev->n_pds_lock);
1638 /* ib_alloc_pd() will initialize pd->ibpd. */
1639 pd->user = udata != NULL;
1647 static int dealloc_pd(struct ib_pd *ibpd)
1649 struct hfi1_pd *pd = to_ipd(ibpd);
1650 struct hfi1_ibdev *dev = to_idev(ibpd->device);
1652 spin_lock(&dev->n_pds_lock);
1653 dev->n_pds_allocated--;
1654 spin_unlock(&dev->n_pds_lock);
1662 * convert ah port,sl to sc
1664 u8 ah_to_sc(struct ib_device *ibdev, struct ib_ah_attr *ah)
1666 struct hfi1_ibport *ibp = to_iport(ibdev, ah->port_num);
1668 return ibp->sl_to_sc[ah->sl];
1671 int hfi1_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
1673 struct hfi1_ibport *ibp;
1674 struct hfi1_pportdata *ppd;
1675 struct hfi1_devdata *dd;
1678 /* A multicast address requires a GRH (see ch. 8.4.1). */
1679 if (ah_attr->dlid >= HFI1_MULTICAST_LID_BASE &&
1680 ah_attr->dlid != HFI1_PERMISSIVE_LID &&
1681 !(ah_attr->ah_flags & IB_AH_GRH))
1683 if ((ah_attr->ah_flags & IB_AH_GRH) &&
1684 ah_attr->grh.sgid_index >= HFI1_GUIDS_PER_PORT)
1686 if (ah_attr->dlid == 0)
1688 if (ah_attr->port_num < 1 ||
1689 ah_attr->port_num > ibdev->phys_port_cnt)
1691 if (ah_attr->static_rate != IB_RATE_PORT_CURRENT &&
1692 ib_rate_to_mbps(ah_attr->static_rate) < 0)
1694 if (ah_attr->sl >= OPA_MAX_SLS)
1696 /* test the mapping for validity */
1697 ibp = to_iport(ibdev, ah_attr->port_num);
1698 ppd = ppd_from_ibp(ibp);
1699 sc5 = ibp->sl_to_sc[ah_attr->sl];
1700 dd = dd_from_ppd(ppd);
1701 if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1709 * create_ah - create an address handle
1710 * @pd: the protection domain
1711 * @ah_attr: the attributes of the AH
1713 * This may be called from interrupt context.
1715 static struct ib_ah *create_ah(struct ib_pd *pd,
1716 struct ib_ah_attr *ah_attr)
1720 struct hfi1_ibdev *dev = to_idev(pd->device);
1721 unsigned long flags;
1723 if (hfi1_check_ah(pd->device, ah_attr)) {
1724 ret = ERR_PTR(-EINVAL);
1728 ah = kmalloc(sizeof(*ah), GFP_ATOMIC);
1730 ret = ERR_PTR(-ENOMEM);
1734 spin_lock_irqsave(&dev->n_ahs_lock, flags);
1735 if (dev->n_ahs_allocated == hfi1_max_ahs) {
1736 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1738 ret = ERR_PTR(-ENOMEM);
1742 dev->n_ahs_allocated++;
1743 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1745 /* ib_create_ah() will initialize ah->ibah. */
1746 ah->attr = *ah_attr;
1747 atomic_set(&ah->refcount, 0);
1755 struct ib_ah *hfi1_create_qp0_ah(struct hfi1_ibport *ibp, u16 dlid)
1757 struct ib_ah_attr attr;
1758 struct ib_ah *ah = ERR_PTR(-EINVAL);
1759 struct hfi1_qp *qp0;
1761 memset(&attr, 0, sizeof(attr));
1763 attr.port_num = ppd_from_ibp(ibp)->port;
1765 qp0 = rcu_dereference(ibp->qp[0]);
1767 ah = ib_create_ah(qp0->ibqp.pd, &attr);
1773 * destroy_ah - destroy an address handle
1774 * @ibah: the AH to destroy
1776 * This may be called from interrupt context.
1778 static int destroy_ah(struct ib_ah *ibah)
1780 struct hfi1_ibdev *dev = to_idev(ibah->device);
1781 struct hfi1_ah *ah = to_iah(ibah);
1782 unsigned long flags;
1784 if (atomic_read(&ah->refcount) != 0)
1787 spin_lock_irqsave(&dev->n_ahs_lock, flags);
1788 dev->n_ahs_allocated--;
1789 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1796 static int modify_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1798 struct hfi1_ah *ah = to_iah(ibah);
1800 if (hfi1_check_ah(ibah->device, ah_attr))
1803 ah->attr = *ah_attr;
1808 static int query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1810 struct hfi1_ah *ah = to_iah(ibah);
1812 *ah_attr = ah->attr;
1818 * hfi1_get_npkeys - return the size of the PKEY table for context 0
1819 * @dd: the hfi1_ib device
1821 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1823 return ARRAY_SIZE(dd->pport[0].pkeys);
1826 static int query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1829 struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1832 if (index >= hfi1_get_npkeys(dd)) {
1837 *pkey = hfi1_get_pkey(to_iport(ibdev, port), index);
1845 * alloc_ucontext - allocate a ucontest
1846 * @ibdev: the infiniband device
1847 * @udata: not used by the driver
1850 static struct ib_ucontext *alloc_ucontext(struct ib_device *ibdev,
1851 struct ib_udata *udata)
1853 struct hfi1_ucontext *context;
1854 struct ib_ucontext *ret;
1856 context = kmalloc(sizeof(*context), GFP_KERNEL);
1858 ret = ERR_PTR(-ENOMEM);
1862 ret = &context->ibucontext;
1868 static int dealloc_ucontext(struct ib_ucontext *context)
1870 kfree(to_iucontext(context));
1874 static void init_ibport(struct hfi1_pportdata *ppd)
1876 struct hfi1_ibport *ibp = &ppd->ibport_data;
1877 size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1880 for (i = 0; i < sz; i++) {
1881 ibp->sl_to_sc[i] = i;
1882 ibp->sc_to_sl[i] = i;
1885 spin_lock_init(&ibp->lock);
1886 /* Set the prefix to the default value (see ch. 4.1.1) */
1887 ibp->gid_prefix = IB_DEFAULT_GID_PREFIX;
1889 /* Below should only set bits defined in OPA PortInfo.CapabilityMask */
1890 ibp->port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1891 IB_PORT_CAP_MASK_NOTICE_SUP;
1892 ibp->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1893 ibp->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1894 ibp->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1895 ibp->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1896 ibp->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1898 RCU_INIT_POINTER(ibp->qp[0], NULL);
1899 RCU_INIT_POINTER(ibp->qp[1], NULL);
1902 static void verbs_txreq_kmem_cache_ctor(void *obj)
1904 struct verbs_txreq *tx = obj;
1906 memset(tx, 0, sizeof(*tx));
1910 * hfi1_register_ib_device - register our device with the infiniband core
1911 * @dd: the device data structure
1912 * Return 0 if successful, errno if unsuccessful.
1914 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1916 struct hfi1_ibdev *dev = &dd->verbs_dev;
1917 struct ib_device *ibdev = &dev->ibdev;
1918 struct hfi1_pportdata *ppd = dd->pport;
1919 unsigned i, lk_tab_size;
1921 size_t lcpysz = IB_DEVICE_NAME_MAX;
1923 char buf[TXREQ_NAME_LEN];
1925 ret = hfi1_qp_init(dev);
1930 for (i = 0; i < dd->num_pports; i++)
1931 init_ibport(ppd + i);
1933 /* Only need to initialize non-zero fields. */
1934 spin_lock_init(&dev->n_pds_lock);
1935 spin_lock_init(&dev->n_ahs_lock);
1936 spin_lock_init(&dev->n_cqs_lock);
1937 spin_lock_init(&dev->n_qps_lock);
1938 spin_lock_init(&dev->n_srqs_lock);
1939 spin_lock_init(&dev->n_mcast_grps_lock);
1940 init_timer(&dev->mem_timer);
1941 dev->mem_timer.function = mem_timer;
1942 dev->mem_timer.data = (unsigned long) dev;
1945 * The top hfi1_lkey_table_size bits are used to index the
1946 * table. The lower 8 bits can be owned by the user (copied from
1947 * the LKEY). The remaining bits act as a generation number or tag.
1949 spin_lock_init(&dev->lk_table.lock);
1950 dev->lk_table.max = 1 << hfi1_lkey_table_size;
1951 /* ensure generation is at least 4 bits (keys.c) */
1952 if (hfi1_lkey_table_size > MAX_LKEY_TABLE_BITS) {
1953 dd_dev_warn(dd, "lkey bits %u too large, reduced to %u\n",
1954 hfi1_lkey_table_size, MAX_LKEY_TABLE_BITS);
1955 hfi1_lkey_table_size = MAX_LKEY_TABLE_BITS;
1957 lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
1958 dev->lk_table.table = (struct hfi1_mregion __rcu **)
1959 vmalloc(lk_tab_size);
1960 if (dev->lk_table.table == NULL) {
1964 RCU_INIT_POINTER(dev->dma_mr, NULL);
1965 for (i = 0; i < dev->lk_table.max; i++)
1966 RCU_INIT_POINTER(dev->lk_table.table[i], NULL);
1967 INIT_LIST_HEAD(&dev->pending_mmaps);
1968 spin_lock_init(&dev->pending_lock);
1969 seqlock_init(&dev->iowait_lock);
1970 dev->mmap_offset = PAGE_SIZE;
1971 spin_lock_init(&dev->mmap_offset_lock);
1972 INIT_LIST_HEAD(&dev->txwait);
1973 INIT_LIST_HEAD(&dev->memwait);
1975 descq_cnt = sdma_get_descq_cnt();
1977 snprintf(buf, sizeof(buf), "hfi1_%u_vtxreq_cache", dd->unit);
1978 /* SLAB_HWCACHE_ALIGN for AHG */
1979 dev->verbs_txreq_cache = kmem_cache_create(buf,
1980 sizeof(struct verbs_txreq),
1981 0, SLAB_HWCACHE_ALIGN,
1982 verbs_txreq_kmem_cache_ctor);
1983 if (!dev->verbs_txreq_cache) {
1985 goto err_verbs_txreq;
1989 * The system image GUID is supposed to be the same for all
1990 * HFIs in a single system but since there can be other
1991 * device types in the system, we can't be sure this is unique.
1993 if (!ib_hfi1_sys_image_guid)
1994 ib_hfi1_sys_image_guid = cpu_to_be64(ppd->guid);
1995 lcpysz = strlcpy(ibdev->name, class_name(), lcpysz);
1996 strlcpy(ibdev->name + lcpysz, "_%d", IB_DEVICE_NAME_MAX - lcpysz);
1997 ibdev->owner = THIS_MODULE;
1998 ibdev->node_guid = cpu_to_be64(ppd->guid);
1999 ibdev->uverbs_abi_ver = HFI1_UVERBS_ABI_VERSION;
2000 ibdev->uverbs_cmd_mask =
2001 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
2002 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
2003 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
2004 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
2005 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
2006 (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
2007 (1ull << IB_USER_VERBS_CMD_MODIFY_AH) |
2008 (1ull << IB_USER_VERBS_CMD_QUERY_AH) |
2009 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) |
2010 (1ull << IB_USER_VERBS_CMD_REG_MR) |
2011 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
2012 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2013 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
2014 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
2015 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
2016 (1ull << IB_USER_VERBS_CMD_POLL_CQ) |
2017 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) |
2018 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
2019 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
2020 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
2021 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
2022 (1ull << IB_USER_VERBS_CMD_POST_SEND) |
2023 (1ull << IB_USER_VERBS_CMD_POST_RECV) |
2024 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
2025 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
2026 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
2027 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
2028 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
2029 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
2030 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2031 ibdev->node_type = RDMA_NODE_IB_CA;
2032 ibdev->phys_port_cnt = dd->num_pports;
2033 ibdev->num_comp_vectors = 1;
2034 ibdev->dma_device = &dd->pcidev->dev;
2035 ibdev->query_device = query_device;
2036 ibdev->modify_device = modify_device;
2037 ibdev->query_port = query_port;
2038 ibdev->modify_port = modify_port;
2039 ibdev->query_pkey = query_pkey;
2040 ibdev->query_gid = query_gid;
2041 ibdev->alloc_ucontext = alloc_ucontext;
2042 ibdev->dealloc_ucontext = dealloc_ucontext;
2043 ibdev->alloc_pd = alloc_pd;
2044 ibdev->dealloc_pd = dealloc_pd;
2045 ibdev->create_ah = create_ah;
2046 ibdev->destroy_ah = destroy_ah;
2047 ibdev->modify_ah = modify_ah;
2048 ibdev->query_ah = query_ah;
2049 ibdev->create_srq = hfi1_create_srq;
2050 ibdev->modify_srq = hfi1_modify_srq;
2051 ibdev->query_srq = hfi1_query_srq;
2052 ibdev->destroy_srq = hfi1_destroy_srq;
2053 ibdev->create_qp = hfi1_create_qp;
2054 ibdev->modify_qp = hfi1_modify_qp;
2055 ibdev->query_qp = hfi1_query_qp;
2056 ibdev->destroy_qp = hfi1_destroy_qp;
2057 ibdev->post_send = post_send;
2058 ibdev->post_recv = post_receive;
2059 ibdev->post_srq_recv = hfi1_post_srq_receive;
2060 ibdev->create_cq = hfi1_create_cq;
2061 ibdev->destroy_cq = hfi1_destroy_cq;
2062 ibdev->resize_cq = hfi1_resize_cq;
2063 ibdev->poll_cq = hfi1_poll_cq;
2064 ibdev->req_notify_cq = hfi1_req_notify_cq;
2065 ibdev->get_dma_mr = hfi1_get_dma_mr;
2066 ibdev->reg_phys_mr = hfi1_reg_phys_mr;
2067 ibdev->reg_user_mr = hfi1_reg_user_mr;
2068 ibdev->dereg_mr = hfi1_dereg_mr;
2069 ibdev->alloc_mr = hfi1_alloc_mr;
2070 ibdev->alloc_fmr = hfi1_alloc_fmr;
2071 ibdev->map_phys_fmr = hfi1_map_phys_fmr;
2072 ibdev->unmap_fmr = hfi1_unmap_fmr;
2073 ibdev->dealloc_fmr = hfi1_dealloc_fmr;
2074 ibdev->attach_mcast = hfi1_multicast_attach;
2075 ibdev->detach_mcast = hfi1_multicast_detach;
2076 ibdev->process_mad = hfi1_process_mad;
2077 ibdev->mmap = hfi1_mmap;
2078 ibdev->dma_ops = &hfi1_dma_mapping_ops;
2079 ibdev->get_port_immutable = port_immutable;
2081 strncpy(ibdev->node_desc, init_utsname()->nodename,
2082 sizeof(ibdev->node_desc));
2084 ret = ib_register_device(ibdev, hfi1_create_port_files);
2088 ret = hfi1_create_agents(dev);
2092 ret = hfi1_verbs_register_sysfs(dd);
2099 hfi1_free_agents(dev);
2101 ib_unregister_device(ibdev);
2104 kmem_cache_destroy(dev->verbs_txreq_cache);
2105 vfree(dev->lk_table.table);
2109 dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2114 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
2116 struct hfi1_ibdev *dev = &dd->verbs_dev;
2117 struct ib_device *ibdev = &dev->ibdev;
2119 hfi1_verbs_unregister_sysfs(dd);
2121 hfi1_free_agents(dev);
2123 ib_unregister_device(ibdev);
2125 if (!list_empty(&dev->txwait))
2126 dd_dev_err(dd, "txwait list not empty!\n");
2127 if (!list_empty(&dev->memwait))
2128 dd_dev_err(dd, "memwait list not empty!\n");
2130 dd_dev_err(dd, "DMA MR not NULL!\n");
2133 del_timer_sync(&dev->mem_timer);
2134 kmem_cache_destroy(dev->verbs_txreq_cache);
2135 vfree(dev->lk_table.table);
2139 * This must be called with s_lock held.
2141 void hfi1_schedule_send(struct hfi1_qp *qp)
2143 if (hfi1_send_ok(qp)) {
2144 struct hfi1_ibport *ibp =
2145 to_iport(qp->ibqp.device, qp->port_num);
2146 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
2148 iowait_schedule(&qp->s_iowait, ppd->hfi1_wq);
2152 void hfi1_cnp_rcv(struct hfi1_packet *packet)
2154 struct hfi1_ibport *ibp = &packet->rcd->ppd->ibport_data;
2156 if (packet->qp->ibqp.qp_type == IB_QPT_UC)
2157 hfi1_uc_rcv(packet);
2158 else if (packet->qp->ibqp.qp_type == IB_QPT_UD)
2159 hfi1_ud_rcv(packet);