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 #include <rdma/rdma_vt.h>
45 
46 #include "qib.h"
47 #include "qib_common.h"
48 
49 static unsigned int ib_qib_qp_table_size = 256;
50 module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO);
51 MODULE_PARM_DESC(qp_table_size, "QP table size");
52 
53 static unsigned int qib_lkey_table_size = 16;
54 module_param_named(lkey_table_size, qib_lkey_table_size, uint,
55 		   S_IRUGO);
56 MODULE_PARM_DESC(lkey_table_size,
57 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
58 
59 static unsigned int ib_qib_max_pds = 0xFFFF;
60 module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO);
61 MODULE_PARM_DESC(max_pds,
62 		 "Maximum number of protection domains to support");
63 
64 static unsigned int ib_qib_max_ahs = 0xFFFF;
65 module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO);
66 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
67 
68 unsigned int ib_qib_max_cqes = 0x2FFFF;
69 module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO);
70 MODULE_PARM_DESC(max_cqes,
71 		 "Maximum number of completion queue entries to support");
72 
73 unsigned int ib_qib_max_cqs = 0x1FFFF;
74 module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO);
75 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
76 
77 unsigned int ib_qib_max_qp_wrs = 0x3FFF;
78 module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO);
79 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
80 
81 unsigned int ib_qib_max_qps = 16384;
82 module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO);
83 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
84 
85 unsigned int ib_qib_max_sges = 0x60;
86 module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO);
87 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
88 
89 unsigned int ib_qib_max_mcast_grps = 16384;
90 module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO);
91 MODULE_PARM_DESC(max_mcast_grps,
92 		 "Maximum number of multicast groups to support");
93 
94 unsigned int ib_qib_max_mcast_qp_attached = 16;
95 module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached,
96 		   uint, S_IRUGO);
97 MODULE_PARM_DESC(max_mcast_qp_attached,
98 		 "Maximum number of attached QPs to support");
99 
100 unsigned int ib_qib_max_srqs = 1024;
101 module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO);
102 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
103 
104 unsigned int ib_qib_max_srq_sges = 128;
105 module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO);
106 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
107 
108 unsigned int ib_qib_max_srq_wrs = 0x1FFFF;
109 module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO);
110 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
111 
112 static unsigned int ib_qib_disable_sma;
113 module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO);
114 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
115 
116 /*
117  * Translate ib_wr_opcode into ib_wc_opcode.
118  */
119 const enum ib_wc_opcode ib_qib_wc_opcode[] = {
120 	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
121 	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
122 	[IB_WR_SEND] = IB_WC_SEND,
123 	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
124 	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
125 	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
126 	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
127 };
128 
129 /*
130  * System image GUID.
131  */
132 __be64 ib_qib_sys_image_guid;
133 
134 /**
135  * qib_copy_sge - copy data to SGE memory
136  * @ss: the SGE state
137  * @data: the data to copy
138  * @length: the length of the data
139  */
140 void qib_copy_sge(struct rvt_sge_state *ss, void *data, u32 length, int release)
141 {
142 	struct rvt_sge *sge = &ss->sge;
143 
144 	while (length) {
145 		u32 len = rvt_get_sge_length(sge, length);
146 
147 		WARN_ON_ONCE(len == 0);
148 		memcpy(sge->vaddr, data, len);
149 		rvt_update_sge(ss, len, release);
150 		data += len;
151 		length -= len;
152 	}
153 }
154 
155 /*
156  * Count the number of DMA descriptors needed to send length bytes of data.
157  * Don't modify the qib_sge_state to get the count.
158  * Return zero if any of the segments is not aligned.
159  */
160 static u32 qib_count_sge(struct rvt_sge_state *ss, u32 length)
161 {
162 	struct rvt_sge *sg_list = ss->sg_list;
163 	struct rvt_sge sge = ss->sge;
164 	u8 num_sge = ss->num_sge;
165 	u32 ndesc = 1;  /* count the header */
166 
167 	while (length) {
168 		u32 len = sge.length;
169 
170 		if (len > length)
171 			len = length;
172 		if (len > sge.sge_length)
173 			len = sge.sge_length;
174 		BUG_ON(len == 0);
175 		if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
176 		    (len != length && (len & (sizeof(u32) - 1)))) {
177 			ndesc = 0;
178 			break;
179 		}
180 		ndesc++;
181 		sge.vaddr += len;
182 		sge.length -= len;
183 		sge.sge_length -= len;
184 		if (sge.sge_length == 0) {
185 			if (--num_sge)
186 				sge = *sg_list++;
187 		} else if (sge.length == 0 && sge.mr->lkey) {
188 			if (++sge.n >= RVT_SEGSZ) {
189 				if (++sge.m >= sge.mr->mapsz)
190 					break;
191 				sge.n = 0;
192 			}
193 			sge.vaddr =
194 				sge.mr->map[sge.m]->segs[sge.n].vaddr;
195 			sge.length =
196 				sge.mr->map[sge.m]->segs[sge.n].length;
197 		}
198 		length -= len;
199 	}
200 	return ndesc;
201 }
202 
203 /*
204  * Copy from the SGEs to the data buffer.
205  */
206 static void qib_copy_from_sge(void *data, struct rvt_sge_state *ss, u32 length)
207 {
208 	struct rvt_sge *sge = &ss->sge;
209 
210 	while (length) {
211 		u32 len = sge->length;
212 
213 		if (len > length)
214 			len = length;
215 		if (len > sge->sge_length)
216 			len = sge->sge_length;
217 		BUG_ON(len == 0);
218 		memcpy(data, sge->vaddr, len);
219 		sge->vaddr += len;
220 		sge->length -= len;
221 		sge->sge_length -= len;
222 		if (sge->sge_length == 0) {
223 			if (--ss->num_sge)
224 				*sge = *ss->sg_list++;
225 		} else if (sge->length == 0 && sge->mr->lkey) {
226 			if (++sge->n >= RVT_SEGSZ) {
227 				if (++sge->m >= sge->mr->mapsz)
228 					break;
229 				sge->n = 0;
230 			}
231 			sge->vaddr =
232 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
233 			sge->length =
234 				sge->mr->map[sge->m]->segs[sge->n].length;
235 		}
236 		data += len;
237 		length -= len;
238 	}
239 }
240 
241 /**
242  * qib_qp_rcv - processing an incoming packet on a QP
243  * @rcd: the context pointer
244  * @hdr: the packet header
245  * @has_grh: true if the packet has a GRH
246  * @data: the packet data
247  * @tlen: the packet length
248  * @qp: the QP the packet came on
249  *
250  * This is called from qib_ib_rcv() to process an incoming packet
251  * for the given QP.
252  * Called at interrupt level.
253  */
254 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
255 		       int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
256 {
257 	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
258 
259 	spin_lock(&qp->r_lock);
260 
261 	/* Check for valid receive state. */
262 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
263 		ibp->rvp.n_pkt_drops++;
264 		goto unlock;
265 	}
266 
267 	switch (qp->ibqp.qp_type) {
268 	case IB_QPT_SMI:
269 	case IB_QPT_GSI:
270 		if (ib_qib_disable_sma)
271 			break;
272 		/* FALLTHROUGH */
273 	case IB_QPT_UD:
274 		qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
275 		break;
276 
277 	case IB_QPT_RC:
278 		qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
279 		break;
280 
281 	case IB_QPT_UC:
282 		qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
283 		break;
284 
285 	default:
286 		break;
287 	}
288 
289 unlock:
290 	spin_unlock(&qp->r_lock);
291 }
292 
293 /**
294  * qib_ib_rcv - process an incoming packet
295  * @rcd: the context pointer
296  * @rhdr: the header of the packet
297  * @data: the packet payload
298  * @tlen: the packet length
299  *
300  * This is called from qib_kreceive() to process an incoming packet at
301  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
302  */
303 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
304 {
305 	struct qib_pportdata *ppd = rcd->ppd;
306 	struct qib_ibport *ibp = &ppd->ibport_data;
307 	struct ib_header *hdr = rhdr;
308 	struct qib_devdata *dd = ppd->dd;
309 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
310 	struct ib_other_headers *ohdr;
311 	struct rvt_qp *qp;
312 	u32 qp_num;
313 	int lnh;
314 	u8 opcode;
315 	u16 lid;
316 
317 	/* 24 == LRH+BTH+CRC */
318 	if (unlikely(tlen < 24))
319 		goto drop;
320 
321 	/* Check for a valid destination LID (see ch. 7.11.1). */
322 	lid = be16_to_cpu(hdr->lrh[1]);
323 	if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
324 		lid &= ~((1 << ppd->lmc) - 1);
325 		if (unlikely(lid != ppd->lid))
326 			goto drop;
327 	}
328 
329 	/* Check for GRH */
330 	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
331 	if (lnh == QIB_LRH_BTH)
332 		ohdr = &hdr->u.oth;
333 	else if (lnh == QIB_LRH_GRH) {
334 		u32 vtf;
335 
336 		ohdr = &hdr->u.l.oth;
337 		if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
338 			goto drop;
339 		vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
340 		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
341 			goto drop;
342 	} else
343 		goto drop;
344 
345 	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
346 #ifdef CONFIG_DEBUG_FS
347 	rcd->opstats->stats[opcode].n_bytes += tlen;
348 	rcd->opstats->stats[opcode].n_packets++;
349 #endif
350 
351 	/* Get the destination QP number. */
352 	qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
353 	if (qp_num == QIB_MULTICAST_QPN) {
354 		struct rvt_mcast *mcast;
355 		struct rvt_mcast_qp *p;
356 
357 		if (lnh != QIB_LRH_GRH)
358 			goto drop;
359 		mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid, lid);
360 		if (mcast == NULL)
361 			goto drop;
362 		this_cpu_inc(ibp->pmastats->n_multicast_rcv);
363 		list_for_each_entry_rcu(p, &mcast->qp_list, list)
364 			qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
365 		/*
366 		 * Notify rvt_multicast_detach() if it is waiting for us
367 		 * to finish.
368 		 */
369 		if (atomic_dec_return(&mcast->refcount) <= 1)
370 			wake_up(&mcast->wait);
371 	} else {
372 		rcu_read_lock();
373 		qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
374 		if (!qp) {
375 			rcu_read_unlock();
376 			goto drop;
377 		}
378 		this_cpu_inc(ibp->pmastats->n_unicast_rcv);
379 		qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
380 		rcu_read_unlock();
381 	}
382 	return;
383 
384 drop:
385 	ibp->rvp.n_pkt_drops++;
386 }
387 
388 /*
389  * This is called from a timer to check for QPs
390  * which need kernel memory in order to send a packet.
391  */
392 static void mem_timer(struct timer_list *t)
393 {
394 	struct qib_ibdev *dev = from_timer(dev, t, mem_timer);
395 	struct list_head *list = &dev->memwait;
396 	struct rvt_qp *qp = NULL;
397 	struct qib_qp_priv *priv = NULL;
398 	unsigned long flags;
399 
400 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
401 	if (!list_empty(list)) {
402 		priv = list_entry(list->next, struct qib_qp_priv, iowait);
403 		qp = priv->owner;
404 		list_del_init(&priv->iowait);
405 		rvt_get_qp(qp);
406 		if (!list_empty(list))
407 			mod_timer(&dev->mem_timer, jiffies + 1);
408 	}
409 	spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
410 
411 	if (qp) {
412 		spin_lock_irqsave(&qp->s_lock, flags);
413 		if (qp->s_flags & RVT_S_WAIT_KMEM) {
414 			qp->s_flags &= ~RVT_S_WAIT_KMEM;
415 			qib_schedule_send(qp);
416 		}
417 		spin_unlock_irqrestore(&qp->s_lock, flags);
418 		rvt_put_qp(qp);
419 	}
420 }
421 
422 #ifdef __LITTLE_ENDIAN
423 static inline u32 get_upper_bits(u32 data, u32 shift)
424 {
425 	return data >> shift;
426 }
427 
428 static inline u32 set_upper_bits(u32 data, u32 shift)
429 {
430 	return data << shift;
431 }
432 
433 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
434 {
435 	data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
436 	data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
437 	return data;
438 }
439 #else
440 static inline u32 get_upper_bits(u32 data, u32 shift)
441 {
442 	return data << shift;
443 }
444 
445 static inline u32 set_upper_bits(u32 data, u32 shift)
446 {
447 	return data >> shift;
448 }
449 
450 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
451 {
452 	data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
453 	data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
454 	return data;
455 }
456 #endif
457 
458 static void copy_io(u32 __iomem *piobuf, struct rvt_sge_state *ss,
459 		    u32 length, unsigned flush_wc)
460 {
461 	u32 extra = 0;
462 	u32 data = 0;
463 	u32 last;
464 
465 	while (1) {
466 		u32 len = ss->sge.length;
467 		u32 off;
468 
469 		if (len > length)
470 			len = length;
471 		if (len > ss->sge.sge_length)
472 			len = ss->sge.sge_length;
473 		BUG_ON(len == 0);
474 		/* If the source address is not aligned, try to align it. */
475 		off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
476 		if (off) {
477 			u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
478 					    ~(sizeof(u32) - 1));
479 			u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
480 			u32 y;
481 
482 			y = sizeof(u32) - off;
483 			if (len > y)
484 				len = y;
485 			if (len + extra >= sizeof(u32)) {
486 				data |= set_upper_bits(v, extra *
487 						       BITS_PER_BYTE);
488 				len = sizeof(u32) - extra;
489 				if (len == length) {
490 					last = data;
491 					break;
492 				}
493 				__raw_writel(data, piobuf);
494 				piobuf++;
495 				extra = 0;
496 				data = 0;
497 			} else {
498 				/* Clear unused upper bytes */
499 				data |= clear_upper_bytes(v, len, extra);
500 				if (len == length) {
501 					last = data;
502 					break;
503 				}
504 				extra += len;
505 			}
506 		} else if (extra) {
507 			/* Source address is aligned. */
508 			u32 *addr = (u32 *) ss->sge.vaddr;
509 			int shift = extra * BITS_PER_BYTE;
510 			int ushift = 32 - shift;
511 			u32 l = len;
512 
513 			while (l >= sizeof(u32)) {
514 				u32 v = *addr;
515 
516 				data |= set_upper_bits(v, shift);
517 				__raw_writel(data, piobuf);
518 				data = get_upper_bits(v, ushift);
519 				piobuf++;
520 				addr++;
521 				l -= sizeof(u32);
522 			}
523 			/*
524 			 * We still have 'extra' number of bytes leftover.
525 			 */
526 			if (l) {
527 				u32 v = *addr;
528 
529 				if (l + extra >= sizeof(u32)) {
530 					data |= set_upper_bits(v, shift);
531 					len -= l + extra - sizeof(u32);
532 					if (len == length) {
533 						last = data;
534 						break;
535 					}
536 					__raw_writel(data, piobuf);
537 					piobuf++;
538 					extra = 0;
539 					data = 0;
540 				} else {
541 					/* Clear unused upper bytes */
542 					data |= clear_upper_bytes(v, l, extra);
543 					if (len == length) {
544 						last = data;
545 						break;
546 					}
547 					extra += l;
548 				}
549 			} else if (len == length) {
550 				last = data;
551 				break;
552 			}
553 		} else if (len == length) {
554 			u32 w;
555 
556 			/*
557 			 * Need to round up for the last dword in the
558 			 * packet.
559 			 */
560 			w = (len + 3) >> 2;
561 			qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
562 			piobuf += w - 1;
563 			last = ((u32 *) ss->sge.vaddr)[w - 1];
564 			break;
565 		} else {
566 			u32 w = len >> 2;
567 
568 			qib_pio_copy(piobuf, ss->sge.vaddr, w);
569 			piobuf += w;
570 
571 			extra = len & (sizeof(u32) - 1);
572 			if (extra) {
573 				u32 v = ((u32 *) ss->sge.vaddr)[w];
574 
575 				/* Clear unused upper bytes */
576 				data = clear_upper_bytes(v, extra, 0);
577 			}
578 		}
579 		rvt_update_sge(ss, len, false);
580 		length -= len;
581 	}
582 	/* Update address before sending packet. */
583 	rvt_update_sge(ss, length, false);
584 	if (flush_wc) {
585 		/* must flush early everything before trigger word */
586 		qib_flush_wc();
587 		__raw_writel(last, piobuf);
588 		/* be sure trigger word is written */
589 		qib_flush_wc();
590 	} else
591 		__raw_writel(last, piobuf);
592 }
593 
594 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
595 					   struct rvt_qp *qp)
596 {
597 	struct qib_qp_priv *priv = qp->priv;
598 	struct qib_verbs_txreq *tx;
599 	unsigned long flags;
600 
601 	spin_lock_irqsave(&qp->s_lock, flags);
602 	spin_lock(&dev->rdi.pending_lock);
603 
604 	if (!list_empty(&dev->txreq_free)) {
605 		struct list_head *l = dev->txreq_free.next;
606 
607 		list_del(l);
608 		spin_unlock(&dev->rdi.pending_lock);
609 		spin_unlock_irqrestore(&qp->s_lock, flags);
610 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
611 	} else {
612 		if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK &&
613 		    list_empty(&priv->iowait)) {
614 			dev->n_txwait++;
615 			qp->s_flags |= RVT_S_WAIT_TX;
616 			list_add_tail(&priv->iowait, &dev->txwait);
617 		}
618 		qp->s_flags &= ~RVT_S_BUSY;
619 		spin_unlock(&dev->rdi.pending_lock);
620 		spin_unlock_irqrestore(&qp->s_lock, flags);
621 		tx = ERR_PTR(-EBUSY);
622 	}
623 	return tx;
624 }
625 
626 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
627 					 struct rvt_qp *qp)
628 {
629 	struct qib_verbs_txreq *tx;
630 	unsigned long flags;
631 
632 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
633 	/* assume the list non empty */
634 	if (likely(!list_empty(&dev->txreq_free))) {
635 		struct list_head *l = dev->txreq_free.next;
636 
637 		list_del(l);
638 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
639 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
640 	} else {
641 		/* call slow path to get the extra lock */
642 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
643 		tx =  __get_txreq(dev, qp);
644 	}
645 	return tx;
646 }
647 
648 void qib_put_txreq(struct qib_verbs_txreq *tx)
649 {
650 	struct qib_ibdev *dev;
651 	struct rvt_qp *qp;
652 	struct qib_qp_priv *priv;
653 	unsigned long flags;
654 
655 	qp = tx->qp;
656 	dev = to_idev(qp->ibqp.device);
657 
658 	if (tx->mr) {
659 		rvt_put_mr(tx->mr);
660 		tx->mr = NULL;
661 	}
662 	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
663 		tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
664 		dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
665 				 tx->txreq.addr, tx->hdr_dwords << 2,
666 				 DMA_TO_DEVICE);
667 		kfree(tx->align_buf);
668 	}
669 
670 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
671 
672 	/* Put struct back on free list */
673 	list_add(&tx->txreq.list, &dev->txreq_free);
674 
675 	if (!list_empty(&dev->txwait)) {
676 		/* Wake up first QP wanting a free struct */
677 		priv = list_entry(dev->txwait.next, struct qib_qp_priv,
678 				  iowait);
679 		qp = priv->owner;
680 		list_del_init(&priv->iowait);
681 		rvt_get_qp(qp);
682 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
683 
684 		spin_lock_irqsave(&qp->s_lock, flags);
685 		if (qp->s_flags & RVT_S_WAIT_TX) {
686 			qp->s_flags &= ~RVT_S_WAIT_TX;
687 			qib_schedule_send(qp);
688 		}
689 		spin_unlock_irqrestore(&qp->s_lock, flags);
690 
691 		rvt_put_qp(qp);
692 	} else
693 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
694 }
695 
696 /*
697  * This is called when there are send DMA descriptors that might be
698  * available.
699  *
700  * This is called with ppd->sdma_lock held.
701  */
702 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
703 {
704 	struct rvt_qp *qp;
705 	struct qib_qp_priv *qpp, *nqpp;
706 	struct rvt_qp *qps[20];
707 	struct qib_ibdev *dev;
708 	unsigned i, n;
709 
710 	n = 0;
711 	dev = &ppd->dd->verbs_dev;
712 	spin_lock(&dev->rdi.pending_lock);
713 
714 	/* Search wait list for first QP wanting DMA descriptors. */
715 	list_for_each_entry_safe(qpp, nqpp, &dev->dmawait, iowait) {
716 		qp = qpp->owner;
717 		if (qp->port_num != ppd->port)
718 			continue;
719 		if (n == ARRAY_SIZE(qps))
720 			break;
721 		if (qpp->s_tx->txreq.sg_count > avail)
722 			break;
723 		avail -= qpp->s_tx->txreq.sg_count;
724 		list_del_init(&qpp->iowait);
725 		rvt_get_qp(qp);
726 		qps[n++] = qp;
727 	}
728 
729 	spin_unlock(&dev->rdi.pending_lock);
730 
731 	for (i = 0; i < n; i++) {
732 		qp = qps[i];
733 		spin_lock(&qp->s_lock);
734 		if (qp->s_flags & RVT_S_WAIT_DMA_DESC) {
735 			qp->s_flags &= ~RVT_S_WAIT_DMA_DESC;
736 			qib_schedule_send(qp);
737 		}
738 		spin_unlock(&qp->s_lock);
739 		rvt_put_qp(qp);
740 	}
741 }
742 
743 /*
744  * This is called with ppd->sdma_lock held.
745  */
746 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
747 {
748 	struct qib_verbs_txreq *tx =
749 		container_of(cookie, struct qib_verbs_txreq, txreq);
750 	struct rvt_qp *qp = tx->qp;
751 	struct qib_qp_priv *priv = qp->priv;
752 
753 	spin_lock(&qp->s_lock);
754 	if (tx->wqe)
755 		qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
756 	else if (qp->ibqp.qp_type == IB_QPT_RC) {
757 		struct ib_header *hdr;
758 
759 		if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
760 			hdr = &tx->align_buf->hdr;
761 		else {
762 			struct qib_ibdev *dev = to_idev(qp->ibqp.device);
763 
764 			hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
765 		}
766 		qib_rc_send_complete(qp, hdr);
767 	}
768 	if (atomic_dec_and_test(&priv->s_dma_busy)) {
769 		if (qp->state == IB_QPS_RESET)
770 			wake_up(&priv->wait_dma);
771 		else if (qp->s_flags & RVT_S_WAIT_DMA) {
772 			qp->s_flags &= ~RVT_S_WAIT_DMA;
773 			qib_schedule_send(qp);
774 		}
775 	}
776 	spin_unlock(&qp->s_lock);
777 
778 	qib_put_txreq(tx);
779 }
780 
781 static int wait_kmem(struct qib_ibdev *dev, struct rvt_qp *qp)
782 {
783 	struct qib_qp_priv *priv = qp->priv;
784 	unsigned long flags;
785 	int ret = 0;
786 
787 	spin_lock_irqsave(&qp->s_lock, flags);
788 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
789 		spin_lock(&dev->rdi.pending_lock);
790 		if (list_empty(&priv->iowait)) {
791 			if (list_empty(&dev->memwait))
792 				mod_timer(&dev->mem_timer, jiffies + 1);
793 			qp->s_flags |= RVT_S_WAIT_KMEM;
794 			list_add_tail(&priv->iowait, &dev->memwait);
795 		}
796 		spin_unlock(&dev->rdi.pending_lock);
797 		qp->s_flags &= ~RVT_S_BUSY;
798 		ret = -EBUSY;
799 	}
800 	spin_unlock_irqrestore(&qp->s_lock, flags);
801 
802 	return ret;
803 }
804 
805 static int qib_verbs_send_dma(struct rvt_qp *qp, struct ib_header *hdr,
806 			      u32 hdrwords, struct rvt_sge_state *ss, u32 len,
807 			      u32 plen, u32 dwords)
808 {
809 	struct qib_qp_priv *priv = qp->priv;
810 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
811 	struct qib_devdata *dd = dd_from_dev(dev);
812 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
813 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
814 	struct qib_verbs_txreq *tx;
815 	struct qib_pio_header *phdr;
816 	u32 control;
817 	u32 ndesc;
818 	int ret;
819 
820 	tx = priv->s_tx;
821 	if (tx) {
822 		priv->s_tx = NULL;
823 		/* resend previously constructed packet */
824 		ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
825 		goto bail;
826 	}
827 
828 	tx = get_txreq(dev, qp);
829 	if (IS_ERR(tx))
830 		goto bail_tx;
831 
832 	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
833 				       be16_to_cpu(hdr->lrh[0]) >> 12);
834 	tx->qp = qp;
835 	tx->wqe = qp->s_wqe;
836 	tx->mr = qp->s_rdma_mr;
837 	if (qp->s_rdma_mr)
838 		qp->s_rdma_mr = NULL;
839 	tx->txreq.callback = sdma_complete;
840 	if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
841 		tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
842 	else
843 		tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
844 	if (plen + 1 > dd->piosize2kmax_dwords)
845 		tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
846 
847 	if (len) {
848 		/*
849 		 * Don't try to DMA if it takes more descriptors than
850 		 * the queue holds.
851 		 */
852 		ndesc = qib_count_sge(ss, len);
853 		if (ndesc >= ppd->sdma_descq_cnt)
854 			ndesc = 0;
855 	} else
856 		ndesc = 1;
857 	if (ndesc) {
858 		phdr = &dev->pio_hdrs[tx->hdr_inx];
859 		phdr->pbc[0] = cpu_to_le32(plen);
860 		phdr->pbc[1] = cpu_to_le32(control);
861 		memcpy(&phdr->hdr, hdr, hdrwords << 2);
862 		tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
863 		tx->txreq.sg_count = ndesc;
864 		tx->txreq.addr = dev->pio_hdrs_phys +
865 			tx->hdr_inx * sizeof(struct qib_pio_header);
866 		tx->hdr_dwords = hdrwords + 2; /* add PBC length */
867 		ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
868 		goto bail;
869 	}
870 
871 	/* Allocate a buffer and copy the header and payload to it. */
872 	tx->hdr_dwords = plen + 1;
873 	phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
874 	if (!phdr)
875 		goto err_tx;
876 	phdr->pbc[0] = cpu_to_le32(plen);
877 	phdr->pbc[1] = cpu_to_le32(control);
878 	memcpy(&phdr->hdr, hdr, hdrwords << 2);
879 	qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
880 
881 	tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
882 					tx->hdr_dwords << 2, DMA_TO_DEVICE);
883 	if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
884 		goto map_err;
885 	tx->align_buf = phdr;
886 	tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
887 	tx->txreq.sg_count = 1;
888 	ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
889 	goto unaligned;
890 
891 map_err:
892 	kfree(phdr);
893 err_tx:
894 	qib_put_txreq(tx);
895 	ret = wait_kmem(dev, qp);
896 unaligned:
897 	ibp->rvp.n_unaligned++;
898 bail:
899 	return ret;
900 bail_tx:
901 	ret = PTR_ERR(tx);
902 	goto bail;
903 }
904 
905 /*
906  * If we are now in the error state, return zero to flush the
907  * send work request.
908  */
909 static int no_bufs_available(struct rvt_qp *qp)
910 {
911 	struct qib_qp_priv *priv = qp->priv;
912 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
913 	struct qib_devdata *dd;
914 	unsigned long flags;
915 	int ret = 0;
916 
917 	/*
918 	 * Note that as soon as want_buffer() is called and
919 	 * possibly before it returns, qib_ib_piobufavail()
920 	 * could be called. Therefore, put QP on the I/O wait list before
921 	 * enabling the PIO avail interrupt.
922 	 */
923 	spin_lock_irqsave(&qp->s_lock, flags);
924 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
925 		spin_lock(&dev->rdi.pending_lock);
926 		if (list_empty(&priv->iowait)) {
927 			dev->n_piowait++;
928 			qp->s_flags |= RVT_S_WAIT_PIO;
929 			list_add_tail(&priv->iowait, &dev->piowait);
930 			dd = dd_from_dev(dev);
931 			dd->f_wantpiobuf_intr(dd, 1);
932 		}
933 		spin_unlock(&dev->rdi.pending_lock);
934 		qp->s_flags &= ~RVT_S_BUSY;
935 		ret = -EBUSY;
936 	}
937 	spin_unlock_irqrestore(&qp->s_lock, flags);
938 	return ret;
939 }
940 
941 static int qib_verbs_send_pio(struct rvt_qp *qp, struct ib_header *ibhdr,
942 			      u32 hdrwords, struct rvt_sge_state *ss, u32 len,
943 			      u32 plen, u32 dwords)
944 {
945 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
946 	struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
947 	u32 *hdr = (u32 *) ibhdr;
948 	u32 __iomem *piobuf_orig;
949 	u32 __iomem *piobuf;
950 	u64 pbc;
951 	unsigned long flags;
952 	unsigned flush_wc;
953 	u32 control;
954 	u32 pbufn;
955 
956 	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
957 		be16_to_cpu(ibhdr->lrh[0]) >> 12);
958 	pbc = ((u64) control << 32) | plen;
959 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
960 	if (unlikely(piobuf == NULL))
961 		return no_bufs_available(qp);
962 
963 	/*
964 	 * Write the pbc.
965 	 * We have to flush after the PBC for correctness on some cpus
966 	 * or WC buffer can be written out of order.
967 	 */
968 	writeq(pbc, piobuf);
969 	piobuf_orig = piobuf;
970 	piobuf += 2;
971 
972 	flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
973 	if (len == 0) {
974 		/*
975 		 * If there is just the header portion, must flush before
976 		 * writing last word of header for correctness, and after
977 		 * the last header word (trigger word).
978 		 */
979 		if (flush_wc) {
980 			qib_flush_wc();
981 			qib_pio_copy(piobuf, hdr, hdrwords - 1);
982 			qib_flush_wc();
983 			__raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
984 			qib_flush_wc();
985 		} else
986 			qib_pio_copy(piobuf, hdr, hdrwords);
987 		goto done;
988 	}
989 
990 	if (flush_wc)
991 		qib_flush_wc();
992 	qib_pio_copy(piobuf, hdr, hdrwords);
993 	piobuf += hdrwords;
994 
995 	/* The common case is aligned and contained in one segment. */
996 	if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
997 		   !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
998 		u32 *addr = (u32 *) ss->sge.vaddr;
999 
1000 		/* Update address before sending packet. */
1001 		rvt_update_sge(ss, len, false);
1002 		if (flush_wc) {
1003 			qib_pio_copy(piobuf, addr, dwords - 1);
1004 			/* must flush early everything before trigger word */
1005 			qib_flush_wc();
1006 			__raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1007 			/* be sure trigger word is written */
1008 			qib_flush_wc();
1009 		} else
1010 			qib_pio_copy(piobuf, addr, dwords);
1011 		goto done;
1012 	}
1013 	copy_io(piobuf, ss, len, flush_wc);
1014 done:
1015 	if (dd->flags & QIB_USE_SPCL_TRIG) {
1016 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
1017 
1018 		qib_flush_wc();
1019 		__raw_writel(0xaebecede, piobuf_orig + spcl_off);
1020 	}
1021 	qib_sendbuf_done(dd, pbufn);
1022 	if (qp->s_rdma_mr) {
1023 		rvt_put_mr(qp->s_rdma_mr);
1024 		qp->s_rdma_mr = NULL;
1025 	}
1026 	if (qp->s_wqe) {
1027 		spin_lock_irqsave(&qp->s_lock, flags);
1028 		qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1029 		spin_unlock_irqrestore(&qp->s_lock, flags);
1030 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1031 		spin_lock_irqsave(&qp->s_lock, flags);
1032 		qib_rc_send_complete(qp, ibhdr);
1033 		spin_unlock_irqrestore(&qp->s_lock, flags);
1034 	}
1035 	return 0;
1036 }
1037 
1038 /**
1039  * qib_verbs_send - send a packet
1040  * @qp: the QP to send on
1041  * @hdr: the packet header
1042  * @hdrwords: the number of 32-bit words in the header
1043  * @ss: the SGE to send
1044  * @len: the length of the packet in bytes
1045  *
1046  * Return zero if packet is sent or queued OK.
1047  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1048  */
1049 int qib_verbs_send(struct rvt_qp *qp, struct ib_header *hdr,
1050 		   u32 hdrwords, struct rvt_sge_state *ss, u32 len)
1051 {
1052 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1053 	u32 plen;
1054 	int ret;
1055 	u32 dwords = (len + 3) >> 2;
1056 
1057 	/*
1058 	 * Calculate the send buffer trigger address.
1059 	 * The +1 counts for the pbc control dword following the pbc length.
1060 	 */
1061 	plen = hdrwords + dwords + 1;
1062 
1063 	/*
1064 	 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1065 	 * can defer SDMA restart until link goes ACTIVE without
1066 	 * worrying about just how we got there.
1067 	 */
1068 	if (qp->ibqp.qp_type == IB_QPT_SMI ||
1069 	    !(dd->flags & QIB_HAS_SEND_DMA))
1070 		ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1071 					 plen, dwords);
1072 	else
1073 		ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1074 					 plen, dwords);
1075 
1076 	return ret;
1077 }
1078 
1079 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1080 			  u64 *rwords, u64 *spkts, u64 *rpkts,
1081 			  u64 *xmit_wait)
1082 {
1083 	int ret;
1084 	struct qib_devdata *dd = ppd->dd;
1085 
1086 	if (!(dd->flags & QIB_PRESENT)) {
1087 		/* no hardware, freeze, etc. */
1088 		ret = -EINVAL;
1089 		goto bail;
1090 	}
1091 	*swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1092 	*rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1093 	*spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1094 	*rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1095 	*xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1096 
1097 	ret = 0;
1098 
1099 bail:
1100 	return ret;
1101 }
1102 
1103 /**
1104  * qib_get_counters - get various chip counters
1105  * @dd: the qlogic_ib device
1106  * @cntrs: counters are placed here
1107  *
1108  * Return the counters needed by recv_pma_get_portcounters().
1109  */
1110 int qib_get_counters(struct qib_pportdata *ppd,
1111 		     struct qib_verbs_counters *cntrs)
1112 {
1113 	int ret;
1114 
1115 	if (!(ppd->dd->flags & QIB_PRESENT)) {
1116 		/* no hardware, freeze, etc. */
1117 		ret = -EINVAL;
1118 		goto bail;
1119 	}
1120 	cntrs->symbol_error_counter =
1121 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1122 	cntrs->link_error_recovery_counter =
1123 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1124 	/*
1125 	 * The link downed counter counts when the other side downs the
1126 	 * connection.  We add in the number of times we downed the link
1127 	 * due to local link integrity errors to compensate.
1128 	 */
1129 	cntrs->link_downed_counter =
1130 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1131 	cntrs->port_rcv_errors =
1132 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1133 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1134 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1135 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1136 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1137 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1138 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1139 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1140 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1141 	cntrs->port_rcv_errors +=
1142 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1143 	cntrs->port_rcv_errors +=
1144 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1145 	cntrs->port_rcv_remphys_errors =
1146 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1147 	cntrs->port_xmit_discards =
1148 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1149 	cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1150 			QIBPORTCNTR_WORDSEND);
1151 	cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1152 			QIBPORTCNTR_WORDRCV);
1153 	cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1154 			QIBPORTCNTR_PKTSEND);
1155 	cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1156 			QIBPORTCNTR_PKTRCV);
1157 	cntrs->local_link_integrity_errors =
1158 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1159 	cntrs->excessive_buffer_overrun_errors =
1160 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1161 	cntrs->vl15_dropped =
1162 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1163 
1164 	ret = 0;
1165 
1166 bail:
1167 	return ret;
1168 }
1169 
1170 /**
1171  * qib_ib_piobufavail - callback when a PIO buffer is available
1172  * @dd: the device pointer
1173  *
1174  * This is called from qib_intr() at interrupt level when a PIO buffer is
1175  * available after qib_verbs_send() returned an error that no buffers were
1176  * available. Disable the interrupt if there are no more QPs waiting.
1177  */
1178 void qib_ib_piobufavail(struct qib_devdata *dd)
1179 {
1180 	struct qib_ibdev *dev = &dd->verbs_dev;
1181 	struct list_head *list;
1182 	struct rvt_qp *qps[5];
1183 	struct rvt_qp *qp;
1184 	unsigned long flags;
1185 	unsigned i, n;
1186 	struct qib_qp_priv *priv;
1187 
1188 	list = &dev->piowait;
1189 	n = 0;
1190 
1191 	/*
1192 	 * Note: checking that the piowait list is empty and clearing
1193 	 * the buffer available interrupt needs to be atomic or we
1194 	 * could end up with QPs on the wait list with the interrupt
1195 	 * disabled.
1196 	 */
1197 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
1198 	while (!list_empty(list)) {
1199 		if (n == ARRAY_SIZE(qps))
1200 			goto full;
1201 		priv = list_entry(list->next, struct qib_qp_priv, iowait);
1202 		qp = priv->owner;
1203 		list_del_init(&priv->iowait);
1204 		rvt_get_qp(qp);
1205 		qps[n++] = qp;
1206 	}
1207 	dd->f_wantpiobuf_intr(dd, 0);
1208 full:
1209 	spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
1210 
1211 	for (i = 0; i < n; i++) {
1212 		qp = qps[i];
1213 
1214 		spin_lock_irqsave(&qp->s_lock, flags);
1215 		if (qp->s_flags & RVT_S_WAIT_PIO) {
1216 			qp->s_flags &= ~RVT_S_WAIT_PIO;
1217 			qib_schedule_send(qp);
1218 		}
1219 		spin_unlock_irqrestore(&qp->s_lock, flags);
1220 
1221 		/* Notify qib_destroy_qp() if it is waiting. */
1222 		rvt_put_qp(qp);
1223 	}
1224 }
1225 
1226 static int qib_query_port(struct rvt_dev_info *rdi, u8 port_num,
1227 			  struct ib_port_attr *props)
1228 {
1229 	struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1230 	struct qib_devdata *dd = dd_from_dev(ibdev);
1231 	struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1232 	enum ib_mtu mtu;
1233 	u16 lid = ppd->lid;
1234 
1235 	/* props being zeroed by the caller, avoid zeroing it here */
1236 	props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1237 	props->lmc = ppd->lmc;
1238 	props->state = dd->f_iblink_state(ppd->lastibcstat);
1239 	props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1240 	props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1241 	props->active_width = ppd->link_width_active;
1242 	/* See rate_show() */
1243 	props->active_speed = ppd->link_speed_active;
1244 	props->max_vl_num = qib_num_vls(ppd->vls_supported);
1245 
1246 	props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1247 	switch (ppd->ibmtu) {
1248 	case 4096:
1249 		mtu = IB_MTU_4096;
1250 		break;
1251 	case 2048:
1252 		mtu = IB_MTU_2048;
1253 		break;
1254 	case 1024:
1255 		mtu = IB_MTU_1024;
1256 		break;
1257 	case 512:
1258 		mtu = IB_MTU_512;
1259 		break;
1260 	case 256:
1261 		mtu = IB_MTU_256;
1262 		break;
1263 	default:
1264 		mtu = IB_MTU_2048;
1265 	}
1266 	props->active_mtu = mtu;
1267 
1268 	return 0;
1269 }
1270 
1271 static int qib_modify_device(struct ib_device *device,
1272 			     int device_modify_mask,
1273 			     struct ib_device_modify *device_modify)
1274 {
1275 	struct qib_devdata *dd = dd_from_ibdev(device);
1276 	unsigned i;
1277 	int ret;
1278 
1279 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1280 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1281 		ret = -EOPNOTSUPP;
1282 		goto bail;
1283 	}
1284 
1285 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1286 		memcpy(device->node_desc, device_modify->node_desc,
1287 		       IB_DEVICE_NODE_DESC_MAX);
1288 		for (i = 0; i < dd->num_pports; i++) {
1289 			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1290 
1291 			qib_node_desc_chg(ibp);
1292 		}
1293 	}
1294 
1295 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1296 		ib_qib_sys_image_guid =
1297 			cpu_to_be64(device_modify->sys_image_guid);
1298 		for (i = 0; i < dd->num_pports; i++) {
1299 			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1300 
1301 			qib_sys_guid_chg(ibp);
1302 		}
1303 	}
1304 
1305 	ret = 0;
1306 
1307 bail:
1308 	return ret;
1309 }
1310 
1311 static int qib_shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1312 {
1313 	struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1314 	struct qib_devdata *dd = dd_from_dev(ibdev);
1315 	struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1316 
1317 	qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1318 
1319 	return 0;
1320 }
1321 
1322 static int qib_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1323 			   int guid_index, __be64 *guid)
1324 {
1325 	struct qib_ibport *ibp = container_of(rvp, struct qib_ibport, rvp);
1326 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1327 
1328 	if (guid_index == 0)
1329 		*guid = ppd->guid;
1330 	else if (guid_index < QIB_GUIDS_PER_PORT)
1331 		*guid = ibp->guids[guid_index - 1];
1332 	else
1333 		return -EINVAL;
1334 
1335 	return 0;
1336 }
1337 
1338 int qib_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1339 {
1340 	if (rdma_ah_get_sl(ah_attr) > 15)
1341 		return -EINVAL;
1342 
1343 	if (rdma_ah_get_dlid(ah_attr) == 0)
1344 		return -EINVAL;
1345 	if (rdma_ah_get_dlid(ah_attr) >=
1346 		be16_to_cpu(IB_MULTICAST_LID_BASE) &&
1347 	    rdma_ah_get_dlid(ah_attr) !=
1348 		be16_to_cpu(IB_LID_PERMISSIVE) &&
1349 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1350 		return -EINVAL;
1351 
1352 	return 0;
1353 }
1354 
1355 static void qib_notify_new_ah(struct ib_device *ibdev,
1356 			      struct rdma_ah_attr *ah_attr,
1357 			      struct rvt_ah *ah)
1358 {
1359 	struct qib_ibport *ibp;
1360 	struct qib_pportdata *ppd;
1361 
1362 	/*
1363 	 * Do not trust reading anything from rvt_ah at this point as it is not
1364 	 * done being setup. We can however modify things which we need to set.
1365 	 */
1366 
1367 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1368 	ppd = ppd_from_ibp(ibp);
1369 	ah->vl = ibp->sl_to_vl[rdma_ah_get_sl(&ah->attr)];
1370 	ah->log_pmtu = ilog2(ppd->ibmtu);
1371 }
1372 
1373 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1374 {
1375 	struct rdma_ah_attr attr;
1376 	struct ib_ah *ah = ERR_PTR(-EINVAL);
1377 	struct rvt_qp *qp0;
1378 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1379 	struct qib_devdata *dd = dd_from_ppd(ppd);
1380 	u8 port_num = ppd->port;
1381 
1382 	memset(&attr, 0, sizeof(attr));
1383 	attr.type = rdma_ah_find_type(&dd->verbs_dev.rdi.ibdev, port_num);
1384 	rdma_ah_set_dlid(&attr, dlid);
1385 	rdma_ah_set_port_num(&attr, port_num);
1386 	rcu_read_lock();
1387 	qp0 = rcu_dereference(ibp->rvp.qp[0]);
1388 	if (qp0)
1389 		ah = rdma_create_ah(qp0->ibqp.pd, &attr);
1390 	rcu_read_unlock();
1391 	return ah;
1392 }
1393 
1394 /**
1395  * qib_get_npkeys - return the size of the PKEY table for context 0
1396  * @dd: the qlogic_ib device
1397  */
1398 unsigned qib_get_npkeys(struct qib_devdata *dd)
1399 {
1400 	return ARRAY_SIZE(dd->rcd[0]->pkeys);
1401 }
1402 
1403 /*
1404  * Return the indexed PKEY from the port PKEY table.
1405  * No need to validate rcd[ctxt]; the port is setup if we are here.
1406  */
1407 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1408 {
1409 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1410 	struct qib_devdata *dd = ppd->dd;
1411 	unsigned ctxt = ppd->hw_pidx;
1412 	unsigned ret;
1413 
1414 	/* dd->rcd null if mini_init or some init failures */
1415 	if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1416 		ret = 0;
1417 	else
1418 		ret = dd->rcd[ctxt]->pkeys[index];
1419 
1420 	return ret;
1421 }
1422 
1423 static void init_ibport(struct qib_pportdata *ppd)
1424 {
1425 	struct qib_verbs_counters cntrs;
1426 	struct qib_ibport *ibp = &ppd->ibport_data;
1427 
1428 	spin_lock_init(&ibp->rvp.lock);
1429 	/* Set the prefix to the default value (see ch. 4.1.1) */
1430 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1431 	ibp->rvp.sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
1432 	ibp->rvp.port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
1433 		IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
1434 		IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
1435 		IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
1436 		IB_PORT_OTHER_LOCAL_CHANGES_SUP;
1437 	if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
1438 		ibp->rvp.port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
1439 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1440 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1441 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1442 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1443 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1444 
1445 	/* Snapshot current HW counters to "clear" them. */
1446 	qib_get_counters(ppd, &cntrs);
1447 	ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
1448 	ibp->z_link_error_recovery_counter =
1449 		cntrs.link_error_recovery_counter;
1450 	ibp->z_link_downed_counter = cntrs.link_downed_counter;
1451 	ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
1452 	ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
1453 	ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
1454 	ibp->z_port_xmit_data = cntrs.port_xmit_data;
1455 	ibp->z_port_rcv_data = cntrs.port_rcv_data;
1456 	ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
1457 	ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
1458 	ibp->z_local_link_integrity_errors =
1459 		cntrs.local_link_integrity_errors;
1460 	ibp->z_excessive_buffer_overrun_errors =
1461 		cntrs.excessive_buffer_overrun_errors;
1462 	ibp->z_vl15_dropped = cntrs.vl15_dropped;
1463 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1464 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1465 }
1466 
1467 /**
1468  * qib_fill_device_attr - Fill in rvt dev info device attributes.
1469  * @dd: the device data structure
1470  */
1471 static void qib_fill_device_attr(struct qib_devdata *dd)
1472 {
1473 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1474 
1475 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1476 
1477 	rdi->dparms.props.max_pd = ib_qib_max_pds;
1478 	rdi->dparms.props.max_ah = ib_qib_max_ahs;
1479 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1480 		IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1481 		IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1482 		IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1483 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1484 	rdi->dparms.props.vendor_id =
1485 		QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1486 	rdi->dparms.props.vendor_part_id = dd->deviceid;
1487 	rdi->dparms.props.hw_ver = dd->minrev;
1488 	rdi->dparms.props.sys_image_guid = ib_qib_sys_image_guid;
1489 	rdi->dparms.props.max_mr_size = ~0ULL;
1490 	rdi->dparms.props.max_qp = ib_qib_max_qps;
1491 	rdi->dparms.props.max_qp_wr = ib_qib_max_qp_wrs;
1492 	rdi->dparms.props.max_sge = ib_qib_max_sges;
1493 	rdi->dparms.props.max_sge_rd = ib_qib_max_sges;
1494 	rdi->dparms.props.max_cq = ib_qib_max_cqs;
1495 	rdi->dparms.props.max_cqe = ib_qib_max_cqes;
1496 	rdi->dparms.props.max_ah = ib_qib_max_ahs;
1497 	rdi->dparms.props.max_mr = rdi->lkey_table.max;
1498 	rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1499 	rdi->dparms.props.max_map_per_fmr = 32767;
1500 	rdi->dparms.props.max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1501 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1502 	rdi->dparms.props.max_srq = ib_qib_max_srqs;
1503 	rdi->dparms.props.max_srq_wr = ib_qib_max_srq_wrs;
1504 	rdi->dparms.props.max_srq_sge = ib_qib_max_srq_sges;
1505 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1506 	rdi->dparms.props.max_pkeys = qib_get_npkeys(dd);
1507 	rdi->dparms.props.max_mcast_grp = ib_qib_max_mcast_grps;
1508 	rdi->dparms.props.max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1509 	rdi->dparms.props.max_total_mcast_qp_attach =
1510 					rdi->dparms.props.max_mcast_qp_attach *
1511 					rdi->dparms.props.max_mcast_grp;
1512 	/* post send table */
1513 	dd->verbs_dev.rdi.post_parms = qib_post_parms;
1514 }
1515 
1516 /**
1517  * qib_register_ib_device - register our device with the infiniband core
1518  * @dd: the device data structure
1519  * Return the allocated qib_ibdev pointer or NULL on error.
1520  */
1521 int qib_register_ib_device(struct qib_devdata *dd)
1522 {
1523 	struct qib_ibdev *dev = &dd->verbs_dev;
1524 	struct ib_device *ibdev = &dev->rdi.ibdev;
1525 	struct qib_pportdata *ppd = dd->pport;
1526 	unsigned i, ctxt;
1527 	int ret;
1528 
1529 	get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
1530 	for (i = 0; i < dd->num_pports; i++)
1531 		init_ibport(ppd + i);
1532 
1533 	/* Only need to initialize non-zero fields. */
1534 	timer_setup(&dev->mem_timer, mem_timer, 0);
1535 
1536 	INIT_LIST_HEAD(&dev->piowait);
1537 	INIT_LIST_HEAD(&dev->dmawait);
1538 	INIT_LIST_HEAD(&dev->txwait);
1539 	INIT_LIST_HEAD(&dev->memwait);
1540 	INIT_LIST_HEAD(&dev->txreq_free);
1541 
1542 	if (ppd->sdma_descq_cnt) {
1543 		dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
1544 						ppd->sdma_descq_cnt *
1545 						sizeof(struct qib_pio_header),
1546 						&dev->pio_hdrs_phys,
1547 						GFP_KERNEL);
1548 		if (!dev->pio_hdrs) {
1549 			ret = -ENOMEM;
1550 			goto err_hdrs;
1551 		}
1552 	}
1553 
1554 	for (i = 0; i < ppd->sdma_descq_cnt; i++) {
1555 		struct qib_verbs_txreq *tx;
1556 
1557 		tx = kzalloc(sizeof(*tx), GFP_KERNEL);
1558 		if (!tx) {
1559 			ret = -ENOMEM;
1560 			goto err_tx;
1561 		}
1562 		tx->hdr_inx = i;
1563 		list_add(&tx->txreq.list, &dev->txreq_free);
1564 	}
1565 
1566 	/*
1567 	 * The system image GUID is supposed to be the same for all
1568 	 * IB HCAs in a single system but since there can be other
1569 	 * device types in the system, we can't be sure this is unique.
1570 	 */
1571 	if (!ib_qib_sys_image_guid)
1572 		ib_qib_sys_image_guid = ppd->guid;
1573 
1574 	strlcpy(ibdev->name, "qib%d", IB_DEVICE_NAME_MAX);
1575 	ibdev->owner = THIS_MODULE;
1576 	ibdev->node_guid = ppd->guid;
1577 	ibdev->phys_port_cnt = dd->num_pports;
1578 	ibdev->dev.parent = &dd->pcidev->dev;
1579 	ibdev->modify_device = qib_modify_device;
1580 	ibdev->process_mad = qib_process_mad;
1581 
1582 	snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
1583 		 "Intel Infiniband HCA %s", init_utsname()->nodename);
1584 
1585 	/*
1586 	 * Fill in rvt info object.
1587 	 */
1588 	dd->verbs_dev.rdi.driver_f.port_callback = qib_create_port_files;
1589 	dd->verbs_dev.rdi.driver_f.get_card_name = qib_get_card_name;
1590 	dd->verbs_dev.rdi.driver_f.get_pci_dev = qib_get_pci_dev;
1591 	dd->verbs_dev.rdi.driver_f.check_ah = qib_check_ah;
1592 	dd->verbs_dev.rdi.driver_f.check_send_wqe = qib_check_send_wqe;
1593 	dd->verbs_dev.rdi.driver_f.notify_new_ah = qib_notify_new_ah;
1594 	dd->verbs_dev.rdi.driver_f.alloc_qpn = qib_alloc_qpn;
1595 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qib_qp_priv_alloc;
1596 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qib_qp_priv_free;
1597 	dd->verbs_dev.rdi.driver_f.free_all_qps = qib_free_all_qps;
1598 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = qib_notify_qp_reset;
1599 	dd->verbs_dev.rdi.driver_f.do_send = qib_do_send;
1600 	dd->verbs_dev.rdi.driver_f.schedule_send = qib_schedule_send;
1601 	dd->verbs_dev.rdi.driver_f.quiesce_qp = qib_quiesce_qp;
1602 	dd->verbs_dev.rdi.driver_f.stop_send_queue = qib_stop_send_queue;
1603 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = qib_flush_qp_waiters;
1604 	dd->verbs_dev.rdi.driver_f.notify_error_qp = qib_notify_error_qp;
1605 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = qib_restart_rc;
1606 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = qib_mtu_to_path_mtu;
1607 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = qib_mtu_from_qp;
1608 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = qib_get_pmtu_from_attr;
1609 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _qib_schedule_send;
1610 	dd->verbs_dev.rdi.driver_f.query_port_state = qib_query_port;
1611 	dd->verbs_dev.rdi.driver_f.shut_down_port = qib_shut_down_port;
1612 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = qib_cap_mask_chg;
1613 	dd->verbs_dev.rdi.driver_f.notify_create_mad_agent =
1614 						qib_notify_create_mad_agent;
1615 	dd->verbs_dev.rdi.driver_f.notify_free_mad_agent =
1616 						qib_notify_free_mad_agent;
1617 
1618 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = QIB_MAX_RDMA_ATOMIC;
1619 	dd->verbs_dev.rdi.driver_f.get_guid_be = qib_get_guid_be;
1620 	dd->verbs_dev.rdi.dparms.lkey_table_size = qib_lkey_table_size;
1621 	dd->verbs_dev.rdi.dparms.qp_table_size = ib_qib_qp_table_size;
1622 	dd->verbs_dev.rdi.dparms.qpn_start = 1;
1623 	dd->verbs_dev.rdi.dparms.qpn_res_start = QIB_KD_QP;
1624 	dd->verbs_dev.rdi.dparms.qpn_res_end = QIB_KD_QP; /* Reserve one QP */
1625 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1626 	dd->verbs_dev.rdi.dparms.qos_shift = 1;
1627 	dd->verbs_dev.rdi.dparms.psn_mask = QIB_PSN_MASK;
1628 	dd->verbs_dev.rdi.dparms.psn_shift = QIB_PSN_SHIFT;
1629 	dd->verbs_dev.rdi.dparms.psn_modify_mask = QIB_PSN_MASK;
1630 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1631 	dd->verbs_dev.rdi.dparms.npkeys = qib_get_npkeys(dd);
1632 	dd->verbs_dev.rdi.dparms.node = dd->assigned_node_id;
1633 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1634 	dd->verbs_dev.rdi.dparms.max_mad_size = IB_MGMT_MAD_SIZE;
1635 
1636 	snprintf(dd->verbs_dev.rdi.dparms.cq_name,
1637 		 sizeof(dd->verbs_dev.rdi.dparms.cq_name),
1638 		 "qib_cq%d", dd->unit);
1639 
1640 	qib_fill_device_attr(dd);
1641 
1642 	ppd = dd->pport;
1643 	for (i = 0; i < dd->num_pports; i++, ppd++) {
1644 		ctxt = ppd->hw_pidx;
1645 		rvt_init_port(&dd->verbs_dev.rdi,
1646 			      &ppd->ibport_data.rvp,
1647 			      i,
1648 			      dd->rcd[ctxt]->pkeys);
1649 	}
1650 
1651 	ret = rvt_register_device(&dd->verbs_dev.rdi);
1652 	if (ret)
1653 		goto err_tx;
1654 
1655 	ret = qib_verbs_register_sysfs(dd);
1656 	if (ret)
1657 		goto err_class;
1658 
1659 	return ret;
1660 
1661 err_class:
1662 	rvt_unregister_device(&dd->verbs_dev.rdi);
1663 err_tx:
1664 	while (!list_empty(&dev->txreq_free)) {
1665 		struct list_head *l = dev->txreq_free.next;
1666 		struct qib_verbs_txreq *tx;
1667 
1668 		list_del(l);
1669 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1670 		kfree(tx);
1671 	}
1672 	if (ppd->sdma_descq_cnt)
1673 		dma_free_coherent(&dd->pcidev->dev,
1674 				  ppd->sdma_descq_cnt *
1675 					sizeof(struct qib_pio_header),
1676 				  dev->pio_hdrs, dev->pio_hdrs_phys);
1677 err_hdrs:
1678 	qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1679 	return ret;
1680 }
1681 
1682 void qib_unregister_ib_device(struct qib_devdata *dd)
1683 {
1684 	struct qib_ibdev *dev = &dd->verbs_dev;
1685 
1686 	qib_verbs_unregister_sysfs(dd);
1687 
1688 	rvt_unregister_device(&dd->verbs_dev.rdi);
1689 
1690 	if (!list_empty(&dev->piowait))
1691 		qib_dev_err(dd, "piowait list not empty!\n");
1692 	if (!list_empty(&dev->dmawait))
1693 		qib_dev_err(dd, "dmawait list not empty!\n");
1694 	if (!list_empty(&dev->txwait))
1695 		qib_dev_err(dd, "txwait list not empty!\n");
1696 	if (!list_empty(&dev->memwait))
1697 		qib_dev_err(dd, "memwait list not empty!\n");
1698 
1699 	del_timer_sync(&dev->mem_timer);
1700 	while (!list_empty(&dev->txreq_free)) {
1701 		struct list_head *l = dev->txreq_free.next;
1702 		struct qib_verbs_txreq *tx;
1703 
1704 		list_del(l);
1705 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1706 		kfree(tx);
1707 	}
1708 	if (dd->pport->sdma_descq_cnt)
1709 		dma_free_coherent(&dd->pcidev->dev,
1710 				  dd->pport->sdma_descq_cnt *
1711 					sizeof(struct qib_pio_header),
1712 				  dev->pio_hdrs, dev->pio_hdrs_phys);
1713 }
1714 
1715 /**
1716  * _qib_schedule_send - schedule progress
1717  * @qp - the qp
1718  *
1719  * This schedules progress w/o regard to the s_flags.
1720  *
1721  * It is only used in post send, which doesn't hold
1722  * the s_lock.
1723  */
1724 void _qib_schedule_send(struct rvt_qp *qp)
1725 {
1726 	struct qib_ibport *ibp =
1727 		to_iport(qp->ibqp.device, qp->port_num);
1728 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1729 	struct qib_qp_priv *priv = qp->priv;
1730 
1731 	queue_work(ppd->qib_wq, &priv->s_work);
1732 }
1733 
1734 /**
1735  * qib_schedule_send - schedule progress
1736  * @qp - the qp
1737  *
1738  * This schedules qp progress.  The s_lock
1739  * should be held.
1740  */
1741 void qib_schedule_send(struct rvt_qp *qp)
1742 {
1743 	if (qib_send_ok(qp))
1744 		_qib_schedule_send(qp);
1745 }
1746