GNU Linux-libre 4.9.287-gnu1
[releases.git] / drivers / infiniband / hw / hfi1 / ruc.c
1 /*
2  * Copyright(c) 2015 - 2017 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47
48 #include <linux/spinlock.h>
49
50 #include "hfi.h"
51 #include "mad.h"
52 #include "qp.h"
53 #include "verbs_txreq.h"
54 #include "trace.h"
55
56 /*
57  * Convert the AETH RNR timeout code into the number of microseconds.
58  */
59 const u32 ib_hfi1_rnr_table[32] = {
60         655360, /* 00: 655.36 */
61         10,     /* 01:    .01 */
62         20,     /* 02     .02 */
63         30,     /* 03:    .03 */
64         40,     /* 04:    .04 */
65         60,     /* 05:    .06 */
66         80,     /* 06:    .08 */
67         120,    /* 07:    .12 */
68         160,    /* 08:    .16 */
69         240,    /* 09:    .24 */
70         320,    /* 0A:    .32 */
71         480,    /* 0B:    .48 */
72         640,    /* 0C:    .64 */
73         960,    /* 0D:    .96 */
74         1280,   /* 0E:   1.28 */
75         1920,   /* 0F:   1.92 */
76         2560,   /* 10:   2.56 */
77         3840,   /* 11:   3.84 */
78         5120,   /* 12:   5.12 */
79         7680,   /* 13:   7.68 */
80         10240,  /* 14:  10.24 */
81         15360,  /* 15:  15.36 */
82         20480,  /* 16:  20.48 */
83         30720,  /* 17:  30.72 */
84         40960,  /* 18:  40.96 */
85         61440,  /* 19:  61.44 */
86         81920,  /* 1A:  81.92 */
87         122880, /* 1B: 122.88 */
88         163840, /* 1C: 163.84 */
89         245760, /* 1D: 245.76 */
90         327680, /* 1E: 327.68 */
91         491520  /* 1F: 491.52 */
92 };
93
94 /*
95  * Validate a RWQE and fill in the SGE state.
96  * Return 1 if OK.
97  */
98 static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
99 {
100         int i, j, ret;
101         struct ib_wc wc;
102         struct rvt_lkey_table *rkt;
103         struct rvt_pd *pd;
104         struct rvt_sge_state *ss;
105
106         rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table;
107         pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
108         ss = &qp->r_sge;
109         ss->sg_list = qp->r_sg_list;
110         qp->r_len = 0;
111         for (i = j = 0; i < wqe->num_sge; i++) {
112                 if (wqe->sg_list[i].length == 0)
113                         continue;
114                 /* Check LKEY */
115                 if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
116                                  &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
117                         goto bad_lkey;
118                 qp->r_len += wqe->sg_list[i].length;
119                 j++;
120         }
121         ss->num_sge = j;
122         ss->total_len = qp->r_len;
123         ret = 1;
124         goto bail;
125
126 bad_lkey:
127         while (j) {
128                 struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
129
130                 rvt_put_mr(sge->mr);
131         }
132         ss->num_sge = 0;
133         memset(&wc, 0, sizeof(wc));
134         wc.wr_id = wqe->wr_id;
135         wc.status = IB_WC_LOC_PROT_ERR;
136         wc.opcode = IB_WC_RECV;
137         wc.qp = &qp->ibqp;
138         /* Signal solicited completion event. */
139         rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
140         ret = 0;
141 bail:
142         return ret;
143 }
144
145 /**
146  * hfi1_rvt_get_rwqe - copy the next RWQE into the QP's RWQE
147  * @qp: the QP
148  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
149  *
150  * Return -1 if there is a local error, 0 if no RWQE is available,
151  * otherwise return 1.
152  *
153  * Can be called from interrupt level.
154  */
155 int hfi1_rvt_get_rwqe(struct rvt_qp *qp, int wr_id_only)
156 {
157         unsigned long flags;
158         struct rvt_rq *rq;
159         struct rvt_rwq *wq;
160         struct rvt_srq *srq;
161         struct rvt_rwqe *wqe;
162         void (*handler)(struct ib_event *, void *);
163         u32 tail;
164         int ret;
165
166         if (qp->ibqp.srq) {
167                 srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
168                 handler = srq->ibsrq.event_handler;
169                 rq = &srq->rq;
170         } else {
171                 srq = NULL;
172                 handler = NULL;
173                 rq = &qp->r_rq;
174         }
175
176         spin_lock_irqsave(&rq->lock, flags);
177         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
178                 ret = 0;
179                 goto unlock;
180         }
181
182         wq = rq->wq;
183         tail = wq->tail;
184         /* Validate tail before using it since it is user writable. */
185         if (tail >= rq->size)
186                 tail = 0;
187         if (unlikely(tail == wq->head)) {
188                 ret = 0;
189                 goto unlock;
190         }
191         /* Make sure entry is read after head index is read. */
192         smp_rmb();
193         wqe = rvt_get_rwqe_ptr(rq, tail);
194         /*
195          * Even though we update the tail index in memory, the verbs
196          * consumer is not supposed to post more entries until a
197          * completion is generated.
198          */
199         if (++tail >= rq->size)
200                 tail = 0;
201         wq->tail = tail;
202         if (!wr_id_only && !init_sge(qp, wqe)) {
203                 ret = -1;
204                 goto unlock;
205         }
206         qp->r_wr_id = wqe->wr_id;
207
208         ret = 1;
209         set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
210         if (handler) {
211                 u32 n;
212
213                 /*
214                  * Validate head pointer value and compute
215                  * the number of remaining WQEs.
216                  */
217                 n = wq->head;
218                 if (n >= rq->size)
219                         n = 0;
220                 if (n < tail)
221                         n += rq->size - tail;
222                 else
223                         n -= tail;
224                 if (n < srq->limit) {
225                         struct ib_event ev;
226
227                         srq->limit = 0;
228                         spin_unlock_irqrestore(&rq->lock, flags);
229                         ev.device = qp->ibqp.device;
230                         ev.element.srq = qp->ibqp.srq;
231                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
232                         handler(&ev, srq->ibsrq.srq_context);
233                         goto bail;
234                 }
235         }
236 unlock:
237         spin_unlock_irqrestore(&rq->lock, flags);
238 bail:
239         return ret;
240 }
241
242 static __be64 get_sguid(struct hfi1_ibport *ibp, unsigned index)
243 {
244         if (!index) {
245                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
246
247                 return cpu_to_be64(ppd->guid);
248         }
249         return ibp->guids[index - 1];
250 }
251
252 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
253 {
254         return (gid->global.interface_id == id &&
255                 (gid->global.subnet_prefix == gid_prefix ||
256                  gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
257 }
258
259 /*
260  *
261  * This should be called with the QP r_lock held.
262  *
263  * The s_lock will be acquired around the hfi1_migrate_qp() call.
264  */
265 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct ib_header *hdr,
266                        int has_grh, struct rvt_qp *qp, u32 bth0)
267 {
268         __be64 guid;
269         unsigned long flags;
270         u8 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
271
272         if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
273                 if (!has_grh) {
274                         if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
275                                 goto err;
276                 } else {
277                         if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
278                                 goto err;
279                         guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
280                         if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
281                                     guid))
282                                 goto err;
283                         if (!gid_ok(
284                                 &hdr->u.l.grh.sgid,
285                                 qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
286                                 qp->alt_ah_attr.grh.dgid.global.interface_id))
287                                 goto err;
288                 }
289                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
290                                             sc5, be16_to_cpu(hdr->lrh[3])))) {
291                         hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
292                                        (u16)bth0,
293                                        (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
294                                        0, qp->ibqp.qp_num,
295                                        be16_to_cpu(hdr->lrh[3]),
296                                        be16_to_cpu(hdr->lrh[1]));
297                         goto err;
298                 }
299                 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
300                 if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
301                     ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
302                         goto err;
303                 spin_lock_irqsave(&qp->s_lock, flags);
304                 hfi1_migrate_qp(qp);
305                 spin_unlock_irqrestore(&qp->s_lock, flags);
306         } else {
307                 if (!has_grh) {
308                         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
309                                 goto err;
310                 } else {
311                         if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
312                                 goto err;
313                         guid = get_sguid(ibp,
314                                          qp->remote_ah_attr.grh.sgid_index);
315                         if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
316                                     guid))
317                                 goto err;
318                         if (!gid_ok(
319                              &hdr->u.l.grh.sgid,
320                              qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
321                              qp->remote_ah_attr.grh.dgid.global.interface_id))
322                                 goto err;
323                 }
324                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
325                                             sc5, be16_to_cpu(hdr->lrh[3])))) {
326                         hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
327                                        (u16)bth0,
328                                        (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
329                                        0, qp->ibqp.qp_num,
330                                        be16_to_cpu(hdr->lrh[3]),
331                                        be16_to_cpu(hdr->lrh[1]));
332                         goto err;
333                 }
334                 /* Validate the SLID. See Ch. 9.6.1.5 */
335                 if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
336                     ppd_from_ibp(ibp)->port != qp->port_num)
337                         goto err;
338                 if (qp->s_mig_state == IB_MIG_REARM &&
339                     !(bth0 & IB_BTH_MIG_REQ))
340                         qp->s_mig_state = IB_MIG_ARMED;
341         }
342
343         return 0;
344
345 err:
346         return 1;
347 }
348
349 /**
350  * ruc_loopback - handle UC and RC loopback requests
351  * @sqp: the sending QP
352  *
353  * This is called from hfi1_do_send() to
354  * forward a WQE addressed to the same HFI.
355  * Note that although we are single threaded due to the send engine, we still
356  * have to protect against post_send().  We don't have to worry about
357  * receive interrupts since this is a connected protocol and all packets
358  * will pass through here.
359  */
360 static void ruc_loopback(struct rvt_qp *sqp)
361 {
362         struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
363         struct rvt_qp *qp;
364         struct rvt_swqe *wqe;
365         struct rvt_sge *sge;
366         unsigned long flags;
367         struct ib_wc wc;
368         u64 sdata;
369         atomic64_t *maddr;
370         enum ib_wc_status send_status;
371         int release;
372         int ret;
373         int copy_last = 0;
374         u32 to;
375         int local_ops = 0;
376
377         rcu_read_lock();
378
379         /*
380          * Note that we check the responder QP state after
381          * checking the requester's state.
382          */
383         qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
384                             sqp->remote_qpn);
385
386         spin_lock_irqsave(&sqp->s_lock, flags);
387
388         /* Return if we are already busy processing a work request. */
389         if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
390             !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
391                 goto unlock;
392
393         sqp->s_flags |= RVT_S_BUSY;
394
395 again:
396         smp_read_barrier_depends(); /* see post_one_send() */
397         if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
398                 goto clr_busy;
399         wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
400
401         /* Return if it is not OK to start a new work request. */
402         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
403                 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
404                         goto clr_busy;
405                 /* We are in the error state, flush the work request. */
406                 send_status = IB_WC_WR_FLUSH_ERR;
407                 goto flush_send;
408         }
409
410         /*
411          * We can rely on the entry not changing without the s_lock
412          * being held until we update s_last.
413          * We increment s_cur to indicate s_last is in progress.
414          */
415         if (sqp->s_last == sqp->s_cur) {
416                 if (++sqp->s_cur >= sqp->s_size)
417                         sqp->s_cur = 0;
418         }
419         spin_unlock_irqrestore(&sqp->s_lock, flags);
420
421         if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
422             qp->ibqp.qp_type != sqp->ibqp.qp_type) {
423                 ibp->rvp.n_pkt_drops++;
424                 /*
425                  * For RC, the requester would timeout and retry so
426                  * shortcut the timeouts and just signal too many retries.
427                  */
428                 if (sqp->ibqp.qp_type == IB_QPT_RC)
429                         send_status = IB_WC_RETRY_EXC_ERR;
430                 else
431                         send_status = IB_WC_SUCCESS;
432                 goto serr;
433         }
434
435         memset(&wc, 0, sizeof(wc));
436         send_status = IB_WC_SUCCESS;
437
438         release = 1;
439         sqp->s_sge.sge = wqe->sg_list[0];
440         sqp->s_sge.sg_list = wqe->sg_list + 1;
441         sqp->s_sge.num_sge = wqe->wr.num_sge;
442         sqp->s_len = wqe->length;
443         switch (wqe->wr.opcode) {
444         case IB_WR_REG_MR:
445                 goto send_comp;
446
447         case IB_WR_LOCAL_INV:
448                 if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
449                         if (rvt_invalidate_rkey(sqp,
450                                                 wqe->wr.ex.invalidate_rkey))
451                                 send_status = IB_WC_LOC_PROT_ERR;
452                         local_ops = 1;
453                 }
454                 goto send_comp;
455
456         case IB_WR_SEND_WITH_INV:
457                 if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
458                         wc.wc_flags = IB_WC_WITH_INVALIDATE;
459                         wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
460                 }
461                 goto send;
462
463         case IB_WR_SEND_WITH_IMM:
464                 wc.wc_flags = IB_WC_WITH_IMM;
465                 wc.ex.imm_data = wqe->wr.ex.imm_data;
466                 /* FALLTHROUGH */
467         case IB_WR_SEND:
468 send:
469                 ret = hfi1_rvt_get_rwqe(qp, 0);
470                 if (ret < 0)
471                         goto op_err;
472                 if (!ret)
473                         goto rnr_nak;
474                 if (wqe->length > qp->r_len)
475                         goto inv_err;
476                 break;
477
478         case IB_WR_RDMA_WRITE_WITH_IMM:
479                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
480                         goto inv_err;
481                 wc.wc_flags = IB_WC_WITH_IMM;
482                 wc.ex.imm_data = wqe->wr.ex.imm_data;
483                 ret = hfi1_rvt_get_rwqe(qp, 1);
484                 if (ret < 0)
485                         goto op_err;
486                 if (!ret)
487                         goto rnr_nak;
488                 /* skip copy_last set and qp_access_flags recheck */
489                 goto do_write;
490         case IB_WR_RDMA_WRITE:
491                 copy_last = ibpd_to_rvtpd(qp->ibqp.pd)->user;
492                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
493                         goto inv_err;
494 do_write:
495                 if (wqe->length == 0)
496                         break;
497                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
498                                           wqe->rdma_wr.remote_addr,
499                                           wqe->rdma_wr.rkey,
500                                           IB_ACCESS_REMOTE_WRITE)))
501                         goto acc_err;
502                 qp->r_sge.sg_list = NULL;
503                 qp->r_sge.num_sge = 1;
504                 qp->r_sge.total_len = wqe->length;
505                 break;
506
507         case IB_WR_RDMA_READ:
508                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
509                         goto inv_err;
510                 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
511                                           wqe->rdma_wr.remote_addr,
512                                           wqe->rdma_wr.rkey,
513                                           IB_ACCESS_REMOTE_READ)))
514                         goto acc_err;
515                 release = 0;
516                 sqp->s_sge.sg_list = NULL;
517                 sqp->s_sge.num_sge = 1;
518                 qp->r_sge.sge = wqe->sg_list[0];
519                 qp->r_sge.sg_list = wqe->sg_list + 1;
520                 qp->r_sge.num_sge = wqe->wr.num_sge;
521                 qp->r_sge.total_len = wqe->length;
522                 break;
523
524         case IB_WR_ATOMIC_CMP_AND_SWP:
525         case IB_WR_ATOMIC_FETCH_AND_ADD:
526                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
527                         goto inv_err;
528                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
529                                           wqe->atomic_wr.remote_addr,
530                                           wqe->atomic_wr.rkey,
531                                           IB_ACCESS_REMOTE_ATOMIC)))
532                         goto acc_err;
533                 /* Perform atomic OP and save result. */
534                 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
535                 sdata = wqe->atomic_wr.compare_add;
536                 *(u64 *)sqp->s_sge.sge.vaddr =
537                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
538                         (u64)atomic64_add_return(sdata, maddr) - sdata :
539                         (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
540                                       sdata, wqe->atomic_wr.swap);
541                 rvt_put_mr(qp->r_sge.sge.mr);
542                 qp->r_sge.num_sge = 0;
543                 goto send_comp;
544
545         default:
546                 send_status = IB_WC_LOC_QP_OP_ERR;
547                 goto serr;
548         }
549
550         sge = &sqp->s_sge.sge;
551         while (sqp->s_len) {
552                 u32 len = sqp->s_len;
553
554                 if (len > sge->length)
555                         len = sge->length;
556                 if (len > sge->sge_length)
557                         len = sge->sge_length;
558                 WARN_ON_ONCE(len == 0);
559                 hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
560                 sge->vaddr += len;
561                 sge->length -= len;
562                 sge->sge_length -= len;
563                 if (sge->sge_length == 0) {
564                         if (!release)
565                                 rvt_put_mr(sge->mr);
566                         if (--sqp->s_sge.num_sge)
567                                 *sge = *sqp->s_sge.sg_list++;
568                 } else if (sge->length == 0 && sge->mr->lkey) {
569                         if (++sge->n >= RVT_SEGSZ) {
570                                 if (++sge->m >= sge->mr->mapsz)
571                                         break;
572                                 sge->n = 0;
573                         }
574                         sge->vaddr =
575                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
576                         sge->length =
577                                 sge->mr->map[sge->m]->segs[sge->n].length;
578                 }
579                 sqp->s_len -= len;
580         }
581         if (release)
582                 rvt_put_ss(&qp->r_sge);
583
584         if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
585                 goto send_comp;
586
587         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
588                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
589         else
590                 wc.opcode = IB_WC_RECV;
591         wc.wr_id = qp->r_wr_id;
592         wc.status = IB_WC_SUCCESS;
593         wc.byte_len = wqe->length;
594         wc.qp = &qp->ibqp;
595         wc.src_qp = qp->remote_qpn;
596         wc.slid = qp->remote_ah_attr.dlid;
597         wc.sl = qp->remote_ah_attr.sl;
598         wc.port_num = 1;
599         /* Signal completion event if the solicited bit is set. */
600         rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
601                      wqe->wr.send_flags & IB_SEND_SOLICITED);
602
603 send_comp:
604         spin_lock_irqsave(&sqp->s_lock, flags);
605         ibp->rvp.n_loop_pkts++;
606 flush_send:
607         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
608         hfi1_send_complete(sqp, wqe, send_status);
609         if (local_ops) {
610                 atomic_dec(&sqp->local_ops_pending);
611                 local_ops = 0;
612         }
613         goto again;
614
615 rnr_nak:
616         /* Handle RNR NAK */
617         if (qp->ibqp.qp_type == IB_QPT_UC)
618                 goto send_comp;
619         ibp->rvp.n_rnr_naks++;
620         /*
621          * Note: we don't need the s_lock held since the BUSY flag
622          * makes this single threaded.
623          */
624         if (sqp->s_rnr_retry == 0) {
625                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
626                 goto serr;
627         }
628         if (sqp->s_rnr_retry_cnt < 7)
629                 sqp->s_rnr_retry--;
630         spin_lock_irqsave(&sqp->s_lock, flags);
631         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
632                 goto clr_busy;
633         to = ib_hfi1_rnr_table[qp->r_min_rnr_timer];
634         hfi1_add_rnr_timer(sqp, to);
635         goto clr_busy;
636
637 op_err:
638         send_status = IB_WC_REM_OP_ERR;
639         wc.status = IB_WC_LOC_QP_OP_ERR;
640         goto err;
641
642 inv_err:
643         send_status =
644                 sqp->ibqp.qp_type == IB_QPT_RC ?
645                         IB_WC_REM_INV_REQ_ERR :
646                         IB_WC_SUCCESS;
647         wc.status = IB_WC_LOC_QP_OP_ERR;
648         goto err;
649
650 acc_err:
651         send_status = IB_WC_REM_ACCESS_ERR;
652         wc.status = IB_WC_LOC_PROT_ERR;
653 err:
654         /* responder goes to error state */
655         hfi1_rc_error(qp, wc.status);
656
657 serr:
658         spin_lock_irqsave(&sqp->s_lock, flags);
659         hfi1_send_complete(sqp, wqe, send_status);
660         if (sqp->ibqp.qp_type == IB_QPT_RC) {
661                 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
662
663                 sqp->s_flags &= ~RVT_S_BUSY;
664                 spin_unlock_irqrestore(&sqp->s_lock, flags);
665                 if (lastwqe) {
666                         struct ib_event ev;
667
668                         ev.device = sqp->ibqp.device;
669                         ev.element.qp = &sqp->ibqp;
670                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
671                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
672                 }
673                 goto done;
674         }
675 clr_busy:
676         sqp->s_flags &= ~RVT_S_BUSY;
677 unlock:
678         spin_unlock_irqrestore(&sqp->s_lock, flags);
679 done:
680         rcu_read_unlock();
681 }
682
683 /**
684  * hfi1_make_grh - construct a GRH header
685  * @ibp: a pointer to the IB port
686  * @hdr: a pointer to the GRH header being constructed
687  * @grh: the global route address to send to
688  * @hwords: the number of 32 bit words of header being sent
689  * @nwords: the number of 32 bit words of data being sent
690  *
691  * Return the size of the header in 32 bit words.
692  */
693 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
694                   struct ib_global_route *grh, u32 hwords, u32 nwords)
695 {
696         hdr->version_tclass_flow =
697                 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
698                             (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
699                             (grh->flow_label << IB_GRH_FLOW_SHIFT));
700         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
701         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
702         hdr->next_hdr = IB_GRH_NEXT_HDR;
703         hdr->hop_limit = grh->hop_limit;
704         /* The SGID is 32-bit aligned. */
705         hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
706         hdr->sgid.global.interface_id =
707                 grh->sgid_index && grh->sgid_index < ARRAY_SIZE(ibp->guids) ?
708                 ibp->guids[grh->sgid_index - 1] :
709                         cpu_to_be64(ppd_from_ibp(ibp)->guid);
710         hdr->dgid = grh->dgid;
711
712         /* GRH header size in 32-bit words. */
713         return sizeof(struct ib_grh) / sizeof(u32);
714 }
715
716 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, hdr.u.oth.bth[2]) / 4)
717
718 /**
719  * build_ahg - create ahg in s_ahg
720  * @qp: a pointer to QP
721  * @npsn: the next PSN for the request/response
722  *
723  * This routine handles the AHG by allocating an ahg entry and causing the
724  * copy of the first middle.
725  *
726  * Subsequent middles use the copied entry, editing the
727  * PSN with 1 or 2 edits.
728  */
729 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
730 {
731         struct hfi1_qp_priv *priv = qp->priv;
732
733         if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR))
734                 clear_ahg(qp);
735         if (!(qp->s_flags & RVT_S_AHG_VALID)) {
736                 /* first middle that needs copy  */
737                 if (qp->s_ahgidx < 0)
738                         qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
739                 if (qp->s_ahgidx >= 0) {
740                         qp->s_ahgpsn = npsn;
741                         priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
742                         /* save to protect a change in another thread */
743                         priv->s_ahg->ahgidx = qp->s_ahgidx;
744                         qp->s_flags |= RVT_S_AHG_VALID;
745                 }
746         } else {
747                 /* subsequent middle after valid */
748                 if (qp->s_ahgidx >= 0) {
749                         priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
750                         priv->s_ahg->ahgidx = qp->s_ahgidx;
751                         priv->s_ahg->ahgcount++;
752                         priv->s_ahg->ahgdesc[0] =
753                                 sdma_build_ahg_descriptor(
754                                         (__force u16)cpu_to_be16((u16)npsn),
755                                         BTH2_OFFSET,
756                                         16,
757                                         16);
758                         if ((npsn & 0xffff0000) !=
759                                         (qp->s_ahgpsn & 0xffff0000)) {
760                                 priv->s_ahg->ahgcount++;
761                                 priv->s_ahg->ahgdesc[1] =
762                                         sdma_build_ahg_descriptor(
763                                                 (__force u16)cpu_to_be16(
764                                                         (u16)(npsn >> 16)),
765                                                 BTH2_OFFSET,
766                                                 0,
767                                                 16);
768                         }
769                 }
770         }
771 }
772
773 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
774                           u32 bth0, u32 bth2, int middle,
775                           struct hfi1_pkt_state *ps)
776 {
777         struct hfi1_qp_priv *priv = qp->priv;
778         struct hfi1_ibport *ibp = ps->ibp;
779         u16 lrh0;
780         u32 nwords;
781         u32 extra_bytes;
782         u32 bth1;
783
784         /* Construct the header. */
785         extra_bytes = -qp->s_cur_size & 3;
786         nwords = (qp->s_cur_size + extra_bytes) >> 2;
787         lrh0 = HFI1_LRH_BTH;
788         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
789                 qp->s_hdrwords += hfi1_make_grh(ibp,
790                                                 &ps->s_txreq->phdr.hdr.u.l.grh,
791                                                 &qp->remote_ah_attr.grh,
792                                                 qp->s_hdrwords, nwords);
793                 lrh0 = HFI1_LRH_GRH;
794                 middle = 0;
795         }
796         lrh0 |= (priv->s_sc & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
797         /*
798          * reset s_ahg/AHG fields
799          *
800          * This insures that the ahgentry/ahgcount
801          * are at a non-AHG default to protect
802          * build_verbs_tx_desc() from using
803          * an include ahgidx.
804          *
805          * build_ahg() will modify as appropriate
806          * to use the AHG feature.
807          */
808         priv->s_ahg->tx_flags = 0;
809         priv->s_ahg->ahgcount = 0;
810         priv->s_ahg->ahgidx = 0;
811         if (qp->s_mig_state == IB_MIG_MIGRATED)
812                 bth0 |= IB_BTH_MIG_REQ;
813         else
814                 middle = 0;
815         if (middle)
816                 build_ahg(qp, bth2);
817         else
818                 qp->s_flags &= ~RVT_S_AHG_VALID;
819         ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
820         ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
821         ps->s_txreq->phdr.hdr.lrh[2] =
822                 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
823         ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
824                                        qp->remote_ah_attr.src_path_bits);
825         bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
826         bth0 |= extra_bytes << 20;
827         ohdr->bth[0] = cpu_to_be32(bth0);
828         bth1 = qp->remote_qpn;
829         if (qp->s_flags & RVT_S_ECN) {
830                 qp->s_flags &= ~RVT_S_ECN;
831                 /* we recently received a FECN, so return a BECN */
832                 bth1 |= (HFI1_BECN_MASK << HFI1_BECN_SHIFT);
833         }
834         ohdr->bth[1] = cpu_to_be32(bth1);
835         ohdr->bth[2] = cpu_to_be32(bth2);
836 }
837
838 /* when sending, force a reschedule every one of these periods */
839 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
840
841 void hfi1_do_send_from_rvt(struct rvt_qp *qp)
842 {
843         hfi1_do_send(qp, false);
844 }
845
846 void _hfi1_do_send(struct work_struct *work)
847 {
848         struct iowait *wait = container_of(work, struct iowait, iowork);
849         struct rvt_qp *qp = iowait_to_qp(wait);
850
851         hfi1_do_send(qp, true);
852 }
853
854 /**
855  * hfi1_do_send - perform a send on a QP
856  * @work: contains a pointer to the QP
857  * @in_thread: true if in a workqueue thread
858  *
859  * Process entries in the send work queue until credit or queue is
860  * exhausted.  Only allow one CPU to send a packet per QP.
861  * Otherwise, two threads could send packets out of order.
862  */
863 void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
864 {
865         struct hfi1_pkt_state ps;
866         struct hfi1_qp_priv *priv = qp->priv;
867         int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
868         unsigned long timeout;
869         unsigned long timeout_int;
870         int cpu;
871
872         ps.dev = to_idev(qp->ibqp.device);
873         ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
874         ps.ppd = ppd_from_ibp(ps.ibp);
875
876         switch (qp->ibqp.qp_type) {
877         case IB_QPT_RC:
878                 if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
879                                                                 ) - 1)) ==
880                                  ps.ppd->lid)) {
881                         ruc_loopback(qp);
882                         return;
883                 }
884                 make_req = hfi1_make_rc_req;
885                 timeout_int = (qp->timeout_jiffies);
886                 break;
887         case IB_QPT_UC:
888                 if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
889                                                                 ) - 1)) ==
890                                  ps.ppd->lid)) {
891                         ruc_loopback(qp);
892                         return;
893                 }
894                 make_req = hfi1_make_uc_req;
895                 timeout_int = SEND_RESCHED_TIMEOUT;
896                 break;
897         default:
898                 make_req = hfi1_make_ud_req;
899                 timeout_int = SEND_RESCHED_TIMEOUT;
900         }
901
902         spin_lock_irqsave(&qp->s_lock, ps.flags);
903
904         /* Return if we are already busy processing a work request. */
905         if (!hfi1_send_ok(qp)) {
906                 spin_unlock_irqrestore(&qp->s_lock, ps.flags);
907                 return;
908         }
909
910         qp->s_flags |= RVT_S_BUSY;
911
912         timeout = jiffies + (timeout_int) / 8;
913         cpu = priv->s_sde ? priv->s_sde->cpu :
914                         cpumask_first(cpumask_of_node(ps.ppd->dd->node));
915         /* insure a pre-built packet is handled  */
916         ps.s_txreq = get_waiting_verbs_txreq(qp);
917         do {
918                 /* Check for a constructed packet to be sent. */
919                 if (qp->s_hdrwords != 0) {
920                         spin_unlock_irqrestore(&qp->s_lock, ps.flags);
921                         /*
922                          * If the packet cannot be sent now, return and
923                          * the send engine will be woken up later.
924                          */
925                         if (hfi1_verbs_send(qp, &ps))
926                                 return;
927                         /* Record that s_ahg is empty. */
928                         qp->s_hdrwords = 0;
929                         /* allow other tasks to run */
930                         if (unlikely(time_after(jiffies, timeout))) {
931                                 if (!in_thread ||
932                                     workqueue_congested(
933                                                 cpu,
934                                                 ps.ppd->hfi1_wq)) {
935                                         spin_lock_irqsave(
936                                                 &qp->s_lock,
937                                                 ps.flags);
938                                         qp->s_flags &= ~RVT_S_BUSY;
939                                         hfi1_schedule_send(qp);
940                                         spin_unlock_irqrestore(
941                                                 &qp->s_lock,
942                                                 ps.flags);
943                                         this_cpu_inc(
944                                                 *ps.ppd->dd->send_schedule);
945                                         return;
946                                 }
947                                 cond_resched();
948                                 this_cpu_inc(
949                                         *ps.ppd->dd->send_schedule);
950                                 timeout = jiffies + (timeout_int) / 8;
951                         }
952                         spin_lock_irqsave(&qp->s_lock, ps.flags);
953                 }
954         } while (make_req(qp, &ps));
955
956         spin_unlock_irqrestore(&qp->s_lock, ps.flags);
957 }
958
959 /*
960  * This should be called with s_lock held.
961  */
962 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
963                         enum ib_wc_status status)
964 {
965         u32 old_last, last;
966         unsigned i;
967
968         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
969                 return;
970
971         last = qp->s_last;
972         old_last = last;
973         if (++last >= qp->s_size)
974                 last = 0;
975         qp->s_last = last;
976         /* See post_send() */
977         barrier();
978         for (i = 0; i < wqe->wr.num_sge; i++) {
979                 struct rvt_sge *sge = &wqe->sg_list[i];
980
981                 rvt_put_mr(sge->mr);
982         }
983         if (qp->ibqp.qp_type == IB_QPT_UD ||
984             qp->ibqp.qp_type == IB_QPT_SMI ||
985             qp->ibqp.qp_type == IB_QPT_GSI)
986                 atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
987
988         /* See ch. 11.2.4.1 and 10.7.3.1 */
989         if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
990             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
991             status != IB_WC_SUCCESS) {
992                 struct ib_wc wc;
993
994                 memset(&wc, 0, sizeof(wc));
995                 wc.wr_id = wqe->wr.wr_id;
996                 wc.status = status;
997                 wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode];
998                 wc.qp = &qp->ibqp;
999                 if (status == IB_WC_SUCCESS)
1000                         wc.byte_len = wqe->length;
1001                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
1002                              status != IB_WC_SUCCESS);
1003         }
1004
1005         if (qp->s_acked == old_last)
1006                 qp->s_acked = last;
1007         if (qp->s_cur == old_last)
1008                 qp->s_cur = last;
1009         if (qp->s_tail == old_last)
1010                 qp->s_tail = last;
1011         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
1012                 qp->s_draining = 0;
1013 }