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 
48 static unsigned int ib_qib_qp_table_size = 256;
49 module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO);
50 MODULE_PARM_DESC(qp_table_size, "QP table size");
51 
52 unsigned int ib_qib_lkey_table_size = 16;
53 module_param_named(lkey_table_size, ib_qib_lkey_table_size, uint,
54 		   S_IRUGO);
55 MODULE_PARM_DESC(lkey_table_size,
56 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
57 
58 static unsigned int ib_qib_max_pds = 0xFFFF;
59 module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO);
60 MODULE_PARM_DESC(max_pds,
61 		 "Maximum number of protection domains to support");
62 
63 static unsigned int ib_qib_max_ahs = 0xFFFF;
64 module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO);
65 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
66 
67 unsigned int ib_qib_max_cqes = 0x2FFFF;
68 module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO);
69 MODULE_PARM_DESC(max_cqes,
70 		 "Maximum number of completion queue entries to support");
71 
72 unsigned int ib_qib_max_cqs = 0x1FFFF;
73 module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO);
74 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
75 
76 unsigned int ib_qib_max_qp_wrs = 0x3FFF;
77 module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO);
78 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
79 
80 unsigned int ib_qib_max_qps = 16384;
81 module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO);
82 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
83 
84 unsigned int ib_qib_max_sges = 0x60;
85 module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO);
86 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
87 
88 unsigned int ib_qib_max_mcast_grps = 16384;
89 module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO);
90 MODULE_PARM_DESC(max_mcast_grps,
91 		 "Maximum number of multicast groups to support");
92 
93 unsigned int ib_qib_max_mcast_qp_attached = 16;
94 module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached,
95 		   uint, S_IRUGO);
96 MODULE_PARM_DESC(max_mcast_qp_attached,
97 		 "Maximum number of attached QPs to support");
98 
99 unsigned int ib_qib_max_srqs = 1024;
100 module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO);
101 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
102 
103 unsigned int ib_qib_max_srq_sges = 128;
104 module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO);
105 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
106 
107 unsigned int ib_qib_max_srq_wrs = 0x1FFFF;
108 module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO);
109 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
110 
111 static unsigned int ib_qib_disable_sma;
112 module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO);
113 MODULE_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  */
120 const 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 
135 struct qib_ucontext {
136 	struct ib_ucontext ibucontext;
137 };
138 
139 static 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  */
148 const 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  */
169 void 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  */
211 void 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  */
251 static 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  */
297 static 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  */
337 static 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 	int avoid_schedule = 0;
350 
351 	spin_lock_irqsave(&qp->s_lock, flags);
352 
353 	/* Check that state is OK to post send. */
354 	if (unlikely(!(ib_qib_state_ops[qp->state] & QIB_POST_SEND_OK)))
355 		goto bail_inval;
356 
357 	/* IB spec says that num_sge == 0 is OK. */
358 	if (wr->num_sge > qp->s_max_sge)
359 		goto bail_inval;
360 
361 	/*
362 	 * Don't allow RDMA reads or atomic operations on UC or
363 	 * undefined operations.
364 	 * Make sure buffer is large enough to hold the result for atomics.
365 	 */
366 	if (wr->opcode == IB_WR_REG_MR) {
367 		if (qib_reg_mr(qp, reg_wr(wr)))
368 			goto bail_inval;
369 	} else if (qp->ibqp.qp_type == IB_QPT_UC) {
370 		if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
371 			goto bail_inval;
372 	} else if (qp->ibqp.qp_type != IB_QPT_RC) {
373 		/* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
374 		if (wr->opcode != IB_WR_SEND &&
375 		    wr->opcode != IB_WR_SEND_WITH_IMM)
376 			goto bail_inval;
377 		/* Check UD destination address PD */
378 		if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
379 			goto bail_inval;
380 	} else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
381 		goto bail_inval;
382 	else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
383 		   (wr->num_sge == 0 ||
384 		    wr->sg_list[0].length < sizeof(u64) ||
385 		    wr->sg_list[0].addr & (sizeof(u64) - 1)))
386 		goto bail_inval;
387 	else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
388 		goto bail_inval;
389 
390 	next = qp->s_head + 1;
391 	if (next >= qp->s_size)
392 		next = 0;
393 	if (next == qp->s_last) {
394 		ret = -ENOMEM;
395 		goto bail;
396 	}
397 
398 	rkt = &to_idev(qp->ibqp.device)->lk_table;
399 	pd = to_ipd(qp->ibqp.pd);
400 	wqe = get_swqe_ptr(qp, qp->s_head);
401 
402 	if (qp->ibqp.qp_type != IB_QPT_UC &&
403 	    qp->ibqp.qp_type != IB_QPT_RC)
404 		memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
405 	else if (wr->opcode == IB_WR_REG_MR)
406 		memcpy(&wqe->reg_wr, reg_wr(wr),
407 			sizeof(wqe->reg_wr));
408 	else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
409 		 wr->opcode == IB_WR_RDMA_WRITE ||
410 		 wr->opcode == IB_WR_RDMA_READ)
411 		memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
412 	else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
413 		 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
414 		memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
415 	else
416 		memcpy(&wqe->wr, wr, sizeof(wqe->wr));
417 
418 	wqe->length = 0;
419 	j = 0;
420 	if (wr->num_sge) {
421 		acc = wr->opcode >= IB_WR_RDMA_READ ?
422 			IB_ACCESS_LOCAL_WRITE : 0;
423 		for (i = 0; i < wr->num_sge; i++) {
424 			u32 length = wr->sg_list[i].length;
425 			int ok;
426 
427 			if (length == 0)
428 				continue;
429 			ok = qib_lkey_ok(rkt, pd, &wqe->sg_list[j],
430 					 &wr->sg_list[i], acc);
431 			if (!ok)
432 				goto bail_inval_free;
433 			wqe->length += length;
434 			j++;
435 		}
436 		wqe->wr.num_sge = j;
437 	}
438 	if (qp->ibqp.qp_type == IB_QPT_UC ||
439 	    qp->ibqp.qp_type == IB_QPT_RC) {
440 		if (wqe->length > 0x80000000U)
441 			goto bail_inval_free;
442 		if (wqe->length <= qp->pmtu)
443 			avoid_schedule = 1;
444 	} else if (wqe->length > (dd_from_ibdev(qp->ibqp.device)->pport +
445 				  qp->port_num - 1)->ibmtu) {
446 		goto bail_inval_free;
447 	} else {
448 		atomic_inc(&to_iah(ud_wr(wr)->ah)->refcount);
449 		avoid_schedule = 1;
450 	}
451 	wqe->ssn = qp->s_ssn++;
452 	qp->s_head = next;
453 
454 	ret = 0;
455 	goto bail;
456 
457 bail_inval_free:
458 	while (j) {
459 		struct qib_sge *sge = &wqe->sg_list[--j];
460 
461 		qib_put_mr(sge->mr);
462 	}
463 bail_inval:
464 	ret = -EINVAL;
465 bail:
466 	if (!ret && !wr->next && !avoid_schedule &&
467 	 !qib_sdma_empty(
468 	   dd_from_ibdev(qp->ibqp.device)->pport + qp->port_num - 1)) {
469 		qib_schedule_send(qp);
470 		*scheduled = 1;
471 	}
472 	spin_unlock_irqrestore(&qp->s_lock, flags);
473 	return ret;
474 }
475 
476 /**
477  * qib_post_send - post a send on a QP
478  * @ibqp: the QP to post the send on
479  * @wr: the list of work requests to post
480  * @bad_wr: the first bad WR is put here
481  *
482  * This may be called from interrupt context.
483  */
484 static int qib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
485 			 struct ib_send_wr **bad_wr)
486 {
487 	struct qib_qp *qp = to_iqp(ibqp);
488 	int err = 0;
489 	int scheduled = 0;
490 
491 	for (; wr; wr = wr->next) {
492 		err = qib_post_one_send(qp, wr, &scheduled);
493 		if (err) {
494 			*bad_wr = wr;
495 			goto bail;
496 		}
497 	}
498 
499 	/* Try to do the send work in the caller's context. */
500 	if (!scheduled)
501 		qib_do_send(&qp->s_work);
502 
503 bail:
504 	return err;
505 }
506 
507 /**
508  * qib_post_receive - post a receive on a QP
509  * @ibqp: the QP to post the receive on
510  * @wr: the WR to post
511  * @bad_wr: the first bad WR is put here
512  *
513  * This may be called from interrupt context.
514  */
515 static int qib_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
516 			    struct ib_recv_wr **bad_wr)
517 {
518 	struct qib_qp *qp = to_iqp(ibqp);
519 	struct qib_rwq *wq = qp->r_rq.wq;
520 	unsigned long flags;
521 	int ret;
522 
523 	/* Check that state is OK to post receive. */
524 	if (!(ib_qib_state_ops[qp->state] & QIB_POST_RECV_OK) || !wq) {
525 		*bad_wr = wr;
526 		ret = -EINVAL;
527 		goto bail;
528 	}
529 
530 	for (; wr; wr = wr->next) {
531 		struct qib_rwqe *wqe;
532 		u32 next;
533 		int i;
534 
535 		if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
536 			*bad_wr = wr;
537 			ret = -EINVAL;
538 			goto bail;
539 		}
540 
541 		spin_lock_irqsave(&qp->r_rq.lock, flags);
542 		next = wq->head + 1;
543 		if (next >= qp->r_rq.size)
544 			next = 0;
545 		if (next == wq->tail) {
546 			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
547 			*bad_wr = wr;
548 			ret = -ENOMEM;
549 			goto bail;
550 		}
551 
552 		wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
553 		wqe->wr_id = wr->wr_id;
554 		wqe->num_sge = wr->num_sge;
555 		for (i = 0; i < wr->num_sge; i++)
556 			wqe->sg_list[i] = wr->sg_list[i];
557 		/* Make sure queue entry is written before the head index. */
558 		smp_wmb();
559 		wq->head = next;
560 		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
561 	}
562 	ret = 0;
563 
564 bail:
565 	return ret;
566 }
567 
568 /**
569  * qib_qp_rcv - processing an incoming packet on a QP
570  * @rcd: the context pointer
571  * @hdr: the packet header
572  * @has_grh: true if the packet has a GRH
573  * @data: the packet data
574  * @tlen: the packet length
575  * @qp: the QP the packet came on
576  *
577  * This is called from qib_ib_rcv() to process an incoming packet
578  * for the given QP.
579  * Called at interrupt level.
580  */
581 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
582 		       int has_grh, void *data, u32 tlen, struct qib_qp *qp)
583 {
584 	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
585 
586 	spin_lock(&qp->r_lock);
587 
588 	/* Check for valid receive state. */
589 	if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) {
590 		ibp->n_pkt_drops++;
591 		goto unlock;
592 	}
593 
594 	switch (qp->ibqp.qp_type) {
595 	case IB_QPT_SMI:
596 	case IB_QPT_GSI:
597 		if (ib_qib_disable_sma)
598 			break;
599 		/* FALLTHROUGH */
600 	case IB_QPT_UD:
601 		qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
602 		break;
603 
604 	case IB_QPT_RC:
605 		qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
606 		break;
607 
608 	case IB_QPT_UC:
609 		qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
610 		break;
611 
612 	default:
613 		break;
614 	}
615 
616 unlock:
617 	spin_unlock(&qp->r_lock);
618 }
619 
620 /**
621  * qib_ib_rcv - process an incoming packet
622  * @rcd: the context pointer
623  * @rhdr: the header of the packet
624  * @data: the packet payload
625  * @tlen: the packet length
626  *
627  * This is called from qib_kreceive() to process an incoming packet at
628  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
629  */
630 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
631 {
632 	struct qib_pportdata *ppd = rcd->ppd;
633 	struct qib_ibport *ibp = &ppd->ibport_data;
634 	struct qib_ib_header *hdr = rhdr;
635 	struct qib_other_headers *ohdr;
636 	struct qib_qp *qp;
637 	u32 qp_num;
638 	int lnh;
639 	u8 opcode;
640 	u16 lid;
641 
642 	/* 24 == LRH+BTH+CRC */
643 	if (unlikely(tlen < 24))
644 		goto drop;
645 
646 	/* Check for a valid destination LID (see ch. 7.11.1). */
647 	lid = be16_to_cpu(hdr->lrh[1]);
648 	if (lid < QIB_MULTICAST_LID_BASE) {
649 		lid &= ~((1 << ppd->lmc) - 1);
650 		if (unlikely(lid != ppd->lid))
651 			goto drop;
652 	}
653 
654 	/* Check for GRH */
655 	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
656 	if (lnh == QIB_LRH_BTH)
657 		ohdr = &hdr->u.oth;
658 	else if (lnh == QIB_LRH_GRH) {
659 		u32 vtf;
660 
661 		ohdr = &hdr->u.l.oth;
662 		if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
663 			goto drop;
664 		vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
665 		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
666 			goto drop;
667 	} else
668 		goto drop;
669 
670 	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
671 #ifdef CONFIG_DEBUG_FS
672 	rcd->opstats->stats[opcode].n_bytes += tlen;
673 	rcd->opstats->stats[opcode].n_packets++;
674 #endif
675 
676 	/* Get the destination QP number. */
677 	qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
678 	if (qp_num == QIB_MULTICAST_QPN) {
679 		struct qib_mcast *mcast;
680 		struct qib_mcast_qp *p;
681 
682 		if (lnh != QIB_LRH_GRH)
683 			goto drop;
684 		mcast = qib_mcast_find(ibp, &hdr->u.l.grh.dgid);
685 		if (mcast == NULL)
686 			goto drop;
687 		this_cpu_inc(ibp->pmastats->n_multicast_rcv);
688 		list_for_each_entry_rcu(p, &mcast->qp_list, list)
689 			qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
690 		/*
691 		 * Notify qib_multicast_detach() if it is waiting for us
692 		 * to finish.
693 		 */
694 		if (atomic_dec_return(&mcast->refcount) <= 1)
695 			wake_up(&mcast->wait);
696 	} else {
697 		if (rcd->lookaside_qp) {
698 			if (rcd->lookaside_qpn != qp_num) {
699 				if (atomic_dec_and_test(
700 					&rcd->lookaside_qp->refcount))
701 					wake_up(
702 					 &rcd->lookaside_qp->wait);
703 				rcd->lookaside_qp = NULL;
704 			}
705 		}
706 		if (!rcd->lookaside_qp) {
707 			qp = qib_lookup_qpn(ibp, qp_num);
708 			if (!qp)
709 				goto drop;
710 			rcd->lookaside_qp = qp;
711 			rcd->lookaside_qpn = qp_num;
712 		} else
713 			qp = rcd->lookaside_qp;
714 		this_cpu_inc(ibp->pmastats->n_unicast_rcv);
715 		qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
716 	}
717 	return;
718 
719 drop:
720 	ibp->n_pkt_drops++;
721 }
722 
723 /*
724  * This is called from a timer to check for QPs
725  * which need kernel memory in order to send a packet.
726  */
727 static void mem_timer(unsigned long data)
728 {
729 	struct qib_ibdev *dev = (struct qib_ibdev *) data;
730 	struct list_head *list = &dev->memwait;
731 	struct qib_qp *qp = NULL;
732 	unsigned long flags;
733 
734 	spin_lock_irqsave(&dev->pending_lock, flags);
735 	if (!list_empty(list)) {
736 		qp = list_entry(list->next, struct qib_qp, iowait);
737 		list_del_init(&qp->iowait);
738 		atomic_inc(&qp->refcount);
739 		if (!list_empty(list))
740 			mod_timer(&dev->mem_timer, jiffies + 1);
741 	}
742 	spin_unlock_irqrestore(&dev->pending_lock, flags);
743 
744 	if (qp) {
745 		spin_lock_irqsave(&qp->s_lock, flags);
746 		if (qp->s_flags & QIB_S_WAIT_KMEM) {
747 			qp->s_flags &= ~QIB_S_WAIT_KMEM;
748 			qib_schedule_send(qp);
749 		}
750 		spin_unlock_irqrestore(&qp->s_lock, flags);
751 		if (atomic_dec_and_test(&qp->refcount))
752 			wake_up(&qp->wait);
753 	}
754 }
755 
756 static void update_sge(struct qib_sge_state *ss, u32 length)
757 {
758 	struct qib_sge *sge = &ss->sge;
759 
760 	sge->vaddr += length;
761 	sge->length -= length;
762 	sge->sge_length -= length;
763 	if (sge->sge_length == 0) {
764 		if (--ss->num_sge)
765 			*sge = *ss->sg_list++;
766 	} else if (sge->length == 0 && sge->mr->lkey) {
767 		if (++sge->n >= QIB_SEGSZ) {
768 			if (++sge->m >= sge->mr->mapsz)
769 				return;
770 			sge->n = 0;
771 		}
772 		sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
773 		sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
774 	}
775 }
776 
777 #ifdef __LITTLE_ENDIAN
778 static inline u32 get_upper_bits(u32 data, u32 shift)
779 {
780 	return data >> shift;
781 }
782 
783 static inline u32 set_upper_bits(u32 data, u32 shift)
784 {
785 	return data << shift;
786 }
787 
788 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
789 {
790 	data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
791 	data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
792 	return data;
793 }
794 #else
795 static inline u32 get_upper_bits(u32 data, u32 shift)
796 {
797 	return data << shift;
798 }
799 
800 static inline u32 set_upper_bits(u32 data, u32 shift)
801 {
802 	return data >> shift;
803 }
804 
805 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
806 {
807 	data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
808 	data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
809 	return data;
810 }
811 #endif
812 
813 static void copy_io(u32 __iomem *piobuf, struct qib_sge_state *ss,
814 		    u32 length, unsigned flush_wc)
815 {
816 	u32 extra = 0;
817 	u32 data = 0;
818 	u32 last;
819 
820 	while (1) {
821 		u32 len = ss->sge.length;
822 		u32 off;
823 
824 		if (len > length)
825 			len = length;
826 		if (len > ss->sge.sge_length)
827 			len = ss->sge.sge_length;
828 		BUG_ON(len == 0);
829 		/* If the source address is not aligned, try to align it. */
830 		off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
831 		if (off) {
832 			u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
833 					    ~(sizeof(u32) - 1));
834 			u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
835 			u32 y;
836 
837 			y = sizeof(u32) - off;
838 			if (len > y)
839 				len = y;
840 			if (len + extra >= sizeof(u32)) {
841 				data |= set_upper_bits(v, extra *
842 						       BITS_PER_BYTE);
843 				len = sizeof(u32) - extra;
844 				if (len == length) {
845 					last = data;
846 					break;
847 				}
848 				__raw_writel(data, piobuf);
849 				piobuf++;
850 				extra = 0;
851 				data = 0;
852 			} else {
853 				/* Clear unused upper bytes */
854 				data |= clear_upper_bytes(v, len, extra);
855 				if (len == length) {
856 					last = data;
857 					break;
858 				}
859 				extra += len;
860 			}
861 		} else if (extra) {
862 			/* Source address is aligned. */
863 			u32 *addr = (u32 *) ss->sge.vaddr;
864 			int shift = extra * BITS_PER_BYTE;
865 			int ushift = 32 - shift;
866 			u32 l = len;
867 
868 			while (l >= sizeof(u32)) {
869 				u32 v = *addr;
870 
871 				data |= set_upper_bits(v, shift);
872 				__raw_writel(data, piobuf);
873 				data = get_upper_bits(v, ushift);
874 				piobuf++;
875 				addr++;
876 				l -= sizeof(u32);
877 			}
878 			/*
879 			 * We still have 'extra' number of bytes leftover.
880 			 */
881 			if (l) {
882 				u32 v = *addr;
883 
884 				if (l + extra >= sizeof(u32)) {
885 					data |= set_upper_bits(v, shift);
886 					len -= l + extra - sizeof(u32);
887 					if (len == length) {
888 						last = data;
889 						break;
890 					}
891 					__raw_writel(data, piobuf);
892 					piobuf++;
893 					extra = 0;
894 					data = 0;
895 				} else {
896 					/* Clear unused upper bytes */
897 					data |= clear_upper_bytes(v, l, extra);
898 					if (len == length) {
899 						last = data;
900 						break;
901 					}
902 					extra += l;
903 				}
904 			} else if (len == length) {
905 				last = data;
906 				break;
907 			}
908 		} else if (len == length) {
909 			u32 w;
910 
911 			/*
912 			 * Need to round up for the last dword in the
913 			 * packet.
914 			 */
915 			w = (len + 3) >> 2;
916 			qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
917 			piobuf += w - 1;
918 			last = ((u32 *) ss->sge.vaddr)[w - 1];
919 			break;
920 		} else {
921 			u32 w = len >> 2;
922 
923 			qib_pio_copy(piobuf, ss->sge.vaddr, w);
924 			piobuf += w;
925 
926 			extra = len & (sizeof(u32) - 1);
927 			if (extra) {
928 				u32 v = ((u32 *) ss->sge.vaddr)[w];
929 
930 				/* Clear unused upper bytes */
931 				data = clear_upper_bytes(v, extra, 0);
932 			}
933 		}
934 		update_sge(ss, len);
935 		length -= len;
936 	}
937 	/* Update address before sending packet. */
938 	update_sge(ss, length);
939 	if (flush_wc) {
940 		/* must flush early everything before trigger word */
941 		qib_flush_wc();
942 		__raw_writel(last, piobuf);
943 		/* be sure trigger word is written */
944 		qib_flush_wc();
945 	} else
946 		__raw_writel(last, piobuf);
947 }
948 
949 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
950 					   struct qib_qp *qp)
951 {
952 	struct qib_verbs_txreq *tx;
953 	unsigned long flags;
954 
955 	spin_lock_irqsave(&qp->s_lock, flags);
956 	spin_lock(&dev->pending_lock);
957 
958 	if (!list_empty(&dev->txreq_free)) {
959 		struct list_head *l = dev->txreq_free.next;
960 
961 		list_del(l);
962 		spin_unlock(&dev->pending_lock);
963 		spin_unlock_irqrestore(&qp->s_lock, flags);
964 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
965 	} else {
966 		if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK &&
967 		    list_empty(&qp->iowait)) {
968 			dev->n_txwait++;
969 			qp->s_flags |= QIB_S_WAIT_TX;
970 			list_add_tail(&qp->iowait, &dev->txwait);
971 		}
972 		qp->s_flags &= ~QIB_S_BUSY;
973 		spin_unlock(&dev->pending_lock);
974 		spin_unlock_irqrestore(&qp->s_lock, flags);
975 		tx = ERR_PTR(-EBUSY);
976 	}
977 	return tx;
978 }
979 
980 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
981 					 struct qib_qp *qp)
982 {
983 	struct qib_verbs_txreq *tx;
984 	unsigned long flags;
985 
986 	spin_lock_irqsave(&dev->pending_lock, flags);
987 	/* assume the list non empty */
988 	if (likely(!list_empty(&dev->txreq_free))) {
989 		struct list_head *l = dev->txreq_free.next;
990 
991 		list_del(l);
992 		spin_unlock_irqrestore(&dev->pending_lock, flags);
993 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
994 	} else {
995 		/* call slow path to get the extra lock */
996 		spin_unlock_irqrestore(&dev->pending_lock, flags);
997 		tx =  __get_txreq(dev, qp);
998 	}
999 	return tx;
1000 }
1001 
1002 void qib_put_txreq(struct qib_verbs_txreq *tx)
1003 {
1004 	struct qib_ibdev *dev;
1005 	struct qib_qp *qp;
1006 	unsigned long flags;
1007 
1008 	qp = tx->qp;
1009 	dev = to_idev(qp->ibqp.device);
1010 
1011 	if (atomic_dec_and_test(&qp->refcount))
1012 		wake_up(&qp->wait);
1013 	if (tx->mr) {
1014 		qib_put_mr(tx->mr);
1015 		tx->mr = NULL;
1016 	}
1017 	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
1018 		tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
1019 		dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
1020 				 tx->txreq.addr, tx->hdr_dwords << 2,
1021 				 DMA_TO_DEVICE);
1022 		kfree(tx->align_buf);
1023 	}
1024 
1025 	spin_lock_irqsave(&dev->pending_lock, flags);
1026 
1027 	/* Put struct back on free list */
1028 	list_add(&tx->txreq.list, &dev->txreq_free);
1029 
1030 	if (!list_empty(&dev->txwait)) {
1031 		/* Wake up first QP wanting a free struct */
1032 		qp = list_entry(dev->txwait.next, struct qib_qp, iowait);
1033 		list_del_init(&qp->iowait);
1034 		atomic_inc(&qp->refcount);
1035 		spin_unlock_irqrestore(&dev->pending_lock, flags);
1036 
1037 		spin_lock_irqsave(&qp->s_lock, flags);
1038 		if (qp->s_flags & QIB_S_WAIT_TX) {
1039 			qp->s_flags &= ~QIB_S_WAIT_TX;
1040 			qib_schedule_send(qp);
1041 		}
1042 		spin_unlock_irqrestore(&qp->s_lock, flags);
1043 
1044 		if (atomic_dec_and_test(&qp->refcount))
1045 			wake_up(&qp->wait);
1046 	} else
1047 		spin_unlock_irqrestore(&dev->pending_lock, flags);
1048 }
1049 
1050 /*
1051  * This is called when there are send DMA descriptors that might be
1052  * available.
1053  *
1054  * This is called with ppd->sdma_lock held.
1055  */
1056 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
1057 {
1058 	struct qib_qp *qp, *nqp;
1059 	struct qib_qp *qps[20];
1060 	struct qib_ibdev *dev;
1061 	unsigned i, n;
1062 
1063 	n = 0;
1064 	dev = &ppd->dd->verbs_dev;
1065 	spin_lock(&dev->pending_lock);
1066 
1067 	/* Search wait list for first QP wanting DMA descriptors. */
1068 	list_for_each_entry_safe(qp, nqp, &dev->dmawait, iowait) {
1069 		if (qp->port_num != ppd->port)
1070 			continue;
1071 		if (n == ARRAY_SIZE(qps))
1072 			break;
1073 		if (qp->s_tx->txreq.sg_count > avail)
1074 			break;
1075 		avail -= qp->s_tx->txreq.sg_count;
1076 		list_del_init(&qp->iowait);
1077 		atomic_inc(&qp->refcount);
1078 		qps[n++] = qp;
1079 	}
1080 
1081 	spin_unlock(&dev->pending_lock);
1082 
1083 	for (i = 0; i < n; i++) {
1084 		qp = qps[i];
1085 		spin_lock(&qp->s_lock);
1086 		if (qp->s_flags & QIB_S_WAIT_DMA_DESC) {
1087 			qp->s_flags &= ~QIB_S_WAIT_DMA_DESC;
1088 			qib_schedule_send(qp);
1089 		}
1090 		spin_unlock(&qp->s_lock);
1091 		if (atomic_dec_and_test(&qp->refcount))
1092 			wake_up(&qp->wait);
1093 	}
1094 }
1095 
1096 /*
1097  * This is called with ppd->sdma_lock held.
1098  */
1099 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
1100 {
1101 	struct qib_verbs_txreq *tx =
1102 		container_of(cookie, struct qib_verbs_txreq, txreq);
1103 	struct qib_qp *qp = tx->qp;
1104 
1105 	spin_lock(&qp->s_lock);
1106 	if (tx->wqe)
1107 		qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
1108 	else if (qp->ibqp.qp_type == IB_QPT_RC) {
1109 		struct qib_ib_header *hdr;
1110 
1111 		if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
1112 			hdr = &tx->align_buf->hdr;
1113 		else {
1114 			struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1115 
1116 			hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
1117 		}
1118 		qib_rc_send_complete(qp, hdr);
1119 	}
1120 	if (atomic_dec_and_test(&qp->s_dma_busy)) {
1121 		if (qp->state == IB_QPS_RESET)
1122 			wake_up(&qp->wait_dma);
1123 		else if (qp->s_flags & QIB_S_WAIT_DMA) {
1124 			qp->s_flags &= ~QIB_S_WAIT_DMA;
1125 			qib_schedule_send(qp);
1126 		}
1127 	}
1128 	spin_unlock(&qp->s_lock);
1129 
1130 	qib_put_txreq(tx);
1131 }
1132 
1133 static int wait_kmem(struct qib_ibdev *dev, struct qib_qp *qp)
1134 {
1135 	unsigned long flags;
1136 	int ret = 0;
1137 
1138 	spin_lock_irqsave(&qp->s_lock, flags);
1139 	if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
1140 		spin_lock(&dev->pending_lock);
1141 		if (list_empty(&qp->iowait)) {
1142 			if (list_empty(&dev->memwait))
1143 				mod_timer(&dev->mem_timer, jiffies + 1);
1144 			qp->s_flags |= QIB_S_WAIT_KMEM;
1145 			list_add_tail(&qp->iowait, &dev->memwait);
1146 		}
1147 		spin_unlock(&dev->pending_lock);
1148 		qp->s_flags &= ~QIB_S_BUSY;
1149 		ret = -EBUSY;
1150 	}
1151 	spin_unlock_irqrestore(&qp->s_lock, flags);
1152 
1153 	return ret;
1154 }
1155 
1156 static int qib_verbs_send_dma(struct qib_qp *qp, struct qib_ib_header *hdr,
1157 			      u32 hdrwords, struct qib_sge_state *ss, u32 len,
1158 			      u32 plen, u32 dwords)
1159 {
1160 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1161 	struct qib_devdata *dd = dd_from_dev(dev);
1162 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1163 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1164 	struct qib_verbs_txreq *tx;
1165 	struct qib_pio_header *phdr;
1166 	u32 control;
1167 	u32 ndesc;
1168 	int ret;
1169 
1170 	tx = qp->s_tx;
1171 	if (tx) {
1172 		qp->s_tx = NULL;
1173 		/* resend previously constructed packet */
1174 		ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
1175 		goto bail;
1176 	}
1177 
1178 	tx = get_txreq(dev, qp);
1179 	if (IS_ERR(tx))
1180 		goto bail_tx;
1181 
1182 	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
1183 				       be16_to_cpu(hdr->lrh[0]) >> 12);
1184 	tx->qp = qp;
1185 	atomic_inc(&qp->refcount);
1186 	tx->wqe = qp->s_wqe;
1187 	tx->mr = qp->s_rdma_mr;
1188 	if (qp->s_rdma_mr)
1189 		qp->s_rdma_mr = NULL;
1190 	tx->txreq.callback = sdma_complete;
1191 	if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
1192 		tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
1193 	else
1194 		tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
1195 	if (plen + 1 > dd->piosize2kmax_dwords)
1196 		tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
1197 
1198 	if (len) {
1199 		/*
1200 		 * Don't try to DMA if it takes more descriptors than
1201 		 * the queue holds.
1202 		 */
1203 		ndesc = qib_count_sge(ss, len);
1204 		if (ndesc >= ppd->sdma_descq_cnt)
1205 			ndesc = 0;
1206 	} else
1207 		ndesc = 1;
1208 	if (ndesc) {
1209 		phdr = &dev->pio_hdrs[tx->hdr_inx];
1210 		phdr->pbc[0] = cpu_to_le32(plen);
1211 		phdr->pbc[1] = cpu_to_le32(control);
1212 		memcpy(&phdr->hdr, hdr, hdrwords << 2);
1213 		tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
1214 		tx->txreq.sg_count = ndesc;
1215 		tx->txreq.addr = dev->pio_hdrs_phys +
1216 			tx->hdr_inx * sizeof(struct qib_pio_header);
1217 		tx->hdr_dwords = hdrwords + 2; /* add PBC length */
1218 		ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
1219 		goto bail;
1220 	}
1221 
1222 	/* Allocate a buffer and copy the header and payload to it. */
1223 	tx->hdr_dwords = plen + 1;
1224 	phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
1225 	if (!phdr)
1226 		goto err_tx;
1227 	phdr->pbc[0] = cpu_to_le32(plen);
1228 	phdr->pbc[1] = cpu_to_le32(control);
1229 	memcpy(&phdr->hdr, hdr, hdrwords << 2);
1230 	qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
1231 
1232 	tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
1233 					tx->hdr_dwords << 2, DMA_TO_DEVICE);
1234 	if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
1235 		goto map_err;
1236 	tx->align_buf = phdr;
1237 	tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
1238 	tx->txreq.sg_count = 1;
1239 	ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
1240 	goto unaligned;
1241 
1242 map_err:
1243 	kfree(phdr);
1244 err_tx:
1245 	qib_put_txreq(tx);
1246 	ret = wait_kmem(dev, qp);
1247 unaligned:
1248 	ibp->n_unaligned++;
1249 bail:
1250 	return ret;
1251 bail_tx:
1252 	ret = PTR_ERR(tx);
1253 	goto bail;
1254 }
1255 
1256 /*
1257  * If we are now in the error state, return zero to flush the
1258  * send work request.
1259  */
1260 static int no_bufs_available(struct qib_qp *qp)
1261 {
1262 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
1263 	struct qib_devdata *dd;
1264 	unsigned long flags;
1265 	int ret = 0;
1266 
1267 	/*
1268 	 * Note that as soon as want_buffer() is called and
1269 	 * possibly before it returns, qib_ib_piobufavail()
1270 	 * could be called. Therefore, put QP on the I/O wait list before
1271 	 * enabling the PIO avail interrupt.
1272 	 */
1273 	spin_lock_irqsave(&qp->s_lock, flags);
1274 	if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) {
1275 		spin_lock(&dev->pending_lock);
1276 		if (list_empty(&qp->iowait)) {
1277 			dev->n_piowait++;
1278 			qp->s_flags |= QIB_S_WAIT_PIO;
1279 			list_add_tail(&qp->iowait, &dev->piowait);
1280 			dd = dd_from_dev(dev);
1281 			dd->f_wantpiobuf_intr(dd, 1);
1282 		}
1283 		spin_unlock(&dev->pending_lock);
1284 		qp->s_flags &= ~QIB_S_BUSY;
1285 		ret = -EBUSY;
1286 	}
1287 	spin_unlock_irqrestore(&qp->s_lock, flags);
1288 	return ret;
1289 }
1290 
1291 static int qib_verbs_send_pio(struct qib_qp *qp, struct qib_ib_header *ibhdr,
1292 			      u32 hdrwords, struct qib_sge_state *ss, u32 len,
1293 			      u32 plen, u32 dwords)
1294 {
1295 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1296 	struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
1297 	u32 *hdr = (u32 *) ibhdr;
1298 	u32 __iomem *piobuf_orig;
1299 	u32 __iomem *piobuf;
1300 	u64 pbc;
1301 	unsigned long flags;
1302 	unsigned flush_wc;
1303 	u32 control;
1304 	u32 pbufn;
1305 
1306 	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
1307 		be16_to_cpu(ibhdr->lrh[0]) >> 12);
1308 	pbc = ((u64) control << 32) | plen;
1309 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
1310 	if (unlikely(piobuf == NULL))
1311 		return no_bufs_available(qp);
1312 
1313 	/*
1314 	 * Write the pbc.
1315 	 * We have to flush after the PBC for correctness on some cpus
1316 	 * or WC buffer can be written out of order.
1317 	 */
1318 	writeq(pbc, piobuf);
1319 	piobuf_orig = piobuf;
1320 	piobuf += 2;
1321 
1322 	flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
1323 	if (len == 0) {
1324 		/*
1325 		 * If there is just the header portion, must flush before
1326 		 * writing last word of header for correctness, and after
1327 		 * the last header word (trigger word).
1328 		 */
1329 		if (flush_wc) {
1330 			qib_flush_wc();
1331 			qib_pio_copy(piobuf, hdr, hdrwords - 1);
1332 			qib_flush_wc();
1333 			__raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1334 			qib_flush_wc();
1335 		} else
1336 			qib_pio_copy(piobuf, hdr, hdrwords);
1337 		goto done;
1338 	}
1339 
1340 	if (flush_wc)
1341 		qib_flush_wc();
1342 	qib_pio_copy(piobuf, hdr, hdrwords);
1343 	piobuf += hdrwords;
1344 
1345 	/* The common case is aligned and contained in one segment. */
1346 	if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1347 		   !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1348 		u32 *addr = (u32 *) ss->sge.vaddr;
1349 
1350 		/* Update address before sending packet. */
1351 		update_sge(ss, len);
1352 		if (flush_wc) {
1353 			qib_pio_copy(piobuf, addr, dwords - 1);
1354 			/* must flush early everything before trigger word */
1355 			qib_flush_wc();
1356 			__raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1357 			/* be sure trigger word is written */
1358 			qib_flush_wc();
1359 		} else
1360 			qib_pio_copy(piobuf, addr, dwords);
1361 		goto done;
1362 	}
1363 	copy_io(piobuf, ss, len, flush_wc);
1364 done:
1365 	if (dd->flags & QIB_USE_SPCL_TRIG) {
1366 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
1367 
1368 		qib_flush_wc();
1369 		__raw_writel(0xaebecede, piobuf_orig + spcl_off);
1370 	}
1371 	qib_sendbuf_done(dd, pbufn);
1372 	if (qp->s_rdma_mr) {
1373 		qib_put_mr(qp->s_rdma_mr);
1374 		qp->s_rdma_mr = NULL;
1375 	}
1376 	if (qp->s_wqe) {
1377 		spin_lock_irqsave(&qp->s_lock, flags);
1378 		qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1379 		spin_unlock_irqrestore(&qp->s_lock, flags);
1380 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1381 		spin_lock_irqsave(&qp->s_lock, flags);
1382 		qib_rc_send_complete(qp, ibhdr);
1383 		spin_unlock_irqrestore(&qp->s_lock, flags);
1384 	}
1385 	return 0;
1386 }
1387 
1388 /**
1389  * qib_verbs_send - send a packet
1390  * @qp: the QP to send on
1391  * @hdr: the packet header
1392  * @hdrwords: the number of 32-bit words in the header
1393  * @ss: the SGE to send
1394  * @len: the length of the packet in bytes
1395  *
1396  * Return zero if packet is sent or queued OK.
1397  * Return non-zero and clear qp->s_flags QIB_S_BUSY otherwise.
1398  */
1399 int qib_verbs_send(struct qib_qp *qp, struct qib_ib_header *hdr,
1400 		   u32 hdrwords, struct qib_sge_state *ss, u32 len)
1401 {
1402 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1403 	u32 plen;
1404 	int ret;
1405 	u32 dwords = (len + 3) >> 2;
1406 
1407 	/*
1408 	 * Calculate the send buffer trigger address.
1409 	 * The +1 counts for the pbc control dword following the pbc length.
1410 	 */
1411 	plen = hdrwords + dwords + 1;
1412 
1413 	/*
1414 	 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1415 	 * can defer SDMA restart until link goes ACTIVE without
1416 	 * worrying about just how we got there.
1417 	 */
1418 	if (qp->ibqp.qp_type == IB_QPT_SMI ||
1419 	    !(dd->flags & QIB_HAS_SEND_DMA))
1420 		ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1421 					 plen, dwords);
1422 	else
1423 		ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1424 					 plen, dwords);
1425 
1426 	return ret;
1427 }
1428 
1429 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1430 			  u64 *rwords, u64 *spkts, u64 *rpkts,
1431 			  u64 *xmit_wait)
1432 {
1433 	int ret;
1434 	struct qib_devdata *dd = ppd->dd;
1435 
1436 	if (!(dd->flags & QIB_PRESENT)) {
1437 		/* no hardware, freeze, etc. */
1438 		ret = -EINVAL;
1439 		goto bail;
1440 	}
1441 	*swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1442 	*rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1443 	*spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1444 	*rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1445 	*xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1446 
1447 	ret = 0;
1448 
1449 bail:
1450 	return ret;
1451 }
1452 
1453 /**
1454  * qib_get_counters - get various chip counters
1455  * @dd: the qlogic_ib device
1456  * @cntrs: counters are placed here
1457  *
1458  * Return the counters needed by recv_pma_get_portcounters().
1459  */
1460 int qib_get_counters(struct qib_pportdata *ppd,
1461 		     struct qib_verbs_counters *cntrs)
1462 {
1463 	int ret;
1464 
1465 	if (!(ppd->dd->flags & QIB_PRESENT)) {
1466 		/* no hardware, freeze, etc. */
1467 		ret = -EINVAL;
1468 		goto bail;
1469 	}
1470 	cntrs->symbol_error_counter =
1471 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1472 	cntrs->link_error_recovery_counter =
1473 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1474 	/*
1475 	 * The link downed counter counts when the other side downs the
1476 	 * connection.  We add in the number of times we downed the link
1477 	 * due to local link integrity errors to compensate.
1478 	 */
1479 	cntrs->link_downed_counter =
1480 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1481 	cntrs->port_rcv_errors =
1482 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1483 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1484 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1485 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1486 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1487 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1488 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1489 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1490 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1491 	cntrs->port_rcv_errors +=
1492 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1493 	cntrs->port_rcv_errors +=
1494 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1495 	cntrs->port_rcv_remphys_errors =
1496 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1497 	cntrs->port_xmit_discards =
1498 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1499 	cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1500 			QIBPORTCNTR_WORDSEND);
1501 	cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1502 			QIBPORTCNTR_WORDRCV);
1503 	cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1504 			QIBPORTCNTR_PKTSEND);
1505 	cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1506 			QIBPORTCNTR_PKTRCV);
1507 	cntrs->local_link_integrity_errors =
1508 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1509 	cntrs->excessive_buffer_overrun_errors =
1510 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1511 	cntrs->vl15_dropped =
1512 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1513 
1514 	ret = 0;
1515 
1516 bail:
1517 	return ret;
1518 }
1519 
1520 /**
1521  * qib_ib_piobufavail - callback when a PIO buffer is available
1522  * @dd: the device pointer
1523  *
1524  * This is called from qib_intr() at interrupt level when a PIO buffer is
1525  * available after qib_verbs_send() returned an error that no buffers were
1526  * available. Disable the interrupt if there are no more QPs waiting.
1527  */
1528 void qib_ib_piobufavail(struct qib_devdata *dd)
1529 {
1530 	struct qib_ibdev *dev = &dd->verbs_dev;
1531 	struct list_head *list;
1532 	struct qib_qp *qps[5];
1533 	struct qib_qp *qp;
1534 	unsigned long flags;
1535 	unsigned i, n;
1536 
1537 	list = &dev->piowait;
1538 	n = 0;
1539 
1540 	/*
1541 	 * Note: checking that the piowait list is empty and clearing
1542 	 * the buffer available interrupt needs to be atomic or we
1543 	 * could end up with QPs on the wait list with the interrupt
1544 	 * disabled.
1545 	 */
1546 	spin_lock_irqsave(&dev->pending_lock, flags);
1547 	while (!list_empty(list)) {
1548 		if (n == ARRAY_SIZE(qps))
1549 			goto full;
1550 		qp = list_entry(list->next, struct qib_qp, iowait);
1551 		list_del_init(&qp->iowait);
1552 		atomic_inc(&qp->refcount);
1553 		qps[n++] = qp;
1554 	}
1555 	dd->f_wantpiobuf_intr(dd, 0);
1556 full:
1557 	spin_unlock_irqrestore(&dev->pending_lock, flags);
1558 
1559 	for (i = 0; i < n; i++) {
1560 		qp = qps[i];
1561 
1562 		spin_lock_irqsave(&qp->s_lock, flags);
1563 		if (qp->s_flags & QIB_S_WAIT_PIO) {
1564 			qp->s_flags &= ~QIB_S_WAIT_PIO;
1565 			qib_schedule_send(qp);
1566 		}
1567 		spin_unlock_irqrestore(&qp->s_lock, flags);
1568 
1569 		/* Notify qib_destroy_qp() if it is waiting. */
1570 		if (atomic_dec_and_test(&qp->refcount))
1571 			wake_up(&qp->wait);
1572 	}
1573 }
1574 
1575 static int qib_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
1576 			    struct ib_udata *uhw)
1577 {
1578 	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1579 	struct qib_ibdev *dev = to_idev(ibdev);
1580 
1581 	if (uhw->inlen || uhw->outlen)
1582 		return -EINVAL;
1583 	memset(props, 0, sizeof(*props));
1584 
1585 	props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1586 		IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1587 		IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1588 		IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1589 	props->page_size_cap = PAGE_SIZE;
1590 	props->vendor_id =
1591 		QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1592 	props->vendor_part_id = dd->deviceid;
1593 	props->hw_ver = dd->minrev;
1594 	props->sys_image_guid = ib_qib_sys_image_guid;
1595 	props->max_mr_size = ~0ULL;
1596 	props->max_qp = ib_qib_max_qps;
1597 	props->max_qp_wr = ib_qib_max_qp_wrs;
1598 	props->max_sge = ib_qib_max_sges;
1599 	props->max_sge_rd = ib_qib_max_sges;
1600 	props->max_cq = ib_qib_max_cqs;
1601 	props->max_ah = ib_qib_max_ahs;
1602 	props->max_cqe = ib_qib_max_cqes;
1603 	props->max_mr = dev->lk_table.max;
1604 	props->max_fmr = dev->lk_table.max;
1605 	props->max_map_per_fmr = 32767;
1606 	props->max_pd = ib_qib_max_pds;
1607 	props->max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1608 	props->max_qp_init_rd_atom = 255;
1609 	/* props->max_res_rd_atom */
1610 	props->max_srq = ib_qib_max_srqs;
1611 	props->max_srq_wr = ib_qib_max_srq_wrs;
1612 	props->max_srq_sge = ib_qib_max_srq_sges;
1613 	/* props->local_ca_ack_delay */
1614 	props->atomic_cap = IB_ATOMIC_GLOB;
1615 	props->max_pkeys = qib_get_npkeys(dd);
1616 	props->max_mcast_grp = ib_qib_max_mcast_grps;
1617 	props->max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1618 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1619 		props->max_mcast_grp;
1620 
1621 	return 0;
1622 }
1623 
1624 static int qib_query_port(struct ib_device *ibdev, u8 port,
1625 			  struct ib_port_attr *props)
1626 {
1627 	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1628 	struct qib_ibport *ibp = to_iport(ibdev, port);
1629 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1630 	enum ib_mtu mtu;
1631 	u16 lid = ppd->lid;
1632 
1633 	memset(props, 0, sizeof(*props));
1634 	props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1635 	props->lmc = ppd->lmc;
1636 	props->sm_lid = ibp->sm_lid;
1637 	props->sm_sl = ibp->sm_sl;
1638 	props->state = dd->f_iblink_state(ppd->lastibcstat);
1639 	props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1640 	props->port_cap_flags = ibp->port_cap_flags;
1641 	props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1642 	props->max_msg_sz = 0x80000000;
1643 	props->pkey_tbl_len = qib_get_npkeys(dd);
1644 	props->bad_pkey_cntr = ibp->pkey_violations;
1645 	props->qkey_viol_cntr = ibp->qkey_violations;
1646 	props->active_width = ppd->link_width_active;
1647 	/* See rate_show() */
1648 	props->active_speed = ppd->link_speed_active;
1649 	props->max_vl_num = qib_num_vls(ppd->vls_supported);
1650 	props->init_type_reply = 0;
1651 
1652 	props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1653 	switch (ppd->ibmtu) {
1654 	case 4096:
1655 		mtu = IB_MTU_4096;
1656 		break;
1657 	case 2048:
1658 		mtu = IB_MTU_2048;
1659 		break;
1660 	case 1024:
1661 		mtu = IB_MTU_1024;
1662 		break;
1663 	case 512:
1664 		mtu = IB_MTU_512;
1665 		break;
1666 	case 256:
1667 		mtu = IB_MTU_256;
1668 		break;
1669 	default:
1670 		mtu = IB_MTU_2048;
1671 	}
1672 	props->active_mtu = mtu;
1673 	props->subnet_timeout = ibp->subnet_timeout;
1674 
1675 	return 0;
1676 }
1677 
1678 static int qib_modify_device(struct ib_device *device,
1679 			     int device_modify_mask,
1680 			     struct ib_device_modify *device_modify)
1681 {
1682 	struct qib_devdata *dd = dd_from_ibdev(device);
1683 	unsigned i;
1684 	int ret;
1685 
1686 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1687 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1688 		ret = -EOPNOTSUPP;
1689 		goto bail;
1690 	}
1691 
1692 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1693 		memcpy(device->node_desc, device_modify->node_desc, 64);
1694 		for (i = 0; i < dd->num_pports; i++) {
1695 			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1696 
1697 			qib_node_desc_chg(ibp);
1698 		}
1699 	}
1700 
1701 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1702 		ib_qib_sys_image_guid =
1703 			cpu_to_be64(device_modify->sys_image_guid);
1704 		for (i = 0; i < dd->num_pports; i++) {
1705 			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1706 
1707 			qib_sys_guid_chg(ibp);
1708 		}
1709 	}
1710 
1711 	ret = 0;
1712 
1713 bail:
1714 	return ret;
1715 }
1716 
1717 static int qib_modify_port(struct ib_device *ibdev, u8 port,
1718 			   int port_modify_mask, struct ib_port_modify *props)
1719 {
1720 	struct qib_ibport *ibp = to_iport(ibdev, port);
1721 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1722 
1723 	ibp->port_cap_flags |= props->set_port_cap_mask;
1724 	ibp->port_cap_flags &= ~props->clr_port_cap_mask;
1725 	if (props->set_port_cap_mask || props->clr_port_cap_mask)
1726 		qib_cap_mask_chg(ibp);
1727 	if (port_modify_mask & IB_PORT_SHUTDOWN)
1728 		qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1729 	if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1730 		ibp->qkey_violations = 0;
1731 	return 0;
1732 }
1733 
1734 static int qib_query_gid(struct ib_device *ibdev, u8 port,
1735 			 int index, union ib_gid *gid)
1736 {
1737 	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1738 	int ret = 0;
1739 
1740 	if (!port || port > dd->num_pports)
1741 		ret = -EINVAL;
1742 	else {
1743 		struct qib_ibport *ibp = to_iport(ibdev, port);
1744 		struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1745 
1746 		gid->global.subnet_prefix = ibp->gid_prefix;
1747 		if (index == 0)
1748 			gid->global.interface_id = ppd->guid;
1749 		else if (index < QIB_GUIDS_PER_PORT)
1750 			gid->global.interface_id = ibp->guids[index - 1];
1751 		else
1752 			ret = -EINVAL;
1753 	}
1754 
1755 	return ret;
1756 }
1757 
1758 static struct ib_pd *qib_alloc_pd(struct ib_device *ibdev,
1759 				  struct ib_ucontext *context,
1760 				  struct ib_udata *udata)
1761 {
1762 	struct qib_ibdev *dev = to_idev(ibdev);
1763 	struct qib_pd *pd;
1764 	struct ib_pd *ret;
1765 
1766 	/*
1767 	 * This is actually totally arbitrary.  Some correctness tests
1768 	 * assume there's a maximum number of PDs that can be allocated.
1769 	 * We don't actually have this limit, but we fail the test if
1770 	 * we allow allocations of more than we report for this value.
1771 	 */
1772 
1773 	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1774 	if (!pd) {
1775 		ret = ERR_PTR(-ENOMEM);
1776 		goto bail;
1777 	}
1778 
1779 	spin_lock(&dev->n_pds_lock);
1780 	if (dev->n_pds_allocated == ib_qib_max_pds) {
1781 		spin_unlock(&dev->n_pds_lock);
1782 		kfree(pd);
1783 		ret = ERR_PTR(-ENOMEM);
1784 		goto bail;
1785 	}
1786 
1787 	dev->n_pds_allocated++;
1788 	spin_unlock(&dev->n_pds_lock);
1789 
1790 	/* ib_alloc_pd() will initialize pd->ibpd. */
1791 	pd->user = udata != NULL;
1792 
1793 	ret = &pd->ibpd;
1794 
1795 bail:
1796 	return ret;
1797 }
1798 
1799 static int qib_dealloc_pd(struct ib_pd *ibpd)
1800 {
1801 	struct qib_pd *pd = to_ipd(ibpd);
1802 	struct qib_ibdev *dev = to_idev(ibpd->device);
1803 
1804 	spin_lock(&dev->n_pds_lock);
1805 	dev->n_pds_allocated--;
1806 	spin_unlock(&dev->n_pds_lock);
1807 
1808 	kfree(pd);
1809 
1810 	return 0;
1811 }
1812 
1813 int qib_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
1814 {
1815 	/* A multicast address requires a GRH (see ch. 8.4.1). */
1816 	if (ah_attr->dlid >= QIB_MULTICAST_LID_BASE &&
1817 	    ah_attr->dlid != QIB_PERMISSIVE_LID &&
1818 	    !(ah_attr->ah_flags & IB_AH_GRH))
1819 		goto bail;
1820 	if ((ah_attr->ah_flags & IB_AH_GRH) &&
1821 	    ah_attr->grh.sgid_index >= QIB_GUIDS_PER_PORT)
1822 		goto bail;
1823 	if (ah_attr->dlid == 0)
1824 		goto bail;
1825 	if (ah_attr->port_num < 1 ||
1826 	    ah_attr->port_num > ibdev->phys_port_cnt)
1827 		goto bail;
1828 	if (ah_attr->static_rate != IB_RATE_PORT_CURRENT &&
1829 	    ib_rate_to_mult(ah_attr->static_rate) < 0)
1830 		goto bail;
1831 	if (ah_attr->sl > 15)
1832 		goto bail;
1833 	return 0;
1834 bail:
1835 	return -EINVAL;
1836 }
1837 
1838 /**
1839  * qib_create_ah - create an address handle
1840  * @pd: the protection domain
1841  * @ah_attr: the attributes of the AH
1842  *
1843  * This may be called from interrupt context.
1844  */
1845 static struct ib_ah *qib_create_ah(struct ib_pd *pd,
1846 				   struct ib_ah_attr *ah_attr)
1847 {
1848 	struct qib_ah *ah;
1849 	struct ib_ah *ret;
1850 	struct qib_ibdev *dev = to_idev(pd->device);
1851 	unsigned long flags;
1852 
1853 	if (qib_check_ah(pd->device, ah_attr)) {
1854 		ret = ERR_PTR(-EINVAL);
1855 		goto bail;
1856 	}
1857 
1858 	ah = kmalloc(sizeof(*ah), GFP_ATOMIC);
1859 	if (!ah) {
1860 		ret = ERR_PTR(-ENOMEM);
1861 		goto bail;
1862 	}
1863 
1864 	spin_lock_irqsave(&dev->n_ahs_lock, flags);
1865 	if (dev->n_ahs_allocated == ib_qib_max_ahs) {
1866 		spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1867 		kfree(ah);
1868 		ret = ERR_PTR(-ENOMEM);
1869 		goto bail;
1870 	}
1871 
1872 	dev->n_ahs_allocated++;
1873 	spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1874 
1875 	/* ib_create_ah() will initialize ah->ibah. */
1876 	ah->attr = *ah_attr;
1877 	atomic_set(&ah->refcount, 0);
1878 
1879 	ret = &ah->ibah;
1880 
1881 bail:
1882 	return ret;
1883 }
1884 
1885 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1886 {
1887 	struct ib_ah_attr attr;
1888 	struct ib_ah *ah = ERR_PTR(-EINVAL);
1889 	struct qib_qp *qp0;
1890 
1891 	memset(&attr, 0, sizeof(attr));
1892 	attr.dlid = dlid;
1893 	attr.port_num = ppd_from_ibp(ibp)->port;
1894 	rcu_read_lock();
1895 	qp0 = rcu_dereference(ibp->qp0);
1896 	if (qp0)
1897 		ah = ib_create_ah(qp0->ibqp.pd, &attr);
1898 	rcu_read_unlock();
1899 	return ah;
1900 }
1901 
1902 /**
1903  * qib_destroy_ah - destroy an address handle
1904  * @ibah: the AH to destroy
1905  *
1906  * This may be called from interrupt context.
1907  */
1908 static int qib_destroy_ah(struct ib_ah *ibah)
1909 {
1910 	struct qib_ibdev *dev = to_idev(ibah->device);
1911 	struct qib_ah *ah = to_iah(ibah);
1912 	unsigned long flags;
1913 
1914 	if (atomic_read(&ah->refcount) != 0)
1915 		return -EBUSY;
1916 
1917 	spin_lock_irqsave(&dev->n_ahs_lock, flags);
1918 	dev->n_ahs_allocated--;
1919 	spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1920 
1921 	kfree(ah);
1922 
1923 	return 0;
1924 }
1925 
1926 static int qib_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1927 {
1928 	struct qib_ah *ah = to_iah(ibah);
1929 
1930 	if (qib_check_ah(ibah->device, ah_attr))
1931 		return -EINVAL;
1932 
1933 	ah->attr = *ah_attr;
1934 
1935 	return 0;
1936 }
1937 
1938 static int qib_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1939 {
1940 	struct qib_ah *ah = to_iah(ibah);
1941 
1942 	*ah_attr = ah->attr;
1943 
1944 	return 0;
1945 }
1946 
1947 /**
1948  * qib_get_npkeys - return the size of the PKEY table for context 0
1949  * @dd: the qlogic_ib device
1950  */
1951 unsigned qib_get_npkeys(struct qib_devdata *dd)
1952 {
1953 	return ARRAY_SIZE(dd->rcd[0]->pkeys);
1954 }
1955 
1956 /*
1957  * Return the indexed PKEY from the port PKEY table.
1958  * No need to validate rcd[ctxt]; the port is setup if we are here.
1959  */
1960 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1961 {
1962 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1963 	struct qib_devdata *dd = ppd->dd;
1964 	unsigned ctxt = ppd->hw_pidx;
1965 	unsigned ret;
1966 
1967 	/* dd->rcd null if mini_init or some init failures */
1968 	if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1969 		ret = 0;
1970 	else
1971 		ret = dd->rcd[ctxt]->pkeys[index];
1972 
1973 	return ret;
1974 }
1975 
1976 static int qib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1977 			  u16 *pkey)
1978 {
1979 	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1980 	int ret;
1981 
1982 	if (index >= qib_get_npkeys(dd)) {
1983 		ret = -EINVAL;
1984 		goto bail;
1985 	}
1986 
1987 	*pkey = qib_get_pkey(to_iport(ibdev, port), index);
1988 	ret = 0;
1989 
1990 bail:
1991 	return ret;
1992 }
1993 
1994 /**
1995  * qib_alloc_ucontext - allocate a ucontest
1996  * @ibdev: the infiniband device
1997  * @udata: not used by the QLogic_IB driver
1998  */
1999 
2000 static struct ib_ucontext *qib_alloc_ucontext(struct ib_device *ibdev,
2001 					      struct ib_udata *udata)
2002 {
2003 	struct qib_ucontext *context;
2004 	struct ib_ucontext *ret;
2005 
2006 	context = kmalloc(sizeof(*context), GFP_KERNEL);
2007 	if (!context) {
2008 		ret = ERR_PTR(-ENOMEM);
2009 		goto bail;
2010 	}
2011 
2012 	ret = &context->ibucontext;
2013 
2014 bail:
2015 	return ret;
2016 }
2017 
2018 static int qib_dealloc_ucontext(struct ib_ucontext *context)
2019 {
2020 	kfree(to_iucontext(context));
2021 	return 0;
2022 }
2023 
2024 static void init_ibport(struct qib_pportdata *ppd)
2025 {
2026 	struct qib_verbs_counters cntrs;
2027 	struct qib_ibport *ibp = &ppd->ibport_data;
2028 
2029 	spin_lock_init(&ibp->lock);
2030 	/* Set the prefix to the default value (see ch. 4.1.1) */
2031 	ibp->gid_prefix = IB_DEFAULT_GID_PREFIX;
2032 	ibp->sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
2033 	ibp->port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
2034 		IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
2035 		IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
2036 		IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
2037 		IB_PORT_OTHER_LOCAL_CHANGES_SUP;
2038 	if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
2039 		ibp->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2040 	ibp->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2041 	ibp->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2042 	ibp->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2043 	ibp->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2044 	ibp->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2045 
2046 	/* Snapshot current HW counters to "clear" them. */
2047 	qib_get_counters(ppd, &cntrs);
2048 	ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
2049 	ibp->z_link_error_recovery_counter =
2050 		cntrs.link_error_recovery_counter;
2051 	ibp->z_link_downed_counter = cntrs.link_downed_counter;
2052 	ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
2053 	ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
2054 	ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
2055 	ibp->z_port_xmit_data = cntrs.port_xmit_data;
2056 	ibp->z_port_rcv_data = cntrs.port_rcv_data;
2057 	ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
2058 	ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
2059 	ibp->z_local_link_integrity_errors =
2060 		cntrs.local_link_integrity_errors;
2061 	ibp->z_excessive_buffer_overrun_errors =
2062 		cntrs.excessive_buffer_overrun_errors;
2063 	ibp->z_vl15_dropped = cntrs.vl15_dropped;
2064 	RCU_INIT_POINTER(ibp->qp0, NULL);
2065 	RCU_INIT_POINTER(ibp->qp1, NULL);
2066 }
2067 
2068 static int qib_port_immutable(struct ib_device *ibdev, u8 port_num,
2069 			      struct ib_port_immutable *immutable)
2070 {
2071 	struct ib_port_attr attr;
2072 	int err;
2073 
2074 	err = qib_query_port(ibdev, port_num, &attr);
2075 	if (err)
2076 		return err;
2077 
2078 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
2079 	immutable->gid_tbl_len = attr.gid_tbl_len;
2080 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
2081 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2082 
2083 	return 0;
2084 }
2085 
2086 /**
2087  * qib_register_ib_device - register our device with the infiniband core
2088  * @dd: the device data structure
2089  * Return the allocated qib_ibdev pointer or NULL on error.
2090  */
2091 int qib_register_ib_device(struct qib_devdata *dd)
2092 {
2093 	struct qib_ibdev *dev = &dd->verbs_dev;
2094 	struct ib_device *ibdev = &dev->ibdev;
2095 	struct qib_pportdata *ppd = dd->pport;
2096 	unsigned i, lk_tab_size;
2097 	int ret;
2098 
2099 	dev->qp_table_size = ib_qib_qp_table_size;
2100 	get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
2101 	dev->qp_table = kmalloc_array(
2102 				dev->qp_table_size,
2103 				sizeof(*dev->qp_table),
2104 				GFP_KERNEL);
2105 	if (!dev->qp_table) {
2106 		ret = -ENOMEM;
2107 		goto err_qpt;
2108 	}
2109 	for (i = 0; i < dev->qp_table_size; i++)
2110 		RCU_INIT_POINTER(dev->qp_table[i], NULL);
2111 
2112 	for (i = 0; i < dd->num_pports; i++)
2113 		init_ibport(ppd + i);
2114 
2115 	/* Only need to initialize non-zero fields. */
2116 	spin_lock_init(&dev->qpt_lock);
2117 	spin_lock_init(&dev->n_pds_lock);
2118 	spin_lock_init(&dev->n_ahs_lock);
2119 	spin_lock_init(&dev->n_cqs_lock);
2120 	spin_lock_init(&dev->n_qps_lock);
2121 	spin_lock_init(&dev->n_srqs_lock);
2122 	spin_lock_init(&dev->n_mcast_grps_lock);
2123 	init_timer(&dev->mem_timer);
2124 	dev->mem_timer.function = mem_timer;
2125 	dev->mem_timer.data = (unsigned long) dev;
2126 
2127 	qib_init_qpn_table(dd, &dev->qpn_table);
2128 
2129 	/*
2130 	 * The top ib_qib_lkey_table_size bits are used to index the
2131 	 * table.  The lower 8 bits can be owned by the user (copied from
2132 	 * the LKEY).  The remaining bits act as a generation number or tag.
2133 	 */
2134 	spin_lock_init(&dev->lk_table.lock);
2135 	/* insure generation is at least 4 bits see keys.c */
2136 	if (ib_qib_lkey_table_size > MAX_LKEY_TABLE_BITS) {
2137 		qib_dev_warn(dd, "lkey bits %u too large, reduced to %u\n",
2138 			ib_qib_lkey_table_size, MAX_LKEY_TABLE_BITS);
2139 		ib_qib_lkey_table_size = MAX_LKEY_TABLE_BITS;
2140 	}
2141 	dev->lk_table.max = 1 << ib_qib_lkey_table_size;
2142 	lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
2143 	dev->lk_table.table = (struct qib_mregion __rcu **)
2144 		vmalloc(lk_tab_size);
2145 	if (dev->lk_table.table == NULL) {
2146 		ret = -ENOMEM;
2147 		goto err_lk;
2148 	}
2149 	RCU_INIT_POINTER(dev->dma_mr, NULL);
2150 	for (i = 0; i < dev->lk_table.max; i++)
2151 		RCU_INIT_POINTER(dev->lk_table.table[i], NULL);
2152 	INIT_LIST_HEAD(&dev->pending_mmaps);
2153 	spin_lock_init(&dev->pending_lock);
2154 	dev->mmap_offset = PAGE_SIZE;
2155 	spin_lock_init(&dev->mmap_offset_lock);
2156 	INIT_LIST_HEAD(&dev->piowait);
2157 	INIT_LIST_HEAD(&dev->dmawait);
2158 	INIT_LIST_HEAD(&dev->txwait);
2159 	INIT_LIST_HEAD(&dev->memwait);
2160 	INIT_LIST_HEAD(&dev->txreq_free);
2161 
2162 	if (ppd->sdma_descq_cnt) {
2163 		dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
2164 						ppd->sdma_descq_cnt *
2165 						sizeof(struct qib_pio_header),
2166 						&dev->pio_hdrs_phys,
2167 						GFP_KERNEL);
2168 		if (!dev->pio_hdrs) {
2169 			ret = -ENOMEM;
2170 			goto err_hdrs;
2171 		}
2172 	}
2173 
2174 	for (i = 0; i < ppd->sdma_descq_cnt; i++) {
2175 		struct qib_verbs_txreq *tx;
2176 
2177 		tx = kzalloc(sizeof(*tx), GFP_KERNEL);
2178 		if (!tx) {
2179 			ret = -ENOMEM;
2180 			goto err_tx;
2181 		}
2182 		tx->hdr_inx = i;
2183 		list_add(&tx->txreq.list, &dev->txreq_free);
2184 	}
2185 
2186 	/*
2187 	 * The system image GUID is supposed to be the same for all
2188 	 * IB HCAs in a single system but since there can be other
2189 	 * device types in the system, we can't be sure this is unique.
2190 	 */
2191 	if (!ib_qib_sys_image_guid)
2192 		ib_qib_sys_image_guid = ppd->guid;
2193 
2194 	strlcpy(ibdev->name, "qib%d", IB_DEVICE_NAME_MAX);
2195 	ibdev->owner = THIS_MODULE;
2196 	ibdev->node_guid = ppd->guid;
2197 	ibdev->uverbs_abi_ver = QIB_UVERBS_ABI_VERSION;
2198 	ibdev->uverbs_cmd_mask =
2199 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2200 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2201 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2202 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2203 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2204 		(1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
2205 		(1ull << IB_USER_VERBS_CMD_MODIFY_AH)           |
2206 		(1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
2207 		(1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
2208 		(1ull << IB_USER_VERBS_CMD_REG_MR)              |
2209 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2210 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2211 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2212 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2213 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2214 		(1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
2215 		(1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
2216 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2217 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2218 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2219 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2220 		(1ull << IB_USER_VERBS_CMD_POST_SEND)           |
2221 		(1ull << IB_USER_VERBS_CMD_POST_RECV)           |
2222 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2223 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2224 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2225 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2226 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2227 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2228 		(1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2229 	ibdev->node_type = RDMA_NODE_IB_CA;
2230 	ibdev->phys_port_cnt = dd->num_pports;
2231 	ibdev->num_comp_vectors = 1;
2232 	ibdev->dma_device = &dd->pcidev->dev;
2233 	ibdev->query_device = qib_query_device;
2234 	ibdev->modify_device = qib_modify_device;
2235 	ibdev->query_port = qib_query_port;
2236 	ibdev->modify_port = qib_modify_port;
2237 	ibdev->query_pkey = qib_query_pkey;
2238 	ibdev->query_gid = qib_query_gid;
2239 	ibdev->alloc_ucontext = qib_alloc_ucontext;
2240 	ibdev->dealloc_ucontext = qib_dealloc_ucontext;
2241 	ibdev->alloc_pd = qib_alloc_pd;
2242 	ibdev->dealloc_pd = qib_dealloc_pd;
2243 	ibdev->create_ah = qib_create_ah;
2244 	ibdev->destroy_ah = qib_destroy_ah;
2245 	ibdev->modify_ah = qib_modify_ah;
2246 	ibdev->query_ah = qib_query_ah;
2247 	ibdev->create_srq = qib_create_srq;
2248 	ibdev->modify_srq = qib_modify_srq;
2249 	ibdev->query_srq = qib_query_srq;
2250 	ibdev->destroy_srq = qib_destroy_srq;
2251 	ibdev->create_qp = qib_create_qp;
2252 	ibdev->modify_qp = qib_modify_qp;
2253 	ibdev->query_qp = qib_query_qp;
2254 	ibdev->destroy_qp = qib_destroy_qp;
2255 	ibdev->post_send = qib_post_send;
2256 	ibdev->post_recv = qib_post_receive;
2257 	ibdev->post_srq_recv = qib_post_srq_receive;
2258 	ibdev->create_cq = qib_create_cq;
2259 	ibdev->destroy_cq = qib_destroy_cq;
2260 	ibdev->resize_cq = qib_resize_cq;
2261 	ibdev->poll_cq = qib_poll_cq;
2262 	ibdev->req_notify_cq = qib_req_notify_cq;
2263 	ibdev->get_dma_mr = qib_get_dma_mr;
2264 	ibdev->reg_user_mr = qib_reg_user_mr;
2265 	ibdev->dereg_mr = qib_dereg_mr;
2266 	ibdev->alloc_mr = qib_alloc_mr;
2267 	ibdev->map_mr_sg = qib_map_mr_sg;
2268 	ibdev->alloc_fmr = qib_alloc_fmr;
2269 	ibdev->map_phys_fmr = qib_map_phys_fmr;
2270 	ibdev->unmap_fmr = qib_unmap_fmr;
2271 	ibdev->dealloc_fmr = qib_dealloc_fmr;
2272 	ibdev->attach_mcast = qib_multicast_attach;
2273 	ibdev->detach_mcast = qib_multicast_detach;
2274 	ibdev->process_mad = qib_process_mad;
2275 	ibdev->mmap = qib_mmap;
2276 	ibdev->dma_ops = &qib_dma_mapping_ops;
2277 	ibdev->get_port_immutable = qib_port_immutable;
2278 
2279 	snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
2280 		 "Intel Infiniband HCA %s", init_utsname()->nodename);
2281 
2282 	ret = ib_register_device(ibdev, qib_create_port_files);
2283 	if (ret)
2284 		goto err_reg;
2285 
2286 	ret = qib_create_agents(dev);
2287 	if (ret)
2288 		goto err_agents;
2289 
2290 	ret = qib_verbs_register_sysfs(dd);
2291 	if (ret)
2292 		goto err_class;
2293 
2294 	goto bail;
2295 
2296 err_class:
2297 	qib_free_agents(dev);
2298 err_agents:
2299 	ib_unregister_device(ibdev);
2300 err_reg:
2301 err_tx:
2302 	while (!list_empty(&dev->txreq_free)) {
2303 		struct list_head *l = dev->txreq_free.next;
2304 		struct qib_verbs_txreq *tx;
2305 
2306 		list_del(l);
2307 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
2308 		kfree(tx);
2309 	}
2310 	if (ppd->sdma_descq_cnt)
2311 		dma_free_coherent(&dd->pcidev->dev,
2312 				  ppd->sdma_descq_cnt *
2313 					sizeof(struct qib_pio_header),
2314 				  dev->pio_hdrs, dev->pio_hdrs_phys);
2315 err_hdrs:
2316 	vfree(dev->lk_table.table);
2317 err_lk:
2318 	kfree(dev->qp_table);
2319 err_qpt:
2320 	qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2321 bail:
2322 	return ret;
2323 }
2324 
2325 void qib_unregister_ib_device(struct qib_devdata *dd)
2326 {
2327 	struct qib_ibdev *dev = &dd->verbs_dev;
2328 	struct ib_device *ibdev = &dev->ibdev;
2329 	u32 qps_inuse;
2330 	unsigned lk_tab_size;
2331 
2332 	qib_verbs_unregister_sysfs(dd);
2333 
2334 	qib_free_agents(dev);
2335 
2336 	ib_unregister_device(ibdev);
2337 
2338 	if (!list_empty(&dev->piowait))
2339 		qib_dev_err(dd, "piowait list not empty!\n");
2340 	if (!list_empty(&dev->dmawait))
2341 		qib_dev_err(dd, "dmawait list not empty!\n");
2342 	if (!list_empty(&dev->txwait))
2343 		qib_dev_err(dd, "txwait list not empty!\n");
2344 	if (!list_empty(&dev->memwait))
2345 		qib_dev_err(dd, "memwait list not empty!\n");
2346 	if (dev->dma_mr)
2347 		qib_dev_err(dd, "DMA MR not NULL!\n");
2348 
2349 	qps_inuse = qib_free_all_qps(dd);
2350 	if (qps_inuse)
2351 		qib_dev_err(dd, "QP memory leak! %u still in use\n",
2352 			    qps_inuse);
2353 
2354 	del_timer_sync(&dev->mem_timer);
2355 	qib_free_qpn_table(&dev->qpn_table);
2356 	while (!list_empty(&dev->txreq_free)) {
2357 		struct list_head *l = dev->txreq_free.next;
2358 		struct qib_verbs_txreq *tx;
2359 
2360 		list_del(l);
2361 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
2362 		kfree(tx);
2363 	}
2364 	if (dd->pport->sdma_descq_cnt)
2365 		dma_free_coherent(&dd->pcidev->dev,
2366 				  dd->pport->sdma_descq_cnt *
2367 					sizeof(struct qib_pio_header),
2368 				  dev->pio_hdrs, dev->pio_hdrs_phys);
2369 	lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table);
2370 	vfree(dev->lk_table.table);
2371 	kfree(dev->qp_table);
2372 }
2373 
2374 /*
2375  * This must be called with s_lock held.
2376  */
2377 void qib_schedule_send(struct qib_qp *qp)
2378 {
2379 	if (qib_send_ok(qp)) {
2380 		struct qib_ibport *ibp =
2381 			to_iport(qp->ibqp.device, qp->port_num);
2382 		struct qib_pportdata *ppd = ppd_from_ibp(ibp);
2383 
2384 		queue_work(ppd->qib_wq, &qp->s_work);
2385 	}
2386 }
2387