• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2012 - 2018 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  */
qib_copy_sge(struct rvt_sge_state * ss,void * data,u32 length,int release)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  */
qib_count_sge(struct rvt_sge_state * ss,u32 length)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  */
qib_copy_from_sge(void * data,struct rvt_sge_state * ss,u32 length)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  */
qib_qp_rcv(struct qib_ctxtdata * rcd,struct ib_header * hdr,int has_grh,void * data,u32 tlen,struct rvt_qp * qp)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  */
qib_ib_rcv(struct qib_ctxtdata * rcd,void * rhdr,void * data,u32 tlen)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 		rcu_read_lock();
364 		list_for_each_entry_rcu(p, &mcast->qp_list, list)
365 			qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
366 		rcu_read_unlock();
367 		/*
368 		 * Notify rvt_multicast_detach() if it is waiting for us
369 		 * to finish.
370 		 */
371 		if (atomic_dec_return(&mcast->refcount) <= 1)
372 			wake_up(&mcast->wait);
373 	} else {
374 		rcu_read_lock();
375 		qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
376 		if (!qp) {
377 			rcu_read_unlock();
378 			goto drop;
379 		}
380 		this_cpu_inc(ibp->pmastats->n_unicast_rcv);
381 		qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
382 		rcu_read_unlock();
383 	}
384 	return;
385 
386 drop:
387 	ibp->rvp.n_pkt_drops++;
388 }
389 
390 /*
391  * This is called from a timer to check for QPs
392  * which need kernel memory in order to send a packet.
393  */
mem_timer(struct timer_list * t)394 static void mem_timer(struct timer_list *t)
395 {
396 	struct qib_ibdev *dev = from_timer(dev, t, mem_timer);
397 	struct list_head *list = &dev->memwait;
398 	struct rvt_qp *qp = NULL;
399 	struct qib_qp_priv *priv = NULL;
400 	unsigned long flags;
401 
402 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
403 	if (!list_empty(list)) {
404 		priv = list_entry(list->next, struct qib_qp_priv, iowait);
405 		qp = priv->owner;
406 		list_del_init(&priv->iowait);
407 		rvt_get_qp(qp);
408 		if (!list_empty(list))
409 			mod_timer(&dev->mem_timer, jiffies + 1);
410 	}
411 	spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
412 
413 	if (qp) {
414 		spin_lock_irqsave(&qp->s_lock, flags);
415 		if (qp->s_flags & RVT_S_WAIT_KMEM) {
416 			qp->s_flags &= ~RVT_S_WAIT_KMEM;
417 			qib_schedule_send(qp);
418 		}
419 		spin_unlock_irqrestore(&qp->s_lock, flags);
420 		rvt_put_qp(qp);
421 	}
422 }
423 
424 #ifdef __LITTLE_ENDIAN
get_upper_bits(u32 data,u32 shift)425 static inline u32 get_upper_bits(u32 data, u32 shift)
426 {
427 	return data >> shift;
428 }
429 
set_upper_bits(u32 data,u32 shift)430 static inline u32 set_upper_bits(u32 data, u32 shift)
431 {
432 	return data << shift;
433 }
434 
clear_upper_bytes(u32 data,u32 n,u32 off)435 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
436 {
437 	data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
438 	data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
439 	return data;
440 }
441 #else
get_upper_bits(u32 data,u32 shift)442 static inline u32 get_upper_bits(u32 data, u32 shift)
443 {
444 	return data << shift;
445 }
446 
set_upper_bits(u32 data,u32 shift)447 static inline u32 set_upper_bits(u32 data, u32 shift)
448 {
449 	return data >> shift;
450 }
451 
clear_upper_bytes(u32 data,u32 n,u32 off)452 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
453 {
454 	data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
455 	data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
456 	return data;
457 }
458 #endif
459 
copy_io(u32 __iomem * piobuf,struct rvt_sge_state * ss,u32 length,unsigned flush_wc)460 static void copy_io(u32 __iomem *piobuf, struct rvt_sge_state *ss,
461 		    u32 length, unsigned flush_wc)
462 {
463 	u32 extra = 0;
464 	u32 data = 0;
465 	u32 last;
466 
467 	while (1) {
468 		u32 len = ss->sge.length;
469 		u32 off;
470 
471 		if (len > length)
472 			len = length;
473 		if (len > ss->sge.sge_length)
474 			len = ss->sge.sge_length;
475 		BUG_ON(len == 0);
476 		/* If the source address is not aligned, try to align it. */
477 		off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
478 		if (off) {
479 			u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
480 					    ~(sizeof(u32) - 1));
481 			u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
482 			u32 y;
483 
484 			y = sizeof(u32) - off;
485 			if (len > y)
486 				len = y;
487 			if (len + extra >= sizeof(u32)) {
488 				data |= set_upper_bits(v, extra *
489 						       BITS_PER_BYTE);
490 				len = sizeof(u32) - extra;
491 				if (len == length) {
492 					last = data;
493 					break;
494 				}
495 				__raw_writel(data, piobuf);
496 				piobuf++;
497 				extra = 0;
498 				data = 0;
499 			} else {
500 				/* Clear unused upper bytes */
501 				data |= clear_upper_bytes(v, len, extra);
502 				if (len == length) {
503 					last = data;
504 					break;
505 				}
506 				extra += len;
507 			}
508 		} else if (extra) {
509 			/* Source address is aligned. */
510 			u32 *addr = (u32 *) ss->sge.vaddr;
511 			int shift = extra * BITS_PER_BYTE;
512 			int ushift = 32 - shift;
513 			u32 l = len;
514 
515 			while (l >= sizeof(u32)) {
516 				u32 v = *addr;
517 
518 				data |= set_upper_bits(v, shift);
519 				__raw_writel(data, piobuf);
520 				data = get_upper_bits(v, ushift);
521 				piobuf++;
522 				addr++;
523 				l -= sizeof(u32);
524 			}
525 			/*
526 			 * We still have 'extra' number of bytes leftover.
527 			 */
528 			if (l) {
529 				u32 v = *addr;
530 
531 				if (l + extra >= sizeof(u32)) {
532 					data |= set_upper_bits(v, shift);
533 					len -= l + extra - sizeof(u32);
534 					if (len == length) {
535 						last = data;
536 						break;
537 					}
538 					__raw_writel(data, piobuf);
539 					piobuf++;
540 					extra = 0;
541 					data = 0;
542 				} else {
543 					/* Clear unused upper bytes */
544 					data |= clear_upper_bytes(v, l, extra);
545 					if (len == length) {
546 						last = data;
547 						break;
548 					}
549 					extra += l;
550 				}
551 			} else if (len == length) {
552 				last = data;
553 				break;
554 			}
555 		} else if (len == length) {
556 			u32 w;
557 
558 			/*
559 			 * Need to round up for the last dword in the
560 			 * packet.
561 			 */
562 			w = (len + 3) >> 2;
563 			qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
564 			piobuf += w - 1;
565 			last = ((u32 *) ss->sge.vaddr)[w - 1];
566 			break;
567 		} else {
568 			u32 w = len >> 2;
569 
570 			qib_pio_copy(piobuf, ss->sge.vaddr, w);
571 			piobuf += w;
572 
573 			extra = len & (sizeof(u32) - 1);
574 			if (extra) {
575 				u32 v = ((u32 *) ss->sge.vaddr)[w];
576 
577 				/* Clear unused upper bytes */
578 				data = clear_upper_bytes(v, extra, 0);
579 			}
580 		}
581 		rvt_update_sge(ss, len, false);
582 		length -= len;
583 	}
584 	/* Update address before sending packet. */
585 	rvt_update_sge(ss, length, false);
586 	if (flush_wc) {
587 		/* must flush early everything before trigger word */
588 		qib_flush_wc();
589 		__raw_writel(last, piobuf);
590 		/* be sure trigger word is written */
591 		qib_flush_wc();
592 	} else
593 		__raw_writel(last, piobuf);
594 }
595 
__get_txreq(struct qib_ibdev * dev,struct rvt_qp * qp)596 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
597 					   struct rvt_qp *qp)
598 {
599 	struct qib_qp_priv *priv = qp->priv;
600 	struct qib_verbs_txreq *tx;
601 	unsigned long flags;
602 
603 	spin_lock_irqsave(&qp->s_lock, flags);
604 	spin_lock(&dev->rdi.pending_lock);
605 
606 	if (!list_empty(&dev->txreq_free)) {
607 		struct list_head *l = dev->txreq_free.next;
608 
609 		list_del(l);
610 		spin_unlock(&dev->rdi.pending_lock);
611 		spin_unlock_irqrestore(&qp->s_lock, flags);
612 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
613 	} else {
614 		if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK &&
615 		    list_empty(&priv->iowait)) {
616 			dev->n_txwait++;
617 			qp->s_flags |= RVT_S_WAIT_TX;
618 			list_add_tail(&priv->iowait, &dev->txwait);
619 		}
620 		qp->s_flags &= ~RVT_S_BUSY;
621 		spin_unlock(&dev->rdi.pending_lock);
622 		spin_unlock_irqrestore(&qp->s_lock, flags);
623 		tx = ERR_PTR(-EBUSY);
624 	}
625 	return tx;
626 }
627 
get_txreq(struct qib_ibdev * dev,struct rvt_qp * qp)628 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
629 					 struct rvt_qp *qp)
630 {
631 	struct qib_verbs_txreq *tx;
632 	unsigned long flags;
633 
634 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
635 	/* assume the list non empty */
636 	if (likely(!list_empty(&dev->txreq_free))) {
637 		struct list_head *l = dev->txreq_free.next;
638 
639 		list_del(l);
640 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
641 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
642 	} else {
643 		/* call slow path to get the extra lock */
644 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
645 		tx =  __get_txreq(dev, qp);
646 	}
647 	return tx;
648 }
649 
qib_put_txreq(struct qib_verbs_txreq * tx)650 void qib_put_txreq(struct qib_verbs_txreq *tx)
651 {
652 	struct qib_ibdev *dev;
653 	struct rvt_qp *qp;
654 	struct qib_qp_priv *priv;
655 	unsigned long flags;
656 
657 	qp = tx->qp;
658 	dev = to_idev(qp->ibqp.device);
659 
660 	if (tx->mr) {
661 		rvt_put_mr(tx->mr);
662 		tx->mr = NULL;
663 	}
664 	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
665 		tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
666 		dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
667 				 tx->txreq.addr, tx->hdr_dwords << 2,
668 				 DMA_TO_DEVICE);
669 		kfree(tx->align_buf);
670 	}
671 
672 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
673 
674 	/* Put struct back on free list */
675 	list_add(&tx->txreq.list, &dev->txreq_free);
676 
677 	if (!list_empty(&dev->txwait)) {
678 		/* Wake up first QP wanting a free struct */
679 		priv = list_entry(dev->txwait.next, struct qib_qp_priv,
680 				  iowait);
681 		qp = priv->owner;
682 		list_del_init(&priv->iowait);
683 		rvt_get_qp(qp);
684 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
685 
686 		spin_lock_irqsave(&qp->s_lock, flags);
687 		if (qp->s_flags & RVT_S_WAIT_TX) {
688 			qp->s_flags &= ~RVT_S_WAIT_TX;
689 			qib_schedule_send(qp);
690 		}
691 		spin_unlock_irqrestore(&qp->s_lock, flags);
692 
693 		rvt_put_qp(qp);
694 	} else
695 		spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
696 }
697 
698 /*
699  * This is called when there are send DMA descriptors that might be
700  * available.
701  *
702  * This is called with ppd->sdma_lock held.
703  */
qib_verbs_sdma_desc_avail(struct qib_pportdata * ppd,unsigned avail)704 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
705 {
706 	struct rvt_qp *qp;
707 	struct qib_qp_priv *qpp, *nqpp;
708 	struct rvt_qp *qps[20];
709 	struct qib_ibdev *dev;
710 	unsigned i, n;
711 
712 	n = 0;
713 	dev = &ppd->dd->verbs_dev;
714 	spin_lock(&dev->rdi.pending_lock);
715 
716 	/* Search wait list for first QP wanting DMA descriptors. */
717 	list_for_each_entry_safe(qpp, nqpp, &dev->dmawait, iowait) {
718 		qp = qpp->owner;
719 		if (qp->port_num != ppd->port)
720 			continue;
721 		if (n == ARRAY_SIZE(qps))
722 			break;
723 		if (qpp->s_tx->txreq.sg_count > avail)
724 			break;
725 		avail -= qpp->s_tx->txreq.sg_count;
726 		list_del_init(&qpp->iowait);
727 		rvt_get_qp(qp);
728 		qps[n++] = qp;
729 	}
730 
731 	spin_unlock(&dev->rdi.pending_lock);
732 
733 	for (i = 0; i < n; i++) {
734 		qp = qps[i];
735 		spin_lock(&qp->s_lock);
736 		if (qp->s_flags & RVT_S_WAIT_DMA_DESC) {
737 			qp->s_flags &= ~RVT_S_WAIT_DMA_DESC;
738 			qib_schedule_send(qp);
739 		}
740 		spin_unlock(&qp->s_lock);
741 		rvt_put_qp(qp);
742 	}
743 }
744 
745 /*
746  * This is called with ppd->sdma_lock held.
747  */
sdma_complete(struct qib_sdma_txreq * cookie,int status)748 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
749 {
750 	struct qib_verbs_txreq *tx =
751 		container_of(cookie, struct qib_verbs_txreq, txreq);
752 	struct rvt_qp *qp = tx->qp;
753 	struct qib_qp_priv *priv = qp->priv;
754 
755 	spin_lock(&qp->s_lock);
756 	if (tx->wqe)
757 		qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
758 	else if (qp->ibqp.qp_type == IB_QPT_RC) {
759 		struct ib_header *hdr;
760 
761 		if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
762 			hdr = &tx->align_buf->hdr;
763 		else {
764 			struct qib_ibdev *dev = to_idev(qp->ibqp.device);
765 
766 			hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
767 		}
768 		qib_rc_send_complete(qp, hdr);
769 	}
770 	if (atomic_dec_and_test(&priv->s_dma_busy)) {
771 		if (qp->state == IB_QPS_RESET)
772 			wake_up(&priv->wait_dma);
773 		else if (qp->s_flags & RVT_S_WAIT_DMA) {
774 			qp->s_flags &= ~RVT_S_WAIT_DMA;
775 			qib_schedule_send(qp);
776 		}
777 	}
778 	spin_unlock(&qp->s_lock);
779 
780 	qib_put_txreq(tx);
781 }
782 
wait_kmem(struct qib_ibdev * dev,struct rvt_qp * qp)783 static int wait_kmem(struct qib_ibdev *dev, struct rvt_qp *qp)
784 {
785 	struct qib_qp_priv *priv = qp->priv;
786 	unsigned long flags;
787 	int ret = 0;
788 
789 	spin_lock_irqsave(&qp->s_lock, flags);
790 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
791 		spin_lock(&dev->rdi.pending_lock);
792 		if (list_empty(&priv->iowait)) {
793 			if (list_empty(&dev->memwait))
794 				mod_timer(&dev->mem_timer, jiffies + 1);
795 			qp->s_flags |= RVT_S_WAIT_KMEM;
796 			list_add_tail(&priv->iowait, &dev->memwait);
797 		}
798 		spin_unlock(&dev->rdi.pending_lock);
799 		qp->s_flags &= ~RVT_S_BUSY;
800 		ret = -EBUSY;
801 	}
802 	spin_unlock_irqrestore(&qp->s_lock, flags);
803 
804 	return ret;
805 }
806 
qib_verbs_send_dma(struct rvt_qp * qp,struct ib_header * hdr,u32 hdrwords,struct rvt_sge_state * ss,u32 len,u32 plen,u32 dwords)807 static int qib_verbs_send_dma(struct rvt_qp *qp, struct ib_header *hdr,
808 			      u32 hdrwords, struct rvt_sge_state *ss, u32 len,
809 			      u32 plen, u32 dwords)
810 {
811 	struct qib_qp_priv *priv = qp->priv;
812 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
813 	struct qib_devdata *dd = dd_from_dev(dev);
814 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
815 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
816 	struct qib_verbs_txreq *tx;
817 	struct qib_pio_header *phdr;
818 	u32 control;
819 	u32 ndesc;
820 	int ret;
821 
822 	tx = priv->s_tx;
823 	if (tx) {
824 		priv->s_tx = NULL;
825 		/* resend previously constructed packet */
826 		ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
827 		goto bail;
828 	}
829 
830 	tx = get_txreq(dev, qp);
831 	if (IS_ERR(tx))
832 		goto bail_tx;
833 
834 	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
835 				       be16_to_cpu(hdr->lrh[0]) >> 12);
836 	tx->qp = qp;
837 	tx->wqe = qp->s_wqe;
838 	tx->mr = qp->s_rdma_mr;
839 	if (qp->s_rdma_mr)
840 		qp->s_rdma_mr = NULL;
841 	tx->txreq.callback = sdma_complete;
842 	if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
843 		tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
844 	else
845 		tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
846 	if (plen + 1 > dd->piosize2kmax_dwords)
847 		tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
848 
849 	if (len) {
850 		/*
851 		 * Don't try to DMA if it takes more descriptors than
852 		 * the queue holds.
853 		 */
854 		ndesc = qib_count_sge(ss, len);
855 		if (ndesc >= ppd->sdma_descq_cnt)
856 			ndesc = 0;
857 	} else
858 		ndesc = 1;
859 	if (ndesc) {
860 		phdr = &dev->pio_hdrs[tx->hdr_inx];
861 		phdr->pbc[0] = cpu_to_le32(plen);
862 		phdr->pbc[1] = cpu_to_le32(control);
863 		memcpy(&phdr->hdr, hdr, hdrwords << 2);
864 		tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
865 		tx->txreq.sg_count = ndesc;
866 		tx->txreq.addr = dev->pio_hdrs_phys +
867 			tx->hdr_inx * sizeof(struct qib_pio_header);
868 		tx->hdr_dwords = hdrwords + 2; /* add PBC length */
869 		ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
870 		goto bail;
871 	}
872 
873 	/* Allocate a buffer and copy the header and payload to it. */
874 	tx->hdr_dwords = plen + 1;
875 	phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
876 	if (!phdr)
877 		goto err_tx;
878 	phdr->pbc[0] = cpu_to_le32(plen);
879 	phdr->pbc[1] = cpu_to_le32(control);
880 	memcpy(&phdr->hdr, hdr, hdrwords << 2);
881 	qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
882 
883 	tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
884 					tx->hdr_dwords << 2, DMA_TO_DEVICE);
885 	if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
886 		goto map_err;
887 	tx->align_buf = phdr;
888 	tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
889 	tx->txreq.sg_count = 1;
890 	ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
891 	goto unaligned;
892 
893 map_err:
894 	kfree(phdr);
895 err_tx:
896 	qib_put_txreq(tx);
897 	ret = wait_kmem(dev, qp);
898 unaligned:
899 	ibp->rvp.n_unaligned++;
900 bail:
901 	return ret;
902 bail_tx:
903 	ret = PTR_ERR(tx);
904 	goto bail;
905 }
906 
907 /*
908  * If we are now in the error state, return zero to flush the
909  * send work request.
910  */
no_bufs_available(struct rvt_qp * qp)911 static int no_bufs_available(struct rvt_qp *qp)
912 {
913 	struct qib_qp_priv *priv = qp->priv;
914 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
915 	struct qib_devdata *dd;
916 	unsigned long flags;
917 	int ret = 0;
918 
919 	/*
920 	 * Note that as soon as want_buffer() is called and
921 	 * possibly before it returns, qib_ib_piobufavail()
922 	 * could be called. Therefore, put QP on the I/O wait list before
923 	 * enabling the PIO avail interrupt.
924 	 */
925 	spin_lock_irqsave(&qp->s_lock, flags);
926 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
927 		spin_lock(&dev->rdi.pending_lock);
928 		if (list_empty(&priv->iowait)) {
929 			dev->n_piowait++;
930 			qp->s_flags |= RVT_S_WAIT_PIO;
931 			list_add_tail(&priv->iowait, &dev->piowait);
932 			dd = dd_from_dev(dev);
933 			dd->f_wantpiobuf_intr(dd, 1);
934 		}
935 		spin_unlock(&dev->rdi.pending_lock);
936 		qp->s_flags &= ~RVT_S_BUSY;
937 		ret = -EBUSY;
938 	}
939 	spin_unlock_irqrestore(&qp->s_lock, flags);
940 	return ret;
941 }
942 
qib_verbs_send_pio(struct rvt_qp * qp,struct ib_header * ibhdr,u32 hdrwords,struct rvt_sge_state * ss,u32 len,u32 plen,u32 dwords)943 static int qib_verbs_send_pio(struct rvt_qp *qp, struct ib_header *ibhdr,
944 			      u32 hdrwords, struct rvt_sge_state *ss, u32 len,
945 			      u32 plen, u32 dwords)
946 {
947 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
948 	struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
949 	u32 *hdr = (u32 *) ibhdr;
950 	u32 __iomem *piobuf_orig;
951 	u32 __iomem *piobuf;
952 	u64 pbc;
953 	unsigned long flags;
954 	unsigned flush_wc;
955 	u32 control;
956 	u32 pbufn;
957 
958 	control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
959 		be16_to_cpu(ibhdr->lrh[0]) >> 12);
960 	pbc = ((u64) control << 32) | plen;
961 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
962 	if (unlikely(piobuf == NULL))
963 		return no_bufs_available(qp);
964 
965 	/*
966 	 * Write the pbc.
967 	 * We have to flush after the PBC for correctness on some cpus
968 	 * or WC buffer can be written out of order.
969 	 */
970 	writeq(pbc, piobuf);
971 	piobuf_orig = piobuf;
972 	piobuf += 2;
973 
974 	flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
975 	if (len == 0) {
976 		/*
977 		 * If there is just the header portion, must flush before
978 		 * writing last word of header for correctness, and after
979 		 * the last header word (trigger word).
980 		 */
981 		if (flush_wc) {
982 			qib_flush_wc();
983 			qib_pio_copy(piobuf, hdr, hdrwords - 1);
984 			qib_flush_wc();
985 			__raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
986 			qib_flush_wc();
987 		} else
988 			qib_pio_copy(piobuf, hdr, hdrwords);
989 		goto done;
990 	}
991 
992 	if (flush_wc)
993 		qib_flush_wc();
994 	qib_pio_copy(piobuf, hdr, hdrwords);
995 	piobuf += hdrwords;
996 
997 	/* The common case is aligned and contained in one segment. */
998 	if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
999 		   !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1000 		u32 *addr = (u32 *) ss->sge.vaddr;
1001 
1002 		/* Update address before sending packet. */
1003 		rvt_update_sge(ss, len, false);
1004 		if (flush_wc) {
1005 			qib_pio_copy(piobuf, addr, dwords - 1);
1006 			/* must flush early everything before trigger word */
1007 			qib_flush_wc();
1008 			__raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1009 			/* be sure trigger word is written */
1010 			qib_flush_wc();
1011 		} else
1012 			qib_pio_copy(piobuf, addr, dwords);
1013 		goto done;
1014 	}
1015 	copy_io(piobuf, ss, len, flush_wc);
1016 done:
1017 	if (dd->flags & QIB_USE_SPCL_TRIG) {
1018 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
1019 
1020 		qib_flush_wc();
1021 		__raw_writel(0xaebecede, piobuf_orig + spcl_off);
1022 	}
1023 	qib_sendbuf_done(dd, pbufn);
1024 	if (qp->s_rdma_mr) {
1025 		rvt_put_mr(qp->s_rdma_mr);
1026 		qp->s_rdma_mr = NULL;
1027 	}
1028 	if (qp->s_wqe) {
1029 		spin_lock_irqsave(&qp->s_lock, flags);
1030 		qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1031 		spin_unlock_irqrestore(&qp->s_lock, flags);
1032 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1033 		spin_lock_irqsave(&qp->s_lock, flags);
1034 		qib_rc_send_complete(qp, ibhdr);
1035 		spin_unlock_irqrestore(&qp->s_lock, flags);
1036 	}
1037 	return 0;
1038 }
1039 
1040 /**
1041  * qib_verbs_send - send a packet
1042  * @qp: the QP to send on
1043  * @hdr: the packet header
1044  * @hdrwords: the number of 32-bit words in the header
1045  * @ss: the SGE to send
1046  * @len: the length of the packet in bytes
1047  *
1048  * Return zero if packet is sent or queued OK.
1049  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1050  */
qib_verbs_send(struct rvt_qp * qp,struct ib_header * hdr,u32 hdrwords,struct rvt_sge_state * ss,u32 len)1051 int qib_verbs_send(struct rvt_qp *qp, struct ib_header *hdr,
1052 		   u32 hdrwords, struct rvt_sge_state *ss, u32 len)
1053 {
1054 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1055 	u32 plen;
1056 	int ret;
1057 	u32 dwords = (len + 3) >> 2;
1058 
1059 	/*
1060 	 * Calculate the send buffer trigger address.
1061 	 * The +1 counts for the pbc control dword following the pbc length.
1062 	 */
1063 	plen = hdrwords + dwords + 1;
1064 
1065 	/*
1066 	 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1067 	 * can defer SDMA restart until link goes ACTIVE without
1068 	 * worrying about just how we got there.
1069 	 */
1070 	if (qp->ibqp.qp_type == IB_QPT_SMI ||
1071 	    !(dd->flags & QIB_HAS_SEND_DMA))
1072 		ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1073 					 plen, dwords);
1074 	else
1075 		ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1076 					 plen, dwords);
1077 
1078 	return ret;
1079 }
1080 
qib_snapshot_counters(struct qib_pportdata * ppd,u64 * swords,u64 * rwords,u64 * spkts,u64 * rpkts,u64 * xmit_wait)1081 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1082 			  u64 *rwords, u64 *spkts, u64 *rpkts,
1083 			  u64 *xmit_wait)
1084 {
1085 	int ret;
1086 	struct qib_devdata *dd = ppd->dd;
1087 
1088 	if (!(dd->flags & QIB_PRESENT)) {
1089 		/* no hardware, freeze, etc. */
1090 		ret = -EINVAL;
1091 		goto bail;
1092 	}
1093 	*swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1094 	*rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1095 	*spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1096 	*rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1097 	*xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1098 
1099 	ret = 0;
1100 
1101 bail:
1102 	return ret;
1103 }
1104 
1105 /**
1106  * qib_get_counters - get various chip counters
1107  * @dd: the qlogic_ib device
1108  * @cntrs: counters are placed here
1109  *
1110  * Return the counters needed by recv_pma_get_portcounters().
1111  */
qib_get_counters(struct qib_pportdata * ppd,struct qib_verbs_counters * cntrs)1112 int qib_get_counters(struct qib_pportdata *ppd,
1113 		     struct qib_verbs_counters *cntrs)
1114 {
1115 	int ret;
1116 
1117 	if (!(ppd->dd->flags & QIB_PRESENT)) {
1118 		/* no hardware, freeze, etc. */
1119 		ret = -EINVAL;
1120 		goto bail;
1121 	}
1122 	cntrs->symbol_error_counter =
1123 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1124 	cntrs->link_error_recovery_counter =
1125 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1126 	/*
1127 	 * The link downed counter counts when the other side downs the
1128 	 * connection.  We add in the number of times we downed the link
1129 	 * due to local link integrity errors to compensate.
1130 	 */
1131 	cntrs->link_downed_counter =
1132 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1133 	cntrs->port_rcv_errors =
1134 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1135 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1136 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1137 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1138 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1139 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1140 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1141 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1142 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1143 	cntrs->port_rcv_errors +=
1144 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1145 	cntrs->port_rcv_errors +=
1146 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1147 	cntrs->port_rcv_remphys_errors =
1148 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1149 	cntrs->port_xmit_discards =
1150 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1151 	cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1152 			QIBPORTCNTR_WORDSEND);
1153 	cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1154 			QIBPORTCNTR_WORDRCV);
1155 	cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1156 			QIBPORTCNTR_PKTSEND);
1157 	cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1158 			QIBPORTCNTR_PKTRCV);
1159 	cntrs->local_link_integrity_errors =
1160 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1161 	cntrs->excessive_buffer_overrun_errors =
1162 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1163 	cntrs->vl15_dropped =
1164 		ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1165 
1166 	ret = 0;
1167 
1168 bail:
1169 	return ret;
1170 }
1171 
1172 /**
1173  * qib_ib_piobufavail - callback when a PIO buffer is available
1174  * @dd: the device pointer
1175  *
1176  * This is called from qib_intr() at interrupt level when a PIO buffer is
1177  * available after qib_verbs_send() returned an error that no buffers were
1178  * available. Disable the interrupt if there are no more QPs waiting.
1179  */
qib_ib_piobufavail(struct qib_devdata * dd)1180 void qib_ib_piobufavail(struct qib_devdata *dd)
1181 {
1182 	struct qib_ibdev *dev = &dd->verbs_dev;
1183 	struct list_head *list;
1184 	struct rvt_qp *qps[5];
1185 	struct rvt_qp *qp;
1186 	unsigned long flags;
1187 	unsigned i, n;
1188 	struct qib_qp_priv *priv;
1189 
1190 	list = &dev->piowait;
1191 	n = 0;
1192 
1193 	/*
1194 	 * Note: checking that the piowait list is empty and clearing
1195 	 * the buffer available interrupt needs to be atomic or we
1196 	 * could end up with QPs on the wait list with the interrupt
1197 	 * disabled.
1198 	 */
1199 	spin_lock_irqsave(&dev->rdi.pending_lock, flags);
1200 	while (!list_empty(list)) {
1201 		if (n == ARRAY_SIZE(qps))
1202 			goto full;
1203 		priv = list_entry(list->next, struct qib_qp_priv, iowait);
1204 		qp = priv->owner;
1205 		list_del_init(&priv->iowait);
1206 		rvt_get_qp(qp);
1207 		qps[n++] = qp;
1208 	}
1209 	dd->f_wantpiobuf_intr(dd, 0);
1210 full:
1211 	spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
1212 
1213 	for (i = 0; i < n; i++) {
1214 		qp = qps[i];
1215 
1216 		spin_lock_irqsave(&qp->s_lock, flags);
1217 		if (qp->s_flags & RVT_S_WAIT_PIO) {
1218 			qp->s_flags &= ~RVT_S_WAIT_PIO;
1219 			qib_schedule_send(qp);
1220 		}
1221 		spin_unlock_irqrestore(&qp->s_lock, flags);
1222 
1223 		/* Notify qib_destroy_qp() if it is waiting. */
1224 		rvt_put_qp(qp);
1225 	}
1226 }
1227 
qib_query_port(struct rvt_dev_info * rdi,u8 port_num,struct ib_port_attr * props)1228 static int qib_query_port(struct rvt_dev_info *rdi, u8 port_num,
1229 			  struct ib_port_attr *props)
1230 {
1231 	struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1232 	struct qib_devdata *dd = dd_from_dev(ibdev);
1233 	struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1234 	enum ib_mtu mtu;
1235 	u16 lid = ppd->lid;
1236 
1237 	/* props being zeroed by the caller, avoid zeroing it here */
1238 	props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1239 	props->lmc = ppd->lmc;
1240 	props->state = dd->f_iblink_state(ppd->lastibcstat);
1241 	props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1242 	props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1243 	props->active_width = ppd->link_width_active;
1244 	/* See rate_show() */
1245 	props->active_speed = ppd->link_speed_active;
1246 	props->max_vl_num = qib_num_vls(ppd->vls_supported);
1247 
1248 	props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1249 	switch (ppd->ibmtu) {
1250 	case 4096:
1251 		mtu = IB_MTU_4096;
1252 		break;
1253 	case 2048:
1254 		mtu = IB_MTU_2048;
1255 		break;
1256 	case 1024:
1257 		mtu = IB_MTU_1024;
1258 		break;
1259 	case 512:
1260 		mtu = IB_MTU_512;
1261 		break;
1262 	case 256:
1263 		mtu = IB_MTU_256;
1264 		break;
1265 	default:
1266 		mtu = IB_MTU_2048;
1267 	}
1268 	props->active_mtu = mtu;
1269 
1270 	return 0;
1271 }
1272 
qib_modify_device(struct ib_device * device,int device_modify_mask,struct ib_device_modify * device_modify)1273 static int qib_modify_device(struct ib_device *device,
1274 			     int device_modify_mask,
1275 			     struct ib_device_modify *device_modify)
1276 {
1277 	struct qib_devdata *dd = dd_from_ibdev(device);
1278 	unsigned i;
1279 	int ret;
1280 
1281 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1282 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1283 		ret = -EOPNOTSUPP;
1284 		goto bail;
1285 	}
1286 
1287 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1288 		memcpy(device->node_desc, device_modify->node_desc,
1289 		       IB_DEVICE_NODE_DESC_MAX);
1290 		for (i = 0; i < dd->num_pports; i++) {
1291 			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1292 
1293 			qib_node_desc_chg(ibp);
1294 		}
1295 	}
1296 
1297 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1298 		ib_qib_sys_image_guid =
1299 			cpu_to_be64(device_modify->sys_image_guid);
1300 		for (i = 0; i < dd->num_pports; i++) {
1301 			struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1302 
1303 			qib_sys_guid_chg(ibp);
1304 		}
1305 	}
1306 
1307 	ret = 0;
1308 
1309 bail:
1310 	return ret;
1311 }
1312 
qib_shut_down_port(struct rvt_dev_info * rdi,u8 port_num)1313 static int qib_shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1314 {
1315 	struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1316 	struct qib_devdata *dd = dd_from_dev(ibdev);
1317 	struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1318 
1319 	qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1320 
1321 	return 0;
1322 }
1323 
qib_get_guid_be(struct rvt_dev_info * rdi,struct rvt_ibport * rvp,int guid_index,__be64 * guid)1324 static int qib_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1325 			   int guid_index, __be64 *guid)
1326 {
1327 	struct qib_ibport *ibp = container_of(rvp, struct qib_ibport, rvp);
1328 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1329 
1330 	if (guid_index == 0)
1331 		*guid = ppd->guid;
1332 	else if (guid_index < QIB_GUIDS_PER_PORT)
1333 		*guid = ibp->guids[guid_index - 1];
1334 	else
1335 		return -EINVAL;
1336 
1337 	return 0;
1338 }
1339 
qib_check_ah(struct ib_device * ibdev,struct rdma_ah_attr * ah_attr)1340 int qib_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1341 {
1342 	if (rdma_ah_get_sl(ah_attr) > 15)
1343 		return -EINVAL;
1344 
1345 	if (rdma_ah_get_dlid(ah_attr) == 0)
1346 		return -EINVAL;
1347 	if (rdma_ah_get_dlid(ah_attr) >=
1348 		be16_to_cpu(IB_MULTICAST_LID_BASE) &&
1349 	    rdma_ah_get_dlid(ah_attr) !=
1350 		be16_to_cpu(IB_LID_PERMISSIVE) &&
1351 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1352 		return -EINVAL;
1353 
1354 	return 0;
1355 }
1356 
qib_notify_new_ah(struct ib_device * ibdev,struct rdma_ah_attr * ah_attr,struct rvt_ah * ah)1357 static void qib_notify_new_ah(struct ib_device *ibdev,
1358 			      struct rdma_ah_attr *ah_attr,
1359 			      struct rvt_ah *ah)
1360 {
1361 	struct qib_ibport *ibp;
1362 	struct qib_pportdata *ppd;
1363 
1364 	/*
1365 	 * Do not trust reading anything from rvt_ah at this point as it is not
1366 	 * done being setup. We can however modify things which we need to set.
1367 	 */
1368 
1369 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1370 	ppd = ppd_from_ibp(ibp);
1371 	ah->vl = ibp->sl_to_vl[rdma_ah_get_sl(&ah->attr)];
1372 	ah->log_pmtu = ilog2(ppd->ibmtu);
1373 }
1374 
qib_create_qp0_ah(struct qib_ibport * ibp,u16 dlid)1375 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1376 {
1377 	struct rdma_ah_attr attr;
1378 	struct ib_ah *ah = ERR_PTR(-EINVAL);
1379 	struct rvt_qp *qp0;
1380 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1381 	struct qib_devdata *dd = dd_from_ppd(ppd);
1382 	u8 port_num = ppd->port;
1383 
1384 	memset(&attr, 0, sizeof(attr));
1385 	attr.type = rdma_ah_find_type(&dd->verbs_dev.rdi.ibdev, port_num);
1386 	rdma_ah_set_dlid(&attr, dlid);
1387 	rdma_ah_set_port_num(&attr, port_num);
1388 	rcu_read_lock();
1389 	qp0 = rcu_dereference(ibp->rvp.qp[0]);
1390 	if (qp0)
1391 		ah = rdma_create_ah(qp0->ibqp.pd, &attr);
1392 	rcu_read_unlock();
1393 	return ah;
1394 }
1395 
1396 /**
1397  * qib_get_npkeys - return the size of the PKEY table for context 0
1398  * @dd: the qlogic_ib device
1399  */
qib_get_npkeys(struct qib_devdata * dd)1400 unsigned qib_get_npkeys(struct qib_devdata *dd)
1401 {
1402 	return ARRAY_SIZE(dd->rcd[0]->pkeys);
1403 }
1404 
1405 /*
1406  * Return the indexed PKEY from the port PKEY table.
1407  * No need to validate rcd[ctxt]; the port is setup if we are here.
1408  */
qib_get_pkey(struct qib_ibport * ibp,unsigned index)1409 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1410 {
1411 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1412 	struct qib_devdata *dd = ppd->dd;
1413 	unsigned ctxt = ppd->hw_pidx;
1414 	unsigned ret;
1415 
1416 	/* dd->rcd null if mini_init or some init failures */
1417 	if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1418 		ret = 0;
1419 	else
1420 		ret = dd->rcd[ctxt]->pkeys[index];
1421 
1422 	return ret;
1423 }
1424 
init_ibport(struct qib_pportdata * ppd)1425 static void init_ibport(struct qib_pportdata *ppd)
1426 {
1427 	struct qib_verbs_counters cntrs;
1428 	struct qib_ibport *ibp = &ppd->ibport_data;
1429 
1430 	spin_lock_init(&ibp->rvp.lock);
1431 	/* Set the prefix to the default value (see ch. 4.1.1) */
1432 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1433 	ibp->rvp.sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
1434 	ibp->rvp.port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
1435 		IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
1436 		IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
1437 		IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
1438 		IB_PORT_OTHER_LOCAL_CHANGES_SUP;
1439 	if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
1440 		ibp->rvp.port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
1441 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1442 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1443 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1444 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1445 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1446 
1447 	/* Snapshot current HW counters to "clear" them. */
1448 	qib_get_counters(ppd, &cntrs);
1449 	ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
1450 	ibp->z_link_error_recovery_counter =
1451 		cntrs.link_error_recovery_counter;
1452 	ibp->z_link_downed_counter = cntrs.link_downed_counter;
1453 	ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
1454 	ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
1455 	ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
1456 	ibp->z_port_xmit_data = cntrs.port_xmit_data;
1457 	ibp->z_port_rcv_data = cntrs.port_rcv_data;
1458 	ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
1459 	ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
1460 	ibp->z_local_link_integrity_errors =
1461 		cntrs.local_link_integrity_errors;
1462 	ibp->z_excessive_buffer_overrun_errors =
1463 		cntrs.excessive_buffer_overrun_errors;
1464 	ibp->z_vl15_dropped = cntrs.vl15_dropped;
1465 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1466 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1467 }
1468 
1469 /**
1470  * qib_fill_device_attr - Fill in rvt dev info device attributes.
1471  * @dd: the device data structure
1472  */
qib_fill_device_attr(struct qib_devdata * dd)1473 static void qib_fill_device_attr(struct qib_devdata *dd)
1474 {
1475 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1476 
1477 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1478 
1479 	rdi->dparms.props.max_pd = ib_qib_max_pds;
1480 	rdi->dparms.props.max_ah = ib_qib_max_ahs;
1481 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1482 		IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1483 		IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1484 		IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1485 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1486 	rdi->dparms.props.vendor_id =
1487 		QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1488 	rdi->dparms.props.vendor_part_id = dd->deviceid;
1489 	rdi->dparms.props.hw_ver = dd->minrev;
1490 	rdi->dparms.props.sys_image_guid = ib_qib_sys_image_guid;
1491 	rdi->dparms.props.max_mr_size = ~0ULL;
1492 	rdi->dparms.props.max_qp = ib_qib_max_qps;
1493 	rdi->dparms.props.max_qp_wr = ib_qib_max_qp_wrs;
1494 	rdi->dparms.props.max_send_sge = ib_qib_max_sges;
1495 	rdi->dparms.props.max_recv_sge = ib_qib_max_sges;
1496 	rdi->dparms.props.max_sge_rd = ib_qib_max_sges;
1497 	rdi->dparms.props.max_cq = ib_qib_max_cqs;
1498 	rdi->dparms.props.max_cqe = ib_qib_max_cqes;
1499 	rdi->dparms.props.max_ah = ib_qib_max_ahs;
1500 	rdi->dparms.props.max_map_per_fmr = 32767;
1501 	rdi->dparms.props.max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1502 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1503 	rdi->dparms.props.max_srq = ib_qib_max_srqs;
1504 	rdi->dparms.props.max_srq_wr = ib_qib_max_srq_wrs;
1505 	rdi->dparms.props.max_srq_sge = ib_qib_max_srq_sges;
1506 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1507 	rdi->dparms.props.max_pkeys = qib_get_npkeys(dd);
1508 	rdi->dparms.props.max_mcast_grp = ib_qib_max_mcast_grps;
1509 	rdi->dparms.props.max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1510 	rdi->dparms.props.max_total_mcast_qp_attach =
1511 					rdi->dparms.props.max_mcast_qp_attach *
1512 					rdi->dparms.props.max_mcast_grp;
1513 	/* post send table */
1514 	dd->verbs_dev.rdi.post_parms = qib_post_parms;
1515 }
1516 
1517 /**
1518  * qib_register_ib_device - register our device with the infiniband core
1519  * @dd: the device data structure
1520  * Return the allocated qib_ibdev pointer or NULL on error.
1521  */
qib_register_ib_device(struct qib_devdata * dd)1522 int qib_register_ib_device(struct qib_devdata *dd)
1523 {
1524 	struct qib_ibdev *dev = &dd->verbs_dev;
1525 	struct ib_device *ibdev = &dev->rdi.ibdev;
1526 	struct qib_pportdata *ppd = dd->pport;
1527 	unsigned i, ctxt;
1528 	int ret;
1529 
1530 	get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
1531 	for (i = 0; i < dd->num_pports; i++)
1532 		init_ibport(ppd + i);
1533 
1534 	/* Only need to initialize non-zero fields. */
1535 	timer_setup(&dev->mem_timer, mem_timer, 0);
1536 
1537 	INIT_LIST_HEAD(&dev->piowait);
1538 	INIT_LIST_HEAD(&dev->dmawait);
1539 	INIT_LIST_HEAD(&dev->txwait);
1540 	INIT_LIST_HEAD(&dev->memwait);
1541 	INIT_LIST_HEAD(&dev->txreq_free);
1542 
1543 	if (ppd->sdma_descq_cnt) {
1544 		dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
1545 						ppd->sdma_descq_cnt *
1546 						sizeof(struct qib_pio_header),
1547 						&dev->pio_hdrs_phys,
1548 						GFP_KERNEL);
1549 		if (!dev->pio_hdrs) {
1550 			ret = -ENOMEM;
1551 			goto err_hdrs;
1552 		}
1553 	}
1554 
1555 	for (i = 0; i < ppd->sdma_descq_cnt; i++) {
1556 		struct qib_verbs_txreq *tx;
1557 
1558 		tx = kzalloc(sizeof(*tx), GFP_KERNEL);
1559 		if (!tx) {
1560 			ret = -ENOMEM;
1561 			goto err_tx;
1562 		}
1563 		tx->hdr_inx = i;
1564 		list_add(&tx->txreq.list, &dev->txreq_free);
1565 	}
1566 
1567 	/*
1568 	 * The system image GUID is supposed to be the same for all
1569 	 * IB HCAs in a single system but since there can be other
1570 	 * device types in the system, we can't be sure this is unique.
1571 	 */
1572 	if (!ib_qib_sys_image_guid)
1573 		ib_qib_sys_image_guid = ppd->guid;
1574 
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_pci_dev = qib_get_pci_dev;
1590 	dd->verbs_dev.rdi.driver_f.check_ah = qib_check_ah;
1591 	dd->verbs_dev.rdi.driver_f.check_send_wqe = qib_check_send_wqe;
1592 	dd->verbs_dev.rdi.driver_f.notify_new_ah = qib_notify_new_ah;
1593 	dd->verbs_dev.rdi.driver_f.alloc_qpn = qib_alloc_qpn;
1594 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qib_qp_priv_alloc;
1595 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qib_qp_priv_free;
1596 	dd->verbs_dev.rdi.driver_f.free_all_qps = qib_free_all_qps;
1597 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = qib_notify_qp_reset;
1598 	dd->verbs_dev.rdi.driver_f.do_send = qib_do_send;
1599 	dd->verbs_dev.rdi.driver_f.schedule_send = qib_schedule_send;
1600 	dd->verbs_dev.rdi.driver_f.quiesce_qp = qib_quiesce_qp;
1601 	dd->verbs_dev.rdi.driver_f.stop_send_queue = qib_stop_send_queue;
1602 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = qib_flush_qp_waiters;
1603 	dd->verbs_dev.rdi.driver_f.notify_error_qp = qib_notify_error_qp;
1604 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = qib_restart_rc;
1605 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = qib_mtu_to_path_mtu;
1606 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = qib_mtu_from_qp;
1607 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = qib_get_pmtu_from_attr;
1608 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _qib_schedule_send;
1609 	dd->verbs_dev.rdi.driver_f.query_port_state = qib_query_port;
1610 	dd->verbs_dev.rdi.driver_f.shut_down_port = qib_shut_down_port;
1611 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = qib_cap_mask_chg;
1612 	dd->verbs_dev.rdi.driver_f.notify_create_mad_agent =
1613 						qib_notify_create_mad_agent;
1614 	dd->verbs_dev.rdi.driver_f.notify_free_mad_agent =
1615 						qib_notify_free_mad_agent;
1616 
1617 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = QIB_MAX_RDMA_ATOMIC;
1618 	dd->verbs_dev.rdi.driver_f.get_guid_be = qib_get_guid_be;
1619 	dd->verbs_dev.rdi.dparms.lkey_table_size = qib_lkey_table_size;
1620 	dd->verbs_dev.rdi.dparms.qp_table_size = ib_qib_qp_table_size;
1621 	dd->verbs_dev.rdi.dparms.qpn_start = 1;
1622 	dd->verbs_dev.rdi.dparms.qpn_res_start = QIB_KD_QP;
1623 	dd->verbs_dev.rdi.dparms.qpn_res_end = QIB_KD_QP; /* Reserve one QP */
1624 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1625 	dd->verbs_dev.rdi.dparms.qos_shift = 1;
1626 	dd->verbs_dev.rdi.dparms.psn_mask = QIB_PSN_MASK;
1627 	dd->verbs_dev.rdi.dparms.psn_shift = QIB_PSN_SHIFT;
1628 	dd->verbs_dev.rdi.dparms.psn_modify_mask = QIB_PSN_MASK;
1629 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1630 	dd->verbs_dev.rdi.dparms.npkeys = qib_get_npkeys(dd);
1631 	dd->verbs_dev.rdi.dparms.node = dd->assigned_node_id;
1632 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1633 	dd->verbs_dev.rdi.dparms.max_mad_size = IB_MGMT_MAD_SIZE;
1634 
1635 	qib_fill_device_attr(dd);
1636 
1637 	ppd = dd->pport;
1638 	for (i = 0; i < dd->num_pports; i++, ppd++) {
1639 		ctxt = ppd->hw_pidx;
1640 		rvt_init_port(&dd->verbs_dev.rdi,
1641 			      &ppd->ibport_data.rvp,
1642 			      i,
1643 			      dd->rcd[ctxt]->pkeys);
1644 	}
1645 
1646 	ret = rvt_register_device(&dd->verbs_dev.rdi, RDMA_DRIVER_QIB);
1647 	if (ret)
1648 		goto err_tx;
1649 
1650 	ret = qib_verbs_register_sysfs(dd);
1651 	if (ret)
1652 		goto err_class;
1653 
1654 	return ret;
1655 
1656 err_class:
1657 	rvt_unregister_device(&dd->verbs_dev.rdi);
1658 err_tx:
1659 	while (!list_empty(&dev->txreq_free)) {
1660 		struct list_head *l = dev->txreq_free.next;
1661 		struct qib_verbs_txreq *tx;
1662 
1663 		list_del(l);
1664 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1665 		kfree(tx);
1666 	}
1667 	if (ppd->sdma_descq_cnt)
1668 		dma_free_coherent(&dd->pcidev->dev,
1669 				  ppd->sdma_descq_cnt *
1670 					sizeof(struct qib_pio_header),
1671 				  dev->pio_hdrs, dev->pio_hdrs_phys);
1672 err_hdrs:
1673 	qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1674 	return ret;
1675 }
1676 
qib_unregister_ib_device(struct qib_devdata * dd)1677 void qib_unregister_ib_device(struct qib_devdata *dd)
1678 {
1679 	struct qib_ibdev *dev = &dd->verbs_dev;
1680 
1681 	qib_verbs_unregister_sysfs(dd);
1682 
1683 	rvt_unregister_device(&dd->verbs_dev.rdi);
1684 
1685 	if (!list_empty(&dev->piowait))
1686 		qib_dev_err(dd, "piowait list not empty!\n");
1687 	if (!list_empty(&dev->dmawait))
1688 		qib_dev_err(dd, "dmawait list not empty!\n");
1689 	if (!list_empty(&dev->txwait))
1690 		qib_dev_err(dd, "txwait list not empty!\n");
1691 	if (!list_empty(&dev->memwait))
1692 		qib_dev_err(dd, "memwait list not empty!\n");
1693 
1694 	del_timer_sync(&dev->mem_timer);
1695 	while (!list_empty(&dev->txreq_free)) {
1696 		struct list_head *l = dev->txreq_free.next;
1697 		struct qib_verbs_txreq *tx;
1698 
1699 		list_del(l);
1700 		tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1701 		kfree(tx);
1702 	}
1703 	if (dd->pport->sdma_descq_cnt)
1704 		dma_free_coherent(&dd->pcidev->dev,
1705 				  dd->pport->sdma_descq_cnt *
1706 					sizeof(struct qib_pio_header),
1707 				  dev->pio_hdrs, dev->pio_hdrs_phys);
1708 }
1709 
1710 /**
1711  * _qib_schedule_send - schedule progress
1712  * @qp - the qp
1713  *
1714  * This schedules progress w/o regard to the s_flags.
1715  *
1716  * It is only used in post send, which doesn't hold
1717  * the s_lock.
1718  */
_qib_schedule_send(struct rvt_qp * qp)1719 void _qib_schedule_send(struct rvt_qp *qp)
1720 {
1721 	struct qib_ibport *ibp =
1722 		to_iport(qp->ibqp.device, qp->port_num);
1723 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1724 	struct qib_qp_priv *priv = qp->priv;
1725 
1726 	queue_work(ppd->qib_wq, &priv->s_work);
1727 }
1728 
1729 /**
1730  * qib_schedule_send - schedule progress
1731  * @qp - the qp
1732  *
1733  * This schedules qp progress.  The s_lock
1734  * should be held.
1735  */
qib_schedule_send(struct rvt_qp * qp)1736 void qib_schedule_send(struct rvt_qp *qp)
1737 {
1738 	if (qib_send_ok(qp))
1739 		_qib_schedule_send(qp);
1740 }
1741