[go: nahoru, domu]

1/*
2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses.  You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 *     Redistribution and use in source and binary forms, with or
12 *     without modification, are permitted provided that the following
13 *     conditions are met:
14 *
15 *      - Redistributions of source code must retain the above
16 *        copyright notice, this list of conditions and the following
17 *        disclaimer.
18 *
19 *      - Redistributions in binary form must reproduce the above
20 *        copyright notice, this list of conditions and the following
21 *        disclaimer in the documentation and/or other materials
22 *        provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34#include <rdma/ib_mad.h>
35#include <rdma/ib_user_verbs.h>
36#include <linux/io.h>
37#include <linux/slab.h>
38#include <linux/module.h>
39#include <linux/utsname.h>
40#include <linux/rculist.h>
41
42#include "ipath_kernel.h"
43#include "ipath_verbs.h"
44#include "ipath_common.h"
45
46static unsigned int ib_ipath_qp_table_size = 251;
47module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
48MODULE_PARM_DESC(qp_table_size, "QP table size");
49
50unsigned int ib_ipath_lkey_table_size = 12;
51module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
52		   S_IRUGO);
53MODULE_PARM_DESC(lkey_table_size,
54		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
55
56static unsigned int ib_ipath_max_pds = 0xFFFF;
57module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
58MODULE_PARM_DESC(max_pds,
59		 "Maximum number of protection domains to support");
60
61static unsigned int ib_ipath_max_ahs = 0xFFFF;
62module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
63MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
64
65unsigned int ib_ipath_max_cqes = 0x2FFFF;
66module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
67MODULE_PARM_DESC(max_cqes,
68		 "Maximum number of completion queue entries to support");
69
70unsigned int ib_ipath_max_cqs = 0x1FFFF;
71module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
72MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
73
74unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
75module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
76		   S_IWUSR | S_IRUGO);
77MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
78
79unsigned int ib_ipath_max_qps = 16384;
80module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
81MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
82
83unsigned int ib_ipath_max_sges = 0x60;
84module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
85MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
86
87unsigned int ib_ipath_max_mcast_grps = 16384;
88module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
89		   S_IWUSR | S_IRUGO);
90MODULE_PARM_DESC(max_mcast_grps,
91		 "Maximum number of multicast groups to support");
92
93unsigned int ib_ipath_max_mcast_qp_attached = 16;
94module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
95		   uint, S_IWUSR | S_IRUGO);
96MODULE_PARM_DESC(max_mcast_qp_attached,
97		 "Maximum number of attached QPs to support");
98
99unsigned int ib_ipath_max_srqs = 1024;
100module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
101MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
102
103unsigned int ib_ipath_max_srq_sges = 128;
104module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
105		   uint, S_IWUSR | S_IRUGO);
106MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
107
108unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
109module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
110		   uint, S_IWUSR | S_IRUGO);
111MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
112
113static unsigned int ib_ipath_disable_sma;
114module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
115MODULE_PARM_DESC(disable_sma, "Disable the SMA");
116
117/*
118 * Note that it is OK to post send work requests in the SQE and ERR
119 * states; ipath_do_send() will process them and generate error
120 * completions as per IB 1.2 C10-96.
121 */
122const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
123	[IB_QPS_RESET] = 0,
124	[IB_QPS_INIT] = IPATH_POST_RECV_OK,
125	[IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
126	[IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
127	    IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK |
128	    IPATH_PROCESS_NEXT_SEND_OK,
129	[IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
130	    IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
131	[IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
132	    IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
133	[IB_QPS_ERR] = IPATH_POST_RECV_OK | IPATH_FLUSH_RECV |
134	    IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
135};
136
137struct ipath_ucontext {
138	struct ib_ucontext ibucontext;
139};
140
141static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
142						  *ibucontext)
143{
144	return container_of(ibucontext, struct ipath_ucontext, ibucontext);
145}
146
147/*
148 * Translate ib_wr_opcode into ib_wc_opcode.
149 */
150const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
151	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
152	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
153	[IB_WR_SEND] = IB_WC_SEND,
154	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
155	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
156	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
157	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
158};
159
160/*
161 * System image GUID.
162 */
163static __be64 sys_image_guid;
164
165/**
166 * ipath_copy_sge - copy data to SGE memory
167 * @ss: the SGE state
168 * @data: the data to copy
169 * @length: the length of the data
170 */
171void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
172{
173	struct ipath_sge *sge = &ss->sge;
174
175	while (length) {
176		u32 len = sge->length;
177
178		if (len > length)
179			len = length;
180		if (len > sge->sge_length)
181			len = sge->sge_length;
182		BUG_ON(len == 0);
183		memcpy(sge->vaddr, data, len);
184		sge->vaddr += len;
185		sge->length -= len;
186		sge->sge_length -= len;
187		if (sge->sge_length == 0) {
188			if (--ss->num_sge)
189				*sge = *ss->sg_list++;
190		} else if (sge->length == 0 && sge->mr != NULL) {
191			if (++sge->n >= IPATH_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 * ipath_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 ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
212{
213	struct ipath_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 (--ss->num_sge)
228				*sge = *ss->sg_list++;
229		} else if (sge->length == 0 && sge->mr != NULL) {
230			if (++sge->n >= IPATH_SEGSZ) {
231				if (++sge->m >= sge->mr->mapsz)
232					break;
233				sge->n = 0;
234			}
235			sge->vaddr =
236				sge->mr->map[sge->m]->segs[sge->n].vaddr;
237			sge->length =
238				sge->mr->map[sge->m]->segs[sge->n].length;
239		}
240		length -= len;
241	}
242}
243
244/*
245 * Count the number of DMA descriptors needed to send length bytes of data.
246 * Don't modify the ipath_sge_state to get the count.
247 * Return zero if any of the segments is not aligned.
248 */
249static u32 ipath_count_sge(struct ipath_sge_state *ss, u32 length)
250{
251	struct ipath_sge *sg_list = ss->sg_list;
252	struct ipath_sge sge = ss->sge;
253	u8 num_sge = ss->num_sge;
254	u32 ndesc = 1;	/* count the header */
255
256	while (length) {
257		u32 len = sge.length;
258
259		if (len > length)
260			len = length;
261		if (len > sge.sge_length)
262			len = sge.sge_length;
263		BUG_ON(len == 0);
264		if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
265		    (len != length && (len & (sizeof(u32) - 1)))) {
266			ndesc = 0;
267			break;
268		}
269		ndesc++;
270		sge.vaddr += len;
271		sge.length -= len;
272		sge.sge_length -= len;
273		if (sge.sge_length == 0) {
274			if (--num_sge)
275				sge = *sg_list++;
276		} else if (sge.length == 0 && sge.mr != NULL) {
277			if (++sge.n >= IPATH_SEGSZ) {
278				if (++sge.m >= sge.mr->mapsz)
279					break;
280				sge.n = 0;
281			}
282			sge.vaddr =
283				sge.mr->map[sge.m]->segs[sge.n].vaddr;
284			sge.length =
285				sge.mr->map[sge.m]->segs[sge.n].length;
286		}
287		length -= len;
288	}
289	return ndesc;
290}
291
292/*
293 * Copy from the SGEs to the data buffer.
294 */
295static void ipath_copy_from_sge(void *data, struct ipath_sge_state *ss,
296				u32 length)
297{
298	struct ipath_sge *sge = &ss->sge;
299
300	while (length) {
301		u32 len = sge->length;
302
303		if (len > length)
304			len = length;
305		if (len > sge->sge_length)
306			len = sge->sge_length;
307		BUG_ON(len == 0);
308		memcpy(data, sge->vaddr, len);
309		sge->vaddr += len;
310		sge->length -= len;
311		sge->sge_length -= len;
312		if (sge->sge_length == 0) {
313			if (--ss->num_sge)
314				*sge = *ss->sg_list++;
315		} else if (sge->length == 0 && sge->mr != NULL) {
316			if (++sge->n >= IPATH_SEGSZ) {
317				if (++sge->m >= sge->mr->mapsz)
318					break;
319				sge->n = 0;
320			}
321			sge->vaddr =
322				sge->mr->map[sge->m]->segs[sge->n].vaddr;
323			sge->length =
324				sge->mr->map[sge->m]->segs[sge->n].length;
325		}
326		data += len;
327		length -= len;
328	}
329}
330
331/**
332 * ipath_post_one_send - post one RC, UC, or UD send work request
333 * @qp: the QP to post on
334 * @wr: the work request to send
335 */
336static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
337{
338	struct ipath_swqe *wqe;
339	u32 next;
340	int i;
341	int j;
342	int acc;
343	int ret;
344	unsigned long flags;
345	struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
346
347	spin_lock_irqsave(&qp->s_lock, flags);
348
349	if (qp->ibqp.qp_type != IB_QPT_SMI &&
350	    !(dd->ipath_flags & IPATH_LINKACTIVE)) {
351		ret = -ENETDOWN;
352		goto bail;
353	}
354
355	/* Check that state is OK to post send. */
356	if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)))
357		goto bail_inval;
358
359	/* IB spec says that num_sge == 0 is OK. */
360	if (wr->num_sge > qp->s_max_sge)
361		goto bail_inval;
362
363	/*
364	 * Don't allow RDMA reads or atomic operations on UC or
365	 * undefined operations.
366	 * Make sure buffer is large enough to hold the result for atomics.
367	 */
368	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_UD) {
372		/* Check 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	wqe = get_swqe_ptr(qp, qp->s_head);
398	wqe->wr = *wr;
399	wqe->length = 0;
400	if (wr->num_sge) {
401		acc = wr->opcode >= IB_WR_RDMA_READ ?
402			IB_ACCESS_LOCAL_WRITE : 0;
403		for (i = 0, j = 0; i < wr->num_sge; i++) {
404			u32 length = wr->sg_list[i].length;
405			int ok;
406
407			if (length == 0)
408				continue;
409			ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
410					   &wr->sg_list[i], acc);
411			if (!ok)
412				goto bail_inval;
413			wqe->length += length;
414			j++;
415		}
416		wqe->wr.num_sge = j;
417	}
418	if (qp->ibqp.qp_type == IB_QPT_UC ||
419	    qp->ibqp.qp_type == IB_QPT_RC) {
420		if (wqe->length > 0x80000000U)
421			goto bail_inval;
422	} else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
423		goto bail_inval;
424	wqe->ssn = qp->s_ssn++;
425	qp->s_head = next;
426
427	ret = 0;
428	goto bail;
429
430bail_inval:
431	ret = -EINVAL;
432bail:
433	spin_unlock_irqrestore(&qp->s_lock, flags);
434	return ret;
435}
436
437/**
438 * ipath_post_send - post a send on a QP
439 * @ibqp: the QP to post the send on
440 * @wr: the list of work requests to post
441 * @bad_wr: the first bad WR is put here
442 *
443 * This may be called from interrupt context.
444 */
445static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
446			   struct ib_send_wr **bad_wr)
447{
448	struct ipath_qp *qp = to_iqp(ibqp);
449	int err = 0;
450
451	for (; wr; wr = wr->next) {
452		err = ipath_post_one_send(qp, wr);
453		if (err) {
454			*bad_wr = wr;
455			goto bail;
456		}
457	}
458
459	/* Try to do the send work in the caller's context. */
460	ipath_do_send((unsigned long) qp);
461
462bail:
463	return err;
464}
465
466/**
467 * ipath_post_receive - post a receive on a QP
468 * @ibqp: the QP to post the receive on
469 * @wr: the WR to post
470 * @bad_wr: the first bad WR is put here
471 *
472 * This may be called from interrupt context.
473 */
474static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
475			      struct ib_recv_wr **bad_wr)
476{
477	struct ipath_qp *qp = to_iqp(ibqp);
478	struct ipath_rwq *wq = qp->r_rq.wq;
479	unsigned long flags;
480	int ret;
481
482	/* Check that state is OK to post receive. */
483	if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
484		*bad_wr = wr;
485		ret = -EINVAL;
486		goto bail;
487	}
488
489	for (; wr; wr = wr->next) {
490		struct ipath_rwqe *wqe;
491		u32 next;
492		int i;
493
494		if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
495			*bad_wr = wr;
496			ret = -EINVAL;
497			goto bail;
498		}
499
500		spin_lock_irqsave(&qp->r_rq.lock, flags);
501		next = wq->head + 1;
502		if (next >= qp->r_rq.size)
503			next = 0;
504		if (next == wq->tail) {
505			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
506			*bad_wr = wr;
507			ret = -ENOMEM;
508			goto bail;
509		}
510
511		wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
512		wqe->wr_id = wr->wr_id;
513		wqe->num_sge = wr->num_sge;
514		for (i = 0; i < wr->num_sge; i++)
515			wqe->sg_list[i] = wr->sg_list[i];
516		/* Make sure queue entry is written before the head index. */
517		smp_wmb();
518		wq->head = next;
519		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
520	}
521	ret = 0;
522
523bail:
524	return ret;
525}
526
527/**
528 * ipath_qp_rcv - processing an incoming packet on a QP
529 * @dev: the device the packet came on
530 * @hdr: the packet header
531 * @has_grh: true if the packet has a GRH
532 * @data: the packet data
533 * @tlen: the packet length
534 * @qp: the QP the packet came on
535 *
536 * This is called from ipath_ib_rcv() to process an incoming packet
537 * for the given QP.
538 * Called at interrupt level.
539 */
540static void ipath_qp_rcv(struct ipath_ibdev *dev,
541			 struct ipath_ib_header *hdr, int has_grh,
542			 void *data, u32 tlen, struct ipath_qp *qp)
543{
544	/* Check for valid receive state. */
545	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
546		dev->n_pkt_drops++;
547		return;
548	}
549
550	switch (qp->ibqp.qp_type) {
551	case IB_QPT_SMI:
552	case IB_QPT_GSI:
553		if (ib_ipath_disable_sma)
554			break;
555		/* FALLTHROUGH */
556	case IB_QPT_UD:
557		ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
558		break;
559
560	case IB_QPT_RC:
561		ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
562		break;
563
564	case IB_QPT_UC:
565		ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
566		break;
567
568	default:
569		break;
570	}
571}
572
573/**
574 * ipath_ib_rcv - process an incoming packet
575 * @arg: the device pointer
576 * @rhdr: the header of the packet
577 * @data: the packet data
578 * @tlen: the packet length
579 *
580 * This is called from ipath_kreceive() to process an incoming packet at
581 * interrupt level. Tlen is the length of the header + data + CRC in bytes.
582 */
583void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
584		  u32 tlen)
585{
586	struct ipath_ib_header *hdr = rhdr;
587	struct ipath_other_headers *ohdr;
588	struct ipath_qp *qp;
589	u32 qp_num;
590	int lnh;
591	u8 opcode;
592	u16 lid;
593
594	if (unlikely(dev == NULL))
595		goto bail;
596
597	if (unlikely(tlen < 24)) {	/* LRH+BTH+CRC */
598		dev->rcv_errors++;
599		goto bail;
600	}
601
602	/* Check for a valid destination LID (see ch. 7.11.1). */
603	lid = be16_to_cpu(hdr->lrh[1]);
604	if (lid < IPATH_MULTICAST_LID_BASE) {
605		lid &= ~((1 << dev->dd->ipath_lmc) - 1);
606		if (unlikely(lid != dev->dd->ipath_lid)) {
607			dev->rcv_errors++;
608			goto bail;
609		}
610	}
611
612	/* Check for GRH */
613	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
614	if (lnh == IPATH_LRH_BTH)
615		ohdr = &hdr->u.oth;
616	else if (lnh == IPATH_LRH_GRH)
617		ohdr = &hdr->u.l.oth;
618	else {
619		dev->rcv_errors++;
620		goto bail;
621	}
622
623	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
624	dev->opstats[opcode].n_bytes += tlen;
625	dev->opstats[opcode].n_packets++;
626
627	/* Get the destination QP number. */
628	qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
629	if (qp_num == IPATH_MULTICAST_QPN) {
630		struct ipath_mcast *mcast;
631		struct ipath_mcast_qp *p;
632
633		if (lnh != IPATH_LRH_GRH) {
634			dev->n_pkt_drops++;
635			goto bail;
636		}
637		mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
638		if (mcast == NULL) {
639			dev->n_pkt_drops++;
640			goto bail;
641		}
642		dev->n_multicast_rcv++;
643		list_for_each_entry_rcu(p, &mcast->qp_list, list)
644			ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
645		/*
646		 * Notify ipath_multicast_detach() if it is waiting for us
647		 * to finish.
648		 */
649		if (atomic_dec_return(&mcast->refcount) <= 1)
650			wake_up(&mcast->wait);
651	} else {
652		qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
653		if (qp) {
654			dev->n_unicast_rcv++;
655			ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
656				     tlen, qp);
657			/*
658			 * Notify ipath_destroy_qp() if it is waiting
659			 * for us to finish.
660			 */
661			if (atomic_dec_and_test(&qp->refcount))
662				wake_up(&qp->wait);
663		} else
664			dev->n_pkt_drops++;
665	}
666
667bail:;
668}
669
670/**
671 * ipath_ib_timer - verbs timer
672 * @arg: the device pointer
673 *
674 * This is called from ipath_do_rcv_timer() at interrupt level to check for
675 * QPs which need retransmits and to collect performance numbers.
676 */
677static void ipath_ib_timer(struct ipath_ibdev *dev)
678{
679	struct ipath_qp *resend = NULL;
680	struct ipath_qp *rnr = NULL;
681	struct list_head *last;
682	struct ipath_qp *qp;
683	unsigned long flags;
684
685	if (dev == NULL)
686		return;
687
688	spin_lock_irqsave(&dev->pending_lock, flags);
689	/* Start filling the next pending queue. */
690	if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
691		dev->pending_index = 0;
692	/* Save any requests still in the new queue, they have timed out. */
693	last = &dev->pending[dev->pending_index];
694	while (!list_empty(last)) {
695		qp = list_entry(last->next, struct ipath_qp, timerwait);
696		list_del_init(&qp->timerwait);
697		qp->timer_next = resend;
698		resend = qp;
699		atomic_inc(&qp->refcount);
700	}
701	last = &dev->rnrwait;
702	if (!list_empty(last)) {
703		qp = list_entry(last->next, struct ipath_qp, timerwait);
704		if (--qp->s_rnr_timeout == 0) {
705			do {
706				list_del_init(&qp->timerwait);
707				qp->timer_next = rnr;
708				rnr = qp;
709				atomic_inc(&qp->refcount);
710				if (list_empty(last))
711					break;
712				qp = list_entry(last->next, struct ipath_qp,
713						timerwait);
714			} while (qp->s_rnr_timeout == 0);
715		}
716	}
717	/*
718	 * We should only be in the started state if pma_sample_start != 0
719	 */
720	if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
721	    --dev->pma_sample_start == 0) {
722		dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
723		ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
724					&dev->ipath_rword,
725					&dev->ipath_spkts,
726					&dev->ipath_rpkts,
727					&dev->ipath_xmit_wait);
728	}
729	if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
730		if (dev->pma_sample_interval == 0) {
731			u64 ta, tb, tc, td, te;
732
733			dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
734			ipath_snapshot_counters(dev->dd, &ta, &tb,
735						&tc, &td, &te);
736
737			dev->ipath_sword = ta - dev->ipath_sword;
738			dev->ipath_rword = tb - dev->ipath_rword;
739			dev->ipath_spkts = tc - dev->ipath_spkts;
740			dev->ipath_rpkts = td - dev->ipath_rpkts;
741			dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
742		}
743		else
744			dev->pma_sample_interval--;
745	}
746	spin_unlock_irqrestore(&dev->pending_lock, flags);
747
748	/* XXX What if timer fires again while this is running? */
749	while (resend != NULL) {
750		qp = resend;
751		resend = qp->timer_next;
752
753		spin_lock_irqsave(&qp->s_lock, flags);
754		if (qp->s_last != qp->s_tail &&
755		    ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
756			dev->n_timeouts++;
757			ipath_restart_rc(qp, qp->s_last_psn + 1);
758		}
759		spin_unlock_irqrestore(&qp->s_lock, flags);
760
761		/* Notify ipath_destroy_qp() if it is waiting. */
762		if (atomic_dec_and_test(&qp->refcount))
763			wake_up(&qp->wait);
764	}
765	while (rnr != NULL) {
766		qp = rnr;
767		rnr = qp->timer_next;
768
769		spin_lock_irqsave(&qp->s_lock, flags);
770		if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
771			ipath_schedule_send(qp);
772		spin_unlock_irqrestore(&qp->s_lock, flags);
773
774		/* Notify ipath_destroy_qp() if it is waiting. */
775		if (atomic_dec_and_test(&qp->refcount))
776			wake_up(&qp->wait);
777	}
778}
779
780static void update_sge(struct ipath_sge_state *ss, u32 length)
781{
782	struct ipath_sge *sge = &ss->sge;
783
784	sge->vaddr += length;
785	sge->length -= length;
786	sge->sge_length -= length;
787	if (sge->sge_length == 0) {
788		if (--ss->num_sge)
789			*sge = *ss->sg_list++;
790	} else if (sge->length == 0 && sge->mr != NULL) {
791		if (++sge->n >= IPATH_SEGSZ) {
792			if (++sge->m >= sge->mr->mapsz)
793				return;
794			sge->n = 0;
795		}
796		sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
797		sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
798	}
799}
800
801#ifdef __LITTLE_ENDIAN
802static inline u32 get_upper_bits(u32 data, u32 shift)
803{
804	return data >> shift;
805}
806
807static inline u32 set_upper_bits(u32 data, u32 shift)
808{
809	return data << shift;
810}
811
812static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
813{
814	data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
815	data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
816	return data;
817}
818#else
819static inline u32 get_upper_bits(u32 data, u32 shift)
820{
821	return data << shift;
822}
823
824static inline u32 set_upper_bits(u32 data, u32 shift)
825{
826	return data >> shift;
827}
828
829static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
830{
831	data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
832	data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
833	return data;
834}
835#endif
836
837static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
838		    u32 length, unsigned flush_wc)
839{
840	u32 extra = 0;
841	u32 data = 0;
842	u32 last;
843
844	while (1) {
845		u32 len = ss->sge.length;
846		u32 off;
847
848		if (len > length)
849			len = length;
850		if (len > ss->sge.sge_length)
851			len = ss->sge.sge_length;
852		BUG_ON(len == 0);
853		/* If the source address is not aligned, try to align it. */
854		off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
855		if (off) {
856			u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
857					    ~(sizeof(u32) - 1));
858			u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
859			u32 y;
860
861			y = sizeof(u32) - off;
862			if (len > y)
863				len = y;
864			if (len + extra >= sizeof(u32)) {
865				data |= set_upper_bits(v, extra *
866						       BITS_PER_BYTE);
867				len = sizeof(u32) - extra;
868				if (len == length) {
869					last = data;
870					break;
871				}
872				__raw_writel(data, piobuf);
873				piobuf++;
874				extra = 0;
875				data = 0;
876			} else {
877				/* Clear unused upper bytes */
878				data |= clear_upper_bytes(v, len, extra);
879				if (len == length) {
880					last = data;
881					break;
882				}
883				extra += len;
884			}
885		} else if (extra) {
886			/* Source address is aligned. */
887			u32 *addr = (u32 *) ss->sge.vaddr;
888			int shift = extra * BITS_PER_BYTE;
889			int ushift = 32 - shift;
890			u32 l = len;
891
892			while (l >= sizeof(u32)) {
893				u32 v = *addr;
894
895				data |= set_upper_bits(v, shift);
896				__raw_writel(data, piobuf);
897				data = get_upper_bits(v, ushift);
898				piobuf++;
899				addr++;
900				l -= sizeof(u32);
901			}
902			/*
903			 * We still have 'extra' number of bytes leftover.
904			 */
905			if (l) {
906				u32 v = *addr;
907
908				if (l + extra >= sizeof(u32)) {
909					data |= set_upper_bits(v, shift);
910					len -= l + extra - sizeof(u32);
911					if (len == length) {
912						last = data;
913						break;
914					}
915					__raw_writel(data, piobuf);
916					piobuf++;
917					extra = 0;
918					data = 0;
919				} else {
920					/* Clear unused upper bytes */
921					data |= clear_upper_bytes(v, l,
922								  extra);
923					if (len == length) {
924						last = data;
925						break;
926					}
927					extra += l;
928				}
929			} else if (len == length) {
930				last = data;
931				break;
932			}
933		} else if (len == length) {
934			u32 w;
935
936			/*
937			 * Need to round up for the last dword in the
938			 * packet.
939			 */
940			w = (len + 3) >> 2;
941			__iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
942			piobuf += w - 1;
943			last = ((u32 *) ss->sge.vaddr)[w - 1];
944			break;
945		} else {
946			u32 w = len >> 2;
947
948			__iowrite32_copy(piobuf, ss->sge.vaddr, w);
949			piobuf += w;
950
951			extra = len & (sizeof(u32) - 1);
952			if (extra) {
953				u32 v = ((u32 *) ss->sge.vaddr)[w];
954
955				/* Clear unused upper bytes */
956				data = clear_upper_bytes(v, extra, 0);
957			}
958		}
959		update_sge(ss, len);
960		length -= len;
961	}
962	/* Update address before sending packet. */
963	update_sge(ss, length);
964	if (flush_wc) {
965		/* must flush early everything before trigger word */
966		ipath_flush_wc();
967		__raw_writel(last, piobuf);
968		/* be sure trigger word is written */
969		ipath_flush_wc();
970	} else
971		__raw_writel(last, piobuf);
972}
973
974/*
975 * Convert IB rate to delay multiplier.
976 */
977unsigned ipath_ib_rate_to_mult(enum ib_rate rate)
978{
979	switch (rate) {
980	case IB_RATE_2_5_GBPS: return 8;
981	case IB_RATE_5_GBPS:   return 4;
982	case IB_RATE_10_GBPS:  return 2;
983	case IB_RATE_20_GBPS:  return 1;
984	default:	       return 0;
985	}
986}
987
988/*
989 * Convert delay multiplier to IB rate
990 */
991static enum ib_rate ipath_mult_to_ib_rate(unsigned mult)
992{
993	switch (mult) {
994	case 8:  return IB_RATE_2_5_GBPS;
995	case 4:  return IB_RATE_5_GBPS;
996	case 2:  return IB_RATE_10_GBPS;
997	case 1:  return IB_RATE_20_GBPS;
998	default: return IB_RATE_PORT_CURRENT;
999	}
1000}
1001
1002static inline struct ipath_verbs_txreq *get_txreq(struct ipath_ibdev *dev)
1003{
1004	struct ipath_verbs_txreq *tx = NULL;
1005	unsigned long flags;
1006
1007	spin_lock_irqsave(&dev->pending_lock, flags);
1008	if (!list_empty(&dev->txreq_free)) {
1009		struct list_head *l = dev->txreq_free.next;
1010
1011		list_del(l);
1012		tx = list_entry(l, struct ipath_verbs_txreq, txreq.list);
1013	}
1014	spin_unlock_irqrestore(&dev->pending_lock, flags);
1015	return tx;
1016}
1017
1018static inline void put_txreq(struct ipath_ibdev *dev,
1019			     struct ipath_verbs_txreq *tx)
1020{
1021	unsigned long flags;
1022
1023	spin_lock_irqsave(&dev->pending_lock, flags);
1024	list_add(&tx->txreq.list, &dev->txreq_free);
1025	spin_unlock_irqrestore(&dev->pending_lock, flags);
1026}
1027
1028static void sdma_complete(void *cookie, int status)
1029{
1030	struct ipath_verbs_txreq *tx = cookie;
1031	struct ipath_qp *qp = tx->qp;
1032	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1033	unsigned long flags;
1034	enum ib_wc_status ibs = status == IPATH_SDMA_TXREQ_S_OK ?
1035		IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR;
1036
1037	if (atomic_dec_and_test(&qp->s_dma_busy)) {
1038		spin_lock_irqsave(&qp->s_lock, flags);
1039		if (tx->wqe)
1040			ipath_send_complete(qp, tx->wqe, ibs);
1041		if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1042		     qp->s_last != qp->s_head) ||
1043		    (qp->s_flags & IPATH_S_WAIT_DMA))
1044			ipath_schedule_send(qp);
1045		spin_unlock_irqrestore(&qp->s_lock, flags);
1046		wake_up(&qp->wait_dma);
1047	} else if (tx->wqe) {
1048		spin_lock_irqsave(&qp->s_lock, flags);
1049		ipath_send_complete(qp, tx->wqe, ibs);
1050		spin_unlock_irqrestore(&qp->s_lock, flags);
1051	}
1052
1053	if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
1054		kfree(tx->txreq.map_addr);
1055	put_txreq(dev, tx);
1056
1057	if (atomic_dec_and_test(&qp->refcount))
1058		wake_up(&qp->wait);
1059}
1060
1061static void decrement_dma_busy(struct ipath_qp *qp)
1062{
1063	unsigned long flags;
1064
1065	if (atomic_dec_and_test(&qp->s_dma_busy)) {
1066		spin_lock_irqsave(&qp->s_lock, flags);
1067		if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1068		     qp->s_last != qp->s_head) ||
1069		    (qp->s_flags & IPATH_S_WAIT_DMA))
1070			ipath_schedule_send(qp);
1071		spin_unlock_irqrestore(&qp->s_lock, flags);
1072		wake_up(&qp->wait_dma);
1073	}
1074}
1075
1076/*
1077 * Compute the number of clock cycles of delay before sending the next packet.
1078 * The multipliers reflect the number of clocks for the fastest rate so
1079 * one tick at 4xDDR is 8 ticks at 1xSDR.
1080 * If the destination port will take longer to receive a packet than
1081 * the outgoing link can send it, we need to delay sending the next packet
1082 * by the difference in time it takes the receiver to receive and the sender
1083 * to send this packet.
1084 * Note that this delay is always correct for UC and RC but not always
1085 * optimal for UD. For UD, the destination HCA can be different for each
1086 * packet, in which case, we could send packets to a different destination
1087 * while "waiting" for the delay. The overhead for doing this without
1088 * HW support is more than just paying the cost of delaying some packets
1089 * unnecessarily.
1090 */
1091static inline unsigned ipath_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult)
1092{
1093	return (rcv_mult > snd_mult) ?
1094		(plen * (rcv_mult - snd_mult) + 1) >> 1 : 0;
1095}
1096
1097static int ipath_verbs_send_dma(struct ipath_qp *qp,
1098				struct ipath_ib_header *hdr, u32 hdrwords,
1099				struct ipath_sge_state *ss, u32 len,
1100				u32 plen, u32 dwords)
1101{
1102	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1103	struct ipath_devdata *dd = dev->dd;
1104	struct ipath_verbs_txreq *tx;
1105	u32 *piobuf;
1106	u32 control;
1107	u32 ndesc;
1108	int ret;
1109
1110	tx = qp->s_tx;
1111	if (tx) {
1112		qp->s_tx = NULL;
1113		/* resend previously constructed packet */
1114		atomic_inc(&qp->s_dma_busy);
1115		ret = ipath_sdma_verbs_send(dd, tx->ss, tx->len, tx);
1116		if (ret) {
1117			qp->s_tx = tx;
1118			decrement_dma_busy(qp);
1119		}
1120		goto bail;
1121	}
1122
1123	tx = get_txreq(dev);
1124	if (!tx) {
1125		ret = -EBUSY;
1126		goto bail;
1127	}
1128
1129	/*
1130	 * Get the saved delay count we computed for the previous packet
1131	 * and save the delay count for this packet to be used next time
1132	 * we get here.
1133	 */
1134	control = qp->s_pkt_delay;
1135	qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1136
1137	tx->qp = qp;
1138	atomic_inc(&qp->refcount);
1139	tx->wqe = qp->s_wqe;
1140	tx->txreq.callback = sdma_complete;
1141	tx->txreq.callback_cookie = tx;
1142	tx->txreq.flags = IPATH_SDMA_TXREQ_F_HEADTOHOST |
1143		IPATH_SDMA_TXREQ_F_INTREQ | IPATH_SDMA_TXREQ_F_FREEDESC;
1144	if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1145		tx->txreq.flags |= IPATH_SDMA_TXREQ_F_USELARGEBUF;
1146
1147	/* VL15 packets bypass credit check */
1148	if ((be16_to_cpu(hdr->lrh[0]) >> 12) == 15) {
1149		control |= 1ULL << 31;
1150		tx->txreq.flags |= IPATH_SDMA_TXREQ_F_VL15;
1151	}
1152
1153	if (len) {
1154		/*
1155		 * Don't try to DMA if it takes more descriptors than
1156		 * the queue holds.
1157		 */
1158		ndesc = ipath_count_sge(ss, len);
1159		if (ndesc >= dd->ipath_sdma_descq_cnt)
1160			ndesc = 0;
1161	} else
1162		ndesc = 1;
1163	if (ndesc) {
1164		tx->hdr.pbc[0] = cpu_to_le32(plen);
1165		tx->hdr.pbc[1] = cpu_to_le32(control);
1166		memcpy(&tx->hdr.hdr, hdr, hdrwords << 2);
1167		tx->txreq.sg_count = ndesc;
1168		tx->map_len = (hdrwords + 2) << 2;
1169		tx->txreq.map_addr = &tx->hdr;
1170		atomic_inc(&qp->s_dma_busy);
1171		ret = ipath_sdma_verbs_send(dd, ss, dwords, tx);
1172		if (ret) {
1173			/* save ss and length in dwords */
1174			tx->ss = ss;
1175			tx->len = dwords;
1176			qp->s_tx = tx;
1177			decrement_dma_busy(qp);
1178		}
1179		goto bail;
1180	}
1181
1182	/* Allocate a buffer and copy the header and payload to it. */
1183	tx->map_len = (plen + 1) << 2;
1184	piobuf = kmalloc(tx->map_len, GFP_ATOMIC);
1185	if (unlikely(piobuf == NULL)) {
1186		ret = -EBUSY;
1187		goto err_tx;
1188	}
1189	tx->txreq.map_addr = piobuf;
1190	tx->txreq.flags |= IPATH_SDMA_TXREQ_F_FREEBUF;
1191	tx->txreq.sg_count = 1;
1192
1193	*piobuf++ = (__force u32) cpu_to_le32(plen);
1194	*piobuf++ = (__force u32) cpu_to_le32(control);
1195	memcpy(piobuf, hdr, hdrwords << 2);
1196	ipath_copy_from_sge(piobuf + hdrwords, ss, len);
1197
1198	atomic_inc(&qp->s_dma_busy);
1199	ret = ipath_sdma_verbs_send(dd, NULL, 0, tx);
1200	/*
1201	 * If we couldn't queue the DMA request, save the info
1202	 * and try again later rather than destroying the
1203	 * buffer and undoing the side effects of the copy.
1204	 */
1205	if (ret) {
1206		tx->ss = NULL;
1207		tx->len = 0;
1208		qp->s_tx = tx;
1209		decrement_dma_busy(qp);
1210	}
1211	dev->n_unaligned++;
1212	goto bail;
1213
1214err_tx:
1215	if (atomic_dec_and_test(&qp->refcount))
1216		wake_up(&qp->wait);
1217	put_txreq(dev, tx);
1218bail:
1219	return ret;
1220}
1221
1222static int ipath_verbs_send_pio(struct ipath_qp *qp,
1223				struct ipath_ib_header *ibhdr, u32 hdrwords,
1224				struct ipath_sge_state *ss, u32 len,
1225				u32 plen, u32 dwords)
1226{
1227	struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1228	u32 *hdr = (u32 *) ibhdr;
1229	u32 __iomem *piobuf;
1230	unsigned flush_wc;
1231	u32 control;
1232	int ret;
1233	unsigned long flags;
1234
1235	piobuf = ipath_getpiobuf(dd, plen, NULL);
1236	if (unlikely(piobuf == NULL)) {
1237		ret = -EBUSY;
1238		goto bail;
1239	}
1240
1241	/*
1242	 * Get the saved delay count we computed for the previous packet
1243	 * and save the delay count for this packet to be used next time
1244	 * we get here.
1245	 */
1246	control = qp->s_pkt_delay;
1247	qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1248
1249	/* VL15 packets bypass credit check */
1250	if ((be16_to_cpu(ibhdr->lrh[0]) >> 12) == 15)
1251		control |= 1ULL << 31;
1252
1253	/*
1254	 * Write the length to the control qword plus any needed flags.
1255	 * We have to flush after the PBC for correctness on some cpus
1256	 * or WC buffer can be written out of order.
1257	 */
1258	writeq(((u64) control << 32) | plen, piobuf);
1259	piobuf += 2;
1260
1261	flush_wc = dd->ipath_flags & IPATH_PIO_FLUSH_WC;
1262	if (len == 0) {
1263		/*
1264		 * If there is just the header portion, must flush before
1265		 * writing last word of header for correctness, and after
1266		 * the last header word (trigger word).
1267		 */
1268		if (flush_wc) {
1269			ipath_flush_wc();
1270			__iowrite32_copy(piobuf, hdr, hdrwords - 1);
1271			ipath_flush_wc();
1272			__raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1273			ipath_flush_wc();
1274		} else
1275			__iowrite32_copy(piobuf, hdr, hdrwords);
1276		goto done;
1277	}
1278
1279	if (flush_wc)
1280		ipath_flush_wc();
1281	__iowrite32_copy(piobuf, hdr, hdrwords);
1282	piobuf += hdrwords;
1283
1284	/* The common case is aligned and contained in one segment. */
1285	if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1286		   !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1287		u32 *addr = (u32 *) ss->sge.vaddr;
1288
1289		/* Update address before sending packet. */
1290		update_sge(ss, len);
1291		if (flush_wc) {
1292			__iowrite32_copy(piobuf, addr, dwords - 1);
1293			/* must flush early everything before trigger word */
1294			ipath_flush_wc();
1295			__raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1296			/* be sure trigger word is written */
1297			ipath_flush_wc();
1298		} else
1299			__iowrite32_copy(piobuf, addr, dwords);
1300		goto done;
1301	}
1302	copy_io(piobuf, ss, len, flush_wc);
1303done:
1304	if (qp->s_wqe) {
1305		spin_lock_irqsave(&qp->s_lock, flags);
1306		ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1307		spin_unlock_irqrestore(&qp->s_lock, flags);
1308	}
1309	ret = 0;
1310bail:
1311	return ret;
1312}
1313
1314/**
1315 * ipath_verbs_send - send a packet
1316 * @qp: the QP to send on
1317 * @hdr: the packet header
1318 * @hdrwords: the number of 32-bit words in the header
1319 * @ss: the SGE to send
1320 * @len: the length of the packet in bytes
1321 */
1322int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
1323		     u32 hdrwords, struct ipath_sge_state *ss, u32 len)
1324{
1325	struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1326	u32 plen;
1327	int ret;
1328	u32 dwords = (len + 3) >> 2;
1329
1330	/*
1331	 * Calculate the send buffer trigger address.
1332	 * The +1 counts for the pbc control dword following the pbc length.
1333	 */
1334	plen = hdrwords + dwords + 1;
1335
1336	/*
1337	 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1338	 * can defer SDMA restart until link goes ACTIVE without
1339	 * worrying about just how we got there.
1340	 */
1341	if (qp->ibqp.qp_type == IB_QPT_SMI ||
1342	    !(dd->ipath_flags & IPATH_HAS_SEND_DMA))
1343		ret = ipath_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1344					   plen, dwords);
1345	else
1346		ret = ipath_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1347					   plen, dwords);
1348
1349	return ret;
1350}
1351
1352int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
1353			    u64 *rwords, u64 *spkts, u64 *rpkts,
1354			    u64 *xmit_wait)
1355{
1356	int ret;
1357
1358	if (!(dd->ipath_flags & IPATH_INITTED)) {
1359		/* no hardware, freeze, etc. */
1360		ret = -EINVAL;
1361		goto bail;
1362	}
1363	*swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
1364	*rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
1365	*spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
1366	*rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
1367	*xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
1368
1369	ret = 0;
1370
1371bail:
1372	return ret;
1373}
1374
1375/**
1376 * ipath_get_counters - get various chip counters
1377 * @dd: the infinipath device
1378 * @cntrs: counters are placed here
1379 *
1380 * Return the counters needed by recv_pma_get_portcounters().
1381 */
1382int ipath_get_counters(struct ipath_devdata *dd,
1383		       struct ipath_verbs_counters *cntrs)
1384{
1385	struct ipath_cregs const *crp = dd->ipath_cregs;
1386	int ret;
1387
1388	if (!(dd->ipath_flags & IPATH_INITTED)) {
1389		/* no hardware, freeze, etc. */
1390		ret = -EINVAL;
1391		goto bail;
1392	}
1393	cntrs->symbol_error_counter =
1394		ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
1395	cntrs->link_error_recovery_counter =
1396		ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
1397	/*
1398	 * The link downed counter counts when the other side downs the
1399	 * connection.  We add in the number of times we downed the link
1400	 * due to local link integrity errors to compensate.
1401	 */
1402	cntrs->link_downed_counter =
1403		ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
1404	cntrs->port_rcv_errors =
1405		ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
1406		ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
1407		ipath_snap_cntr(dd, crp->cr_portovflcnt) +
1408		ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
1409		ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
1410		ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
1411		ipath_snap_cntr(dd, crp->cr_erricrccnt) +
1412		ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
1413		ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1414		ipath_snap_cntr(dd, crp->cr_badformatcnt) +
1415		dd->ipath_rxfc_unsupvl_errs;
1416	if (crp->cr_rxotherlocalphyerrcnt)
1417		cntrs->port_rcv_errors +=
1418			ipath_snap_cntr(dd, crp->cr_rxotherlocalphyerrcnt);
1419	if (crp->cr_rxvlerrcnt)
1420		cntrs->port_rcv_errors +=
1421			ipath_snap_cntr(dd, crp->cr_rxvlerrcnt);
1422	cntrs->port_rcv_remphys_errors =
1423		ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
1424	cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
1425	cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
1426	cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
1427	cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
1428	cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
1429	cntrs->local_link_integrity_errors =
1430		crp->cr_locallinkintegrityerrcnt ?
1431		ipath_snap_cntr(dd, crp->cr_locallinkintegrityerrcnt) :
1432		((dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
1433		 dd->ipath_lli_errs : dd->ipath_lli_errors);
1434	cntrs->excessive_buffer_overrun_errors =
1435		crp->cr_excessbufferovflcnt ?
1436		ipath_snap_cntr(dd, crp->cr_excessbufferovflcnt) :
1437		dd->ipath_overrun_thresh_errs;
1438	cntrs->vl15_dropped = crp->cr_vl15droppedpktcnt ?
1439		ipath_snap_cntr(dd, crp->cr_vl15droppedpktcnt) : 0;
1440
1441	ret = 0;
1442
1443bail:
1444	return ret;
1445}
1446
1447/**
1448 * ipath_ib_piobufavail - callback when a PIO buffer is available
1449 * @arg: the device pointer
1450 *
1451 * This is called from ipath_intr() at interrupt level when a PIO buffer is
1452 * available after ipath_verbs_send() returned an error that no buffers were
1453 * available.  Return 1 if we consumed all the PIO buffers and we still have
1454 * QPs waiting for buffers (for now, just restart the send tasklet and
1455 * return zero).
1456 */
1457int ipath_ib_piobufavail(struct ipath_ibdev *dev)
1458{
1459	struct list_head *list;
1460	struct ipath_qp *qplist;
1461	struct ipath_qp *qp;
1462	unsigned long flags;
1463
1464	if (dev == NULL)
1465		goto bail;
1466
1467	list = &dev->piowait;
1468	qplist = NULL;
1469
1470	spin_lock_irqsave(&dev->pending_lock, flags);
1471	while (!list_empty(list)) {
1472		qp = list_entry(list->next, struct ipath_qp, piowait);
1473		list_del_init(&qp->piowait);
1474		qp->pio_next = qplist;
1475		qplist = qp;
1476		atomic_inc(&qp->refcount);
1477	}
1478	spin_unlock_irqrestore(&dev->pending_lock, flags);
1479
1480	while (qplist != NULL) {
1481		qp = qplist;
1482		qplist = qp->pio_next;
1483
1484		spin_lock_irqsave(&qp->s_lock, flags);
1485		if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
1486			ipath_schedule_send(qp);
1487		spin_unlock_irqrestore(&qp->s_lock, flags);
1488
1489		/* Notify ipath_destroy_qp() if it is waiting. */
1490		if (atomic_dec_and_test(&qp->refcount))
1491			wake_up(&qp->wait);
1492	}
1493
1494bail:
1495	return 0;
1496}
1497
1498static int ipath_query_device(struct ib_device *ibdev,
1499			      struct ib_device_attr *props)
1500{
1501	struct ipath_ibdev *dev = to_idev(ibdev);
1502
1503	memset(props, 0, sizeof(*props));
1504
1505	props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1506		IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1507		IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1508		IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1509	props->page_size_cap = PAGE_SIZE;
1510	props->vendor_id =
1511		IPATH_SRC_OUI_1 << 16 | IPATH_SRC_OUI_2 << 8 | IPATH_SRC_OUI_3;
1512	props->vendor_part_id = dev->dd->ipath_deviceid;
1513	props->hw_ver = dev->dd->ipath_pcirev;
1514
1515	props->sys_image_guid = dev->sys_image_guid;
1516
1517	props->max_mr_size = ~0ull;
1518	props->max_qp = ib_ipath_max_qps;
1519	props->max_qp_wr = ib_ipath_max_qp_wrs;
1520	props->max_sge = ib_ipath_max_sges;
1521	props->max_cq = ib_ipath_max_cqs;
1522	props->max_ah = ib_ipath_max_ahs;
1523	props->max_cqe = ib_ipath_max_cqes;
1524	props->max_mr = dev->lk_table.max;
1525	props->max_fmr = dev->lk_table.max;
1526	props->max_map_per_fmr = 32767;
1527	props->max_pd = ib_ipath_max_pds;
1528	props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
1529	props->max_qp_init_rd_atom = 255;
1530	/* props->max_res_rd_atom */
1531	props->max_srq = ib_ipath_max_srqs;
1532	props->max_srq_wr = ib_ipath_max_srq_wrs;
1533	props->max_srq_sge = ib_ipath_max_srq_sges;
1534	/* props->local_ca_ack_delay */
1535	props->atomic_cap = IB_ATOMIC_GLOB;
1536	props->max_pkeys = ipath_get_npkeys(dev->dd);
1537	props->max_mcast_grp = ib_ipath_max_mcast_grps;
1538	props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1539	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1540		props->max_mcast_grp;
1541
1542	return 0;
1543}
1544
1545const u8 ipath_cvt_physportstate[32] = {
1546	[INFINIPATH_IBCS_LT_STATE_DISABLED] = IB_PHYSPORTSTATE_DISABLED,
1547	[INFINIPATH_IBCS_LT_STATE_LINKUP] = IB_PHYSPORTSTATE_LINKUP,
1548	[INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = IB_PHYSPORTSTATE_POLL,
1549	[INFINIPATH_IBCS_LT_STATE_POLLQUIET] = IB_PHYSPORTSTATE_POLL,
1550	[INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = IB_PHYSPORTSTATE_SLEEP,
1551	[INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = IB_PHYSPORTSTATE_SLEEP,
1552	[INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] =
1553		IB_PHYSPORTSTATE_CFG_TRAIN,
1554	[INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] =
1555		IB_PHYSPORTSTATE_CFG_TRAIN,
1556	[INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] =
1557		IB_PHYSPORTSTATE_CFG_TRAIN,
1558	[INFINIPATH_IBCS_LT_STATE_CFGIDLE] = IB_PHYSPORTSTATE_CFG_TRAIN,
1559	[INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] =
1560		IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1561	[INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] =
1562		IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1563	[INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] =
1564		IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1565	[0x10] = IB_PHYSPORTSTATE_CFG_TRAIN,
1566	[0x11] = IB_PHYSPORTSTATE_CFG_TRAIN,
1567	[0x12] = IB_PHYSPORTSTATE_CFG_TRAIN,
1568	[0x13] = IB_PHYSPORTSTATE_CFG_TRAIN,
1569	[0x14] = IB_PHYSPORTSTATE_CFG_TRAIN,
1570	[0x15] = IB_PHYSPORTSTATE_CFG_TRAIN,
1571	[0x16] = IB_PHYSPORTSTATE_CFG_TRAIN,
1572	[0x17] = IB_PHYSPORTSTATE_CFG_TRAIN
1573};
1574
1575u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1576{
1577	return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1578}
1579
1580static int ipath_query_port(struct ib_device *ibdev,
1581			    u8 port, struct ib_port_attr *props)
1582{
1583	struct ipath_ibdev *dev = to_idev(ibdev);
1584	struct ipath_devdata *dd = dev->dd;
1585	enum ib_mtu mtu;
1586	u16 lid = dd->ipath_lid;
1587	u64 ibcstat;
1588
1589	memset(props, 0, sizeof(*props));
1590	props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1591	props->lmc = dd->ipath_lmc;
1592	props->sm_lid = dev->sm_lid;
1593	props->sm_sl = dev->sm_sl;
1594	ibcstat = dd->ipath_lastibcstat;
1595	/* map LinkState to IB portinfo values.  */
1596	props->state = ipath_ib_linkstate(dd, ibcstat) + 1;
1597
1598	/* See phys_state_show() */
1599	props->phys_state = /* MEA: assumes shift == 0 */
1600		ipath_cvt_physportstate[dd->ipath_lastibcstat &
1601		dd->ibcs_lts_mask];
1602	props->port_cap_flags = dev->port_cap_flags;
1603	props->gid_tbl_len = 1;
1604	props->max_msg_sz = 0x80000000;
1605	props->pkey_tbl_len = ipath_get_npkeys(dd);
1606	props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1607		dev->z_pkey_violations;
1608	props->qkey_viol_cntr = dev->qkey_violations;
1609	props->active_width = dd->ipath_link_width_active;
1610	/* See rate_show() */
1611	props->active_speed = dd->ipath_link_speed_active;
1612	props->max_vl_num = 1;		/* VLCap = VL0 */
1613	props->init_type_reply = 0;
1614
1615	props->max_mtu = ipath_mtu4096 ? IB_MTU_4096 : IB_MTU_2048;
1616	switch (dd->ipath_ibmtu) {
1617	case 4096:
1618		mtu = IB_MTU_4096;
1619		break;
1620	case 2048:
1621		mtu = IB_MTU_2048;
1622		break;
1623	case 1024:
1624		mtu = IB_MTU_1024;
1625		break;
1626	case 512:
1627		mtu = IB_MTU_512;
1628		break;
1629	case 256:
1630		mtu = IB_MTU_256;
1631		break;
1632	default:
1633		mtu = IB_MTU_2048;
1634	}
1635	props->active_mtu = mtu;
1636	props->subnet_timeout = dev->subnet_timeout;
1637
1638	return 0;
1639}
1640
1641static int ipath_modify_device(struct ib_device *device,
1642			       int device_modify_mask,
1643			       struct ib_device_modify *device_modify)
1644{
1645	int ret;
1646
1647	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1648				   IB_DEVICE_MODIFY_NODE_DESC)) {
1649		ret = -EOPNOTSUPP;
1650		goto bail;
1651	}
1652
1653	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1654		memcpy(device->node_desc, device_modify->node_desc, 64);
1655
1656	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1657		to_idev(device)->sys_image_guid =
1658			cpu_to_be64(device_modify->sys_image_guid);
1659
1660	ret = 0;
1661
1662bail:
1663	return ret;
1664}
1665
1666static int ipath_modify_port(struct ib_device *ibdev,
1667			     u8 port, int port_modify_mask,
1668			     struct ib_port_modify *props)
1669{
1670	struct ipath_ibdev *dev = to_idev(ibdev);
1671
1672	dev->port_cap_flags |= props->set_port_cap_mask;
1673	dev->port_cap_flags &= ~props->clr_port_cap_mask;
1674	if (port_modify_mask & IB_PORT_SHUTDOWN)
1675		ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1676	if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1677		dev->qkey_violations = 0;
1678	return 0;
1679}
1680
1681static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1682			   int index, union ib_gid *gid)
1683{
1684	struct ipath_ibdev *dev = to_idev(ibdev);
1685	int ret;
1686
1687	if (index >= 1) {
1688		ret = -EINVAL;
1689		goto bail;
1690	}
1691	gid->global.subnet_prefix = dev->gid_prefix;
1692	gid->global.interface_id = dev->dd->ipath_guid;
1693
1694	ret = 0;
1695
1696bail:
1697	return ret;
1698}
1699
1700static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1701				    struct ib_ucontext *context,
1702				    struct ib_udata *udata)
1703{
1704	struct ipath_ibdev *dev = to_idev(ibdev);
1705	struct ipath_pd *pd;
1706	struct ib_pd *ret;
1707
1708	/*
1709	 * This is actually totally arbitrary.	Some correctness tests
1710	 * assume there's a maximum number of PDs that can be allocated.
1711	 * We don't actually have this limit, but we fail the test if
1712	 * we allow allocations of more than we report for this value.
1713	 */
1714
1715	pd = kmalloc(sizeof *pd, GFP_KERNEL);
1716	if (!pd) {
1717		ret = ERR_PTR(-ENOMEM);
1718		goto bail;
1719	}
1720
1721	spin_lock(&dev->n_pds_lock);
1722	if (dev->n_pds_allocated == ib_ipath_max_pds) {
1723		spin_unlock(&dev->n_pds_lock);
1724		kfree(pd);
1725		ret = ERR_PTR(-ENOMEM);
1726		goto bail;
1727	}
1728
1729	dev->n_pds_allocated++;
1730	spin_unlock(&dev->n_pds_lock);
1731
1732	/* ib_alloc_pd() will initialize pd->ibpd. */
1733	pd->user = udata != NULL;
1734
1735	ret = &pd->ibpd;
1736
1737bail:
1738	return ret;
1739}
1740
1741static int ipath_dealloc_pd(struct ib_pd *ibpd)
1742{
1743	struct ipath_pd *pd = to_ipd(ibpd);
1744	struct ipath_ibdev *dev = to_idev(ibpd->device);
1745
1746	spin_lock(&dev->n_pds_lock);
1747	dev->n_pds_allocated--;
1748	spin_unlock(&dev->n_pds_lock);
1749
1750	kfree(pd);
1751
1752	return 0;
1753}
1754
1755/**
1756 * ipath_create_ah - create an address handle
1757 * @pd: the protection domain
1758 * @ah_attr: the attributes of the AH
1759 *
1760 * This may be called from interrupt context.
1761 */
1762static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1763				     struct ib_ah_attr *ah_attr)
1764{
1765	struct ipath_ah *ah;
1766	struct ib_ah *ret;
1767	struct ipath_ibdev *dev = to_idev(pd->device);
1768	unsigned long flags;
1769
1770	/* A multicast address requires a GRH (see ch. 8.4.1). */
1771	if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1772	    ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1773	    !(ah_attr->ah_flags & IB_AH_GRH)) {
1774		ret = ERR_PTR(-EINVAL);
1775		goto bail;
1776	}
1777
1778	if (ah_attr->dlid == 0) {
1779		ret = ERR_PTR(-EINVAL);
1780		goto bail;
1781	}
1782
1783	if (ah_attr->port_num < 1 ||
1784	    ah_attr->port_num > pd->device->phys_port_cnt) {
1785		ret = ERR_PTR(-EINVAL);
1786		goto bail;
1787	}
1788
1789	ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1790	if (!ah) {
1791		ret = ERR_PTR(-ENOMEM);
1792		goto bail;
1793	}
1794
1795	spin_lock_irqsave(&dev->n_ahs_lock, flags);
1796	if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1797		spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1798		kfree(ah);
1799		ret = ERR_PTR(-ENOMEM);
1800		goto bail;
1801	}
1802
1803	dev->n_ahs_allocated++;
1804	spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1805
1806	/* ib_create_ah() will initialize ah->ibah. */
1807	ah->attr = *ah_attr;
1808	ah->attr.static_rate = ipath_ib_rate_to_mult(ah_attr->static_rate);
1809
1810	ret = &ah->ibah;
1811
1812bail:
1813	return ret;
1814}
1815
1816/**
1817 * ipath_destroy_ah - destroy an address handle
1818 * @ibah: the AH to destroy
1819 *
1820 * This may be called from interrupt context.
1821 */
1822static int ipath_destroy_ah(struct ib_ah *ibah)
1823{
1824	struct ipath_ibdev *dev = to_idev(ibah->device);
1825	struct ipath_ah *ah = to_iah(ibah);
1826	unsigned long flags;
1827
1828	spin_lock_irqsave(&dev->n_ahs_lock, flags);
1829	dev->n_ahs_allocated--;
1830	spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1831
1832	kfree(ah);
1833
1834	return 0;
1835}
1836
1837static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1838{
1839	struct ipath_ah *ah = to_iah(ibah);
1840
1841	*ah_attr = ah->attr;
1842	ah_attr->static_rate = ipath_mult_to_ib_rate(ah->attr.static_rate);
1843
1844	return 0;
1845}
1846
1847/**
1848 * ipath_get_npkeys - return the size of the PKEY table for port 0
1849 * @dd: the infinipath device
1850 */
1851unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1852{
1853	return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1854}
1855
1856/**
1857 * ipath_get_pkey - return the indexed PKEY from the port PKEY table
1858 * @dd: the infinipath device
1859 * @index: the PKEY index
1860 */
1861unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1862{
1863	unsigned ret;
1864
1865	/* always a kernel port, no locking needed */
1866	if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1867		ret = 0;
1868	else
1869		ret = dd->ipath_pd[0]->port_pkeys[index];
1870
1871	return ret;
1872}
1873
1874static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1875			    u16 *pkey)
1876{
1877	struct ipath_ibdev *dev = to_idev(ibdev);
1878	int ret;
1879
1880	if (index >= ipath_get_npkeys(dev->dd)) {
1881		ret = -EINVAL;
1882		goto bail;
1883	}
1884
1885	*pkey = ipath_get_pkey(dev->dd, index);
1886	ret = 0;
1887
1888bail:
1889	return ret;
1890}
1891
1892/**
1893 * ipath_alloc_ucontext - allocate a ucontest
1894 * @ibdev: the infiniband device
1895 * @udata: not used by the InfiniPath driver
1896 */
1897
1898static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1899						struct ib_udata *udata)
1900{
1901	struct ipath_ucontext *context;
1902	struct ib_ucontext *ret;
1903
1904	context = kmalloc(sizeof *context, GFP_KERNEL);
1905	if (!context) {
1906		ret = ERR_PTR(-ENOMEM);
1907		goto bail;
1908	}
1909
1910	ret = &context->ibucontext;
1911
1912bail:
1913	return ret;
1914}
1915
1916static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1917{
1918	kfree(to_iucontext(context));
1919	return 0;
1920}
1921
1922static int ipath_verbs_register_sysfs(struct ib_device *dev);
1923
1924static void __verbs_timer(unsigned long arg)
1925{
1926	struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1927
1928	/* Handle verbs layer timeouts. */
1929	ipath_ib_timer(dd->verbs_dev);
1930
1931	mod_timer(&dd->verbs_timer, jiffies + 1);
1932}
1933
1934static int enable_timer(struct ipath_devdata *dd)
1935{
1936	/*
1937	 * Early chips had a design flaw where the chip and kernel idea
1938	 * of the tail register don't always agree, and therefore we won't
1939	 * get an interrupt on the next packet received.
1940	 * If the board supports per packet receive interrupts, use it.
1941	 * Otherwise, the timer function periodically checks for packets
1942	 * to cover this case.
1943	 * Either way, the timer is needed for verbs layer related
1944	 * processing.
1945	 */
1946	if (dd->ipath_flags & IPATH_GPIO_INTR) {
1947		ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1948				 0x2074076542310ULL);
1949		/* Enable GPIO bit 2 interrupt */
1950		dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1951		ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1952				 dd->ipath_gpio_mask);
1953	}
1954
1955	init_timer(&dd->verbs_timer);
1956	dd->verbs_timer.function = __verbs_timer;
1957	dd->verbs_timer.data = (unsigned long)dd;
1958	dd->verbs_timer.expires = jiffies + 1;
1959	add_timer(&dd->verbs_timer);
1960
1961	return 0;
1962}
1963
1964static int disable_timer(struct ipath_devdata *dd)
1965{
1966	/* Disable GPIO bit 2 interrupt */
1967	if (dd->ipath_flags & IPATH_GPIO_INTR) {
1968                /* Disable GPIO bit 2 interrupt */
1969		dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1970		ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1971				 dd->ipath_gpio_mask);
1972		/*
1973		 * We might want to undo changes to debugportselect,
1974		 * but how?
1975		 */
1976	}
1977
1978	del_timer_sync(&dd->verbs_timer);
1979
1980	return 0;
1981}
1982
1983/**
1984 * ipath_register_ib_device - register our device with the infiniband core
1985 * @dd: the device data structure
1986 * Return the allocated ipath_ibdev pointer or NULL on error.
1987 */
1988int ipath_register_ib_device(struct ipath_devdata *dd)
1989{
1990	struct ipath_verbs_counters cntrs;
1991	struct ipath_ibdev *idev;
1992	struct ib_device *dev;
1993	struct ipath_verbs_txreq *tx;
1994	unsigned i;
1995	int ret;
1996
1997	idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1998	if (idev == NULL) {
1999		ret = -ENOMEM;
2000		goto bail;
2001	}
2002
2003	dev = &idev->ibdev;
2004
2005	if (dd->ipath_sdma_descq_cnt) {
2006		tx = kmalloc(dd->ipath_sdma_descq_cnt * sizeof *tx,
2007			     GFP_KERNEL);
2008		if (tx == NULL) {
2009			ret = -ENOMEM;
2010			goto err_tx;
2011		}
2012	} else
2013		tx = NULL;
2014	idev->txreq_bufs = tx;
2015
2016	/* Only need to initialize non-zero fields. */
2017	spin_lock_init(&idev->n_pds_lock);
2018	spin_lock_init(&idev->n_ahs_lock);
2019	spin_lock_init(&idev->n_cqs_lock);
2020	spin_lock_init(&idev->n_qps_lock);
2021	spin_lock_init(&idev->n_srqs_lock);
2022	spin_lock_init(&idev->n_mcast_grps_lock);
2023
2024	spin_lock_init(&idev->qp_table.lock);
2025	spin_lock_init(&idev->lk_table.lock);
2026	idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
2027	/* Set the prefix to the default value (see ch. 4.1.1) */
2028	idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
2029
2030	ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
2031	if (ret)
2032		goto err_qp;
2033
2034	/*
2035	 * The top ib_ipath_lkey_table_size bits are used to index the
2036	 * table.  The lower 8 bits can be owned by the user (copied from
2037	 * the LKEY).  The remaining bits act as a generation number or tag.
2038	 */
2039	idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
2040	idev->lk_table.table = kzalloc(idev->lk_table.max *
2041				       sizeof(*idev->lk_table.table),
2042				       GFP_KERNEL);
2043	if (idev->lk_table.table == NULL) {
2044		ret = -ENOMEM;
2045		goto err_lk;
2046	}
2047	INIT_LIST_HEAD(&idev->pending_mmaps);
2048	spin_lock_init(&idev->pending_lock);
2049	idev->mmap_offset = PAGE_SIZE;
2050	spin_lock_init(&idev->mmap_offset_lock);
2051	INIT_LIST_HEAD(&idev->pending[0]);
2052	INIT_LIST_HEAD(&idev->pending[1]);
2053	INIT_LIST_HEAD(&idev->pending[2]);
2054	INIT_LIST_HEAD(&idev->piowait);
2055	INIT_LIST_HEAD(&idev->rnrwait);
2056	INIT_LIST_HEAD(&idev->txreq_free);
2057	idev->pending_index = 0;
2058	idev->port_cap_flags =
2059		IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
2060	if (dd->ipath_flags & IPATH_HAS_LINK_LATENCY)
2061		idev->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2062	idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2063	idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2064	idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2065	idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2066	idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2067
2068	/* Snapshot current HW counters to "clear" them. */
2069	ipath_get_counters(dd, &cntrs);
2070	idev->z_symbol_error_counter = cntrs.symbol_error_counter;
2071	idev->z_link_error_recovery_counter =
2072		cntrs.link_error_recovery_counter;
2073	idev->z_link_downed_counter = cntrs.link_downed_counter;
2074	idev->z_port_rcv_errors = cntrs.port_rcv_errors;
2075	idev->z_port_rcv_remphys_errors =
2076		cntrs.port_rcv_remphys_errors;
2077	idev->z_port_xmit_discards = cntrs.port_xmit_discards;
2078	idev->z_port_xmit_data = cntrs.port_xmit_data;
2079	idev->z_port_rcv_data = cntrs.port_rcv_data;
2080	idev->z_port_xmit_packets = cntrs.port_xmit_packets;
2081	idev->z_port_rcv_packets = cntrs.port_rcv_packets;
2082	idev->z_local_link_integrity_errors =
2083		cntrs.local_link_integrity_errors;
2084	idev->z_excessive_buffer_overrun_errors =
2085		cntrs.excessive_buffer_overrun_errors;
2086	idev->z_vl15_dropped = cntrs.vl15_dropped;
2087
2088	for (i = 0; i < dd->ipath_sdma_descq_cnt; i++, tx++)
2089		list_add(&tx->txreq.list, &idev->txreq_free);
2090
2091	/*
2092	 * The system image GUID is supposed to be the same for all
2093	 * IB HCAs in a single system but since there can be other
2094	 * device types in the system, we can't be sure this is unique.
2095	 */
2096	if (!sys_image_guid)
2097		sys_image_guid = dd->ipath_guid;
2098	idev->sys_image_guid = sys_image_guid;
2099	idev->ib_unit = dd->ipath_unit;
2100	idev->dd = dd;
2101
2102	strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
2103	dev->owner = THIS_MODULE;
2104	dev->node_guid = dd->ipath_guid;
2105	dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
2106	dev->uverbs_cmd_mask =
2107		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
2108		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
2109		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
2110		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
2111		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
2112		(1ull << IB_USER_VERBS_CMD_CREATE_AH)		|
2113		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)		|
2114		(1ull << IB_USER_VERBS_CMD_QUERY_AH)		|
2115		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
2116		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
2117		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2118		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
2119		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
2120		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
2121		(1ull << IB_USER_VERBS_CMD_POLL_CQ)		|
2122		(1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)	|
2123		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
2124		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
2125		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
2126		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
2127		(1ull << IB_USER_VERBS_CMD_POST_SEND)		|
2128		(1ull << IB_USER_VERBS_CMD_POST_RECV)		|
2129		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
2130		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
2131		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
2132		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
2133		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
2134		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
2135		(1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2136	dev->node_type = RDMA_NODE_IB_CA;
2137	dev->phys_port_cnt = 1;
2138	dev->num_comp_vectors = 1;
2139	dev->dma_device = &dd->pcidev->dev;
2140	dev->query_device = ipath_query_device;
2141	dev->modify_device = ipath_modify_device;
2142	dev->query_port = ipath_query_port;
2143	dev->modify_port = ipath_modify_port;
2144	dev->query_pkey = ipath_query_pkey;
2145	dev->query_gid = ipath_query_gid;
2146	dev->alloc_ucontext = ipath_alloc_ucontext;
2147	dev->dealloc_ucontext = ipath_dealloc_ucontext;
2148	dev->alloc_pd = ipath_alloc_pd;
2149	dev->dealloc_pd = ipath_dealloc_pd;
2150	dev->create_ah = ipath_create_ah;
2151	dev->destroy_ah = ipath_destroy_ah;
2152	dev->query_ah = ipath_query_ah;
2153	dev->create_srq = ipath_create_srq;
2154	dev->modify_srq = ipath_modify_srq;
2155	dev->query_srq = ipath_query_srq;
2156	dev->destroy_srq = ipath_destroy_srq;
2157	dev->create_qp = ipath_create_qp;
2158	dev->modify_qp = ipath_modify_qp;
2159	dev->query_qp = ipath_query_qp;
2160	dev->destroy_qp = ipath_destroy_qp;
2161	dev->post_send = ipath_post_send;
2162	dev->post_recv = ipath_post_receive;
2163	dev->post_srq_recv = ipath_post_srq_receive;
2164	dev->create_cq = ipath_create_cq;
2165	dev->destroy_cq = ipath_destroy_cq;
2166	dev->resize_cq = ipath_resize_cq;
2167	dev->poll_cq = ipath_poll_cq;
2168	dev->req_notify_cq = ipath_req_notify_cq;
2169	dev->get_dma_mr = ipath_get_dma_mr;
2170	dev->reg_phys_mr = ipath_reg_phys_mr;
2171	dev->reg_user_mr = ipath_reg_user_mr;
2172	dev->dereg_mr = ipath_dereg_mr;
2173	dev->alloc_fmr = ipath_alloc_fmr;
2174	dev->map_phys_fmr = ipath_map_phys_fmr;
2175	dev->unmap_fmr = ipath_unmap_fmr;
2176	dev->dealloc_fmr = ipath_dealloc_fmr;
2177	dev->attach_mcast = ipath_multicast_attach;
2178	dev->detach_mcast = ipath_multicast_detach;
2179	dev->process_mad = ipath_process_mad;
2180	dev->mmap = ipath_mmap;
2181	dev->dma_ops = &ipath_dma_mapping_ops;
2182
2183	snprintf(dev->node_desc, sizeof(dev->node_desc),
2184		 IPATH_IDSTR " %s", init_utsname()->nodename);
2185
2186	ret = ib_register_device(dev, NULL);
2187	if (ret)
2188		goto err_reg;
2189
2190	ret = ipath_verbs_register_sysfs(dev);
2191	if (ret)
2192		goto err_class;
2193
2194	enable_timer(dd);
2195
2196	goto bail;
2197
2198err_class:
2199	ib_unregister_device(dev);
2200err_reg:
2201	kfree(idev->lk_table.table);
2202err_lk:
2203	kfree(idev->qp_table.table);
2204err_qp:
2205	kfree(idev->txreq_bufs);
2206err_tx:
2207	ib_dealloc_device(dev);
2208	ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2209	idev = NULL;
2210
2211bail:
2212	dd->verbs_dev = idev;
2213	return ret;
2214}
2215
2216void ipath_unregister_ib_device(struct ipath_ibdev *dev)
2217{
2218	struct ib_device *ibdev = &dev->ibdev;
2219	u32 qps_inuse;
2220
2221	ib_unregister_device(ibdev);
2222
2223	disable_timer(dev->dd);
2224
2225	if (!list_empty(&dev->pending[0]) ||
2226	    !list_empty(&dev->pending[1]) ||
2227	    !list_empty(&dev->pending[2]))
2228		ipath_dev_err(dev->dd, "pending list not empty!\n");
2229	if (!list_empty(&dev->piowait))
2230		ipath_dev_err(dev->dd, "piowait list not empty!\n");
2231	if (!list_empty(&dev->rnrwait))
2232		ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
2233	if (!ipath_mcast_tree_empty())
2234		ipath_dev_err(dev->dd, "multicast table memory leak!\n");
2235	/*
2236	 * Note that ipath_unregister_ib_device() can be called before all
2237	 * the QPs are destroyed!
2238	 */
2239	qps_inuse = ipath_free_all_qps(&dev->qp_table);
2240	if (qps_inuse)
2241		ipath_dev_err(dev->dd, "QP memory leak! %u still in use\n",
2242			qps_inuse);
2243	kfree(dev->qp_table.table);
2244	kfree(dev->lk_table.table);
2245	kfree(dev->txreq_bufs);
2246	ib_dealloc_device(ibdev);
2247}
2248
2249static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2250			char *buf)
2251{
2252	struct ipath_ibdev *dev =
2253		container_of(device, struct ipath_ibdev, ibdev.dev);
2254
2255	return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
2256}
2257
2258static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2259			char *buf)
2260{
2261	struct ipath_ibdev *dev =
2262		container_of(device, struct ipath_ibdev, ibdev.dev);
2263	int ret;
2264
2265	ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
2266	if (ret < 0)
2267		goto bail;
2268	strcat(buf, "\n");
2269	ret = strlen(buf);
2270
2271bail:
2272	return ret;
2273}
2274
2275static ssize_t show_stats(struct device *device, struct device_attribute *attr,
2276			  char *buf)
2277{
2278	struct ipath_ibdev *dev =
2279		container_of(device, struct ipath_ibdev, ibdev.dev);
2280	int i;
2281	int len;
2282
2283	len = sprintf(buf,
2284		      "RC resends  %d\n"
2285		      "RC no QACK  %d\n"
2286		      "RC ACKs     %d\n"
2287		      "RC SEQ NAKs %d\n"
2288		      "RC RDMA seq %d\n"
2289		      "RC RNR NAKs %d\n"
2290		      "RC OTH NAKs %d\n"
2291		      "RC timeouts %d\n"
2292		      "RC RDMA dup %d\n"
2293		      "piobuf wait %d\n"
2294		      "unaligned   %d\n"
2295		      "PKT drops   %d\n"
2296		      "WQE errs    %d\n",
2297		      dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
2298		      dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
2299		      dev->n_other_naks, dev->n_timeouts,
2300		      dev->n_rdma_dup_busy, dev->n_piowait, dev->n_unaligned,
2301		      dev->n_pkt_drops, dev->n_wqe_errs);
2302	for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
2303		const struct ipath_opcode_stats *si = &dev->opstats[i];
2304
2305		if (!si->n_packets && !si->n_bytes)
2306			continue;
2307		len += sprintf(buf + len, "%02x %llu/%llu\n", i,
2308			       (unsigned long long) si->n_packets,
2309			       (unsigned long long) si->n_bytes);
2310	}
2311	return len;
2312}
2313
2314static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2315static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2316static DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
2317static DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
2318
2319static struct device_attribute *ipath_class_attributes[] = {
2320	&dev_attr_hw_rev,
2321	&dev_attr_hca_type,
2322	&dev_attr_board_id,
2323	&dev_attr_stats
2324};
2325
2326static int ipath_verbs_register_sysfs(struct ib_device *dev)
2327{
2328	int i;
2329	int ret;
2330
2331	for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i) {
2332		ret = device_create_file(&dev->dev,
2333				       ipath_class_attributes[i]);
2334		if (ret)
2335			goto bail;
2336	}
2337	return 0;
2338bail:
2339	for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
2340		device_remove_file(&dev->dev, ipath_class_attributes[i]);
2341	return ret;
2342}
2343