1/*
2 * Copyright (c) 2012, 2013 Intel Corporation.  All rights reserved.
3 * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
4 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses.  You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
11 *
12 *     Redistribution and use in source and binary forms, with or
13 *     without modification, are permitted provided that the following
14 *     conditions are met:
15 *
16 *      - Redistributions of source code must retain the above
17 *        copyright notice, this list of conditions and the following
18 *        disclaimer.
19 *
20 *      - Redistributions in binary form must reproduce the above
21 *        copyright notice, this list of conditions and the following
22 *        disclaimer in the documentation and/or other materials
23 *        provided with the distribution.
24 *
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * SOFTWARE.
33 */
34
35#include <rdma/ib_mad.h>
36#include <rdma/ib_user_verbs.h>
37#include <linux/io.h>
38#include <linux/module.h>
39#include <linux/utsname.h>
40#include <linux/rculist.h>
41#include <linux/mm.h>
42#include <linux/random.h>
43#include <linux/vmalloc.h>
44
45#include "qib.h"
46#include "qib_common.h"
47
48static unsigned int ib_qib_qp_table_size = 256;
49module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO);
50MODULE_PARM_DESC(qp_table_size, "QP table size");
51
52unsigned int ib_qib_lkey_table_size = 16;
53module_param_named(lkey_table_size, ib_qib_lkey_table_size, uint,
54		   S_IRUGO);
55MODULE_PARM_DESC(lkey_table_size,
56		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
57
58static unsigned int ib_qib_max_pds = 0xFFFF;
59module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO);
60MODULE_PARM_DESC(max_pds,
61		 "Maximum number of protection domains to support");
62
63static unsigned int ib_qib_max_ahs = 0xFFFF;
64module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO);
65MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
66
67unsigned int ib_qib_max_cqes = 0x2FFFF;
68module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO);
69MODULE_PARM_DESC(max_cqes,
70		 "Maximum number of completion queue entries to support");
71
72unsigned int ib_qib_max_cqs = 0x1FFFF;
73module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO);
74MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
75
76unsigned int ib_qib_max_qp_wrs = 0x3FFF;
77module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO);
78MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
79
80unsigned int ib_qib_max_qps = 16384;
81module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO);
82MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
83
84unsigned int ib_qib_max_sges = 0x60;
85module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO);
86MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
87
88unsigned int ib_qib_max_mcast_grps = 16384;
89module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO);
90MODULE_PARM_DESC(max_mcast_grps,
91		 "Maximum number of multicast groups to support");
92
93unsigned int ib_qib_max_mcast_qp_attached = 16;
94module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached,
95		   uint, S_IRUGO);
96MODULE_PARM_DESC(max_mcast_qp_attached,
97		 "Maximum number of attached QPs to support");
98
99unsigned int ib_qib_max_srqs = 1024;
100module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO);
101MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
102
103unsigned int ib_qib_max_srq_sges = 128;
104module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO);
105MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
106
107unsigned int ib_qib_max_srq_wrs = 0x1FFFF;
108module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO);
109MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
110
111static unsigned int ib_qib_disable_sma;
112module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO);
113MODULE_PARM_DESC(disable_sma, "Disable the SMA");
114
115/*
116 * Note that it is OK to post send work requests in the SQE and ERR
117 * states; qib_do_send() will process them and generate error
118 * completions as per IB 1.2 C10-96.
119 */
120const int ib_qib_state_ops[IB_QPS_ERR + 1] = {
121	[IB_QPS_RESET] = 0,
122	[IB_QPS_INIT] = QIB_POST_RECV_OK,
123	[IB_QPS_RTR] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK,
124	[IB_QPS_RTS] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK |
125	    QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK |
126	    QIB_PROCESS_NEXT_SEND_OK,
127	[IB_QPS_SQD] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK |
128	    QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK,
129	[IB_QPS_SQE] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK |
130	    QIB_POST_SEND_OK | QIB_FLUSH_SEND,
131	[IB_QPS_ERR] = QIB_POST_RECV_OK | QIB_FLUSH_RECV |
132	    QIB_POST_SEND_OK | QIB_FLUSH_SEND,
133};
134
135struct qib_ucontext {
136	struct ib_ucontext ibucontext;
137};
138
139static inline struct qib_ucontext *to_iucontext(struct ib_ucontext
140						  *ibucontext)
141{
142	return container_of(ibucontext, struct qib_ucontext, ibucontext);
143}
144
145/*
146 * Translate ib_wr_opcode into ib_wc_opcode.
147 */
148const enum ib_wc_opcode ib_qib_wc_opcode[] = {
149	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
150	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
151	[IB_WR_SEND] = IB_WC_SEND,
152	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
153	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
154	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
155	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
156};
157
158/*
159 * System image GUID.
160 */
161__be64 ib_qib_sys_image_guid;
162
163/**
164 * qib_copy_sge - copy data to SGE memory
165 * @ss: the SGE state
166 * @data: the data to copy
167 * @length: the length of the data
168 */
169void qib_copy_sge(struct qib_sge_state *ss, void *data, u32 length, int release)
170{
171	struct qib_sge *sge = &ss->sge;
172
173	while (length) {
174		u32 len = sge->length;
175
176		if (len > length)
177			len = length;
178		if (len > sge->sge_length)
179			len = sge->sge_length;
180		BUG_ON(len == 0);
181		memcpy(sge->vaddr, data, len);
182		sge->vaddr += len;
183		sge->length -= len;
184		sge->sge_length -= len;
185		if (sge->sge_length == 0) {
186			if (release)
187				qib_put_mr(sge->mr);
188			if (--ss->num_sge)
189				*sge = *ss->sg_list++;
190		} else if (sge->length == 0 && sge->mr->lkey) {
191			if (++sge->n >= QIB_SEGSZ) {
192				if (++sge->m >= sge->mr->mapsz)
193					break;
194				sge->n = 0;
195			}
196			sge->vaddr =
197				sge->mr->map[sge->m]->segs[sge->n].vaddr;
198			sge->length =
199				sge->mr->map[sge->m]->segs[sge->n].length;
200		}
201		data += len;
202		length -= len;
203	}
204}
205
206/**
207 * qib_skip_sge - skip over SGE memory - XXX almost dup of prev func
208 * @ss: the SGE state
209 * @length: the number of bytes to skip
210 */
211void qib_skip_sge(struct qib_sge_state *ss, u32 length, int release)
212{
213	struct qib_sge *sge = &ss->sge;
214
215	while (length) {
216		u32 len = sge->length;
217
218		if (len > length)
219			len = length;
220		if (len > sge->sge_length)
221			len = sge->sge_length;
222		BUG_ON(len == 0);
223		sge->vaddr += len;
224		sge->length -= len;
225		sge->sge_length -= len;
226		if (sge->sge_length == 0) {
227			if (release)
228				qib_put_mr(sge->mr);
229			if (--ss->num_sge)
230				*sge = *ss->sg_list++;
231		} else if (sge->length == 0 && sge->mr->lkey) {
232			if (++sge->n >= QIB_SEGSZ) {
233				if (++sge->m >= sge->mr->mapsz)
234					break;
235				sge->n = 0;
236			}
237			sge->vaddr =
238				sge->mr->map[sge->m]->segs[sge->n].vaddr;
239			sge->length =
240				sge->mr->map[sge->m]->segs[sge->n].length;
241		}
242		length -= len;
243	}
244}
245
246/*
247 * Count the number of DMA descriptors needed to send length bytes of data.
248 * Don't modify the qib_sge_state to get the count.
249 * Return zero if any of the segments is not aligned.
250 */
251static u32 qib_count_sge(struct qib_sge_state *ss, u32 length)
252{
253	struct qib_sge *sg_list = ss->sg_list;
254	struct qib_sge sge = ss->sge;
255	u8 num_sge = ss->num_sge;
256	u32 ndesc = 1;  /* count the header */
257
258	while (length) {
259		u32 len = sge.length;
260
261		if (len > length)
262			len = length;
263		if (len > sge.sge_length)
264			len = sge.sge_length;
265		BUG_ON(len == 0);
266		if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
267		    (len != length && (len & (sizeof(u32) - 1)))) {
268			ndesc = 0;
269			break;
270		}
271		ndesc++;
272		sge.vaddr += len;
273		sge.length -= len;
274		sge.sge_length -= len;
275		if (sge.sge_length == 0) {
276			if (--num_sge)
277				sge = *sg_list++;
278		} else if (sge.length == 0 && sge.mr->lkey) {
279			if (++sge.n >= QIB_SEGSZ) {
280				if (++sge.m >= sge.mr->mapsz)
281					break;
282				sge.n = 0;
283			}
284			sge.vaddr =
285				sge.mr->map[sge.m]->segs[sge.n].vaddr;
286			sge.length =
287				sge.mr->map[sge.m]->segs[sge.n].length;
288		}
289		length -= len;
290	}
291	return ndesc;
292}
293
294/*
295 * Copy from the SGEs to the data buffer.
296 */
297static void qib_copy_from_sge(void *data, struct qib_sge_state *ss, u32 length)
298{
299	struct qib_sge *sge = &ss->sge;
300
301	while (length) {
302		u32 len = sge->length;
303
304		if (len > length)
305			len = length;
306		if (len > sge->sge_length)
307			len = sge->sge_length;
308		BUG_ON(len == 0);
309		memcpy(data, sge->vaddr, len);
310		sge->vaddr += len;
311		sge->length -= len;
312		sge->sge_length -= len;
313		if (sge->sge_length == 0) {
314			if (--ss->num_sge)
315				*sge = *ss->sg_list++;
316		} else if (sge->length == 0 && sge->mr->lkey) {
317			if (++sge->n >= QIB_SEGSZ) {
318				if (++sge->m >= sge->mr->mapsz)
319					break;
320				sge->n = 0;
321			}
322			sge->vaddr =
323				sge->mr->map[sge->m]->segs[sge->n].vaddr;
324			sge->length =
325				sge->mr->map[sge->m]->segs[sge->n].length;
326		}
327		data += len;
328		length -= len;
329	}
330}
331
332/**
333 * qib_post_one_send - post one RC, UC, or UD send work request
334 * @qp: the QP to post on
335 * @wr: the work request to send
336 */
337static int qib_post_one_send(struct qib_qp *qp, struct ib_send_wr *wr,
338	int *scheduled)
339{
340	struct qib_swqe *wqe;
341	u32 next;
342	int i;
343	int j;
344	int acc;
345	int ret;
346	unsigned long flags;
347	struct qib_lkey_table *rkt;
348	struct qib_pd *pd;
349
350	spin_lock_irqsave(&qp->s_lock, flags);
351
352	/* Check that state is OK to post send. */
353	if (unlikely(!(ib_qib_state_ops[qp->state] & QIB_POST_SEND_OK)))
354		goto bail_inval;
355
356	/* IB spec says that num_sge == 0 is OK. */
357	if (wr->num_sge > qp->s_max_sge)
358		goto bail_inval;
359
360	/*
361	 * Don't allow RDMA reads or atomic operations on UC or
362	 * undefined operations.
363	 * Make sure buffer is large enough to hold the result for atomics.
364	 */
365	if (wr->opcode == IB_WR_FAST_REG_MR) {
366		if (qib_fast_reg_mr(qp, wr))
367			goto bail_inval;
368	} else if (qp->ibqp.qp_type == IB_QPT_UC) {
369		if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
370			goto bail_inval;
371	} else if (qp->ibqp.qp_type != IB_QPT_RC) {
372		/* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
373		if (wr->opcode != IB_WR_SEND &&
374		    wr->opcode != IB_WR_SEND_WITH_IMM)
375			goto bail_inval;
376		/* Check UD destination address PD */
377		if (qp->ibqp.pd != wr->wr.ud.ah->pd)
378			goto bail_inval;
379	} else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
380		goto bail_inval;
381	else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
382		   (wr->num_sge == 0 ||
383		    wr->sg_list[0].length < sizeof(u64) ||
384		    wr->sg_list[0].addr & (sizeof(u64) - 1)))
385		goto bail_inval;
386	else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
387		goto bail_inval;
388
389	next = qp->s_head + 1;
390	if (next >= qp->s_size)
391		next = 0;
392	if (next == qp->s_last) {
393		ret = -ENOMEM;
394		goto bail;
395	}
396
397	rkt = &to_idev(qp->ibqp.device)->lk_table;
398	pd = to_ipd(qp->ibqp.pd);
399	wqe = get_swqe_ptr(qp, qp->s_head);
400	wqe->wr = *wr;
401	wqe->length = 0;
402	j = 0;
403	if (wr->num_sge) {
404		acc = wr->opcode >= IB_WR_RDMA_READ ?
405			IB_ACCESS_LOCAL_WRITE : 0;
406		for (i = 0; i < wr->num_sge; i++) {
407			u32 length = wr->sg_list[i].length;
408			int ok;
409
410			if (length == 0)
411				continue;
412			ok = qib_lkey_ok(rkt, pd, &wqe->sg_list[j],
413					 &wr->sg_list[i], acc);
414			if (!ok)
415				goto bail_inval_free;
416			wqe->length += length;
417			j++;
418		}
419		wqe->wr.num_sge = j;
420	}
421	if (qp->ibqp.qp_type == IB_QPT_UC ||
422	    qp->ibqp.qp_type == IB_QPT_RC) {
423		if (wqe->length > 0x80000000U)
424			goto bail_inval_free;
425	} else if (wqe->length > (dd_from_ibdev(qp->ibqp.device)->pport +
426				  qp->port_num - 1)->ibmtu)
427		goto bail_inval_free;
428	else
429		atomic_inc(&to_iah(wr->wr.ud.ah)->refcount);
430	wqe->ssn = qp->s_ssn++;
431	qp->s_head = next;
432
433	ret = 0;
434	goto bail;
435
436bail_inval_free:
437	while (j) {
438		struct qib_sge *sge = &wqe->sg_list[--j];
439
440		qib_put_mr(sge->mr);
441	}
442bail_inval:
443	ret = -EINVAL;
444bail:
445	if (!ret && !wr->next &&
446	 !qib_sdma_empty(
447	   dd_from_ibdev(qp->ibqp.device)->pport + qp->port_num - 1)) {
448		qib_schedule_send(qp);
449		*scheduled = 1;
450	}
451	spin_unlock_irqrestore(&qp->s_lock, flags);
452	return ret;
453}
454
455/**
456 * qib_post_send - post a send on a QP
457 * @ibqp: the QP to post the send on
458 * @wr: the list of work requests to post
459 * @bad_wr: the first bad WR is put here
460 *
461 * This may be called from interrupt context.
462 */
463static int qib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
464			 struct ib_send_wr **bad_wr)
465{
466	struct qib_qp *qp = to_iqp(ibqp);
467	int err = 0;
468	int scheduled = 0;
469
470	for (; wr; wr = wr->next) {
471		err = qib_post_one_send(qp, wr, &scheduled);
472		if (err) {
473			*bad_wr = wr;
474			goto bail;
475		}
476	}
477
478	/* Try to do the send work in the caller's context. */
479	if (!scheduled)
480		qib_do_send(&qp->s_work);
481
482bail:
483	return err;
484}
485
486/**
487 * qib_post_receive - post a receive on a QP
488 * @ibqp: the QP to post the receive on
489 * @wr: the WR to post
490 * @bad_wr: the first bad WR is put here
491 *
492 * This may be called from interrupt context.
493 */
494static int qib_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
495			    struct ib_recv_wr **bad_wr)
496{
497	struct qib_qp *qp = to_iqp(ibqp);
498	struct qib_rwq *wq = qp->r_rq.wq;
499	unsigned long flags;
500	int ret;
501
502	/* Check that state is OK to post receive. */
503	if (!(ib_qib_state_ops[qp->state] & QIB_POST_RECV_OK) || !wq) {
504		*bad_wr = wr;
505		ret = -EINVAL;
506		goto bail;
507	}
508
509	for (; wr; wr = wr->next) {
510		struct qib_rwqe *wqe;
511		u32 next;
512		int i;
513
514		if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
515			*bad_wr = wr;
516			ret = -EINVAL;
517			goto bail;
518		}
519
520		spin_lock_irqsave(&qp->r_rq.lock, flags);
521		next = wq->head + 1;
522		if (next >= qp->r_rq.size)
523			next = 0;
524		if (next == wq->tail) {
525			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
526			*bad_wr = wr;
527			ret = -ENOMEM;
528			goto bail;
529		}
530
531		wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
532		wqe->wr_id = wr->wr_id;
533		wqe->num_sge = wr->num_sge;
534		for (i = 0; i < wr->num_sge; i++)
535			wqe->sg_list[i] = wr->sg_list[i];
536		/* Make sure queue entry is written before the head index. */
537		smp_wmb();
538		wq->head = next;
539		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
540	}
541	ret = 0;
542
543bail:
544	return ret;
545}
546
547/**
548 * qib_qp_rcv - processing an incoming packet on a QP
549 * @rcd: the context pointer
550 * @hdr: the packet header
551 * @has_grh: true if the packet has a GRH
552 * @data: the packet data
553 * @tlen: the packet length
554 * @qp: the QP the packet came on
555 *
556 * This is called from qib_ib_rcv() to process an incoming packet
557 * for the given QP.
558 * Called at interrupt level.
559 */
560static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
561		       int has_grh, void *data, u32 tlen, struct qib_qp *qp)
562{
563	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
564
565	spin_lock(&qp->r_lock);
566
567	/* Check for valid receive state. */
568	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) {
569		ibp->n_pkt_drops++;
570		goto unlock;
571	}
572
573	switch (qp->ibqp.qp_type) {
574	case IB_QPT_SMI:
575	case IB_QPT_GSI:
576		if (ib_qib_disable_sma)
577			break;
578		/* FALLTHROUGH */
579	case IB_QPT_UD:
580		qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
581		break;
582
583	case IB_QPT_RC:
584		qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
585		break;
586
587	case IB_QPT_UC:
588		qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
589		break;
590
591	default:
592		break;
593	}
594
595unlock:
596	spin_unlock(&qp->r_lock);
597}
598
599/**
600 * qib_ib_rcv - process an incoming packet
601 * @rcd: the context pointer
602 * @rhdr: the header of the packet
603 * @data: the packet payload
604 * @tlen: the packet length
605 *
606 * This is called from qib_kreceive() to process an incoming packet at
607 * interrupt level. Tlen is the length of the header + data + CRC in bytes.
608 */
609void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
610{
611	struct qib_pportdata *ppd = rcd->ppd;
612	struct qib_ibport *ibp = &ppd->ibport_data;
613	struct qib_ib_header *hdr = rhdr;
614	struct qib_other_headers *ohdr;
615	struct qib_qp *qp;
616	u32 qp_num;
617	int lnh;
618	u8 opcode;
619	u16 lid;
620
621	/* 24 == LRH+BTH+CRC */
622	if (unlikely(tlen < 24))
623		goto drop;
624
625	/* Check for a valid destination LID (see ch. 7.11.1). */
626	lid = be16_to_cpu(hdr->lrh[1]);
627	if (lid < QIB_MULTICAST_LID_BASE) {
628		lid &= ~((1 << ppd->lmc) - 1);
629		if (unlikely(lid != ppd->lid))
630			goto drop;
631	}
632
633	/* Check for GRH */
634	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
635	if (lnh == QIB_LRH_BTH)
636		ohdr = &hdr->u.oth;
637	else if (lnh == QIB_LRH_GRH) {
638		u32 vtf;
639
640		ohdr = &hdr->u.l.oth;
641		if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
642			goto drop;
643		vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
644		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
645			goto drop;
646	} else
647		goto drop;
648
649	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
650#ifdef CONFIG_DEBUG_FS
651	rcd->opstats->stats[opcode].n_bytes += tlen;
652	rcd->opstats->stats[opcode].n_packets++;
653#endif
654
655	/* Get the destination QP number. */
656	qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
657	if (qp_num == QIB_MULTICAST_QPN) {
658		struct qib_mcast *mcast;
659		struct qib_mcast_qp *p;
660
661		if (lnh != QIB_LRH_GRH)
662			goto drop;
663		mcast = qib_mcast_find(ibp, &hdr->u.l.grh.dgid);
664		if (mcast == NULL)
665			goto drop;
666		this_cpu_inc(ibp->pmastats->n_multicast_rcv);
667		list_for_each_entry_rcu(p, &mcast->qp_list, list)
668			qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
669		/*
670		 * Notify qib_multicast_detach() if it is waiting for us
671		 * to finish.
672		 */
673		if (atomic_dec_return(&mcast->refcount) <= 1)
674			wake_up(&mcast->wait);
675	} else {
676		if (rcd->lookaside_qp) {
677			if (rcd->lookaside_qpn != qp_num) {
678				if (atomic_dec_and_test(
679					&rcd->lookaside_qp->refcount))
680					wake_up(
681					 &rcd->lookaside_qp->wait);
682				rcd->lookaside_qp = NULL;
683			}
684		}
685		if (!rcd->lookaside_qp) {
686			qp = qib_lookup_qpn(ibp, qp_num);
687			if (!qp)
688				goto drop;
689			rcd->lookaside_qp = qp;
690			rcd->lookaside_qpn = qp_num;
691		} else
692			qp = rcd->lookaside_qp;
693		this_cpu_inc(ibp->pmastats->n_unicast_rcv);
694		qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
695	}
696	return;
697
698drop:
699	ibp->n_pkt_drops++;
700}
701
702/*
703 * This is called from a timer to check for QPs
704 * which need kernel memory in order to send a packet.
705 */
706static void mem_timer(unsigned long data)
707{
708	struct qib_ibdev *dev = (struct qib_ibdev *) data;
709	struct list_head *list = &dev->memwait;
710	struct qib_qp *qp = NULL;
711	unsigned long flags;
712
713	spin_lock_irqsave(&dev->pending_lock, flags);
714	if (!list_empty(list)) {
715		qp = list_entry(list->next, struct qib_qp, iowait);
716		list_del_init(&qp->iowait);
717		atomic_inc(&qp->refcount);
718		if (!list_empty(list))
719			mod_timer(&dev->mem_timer, jiffies + 1);
720	}
721	spin_unlock_irqrestore(&dev->pending_lock, flags);
722
723	if (qp) {
724		spin_lock_irqsave(&qp->s_lock, flags);
725		if (qp->s_flags & QIB_S_WAIT_KMEM) {
726			qp->s_flags &= ~QIB_S_WAIT_KMEM;
727			qib_schedule_send(qp);
728		}
729		spin_unlock_irqrestore(&qp->s_lock, flags);
730		if (atomic_dec_and_test(&qp->refcount))
731			wake_up(&qp->wait);
732	}
733}
734
735static void update_sge(struct qib_sge_state *ss, u32 length)
736{
737	struct qib_sge *sge = &ss->sge;
738
739	sge->vaddr += length;
740	sge->length -= length;
741	sge->sge_length -= length;
742	if (sge->sge_length == 0) {
743		if (--ss->num_sge)
744			*sge = *ss->sg_list++;
745	} else if (sge->length == 0 && sge->mr->lkey) {
746		if (++sge->n >= QIB_SEGSZ) {
747			if (++sge->m >= sge->mr->mapsz)
748				return;
749			sge->n = 0;
750		}
751		sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
752		sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
753	}
754}
755
756#ifdef __LITTLE_ENDIAN
757static inline u32 get_upper_bits(u32 data, u32 shift)
758{
759	return data >> shift;
760}
761
762static inline u32 set_upper_bits(u32 data, u32 shift)
763{
764	return data << shift;
765}
766
767static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
768{
769	data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
770	data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
771	return data;
772}
773#else
774static inline u32 get_upper_bits(u32 data, u32 shift)
775{
776	return data << shift;
777}
778
779static inline u32 set_upper_bits(u32 data, u32 shift)
780{
781	return data >> shift;
782}
783
784static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
785{
786	data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
787	data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
788	return data;
789}
790#endif
791
792static void copy_io(u32 __iomem *piobuf, struct qib_sge_state *ss,
793		    u32 length, unsigned flush_wc)
794{
795	u32 extra = 0;
796	u32 data = 0;
797	u32 last;
798
799	while (1) {
800		u32 len = ss->sge.length;
801		u32 off;
802
803		if (len > length)
804			len = length;
805		if (len > ss->sge.sge_length)
806			len = ss->sge.sge_length;
807		BUG_ON(len == 0);
808		/* If the source address is not aligned, try to align it. */
809		off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
810		if (off) {
811			u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
812					    ~(sizeof(u32) - 1));
813			u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
814			u32 y;
815
816			y = sizeof(u32) - off;
817			if (len > y)
818				len = y;
819			if (len + extra >= sizeof(u32)) {
820				data |= set_upper_bits(v, extra *
821						       BITS_PER_BYTE);
822				len = sizeof(u32) - extra;
823				if (len == length) {
824					last = data;
825					break;
826				}
827				__raw_writel(data, piobuf);
828				piobuf++;
829				extra = 0;
830				data = 0;
831			} else {
832				/* Clear unused upper bytes */
833				data |= clear_upper_bytes(v, len, extra);
834				if (len == length) {
835					last = data;
836					break;
837				}
838				extra += len;
839			}
840		} else if (extra) {
841			/* Source address is aligned. */
842			u32 *addr = (u32 *) ss->sge.vaddr;
843			int shift = extra * BITS_PER_BYTE;
844			int ushift = 32 - shift;
845			u32 l = len;
846
847			while (l >= sizeof(u32)) {
848				u32 v = *addr;
849
850				data |= set_upper_bits(v, shift);
851				__raw_writel(data, piobuf);
852				data = get_upper_bits(v, ushift);
853				piobuf++;
854				addr++;
855				l -= sizeof(u32);
856			}
857			/*
858			 * We still have 'extra' number of bytes leftover.
859			 */
860			if (l) {
861				u32 v = *addr;
862
863				if (l + extra >= sizeof(u32)) {
864					data |= set_upper_bits(v, shift);
865					len -= l + extra - sizeof(u32);
866					if (len == length) {
867						last = data;
868						break;
869					}
870					__raw_writel(data, piobuf);
871					piobuf++;
872					extra = 0;
873					data = 0;
874				} else {
875					/* Clear unused upper bytes */
876					data |= clear_upper_bytes(v, l, extra);
877					if (len == length) {
878						last = data;
879						break;
880					}
881					extra += l;
882				}
883			} else if (len == length) {
884				last = data;
885				break;
886			}
887		} else if (len == length) {
888			u32 w;
889
890			/*
891			 * Need to round up for the last dword in the
892			 * packet.
893			 */
894			w = (len + 3) >> 2;
895			qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
896			piobuf += w - 1;
897			last = ((u32 *) ss->sge.vaddr)[w - 1];
898			break;
899		} else {
900			u32 w = len >> 2;
901
902			qib_pio_copy(piobuf, ss->sge.vaddr, w);
903			piobuf += w;
904
905			extra = len & (sizeof(u32) - 1);
906			if (extra) {
907				u32 v = ((u32 *) ss->sge.vaddr)[w];
908
909				/* Clear unused upper bytes */
910				data = clear_upper_bytes(v, extra, 0);
911			}
912		}
913		update_sge(ss, len);
914		length -= len;
915	}
916	/* Update address before sending packet. */
917	update_sge(ss, length);
918	if (flush_wc) {
919		/* must flush early everything before trigger word */
920		qib_flush_wc();
921		__raw_writel(last, piobuf);
922		/* be sure trigger word is written */
923		qib_flush_wc();
924	} else
925		__raw_writel(last, piobuf);
926}
927
928static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
929					   struct qib_qp *qp)
930{
931	struct qib_verbs_txreq *tx;
932	unsigned long flags;
933
934	spin_lock_irqsave(&qp->s_lock, flags);
935	spin_lock(&dev->pending_lock);
936
937	if (!list_empty(&dev->txreq_free)) {
938		struct list_head *l = dev->txreq_free.next;
939
940		list_del(l);
941		spin_unlock(&dev->pending_lock);
942		spin_unlock_irqrestore(&qp->s_lock, flags);
943		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
944	} else {
945		if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK &&
946		    list_empty(&qp->iowait)) {
947			dev->n_txwait++;
948			qp->s_flags |= QIB_S_WAIT_TX;
949			list_add_tail(&qp->iowait, &dev->txwait);
950		}
951		qp->s_flags &= ~QIB_S_BUSY;
952		spin_unlock(&dev->pending_lock);
953		spin_unlock_irqrestore(&qp->s_lock, flags);
954		tx = ERR_PTR(-EBUSY);
955	}
956	return tx;
957}
958
959static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
960					 struct qib_qp *qp)
961{
962	struct qib_verbs_txreq *tx;
963	unsigned long flags;
964
965	spin_lock_irqsave(&dev->pending_lock, flags);
966	/* assume the list non empty */
967	if (likely(!list_empty(&dev->txreq_free))) {
968		struct list_head *l = dev->txreq_free.next;
969
970		list_del(l);
971		spin_unlock_irqrestore(&dev->pending_lock, flags);
972		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
973	} else {
974		/* call slow path to get the extra lock */
975		spin_unlock_irqrestore(&dev->pending_lock, flags);
976		tx =  __get_txreq(dev, qp);
977	}
978	return tx;
979}
980
981void qib_put_txreq(struct qib_verbs_txreq *tx)
982{
983	struct qib_ibdev *dev;
984	struct qib_qp *qp;
985	unsigned long flags;
986
987	qp = tx->qp;
988	dev = to_idev(qp->ibqp.device);
989
990	if (atomic_dec_and_test(&qp->refcount))
991		wake_up(&qp->wait);
992	if (tx->mr) {
993		qib_put_mr(tx->mr);
994		tx->mr = NULL;
995	}
996	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
997		tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
998		dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
999				 tx->txreq.addr, tx->hdr_dwords << 2,
1000				 DMA_TO_DEVICE);
1001		kfree(tx->align_buf);
1002	}
1003
1004	spin_lock_irqsave(&dev->pending_lock, flags);
1005
1006	/* Put struct back on free list */
1007	list_add(&tx->txreq.list, &dev->txreq_free);
1008
1009	if (!list_empty(&dev->txwait)) {
1010		/* Wake up first QP wanting a free struct */
1011		qp = list_entry(dev->txwait.next, struct qib_qp, iowait);
1012		list_del_init(&qp->iowait);
1013		atomic_inc(&qp->refcount);
1014		spin_unlock_irqrestore(&dev->pending_lock, flags);
1015
1016		spin_lock_irqsave(&qp->s_lock, flags);
1017		if (qp->s_flags & QIB_S_WAIT_TX) {
1018			qp->s_flags &= ~QIB_S_WAIT_TX;
1019			qib_schedule_send(qp);
1020		}
1021		spin_unlock_irqrestore(&qp->s_lock, flags);
1022
1023		if (atomic_dec_and_test(&qp->refcount))
1024			wake_up(&qp->wait);
1025	} else
1026		spin_unlock_irqrestore(&dev->pending_lock, flags);
1027}
1028
1029/*
1030 * This is called when there are send DMA descriptors that might be
1031 * available.
1032 *
1033 * This is called with ppd->sdma_lock held.
1034 */
1035void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
1036{
1037	struct qib_qp *qp, *nqp;
1038	struct qib_qp *qps[20];
1039	struct qib_ibdev *dev;
1040	unsigned i, n;
1041
1042	n = 0;
1043	dev = &ppd->dd->verbs_dev;
1044	spin_lock(&dev->pending_lock);
1045
1046	/* Search wait list for first QP wanting DMA descriptors. */
1047	list_for_each_entry_safe(qp, nqp, &dev->dmawait, iowait) {
1048		if (qp->port_num != ppd->port)
1049			continue;
1050		if (n == ARRAY_SIZE(qps))
1051			break;
1052		if (qp->s_tx->txreq.sg_count > avail)
1053			break;
1054		avail -= qp->s_tx->txreq.sg_count;
1055		list_del_init(&qp->iowait);
1056		atomic_inc(&qp->refcount);
1057		qps[n++] = qp;
1058	}
1059
1060	spin_unlock(&dev->pending_lock);
1061
1062	for (i = 0; i < n; i++) {
1063		qp = qps[i];
1064		spin_lock(&qp->s_lock);
1065		if (qp->s_flags & QIB_S_WAIT_DMA_DESC) {
1066			qp->s_flags &= ~QIB_S_WAIT_DMA_DESC;
1067			qib_schedule_send(qp);
1068		}
1069		spin_unlock(&qp->s_lock);
1070		if (atomic_dec_and_test(&qp->refcount))
1071			wake_up(&qp->wait);
1072	}
1073}
1074
1075/*
1076 * This is called with ppd->sdma_lock held.
1077 */
1078static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
1079{
1080	struct qib_verbs_txreq *tx =
1081		container_of(cookie, struct qib_verbs_txreq, txreq);
1082	struct qib_qp *qp = tx->qp;
1083
1084	spin_lock(&qp->s_lock);
1085	if (tx->wqe)
1086		qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
1087	else if (qp->ibqp.qp_type == IB_QPT_RC) {
1088		struct qib_ib_header *hdr;
1089
1090		if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
1091			hdr = &tx->align_buf->hdr;
1092		else {
1093			struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1094
1095			hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
1096		}
1097		qib_rc_send_complete(qp, hdr);
1098	}
1099	if (atomic_dec_and_test(&qp->s_dma_busy)) {
1100		if (qp->state == IB_QPS_RESET)
1101			wake_up(&qp->wait_dma);
1102		else if (qp->s_flags & QIB_S_WAIT_DMA) {
1103			qp->s_flags &= ~QIB_S_WAIT_DMA;
1104			qib_schedule_send(qp);
1105		}
1106	}
1107	spin_unlock(&qp->s_lock);
1108
1109	qib_put_txreq(tx);
1110}
1111
1112static int wait_kmem(struct qib_ibdev *dev, struct qib_qp *qp)
1113{
1114	unsigned long flags;
1115	int ret = 0;
1116
1117	spin_lock_irqsave(&qp->s_lock, flags);
1118	if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
1119		spin_lock(&dev->pending_lock);
1120		if (list_empty(&qp->iowait)) {
1121			if (list_empty(&dev->memwait))
1122				mod_timer(&dev->mem_timer, jiffies + 1);
1123			qp->s_flags |= QIB_S_WAIT_KMEM;
1124			list_add_tail(&qp->iowait, &dev->memwait);
1125		}
1126		spin_unlock(&dev->pending_lock);
1127		qp->s_flags &= ~QIB_S_BUSY;
1128		ret = -EBUSY;
1129	}
1130	spin_unlock_irqrestore(&qp->s_lock, flags);
1131
1132	return ret;
1133}
1134
1135static int qib_verbs_send_dma(struct qib_qp *qp, struct qib_ib_header *hdr,
1136			      u32 hdrwords, struct qib_sge_state *ss, u32 len,
1137			      u32 plen, u32 dwords)
1138{
1139	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1140	struct qib_devdata *dd = dd_from_dev(dev);
1141	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1142	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1143	struct qib_verbs_txreq *tx;
1144	struct qib_pio_header *phdr;
1145	u32 control;
1146	u32 ndesc;
1147	int ret;
1148
1149	tx = qp->s_tx;
1150	if (tx) {
1151		qp->s_tx = NULL;
1152		/* resend previously constructed packet */
1153		ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
1154		goto bail;
1155	}
1156
1157	tx = get_txreq(dev, qp);
1158	if (IS_ERR(tx))
1159		goto bail_tx;
1160
1161	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
1162				       be16_to_cpu(hdr->lrh[0]) >> 12);
1163	tx->qp = qp;
1164	atomic_inc(&qp->refcount);
1165	tx->wqe = qp->s_wqe;
1166	tx->mr = qp->s_rdma_mr;
1167	if (qp->s_rdma_mr)
1168		qp->s_rdma_mr = NULL;
1169	tx->txreq.callback = sdma_complete;
1170	if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
1171		tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
1172	else
1173		tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
1174	if (plen + 1 > dd->piosize2kmax_dwords)
1175		tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
1176
1177	if (len) {
1178		/*
1179		 * Don't try to DMA if it takes more descriptors than
1180		 * the queue holds.
1181		 */
1182		ndesc = qib_count_sge(ss, len);
1183		if (ndesc >= ppd->sdma_descq_cnt)
1184			ndesc = 0;
1185	} else
1186		ndesc = 1;
1187	if (ndesc) {
1188		phdr = &dev->pio_hdrs[tx->hdr_inx];
1189		phdr->pbc[0] = cpu_to_le32(plen);
1190		phdr->pbc[1] = cpu_to_le32(control);
1191		memcpy(&phdr->hdr, hdr, hdrwords << 2);
1192		tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
1193		tx->txreq.sg_count = ndesc;
1194		tx->txreq.addr = dev->pio_hdrs_phys +
1195			tx->hdr_inx * sizeof(struct qib_pio_header);
1196		tx->hdr_dwords = hdrwords + 2; /* add PBC length */
1197		ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
1198		goto bail;
1199	}
1200
1201	/* Allocate a buffer and copy the header and payload to it. */
1202	tx->hdr_dwords = plen + 1;
1203	phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
1204	if (!phdr)
1205		goto err_tx;
1206	phdr->pbc[0] = cpu_to_le32(plen);
1207	phdr->pbc[1] = cpu_to_le32(control);
1208	memcpy(&phdr->hdr, hdr, hdrwords << 2);
1209	qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
1210
1211	tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
1212					tx->hdr_dwords << 2, DMA_TO_DEVICE);
1213	if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
1214		goto map_err;
1215	tx->align_buf = phdr;
1216	tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
1217	tx->txreq.sg_count = 1;
1218	ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
1219	goto unaligned;
1220
1221map_err:
1222	kfree(phdr);
1223err_tx:
1224	qib_put_txreq(tx);
1225	ret = wait_kmem(dev, qp);
1226unaligned:
1227	ibp->n_unaligned++;
1228bail:
1229	return ret;
1230bail_tx:
1231	ret = PTR_ERR(tx);
1232	goto bail;
1233}
1234
1235/*
1236 * If we are now in the error state, return zero to flush the
1237 * send work request.
1238 */
1239static int no_bufs_available(struct qib_qp *qp)
1240{
1241	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1242	struct qib_devdata *dd;
1243	unsigned long flags;
1244	int ret = 0;
1245
1246	/*
1247	 * Note that as soon as want_buffer() is called and
1248	 * possibly before it returns, qib_ib_piobufavail()
1249	 * could be called. Therefore, put QP on the I/O wait list before
1250	 * enabling the PIO avail interrupt.
1251	 */
1252	spin_lock_irqsave(&qp->s_lock, flags);
1253	if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
1254		spin_lock(&dev->pending_lock);
1255		if (list_empty(&qp->iowait)) {
1256			dev->n_piowait++;
1257			qp->s_flags |= QIB_S_WAIT_PIO;
1258			list_add_tail(&qp->iowait, &dev->piowait);
1259			dd = dd_from_dev(dev);
1260			dd->f_wantpiobuf_intr(dd, 1);
1261		}
1262		spin_unlock(&dev->pending_lock);
1263		qp->s_flags &= ~QIB_S_BUSY;
1264		ret = -EBUSY;
1265	}
1266	spin_unlock_irqrestore(&qp->s_lock, flags);
1267	return ret;
1268}
1269
1270static int qib_verbs_send_pio(struct qib_qp *qp, struct qib_ib_header *ibhdr,
1271			      u32 hdrwords, struct qib_sge_state *ss, u32 len,
1272			      u32 plen, u32 dwords)
1273{
1274	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1275	struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
1276	u32 *hdr = (u32 *) ibhdr;
1277	u32 __iomem *piobuf_orig;
1278	u32 __iomem *piobuf;
1279	u64 pbc;
1280	unsigned long flags;
1281	unsigned flush_wc;
1282	u32 control;
1283	u32 pbufn;
1284
1285	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
1286		be16_to_cpu(ibhdr->lrh[0]) >> 12);
1287	pbc = ((u64) control << 32) | plen;
1288	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
1289	if (unlikely(piobuf == NULL))
1290		return no_bufs_available(qp);
1291
1292	/*
1293	 * Write the pbc.
1294	 * We have to flush after the PBC for correctness on some cpus
1295	 * or WC buffer can be written out of order.
1296	 */
1297	writeq(pbc, piobuf);
1298	piobuf_orig = piobuf;
1299	piobuf += 2;
1300
1301	flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
1302	if (len == 0) {
1303		/*
1304		 * If there is just the header portion, must flush before
1305		 * writing last word of header for correctness, and after
1306		 * the last header word (trigger word).
1307		 */
1308		if (flush_wc) {
1309			qib_flush_wc();
1310			qib_pio_copy(piobuf, hdr, hdrwords - 1);
1311			qib_flush_wc();
1312			__raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1313			qib_flush_wc();
1314		} else
1315			qib_pio_copy(piobuf, hdr, hdrwords);
1316		goto done;
1317	}
1318
1319	if (flush_wc)
1320		qib_flush_wc();
1321	qib_pio_copy(piobuf, hdr, hdrwords);
1322	piobuf += hdrwords;
1323
1324	/* The common case is aligned and contained in one segment. */
1325	if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1326		   !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1327		u32 *addr = (u32 *) ss->sge.vaddr;
1328
1329		/* Update address before sending packet. */
1330		update_sge(ss, len);
1331		if (flush_wc) {
1332			qib_pio_copy(piobuf, addr, dwords - 1);
1333			/* must flush early everything before trigger word */
1334			qib_flush_wc();
1335			__raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1336			/* be sure trigger word is written */
1337			qib_flush_wc();
1338		} else
1339			qib_pio_copy(piobuf, addr, dwords);
1340		goto done;
1341	}
1342	copy_io(piobuf, ss, len, flush_wc);
1343done:
1344	if (dd->flags & QIB_USE_SPCL_TRIG) {
1345		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
1346
1347		qib_flush_wc();
1348		__raw_writel(0xaebecede, piobuf_orig + spcl_off);
1349	}
1350	qib_sendbuf_done(dd, pbufn);
1351	if (qp->s_rdma_mr) {
1352		qib_put_mr(qp->s_rdma_mr);
1353		qp->s_rdma_mr = NULL;
1354	}
1355	if (qp->s_wqe) {
1356		spin_lock_irqsave(&qp->s_lock, flags);
1357		qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1358		spin_unlock_irqrestore(&qp->s_lock, flags);
1359	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1360		spin_lock_irqsave(&qp->s_lock, flags);
1361		qib_rc_send_complete(qp, ibhdr);
1362		spin_unlock_irqrestore(&qp->s_lock, flags);
1363	}
1364	return 0;
1365}
1366
1367/**
1368 * qib_verbs_send - send a packet
1369 * @qp: the QP to send on
1370 * @hdr: the packet header
1371 * @hdrwords: the number of 32-bit words in the header
1372 * @ss: the SGE to send
1373 * @len: the length of the packet in bytes
1374 *
1375 * Return zero if packet is sent or queued OK.
1376 * Return non-zero and clear qp->s_flags QIB_S_BUSY otherwise.
1377 */
1378int qib_verbs_send(struct qib_qp *qp, struct qib_ib_header *hdr,
1379		   u32 hdrwords, struct qib_sge_state *ss, u32 len)
1380{
1381	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1382	u32 plen;
1383	int ret;
1384	u32 dwords = (len + 3) >> 2;
1385
1386	/*
1387	 * Calculate the send buffer trigger address.
1388	 * The +1 counts for the pbc control dword following the pbc length.
1389	 */
1390	plen = hdrwords + dwords + 1;
1391
1392	/*
1393	 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1394	 * can defer SDMA restart until link goes ACTIVE without
1395	 * worrying about just how we got there.
1396	 */
1397	if (qp->ibqp.qp_type == IB_QPT_SMI ||
1398	    !(dd->flags & QIB_HAS_SEND_DMA))
1399		ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1400					 plen, dwords);
1401	else
1402		ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1403					 plen, dwords);
1404
1405	return ret;
1406}
1407
1408int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1409			  u64 *rwords, u64 *spkts, u64 *rpkts,
1410			  u64 *xmit_wait)
1411{
1412	int ret;
1413	struct qib_devdata *dd = ppd->dd;
1414
1415	if (!(dd->flags & QIB_PRESENT)) {
1416		/* no hardware, freeze, etc. */
1417		ret = -EINVAL;
1418		goto bail;
1419	}
1420	*swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1421	*rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1422	*spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1423	*rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1424	*xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1425
1426	ret = 0;
1427
1428bail:
1429	return ret;
1430}
1431
1432/**
1433 * qib_get_counters - get various chip counters
1434 * @dd: the qlogic_ib device
1435 * @cntrs: counters are placed here
1436 *
1437 * Return the counters needed by recv_pma_get_portcounters().
1438 */
1439int qib_get_counters(struct qib_pportdata *ppd,
1440		     struct qib_verbs_counters *cntrs)
1441{
1442	int ret;
1443
1444	if (!(ppd->dd->flags & QIB_PRESENT)) {
1445		/* no hardware, freeze, etc. */
1446		ret = -EINVAL;
1447		goto bail;
1448	}
1449	cntrs->symbol_error_counter =
1450		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1451	cntrs->link_error_recovery_counter =
1452		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1453	/*
1454	 * The link downed counter counts when the other side downs the
1455	 * connection.  We add in the number of times we downed the link
1456	 * due to local link integrity errors to compensate.
1457	 */
1458	cntrs->link_downed_counter =
1459		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1460	cntrs->port_rcv_errors =
1461		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1462		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1463		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1464		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1465		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1466		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1467		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1468		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1469		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1470	cntrs->port_rcv_errors +=
1471		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1472	cntrs->port_rcv_errors +=
1473		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1474	cntrs->port_rcv_remphys_errors =
1475		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1476	cntrs->port_xmit_discards =
1477		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1478	cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1479			QIBPORTCNTR_WORDSEND);
1480	cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1481			QIBPORTCNTR_WORDRCV);
1482	cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1483			QIBPORTCNTR_PKTSEND);
1484	cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1485			QIBPORTCNTR_PKTRCV);
1486	cntrs->local_link_integrity_errors =
1487		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1488	cntrs->excessive_buffer_overrun_errors =
1489		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1490	cntrs->vl15_dropped =
1491		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1492
1493	ret = 0;
1494
1495bail:
1496	return ret;
1497}
1498
1499/**
1500 * qib_ib_piobufavail - callback when a PIO buffer is available
1501 * @dd: the device pointer
1502 *
1503 * This is called from qib_intr() at interrupt level when a PIO buffer is
1504 * available after qib_verbs_send() returned an error that no buffers were
1505 * available. Disable the interrupt if there are no more QPs waiting.
1506 */
1507void qib_ib_piobufavail(struct qib_devdata *dd)
1508{
1509	struct qib_ibdev *dev = &dd->verbs_dev;
1510	struct list_head *list;
1511	struct qib_qp *qps[5];
1512	struct qib_qp *qp;
1513	unsigned long flags;
1514	unsigned i, n;
1515
1516	list = &dev->piowait;
1517	n = 0;
1518
1519	/*
1520	 * Note: checking that the piowait list is empty and clearing
1521	 * the buffer available interrupt needs to be atomic or we
1522	 * could end up with QPs on the wait list with the interrupt
1523	 * disabled.
1524	 */
1525	spin_lock_irqsave(&dev->pending_lock, flags);
1526	while (!list_empty(list)) {
1527		if (n == ARRAY_SIZE(qps))
1528			goto full;
1529		qp = list_entry(list->next, struct qib_qp, iowait);
1530		list_del_init(&qp->iowait);
1531		atomic_inc(&qp->refcount);
1532		qps[n++] = qp;
1533	}
1534	dd->f_wantpiobuf_intr(dd, 0);
1535full:
1536	spin_unlock_irqrestore(&dev->pending_lock, flags);
1537
1538	for (i = 0; i < n; i++) {
1539		qp = qps[i];
1540
1541		spin_lock_irqsave(&qp->s_lock, flags);
1542		if (qp->s_flags & QIB_S_WAIT_PIO) {
1543			qp->s_flags &= ~QIB_S_WAIT_PIO;
1544			qib_schedule_send(qp);
1545		}
1546		spin_unlock_irqrestore(&qp->s_lock, flags);
1547
1548		/* Notify qib_destroy_qp() if it is waiting. */
1549		if (atomic_dec_and_test(&qp->refcount))
1550			wake_up(&qp->wait);
1551	}
1552}
1553
1554static int qib_query_device(struct ib_device *ibdev,
1555			    struct ib_device_attr *props)
1556{
1557	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1558	struct qib_ibdev *dev = to_idev(ibdev);
1559
1560	memset(props, 0, sizeof(*props));
1561
1562	props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1563		IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1564		IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1565		IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1566	props->page_size_cap = PAGE_SIZE;
1567	props->vendor_id =
1568		QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1569	props->vendor_part_id = dd->deviceid;
1570	props->hw_ver = dd->minrev;
1571	props->sys_image_guid = ib_qib_sys_image_guid;
1572	props->max_mr_size = ~0ULL;
1573	props->max_qp = ib_qib_max_qps;
1574	props->max_qp_wr = ib_qib_max_qp_wrs;
1575	props->max_sge = ib_qib_max_sges;
1576	props->max_cq = ib_qib_max_cqs;
1577	props->max_ah = ib_qib_max_ahs;
1578	props->max_cqe = ib_qib_max_cqes;
1579	props->max_mr = dev->lk_table.max;
1580	props->max_fmr = dev->lk_table.max;
1581	props->max_map_per_fmr = 32767;
1582	props->max_pd = ib_qib_max_pds;
1583	props->max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1584	props->max_qp_init_rd_atom = 255;
1585	/* props->max_res_rd_atom */
1586	props->max_srq = ib_qib_max_srqs;
1587	props->max_srq_wr = ib_qib_max_srq_wrs;
1588	props->max_srq_sge = ib_qib_max_srq_sges;
1589	/* props->local_ca_ack_delay */
1590	props->atomic_cap = IB_ATOMIC_GLOB;
1591	props->max_pkeys = qib_get_npkeys(dd);
1592	props->max_mcast_grp = ib_qib_max_mcast_grps;
1593	props->max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1594	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1595		props->max_mcast_grp;
1596
1597	return 0;
1598}
1599
1600static int qib_query_port(struct ib_device *ibdev, u8 port,
1601			  struct ib_port_attr *props)
1602{
1603	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1604	struct qib_ibport *ibp = to_iport(ibdev, port);
1605	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1606	enum ib_mtu mtu;
1607	u16 lid = ppd->lid;
1608
1609	memset(props, 0, sizeof(*props));
1610	props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1611	props->lmc = ppd->lmc;
1612	props->sm_lid = ibp->sm_lid;
1613	props->sm_sl = ibp->sm_sl;
1614	props->state = dd->f_iblink_state(ppd->lastibcstat);
1615	props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1616	props->port_cap_flags = ibp->port_cap_flags;
1617	props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1618	props->max_msg_sz = 0x80000000;
1619	props->pkey_tbl_len = qib_get_npkeys(dd);
1620	props->bad_pkey_cntr = ibp->pkey_violations;
1621	props->qkey_viol_cntr = ibp->qkey_violations;
1622	props->active_width = ppd->link_width_active;
1623	/* See rate_show() */
1624	props->active_speed = ppd->link_speed_active;
1625	props->max_vl_num = qib_num_vls(ppd->vls_supported);
1626	props->init_type_reply = 0;
1627
1628	props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1629	switch (ppd->ibmtu) {
1630	case 4096:
1631		mtu = IB_MTU_4096;
1632		break;
1633	case 2048:
1634		mtu = IB_MTU_2048;
1635		break;
1636	case 1024:
1637		mtu = IB_MTU_1024;
1638		break;
1639	case 512:
1640		mtu = IB_MTU_512;
1641		break;
1642	case 256:
1643		mtu = IB_MTU_256;
1644		break;
1645	default:
1646		mtu = IB_MTU_2048;
1647	}
1648	props->active_mtu = mtu;
1649	props->subnet_timeout = ibp->subnet_timeout;
1650
1651	return 0;
1652}
1653
1654static int qib_modify_device(struct ib_device *device,
1655			     int device_modify_mask,
1656			     struct ib_device_modify *device_modify)
1657{
1658	struct qib_devdata *dd = dd_from_ibdev(device);
1659	unsigned i;
1660	int ret;
1661
1662	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1663				   IB_DEVICE_MODIFY_NODE_DESC)) {
1664		ret = -EOPNOTSUPP;
1665		goto bail;
1666	}
1667
1668	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1669		memcpy(device->node_desc, device_modify->node_desc, 64);
1670		for (i = 0; i < dd->num_pports; i++) {
1671			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1672
1673			qib_node_desc_chg(ibp);
1674		}
1675	}
1676
1677	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1678		ib_qib_sys_image_guid =
1679			cpu_to_be64(device_modify->sys_image_guid);
1680		for (i = 0; i < dd->num_pports; i++) {
1681			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1682
1683			qib_sys_guid_chg(ibp);
1684		}
1685	}
1686
1687	ret = 0;
1688
1689bail:
1690	return ret;
1691}
1692
1693static int qib_modify_port(struct ib_device *ibdev, u8 port,
1694			   int port_modify_mask, struct ib_port_modify *props)
1695{
1696	struct qib_ibport *ibp = to_iport(ibdev, port);
1697	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1698
1699	ibp->port_cap_flags |= props->set_port_cap_mask;
1700	ibp->port_cap_flags &= ~props->clr_port_cap_mask;
1701	if (props->set_port_cap_mask || props->clr_port_cap_mask)
1702		qib_cap_mask_chg(ibp);
1703	if (port_modify_mask & IB_PORT_SHUTDOWN)
1704		qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1705	if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1706		ibp->qkey_violations = 0;
1707	return 0;
1708}
1709
1710static int qib_query_gid(struct ib_device *ibdev, u8 port,
1711			 int index, union ib_gid *gid)
1712{
1713	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1714	int ret = 0;
1715
1716	if (!port || port > dd->num_pports)
1717		ret = -EINVAL;
1718	else {
1719		struct qib_ibport *ibp = to_iport(ibdev, port);
1720		struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1721
1722		gid->global.subnet_prefix = ibp->gid_prefix;
1723		if (index == 0)
1724			gid->global.interface_id = ppd->guid;
1725		else if (index < QIB_GUIDS_PER_PORT)
1726			gid->global.interface_id = ibp->guids[index - 1];
1727		else
1728			ret = -EINVAL;
1729	}
1730
1731	return ret;
1732}
1733
1734static struct ib_pd *qib_alloc_pd(struct ib_device *ibdev,
1735				  struct ib_ucontext *context,
1736				  struct ib_udata *udata)
1737{
1738	struct qib_ibdev *dev = to_idev(ibdev);
1739	struct qib_pd *pd;
1740	struct ib_pd *ret;
1741
1742	/*
1743	 * This is actually totally arbitrary.  Some correctness tests
1744	 * assume there's a maximum number of PDs that can be allocated.
1745	 * We don't actually have this limit, but we fail the test if
1746	 * we allow allocations of more than we report for this value.
1747	 */
1748
1749	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1750	if (!pd) {
1751		ret = ERR_PTR(-ENOMEM);
1752		goto bail;
1753	}
1754
1755	spin_lock(&dev->n_pds_lock);
1756	if (dev->n_pds_allocated == ib_qib_max_pds) {
1757		spin_unlock(&dev->n_pds_lock);
1758		kfree(pd);
1759		ret = ERR_PTR(-ENOMEM);
1760		goto bail;
1761	}
1762
1763	dev->n_pds_allocated++;
1764	spin_unlock(&dev->n_pds_lock);
1765
1766	/* ib_alloc_pd() will initialize pd->ibpd. */
1767	pd->user = udata != NULL;
1768
1769	ret = &pd->ibpd;
1770
1771bail:
1772	return ret;
1773}
1774
1775static int qib_dealloc_pd(struct ib_pd *ibpd)
1776{
1777	struct qib_pd *pd = to_ipd(ibpd);
1778	struct qib_ibdev *dev = to_idev(ibpd->device);
1779
1780	spin_lock(&dev->n_pds_lock);
1781	dev->n_pds_allocated--;
1782	spin_unlock(&dev->n_pds_lock);
1783
1784	kfree(pd);
1785
1786	return 0;
1787}
1788
1789int qib_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
1790{
1791	/* A multicast address requires a GRH (see ch. 8.4.1). */
1792	if (ah_attr->dlid >= QIB_MULTICAST_LID_BASE &&
1793	    ah_attr->dlid != QIB_PERMISSIVE_LID &&
1794	    !(ah_attr->ah_flags & IB_AH_GRH))
1795		goto bail;
1796	if ((ah_attr->ah_flags & IB_AH_GRH) &&
1797	    ah_attr->grh.sgid_index >= QIB_GUIDS_PER_PORT)
1798		goto bail;
1799	if (ah_attr->dlid == 0)
1800		goto bail;
1801	if (ah_attr->port_num < 1 ||
1802	    ah_attr->port_num > ibdev->phys_port_cnt)
1803		goto bail;
1804	if (ah_attr->static_rate != IB_RATE_PORT_CURRENT &&
1805	    ib_rate_to_mult(ah_attr->static_rate) < 0)
1806		goto bail;
1807	if (ah_attr->sl > 15)
1808		goto bail;
1809	return 0;
1810bail:
1811	return -EINVAL;
1812}
1813
1814/**
1815 * qib_create_ah - create an address handle
1816 * @pd: the protection domain
1817 * @ah_attr: the attributes of the AH
1818 *
1819 * This may be called from interrupt context.
1820 */
1821static struct ib_ah *qib_create_ah(struct ib_pd *pd,
1822				   struct ib_ah_attr *ah_attr)
1823{
1824	struct qib_ah *ah;
1825	struct ib_ah *ret;
1826	struct qib_ibdev *dev = to_idev(pd->device);
1827	unsigned long flags;
1828
1829	if (qib_check_ah(pd->device, ah_attr)) {
1830		ret = ERR_PTR(-EINVAL);
1831		goto bail;
1832	}
1833
1834	ah = kmalloc(sizeof(*ah), GFP_ATOMIC);
1835	if (!ah) {
1836		ret = ERR_PTR(-ENOMEM);
1837		goto bail;
1838	}
1839
1840	spin_lock_irqsave(&dev->n_ahs_lock, flags);
1841	if (dev->n_ahs_allocated == ib_qib_max_ahs) {
1842		spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1843		kfree(ah);
1844		ret = ERR_PTR(-ENOMEM);
1845		goto bail;
1846	}
1847
1848	dev->n_ahs_allocated++;
1849	spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1850
1851	/* ib_create_ah() will initialize ah->ibah. */
1852	ah->attr = *ah_attr;
1853	atomic_set(&ah->refcount, 0);
1854
1855	ret = &ah->ibah;
1856
1857bail:
1858	return ret;
1859}
1860
1861struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1862{
1863	struct ib_ah_attr attr;
1864	struct ib_ah *ah = ERR_PTR(-EINVAL);
1865	struct qib_qp *qp0;
1866
1867	memset(&attr, 0, sizeof(attr));
1868	attr.dlid = dlid;
1869	attr.port_num = ppd_from_ibp(ibp)->port;
1870	rcu_read_lock();
1871	qp0 = rcu_dereference(ibp->qp0);
1872	if (qp0)
1873		ah = ib_create_ah(qp0->ibqp.pd, &attr);
1874	rcu_read_unlock();
1875	return ah;
1876}
1877
1878/**
1879 * qib_destroy_ah - destroy an address handle
1880 * @ibah: the AH to destroy
1881 *
1882 * This may be called from interrupt context.
1883 */
1884static int qib_destroy_ah(struct ib_ah *ibah)
1885{
1886	struct qib_ibdev *dev = to_idev(ibah->device);
1887	struct qib_ah *ah = to_iah(ibah);
1888	unsigned long flags;
1889
1890	if (atomic_read(&ah->refcount) != 0)
1891		return -EBUSY;
1892
1893	spin_lock_irqsave(&dev->n_ahs_lock, flags);
1894	dev->n_ahs_allocated--;
1895	spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1896
1897	kfree(ah);
1898
1899	return 0;
1900}
1901
1902static int qib_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1903{
1904	struct qib_ah *ah = to_iah(ibah);
1905
1906	if (qib_check_ah(ibah->device, ah_attr))
1907		return -EINVAL;
1908
1909	ah->attr = *ah_attr;
1910
1911	return 0;
1912}
1913
1914static int qib_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1915{
1916	struct qib_ah *ah = to_iah(ibah);
1917
1918	*ah_attr = ah->attr;
1919
1920	return 0;
1921}
1922
1923/**
1924 * qib_get_npkeys - return the size of the PKEY table for context 0
1925 * @dd: the qlogic_ib device
1926 */
1927unsigned qib_get_npkeys(struct qib_devdata *dd)
1928{
1929	return ARRAY_SIZE(dd->rcd[0]->pkeys);
1930}
1931
1932/*
1933 * Return the indexed PKEY from the port PKEY table.
1934 * No need to validate rcd[ctxt]; the port is setup if we are here.
1935 */
1936unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1937{
1938	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1939	struct qib_devdata *dd = ppd->dd;
1940	unsigned ctxt = ppd->hw_pidx;
1941	unsigned ret;
1942
1943	/* dd->rcd null if mini_init or some init failures */
1944	if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1945		ret = 0;
1946	else
1947		ret = dd->rcd[ctxt]->pkeys[index];
1948
1949	return ret;
1950}
1951
1952static int qib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1953			  u16 *pkey)
1954{
1955	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1956	int ret;
1957
1958	if (index >= qib_get_npkeys(dd)) {
1959		ret = -EINVAL;
1960		goto bail;
1961	}
1962
1963	*pkey = qib_get_pkey(to_iport(ibdev, port), index);
1964	ret = 0;
1965
1966bail:
1967	return ret;
1968}
1969
1970/**
1971 * qib_alloc_ucontext - allocate a ucontest
1972 * @ibdev: the infiniband device
1973 * @udata: not used by the QLogic_IB driver
1974 */
1975
1976static struct ib_ucontext *qib_alloc_ucontext(struct ib_device *ibdev,
1977					      struct ib_udata *udata)
1978{
1979	struct qib_ucontext *context;
1980	struct ib_ucontext *ret;
1981
1982	context = kmalloc(sizeof(*context), GFP_KERNEL);
1983	if (!context) {
1984		ret = ERR_PTR(-ENOMEM);
1985		goto bail;
1986	}
1987
1988	ret = &context->ibucontext;
1989
1990bail:
1991	return ret;
1992}
1993
1994static int qib_dealloc_ucontext(struct ib_ucontext *context)
1995{
1996	kfree(to_iucontext(context));
1997	return 0;
1998}
1999
2000static void init_ibport(struct qib_pportdata *ppd)
2001{
2002	struct qib_verbs_counters cntrs;
2003	struct qib_ibport *ibp = &ppd->ibport_data;
2004
2005	spin_lock_init(&ibp->lock);
2006	/* Set the prefix to the default value (see ch. 4.1.1) */
2007	ibp->gid_prefix = IB_DEFAULT_GID_PREFIX;
2008	ibp->sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
2009	ibp->port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
2010		IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
2011		IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
2012		IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
2013		IB_PORT_OTHER_LOCAL_CHANGES_SUP;
2014	if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
2015		ibp->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2016	ibp->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2017	ibp->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2018	ibp->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2019	ibp->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2020	ibp->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2021
2022	/* Snapshot current HW counters to "clear" them. */
2023	qib_get_counters(ppd, &cntrs);
2024	ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
2025	ibp->z_link_error_recovery_counter =
2026		cntrs.link_error_recovery_counter;
2027	ibp->z_link_downed_counter = cntrs.link_downed_counter;
2028	ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
2029	ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
2030	ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
2031	ibp->z_port_xmit_data = cntrs.port_xmit_data;
2032	ibp->z_port_rcv_data = cntrs.port_rcv_data;
2033	ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
2034	ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
2035	ibp->z_local_link_integrity_errors =
2036		cntrs.local_link_integrity_errors;
2037	ibp->z_excessive_buffer_overrun_errors =
2038		cntrs.excessive_buffer_overrun_errors;
2039	ibp->z_vl15_dropped = cntrs.vl15_dropped;
2040	RCU_INIT_POINTER(ibp->qp0, NULL);
2041	RCU_INIT_POINTER(ibp->qp1, NULL);
2042}
2043
2044/**
2045 * qib_register_ib_device - register our device with the infiniband core
2046 * @dd: the device data structure
2047 * Return the allocated qib_ibdev pointer or NULL on error.
2048 */
2049int qib_register_ib_device(struct qib_devdata *dd)
2050{
2051	struct qib_ibdev *dev = &dd->verbs_dev;
2052	struct ib_device *ibdev = &dev->ibdev;
2053	struct qib_pportdata *ppd = dd->pport;
2054	unsigned i, lk_tab_size;
2055	int ret;
2056
2057	dev->qp_table_size = ib_qib_qp_table_size;
2058	get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
2059	dev->qp_table = kmalloc_array(
2060				dev->qp_table_size,
2061				sizeof(*dev->qp_table),
2062				GFP_KERNEL);
2063	if (!dev->qp_table) {
2064		ret = -ENOMEM;
2065		goto err_qpt;
2066	}
2067	for (i = 0; i < dev->qp_table_size; i++)
2068		RCU_INIT_POINTER(dev->qp_table[i], NULL);
2069
2070	for (i = 0; i < dd->num_pports; i++)
2071		init_ibport(ppd + i);
2072
2073	/* Only need to initialize non-zero fields. */
2074	spin_lock_init(&dev->qpt_lock);
2075	spin_lock_init(&dev->n_pds_lock);
2076	spin_lock_init(&dev->n_ahs_lock);
2077	spin_lock_init(&dev->n_cqs_lock);
2078	spin_lock_init(&dev->n_qps_lock);
2079	spin_lock_init(&dev->n_srqs_lock);
2080	spin_lock_init(&dev->n_mcast_grps_lock);
2081	init_timer(&dev->mem_timer);
2082	dev->mem_timer.function = mem_timer;
2083	dev->mem_timer.data = (unsigned long) dev;
2084
2085	qib_init_qpn_table(dd, &dev->qpn_table);
2086
2087	/*
2088	 * The top ib_qib_lkey_table_size bits are used to index the
2089	 * table.  The lower 8 bits can be owned by the user (copied from
2090	 * the LKEY).  The remaining bits act as a generation number or tag.
2091	 */
2092	spin_lock_init(&dev->lk_table.lock);
2093	/* insure generation is at least 4 bits see keys.c */
2094	if (ib_qib_lkey_table_size > MAX_LKEY_TABLE_BITS) {
2095		qib_dev_warn(dd, "lkey bits %u too large, reduced to %u\n",
2096			ib_qib_lkey_table_size, MAX_LKEY_TABLE_BITS);
2097		ib_qib_lkey_table_size = MAX_LKEY_TABLE_BITS;
2098	}
2099	dev->lk_table.max = 1 << ib_qib_lkey_table_size;
2100	lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
2101	dev->lk_table.table = (struct qib_mregion __rcu **)
2102		vmalloc(lk_tab_size);
2103	if (dev->lk_table.table == NULL) {
2104		ret = -ENOMEM;
2105		goto err_lk;
2106	}
2107	RCU_INIT_POINTER(dev->dma_mr, NULL);
2108	for (i = 0; i < dev->lk_table.max; i++)
2109		RCU_INIT_POINTER(dev->lk_table.table[i], NULL);
2110	INIT_LIST_HEAD(&dev->pending_mmaps);
2111	spin_lock_init(&dev->pending_lock);
2112	dev->mmap_offset = PAGE_SIZE;
2113	spin_lock_init(&dev->mmap_offset_lock);
2114	INIT_LIST_HEAD(&dev->piowait);
2115	INIT_LIST_HEAD(&dev->dmawait);
2116	INIT_LIST_HEAD(&dev->txwait);
2117	INIT_LIST_HEAD(&dev->memwait);
2118	INIT_LIST_HEAD(&dev->txreq_free);
2119
2120	if (ppd->sdma_descq_cnt) {
2121		dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
2122						ppd->sdma_descq_cnt *
2123						sizeof(struct qib_pio_header),
2124						&dev->pio_hdrs_phys,
2125						GFP_KERNEL);
2126		if (!dev->pio_hdrs) {
2127			ret = -ENOMEM;
2128			goto err_hdrs;
2129		}
2130	}
2131
2132	for (i = 0; i < ppd->sdma_descq_cnt; i++) {
2133		struct qib_verbs_txreq *tx;
2134
2135		tx = kzalloc(sizeof(*tx), GFP_KERNEL);
2136		if (!tx) {
2137			ret = -ENOMEM;
2138			goto err_tx;
2139		}
2140		tx->hdr_inx = i;
2141		list_add(&tx->txreq.list, &dev->txreq_free);
2142	}
2143
2144	/*
2145	 * The system image GUID is supposed to be the same for all
2146	 * IB HCAs in a single system but since there can be other
2147	 * device types in the system, we can't be sure this is unique.
2148	 */
2149	if (!ib_qib_sys_image_guid)
2150		ib_qib_sys_image_guid = ppd->guid;
2151
2152	strlcpy(ibdev->name, "qib%d", IB_DEVICE_NAME_MAX);
2153	ibdev->owner = THIS_MODULE;
2154	ibdev->node_guid = ppd->guid;
2155	ibdev->uverbs_abi_ver = QIB_UVERBS_ABI_VERSION;
2156	ibdev->uverbs_cmd_mask =
2157		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2158		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2159		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2160		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2161		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2162		(1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
2163		(1ull << IB_USER_VERBS_CMD_MODIFY_AH)           |
2164		(1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
2165		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
2166		(1ull << IB_USER_VERBS_CMD_REG_MR)              |
2167		(1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2168		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2169		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2170		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2171		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2172		(1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
2173		(1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
2174		(1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2175		(1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2176		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2177		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2178		(1ull << IB_USER_VERBS_CMD_POST_SEND)           |
2179		(1ull << IB_USER_VERBS_CMD_POST_RECV)           |
2180		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2181		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2182		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2183		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2184		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2185		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2186		(1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2187	ibdev->node_type = RDMA_NODE_IB_CA;
2188	ibdev->phys_port_cnt = dd->num_pports;
2189	ibdev->num_comp_vectors = 1;
2190	ibdev->dma_device = &dd->pcidev->dev;
2191	ibdev->query_device = qib_query_device;
2192	ibdev->modify_device = qib_modify_device;
2193	ibdev->query_port = qib_query_port;
2194	ibdev->modify_port = qib_modify_port;
2195	ibdev->query_pkey = qib_query_pkey;
2196	ibdev->query_gid = qib_query_gid;
2197	ibdev->alloc_ucontext = qib_alloc_ucontext;
2198	ibdev->dealloc_ucontext = qib_dealloc_ucontext;
2199	ibdev->alloc_pd = qib_alloc_pd;
2200	ibdev->dealloc_pd = qib_dealloc_pd;
2201	ibdev->create_ah = qib_create_ah;
2202	ibdev->destroy_ah = qib_destroy_ah;
2203	ibdev->modify_ah = qib_modify_ah;
2204	ibdev->query_ah = qib_query_ah;
2205	ibdev->create_srq = qib_create_srq;
2206	ibdev->modify_srq = qib_modify_srq;
2207	ibdev->query_srq = qib_query_srq;
2208	ibdev->destroy_srq = qib_destroy_srq;
2209	ibdev->create_qp = qib_create_qp;
2210	ibdev->modify_qp = qib_modify_qp;
2211	ibdev->query_qp = qib_query_qp;
2212	ibdev->destroy_qp = qib_destroy_qp;
2213	ibdev->post_send = qib_post_send;
2214	ibdev->post_recv = qib_post_receive;
2215	ibdev->post_srq_recv = qib_post_srq_receive;
2216	ibdev->create_cq = qib_create_cq;
2217	ibdev->destroy_cq = qib_destroy_cq;
2218	ibdev->resize_cq = qib_resize_cq;
2219	ibdev->poll_cq = qib_poll_cq;
2220	ibdev->req_notify_cq = qib_req_notify_cq;
2221	ibdev->get_dma_mr = qib_get_dma_mr;
2222	ibdev->reg_phys_mr = qib_reg_phys_mr;
2223	ibdev->reg_user_mr = qib_reg_user_mr;
2224	ibdev->dereg_mr = qib_dereg_mr;
2225	ibdev->alloc_fast_reg_mr = qib_alloc_fast_reg_mr;
2226	ibdev->alloc_fast_reg_page_list = qib_alloc_fast_reg_page_list;
2227	ibdev->free_fast_reg_page_list = qib_free_fast_reg_page_list;
2228	ibdev->alloc_fmr = qib_alloc_fmr;
2229	ibdev->map_phys_fmr = qib_map_phys_fmr;
2230	ibdev->unmap_fmr = qib_unmap_fmr;
2231	ibdev->dealloc_fmr = qib_dealloc_fmr;
2232	ibdev->attach_mcast = qib_multicast_attach;
2233	ibdev->detach_mcast = qib_multicast_detach;
2234	ibdev->process_mad = qib_process_mad;
2235	ibdev->mmap = qib_mmap;
2236	ibdev->dma_ops = &qib_dma_mapping_ops;
2237
2238	snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
2239		 "Intel Infiniband HCA %s", init_utsname()->nodename);
2240
2241	ret = ib_register_device(ibdev, qib_create_port_files);
2242	if (ret)
2243		goto err_reg;
2244
2245	ret = qib_create_agents(dev);
2246	if (ret)
2247		goto err_agents;
2248
2249	ret = qib_verbs_register_sysfs(dd);
2250	if (ret)
2251		goto err_class;
2252
2253	goto bail;
2254
2255err_class:
2256	qib_free_agents(dev);
2257err_agents:
2258	ib_unregister_device(ibdev);
2259err_reg:
2260err_tx:
2261	while (!list_empty(&dev->txreq_free)) {
2262		struct list_head *l = dev->txreq_free.next;
2263		struct qib_verbs_txreq *tx;
2264
2265		list_del(l);
2266		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
2267		kfree(tx);
2268	}
2269	if (ppd->sdma_descq_cnt)
2270		dma_free_coherent(&dd->pcidev->dev,
2271				  ppd->sdma_descq_cnt *
2272					sizeof(struct qib_pio_header),
2273				  dev->pio_hdrs, dev->pio_hdrs_phys);
2274err_hdrs:
2275	vfree(dev->lk_table.table);
2276err_lk:
2277	kfree(dev->qp_table);
2278err_qpt:
2279	qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2280bail:
2281	return ret;
2282}
2283
2284void qib_unregister_ib_device(struct qib_devdata *dd)
2285{
2286	struct qib_ibdev *dev = &dd->verbs_dev;
2287	struct ib_device *ibdev = &dev->ibdev;
2288	u32 qps_inuse;
2289	unsigned lk_tab_size;
2290
2291	qib_verbs_unregister_sysfs(dd);
2292
2293	qib_free_agents(dev);
2294
2295	ib_unregister_device(ibdev);
2296
2297	if (!list_empty(&dev->piowait))
2298		qib_dev_err(dd, "piowait list not empty!\n");
2299	if (!list_empty(&dev->dmawait))
2300		qib_dev_err(dd, "dmawait list not empty!\n");
2301	if (!list_empty(&dev->txwait))
2302		qib_dev_err(dd, "txwait list not empty!\n");
2303	if (!list_empty(&dev->memwait))
2304		qib_dev_err(dd, "memwait list not empty!\n");
2305	if (dev->dma_mr)
2306		qib_dev_err(dd, "DMA MR not NULL!\n");
2307
2308	qps_inuse = qib_free_all_qps(dd);
2309	if (qps_inuse)
2310		qib_dev_err(dd, "QP memory leak! %u still in use\n",
2311			    qps_inuse);
2312
2313	del_timer_sync(&dev->mem_timer);
2314	qib_free_qpn_table(&dev->qpn_table);
2315	while (!list_empty(&dev->txreq_free)) {
2316		struct list_head *l = dev->txreq_free.next;
2317		struct qib_verbs_txreq *tx;
2318
2319		list_del(l);
2320		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
2321		kfree(tx);
2322	}
2323	if (dd->pport->sdma_descq_cnt)
2324		dma_free_coherent(&dd->pcidev->dev,
2325				  dd->pport->sdma_descq_cnt *
2326					sizeof(struct qib_pio_header),
2327				  dev->pio_hdrs, dev->pio_hdrs_phys);
2328	lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
2329	vfree(dev->lk_table.table);
2330	kfree(dev->qp_table);
2331}
2332
2333/*
2334 * This must be called with s_lock held.
2335 */
2336void qib_schedule_send(struct qib_qp *qp)
2337{
2338	if (qib_send_ok(qp)) {
2339		struct qib_ibport *ibp =
2340			to_iport(qp->ibqp.device, qp->port_num);
2341		struct qib_pportdata *ppd = ppd_from_ibp(ibp);
2342
2343		queue_work(ppd->qib_wq, &qp->s_work);
2344	}
2345}
2346