• Home
  • History
  • Annotate
  • Line#
  • Navigate
  • Raw
  • Download
  • only in /netgear-R7000-V1.0.7.12_1.2.5/components/opensource/linux/linux-2.6.36/drivers/infiniband/hw/ipath/
1/*
2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses.  You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 *     Redistribution and use in source and binary forms, with or
12 *     without modification, are permitted provided that the following
13 *     conditions are met:
14 *
15 *      - Redistributions of source code must retain the above
16 *        copyright notice, this list of conditions and the following
17 *        disclaimer.
18 *
19 *      - Redistributions in binary form must reproduce the above
20 *        copyright notice, this list of conditions and the following
21 *        disclaimer in the documentation and/or other materials
22 *        provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34#include <linux/sched.h>
35#include <rdma/ib_smi.h>
36
37#include "ipath_verbs.h"
38#include "ipath_kernel.h"
39
40/**
41 * ipath_ud_loopback - handle send on loopback QPs
42 * @sqp: the sending QP
43 * @swqe: the send work request
44 *
45 * This is called from ipath_make_ud_req() to forward a WQE addressed
46 * to the same HCA.
47 * Note that the receive interrupt handler may be calling ipath_ud_rcv()
48 * while this is being called.
49 */
50static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
51{
52	struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
53	struct ipath_qp *qp;
54	struct ib_ah_attr *ah_attr;
55	unsigned long flags;
56	struct ipath_rq *rq;
57	struct ipath_srq *srq;
58	struct ipath_sge_state rsge;
59	struct ipath_sge *sge;
60	struct ipath_rwq *wq;
61	struct ipath_rwqe *wqe;
62	void (*handler)(struct ib_event *, void *);
63	struct ib_wc wc;
64	u32 tail;
65	u32 rlen;
66	u32 length;
67
68	qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
69	if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
70		dev->n_pkt_drops++;
71		goto done;
72	}
73
74	/*
75	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
76	 * Qkeys with the high order bit set mean use the
77	 * qkey from the QP context instead of the WR (see 10.2.5).
78	 */
79	if (unlikely(qp->ibqp.qp_num &&
80		     ((int) swqe->wr.wr.ud.remote_qkey < 0 ?
81		      sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
82		dev->qkey_violations++;
83		dev->n_pkt_drops++;
84		goto drop;
85	}
86
87	/*
88	 * A GRH is expected to preceed the data even if not
89	 * present on the wire.
90	 */
91	length = swqe->length;
92	memset(&wc, 0, sizeof wc);
93	wc.byte_len = length + sizeof(struct ib_grh);
94
95	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
96		wc.wc_flags = IB_WC_WITH_IMM;
97		wc.ex.imm_data = swqe->wr.ex.imm_data;
98	}
99
100	/*
101	 * This would be a lot simpler if we could call ipath_get_rwqe()
102	 * but that uses state that the receive interrupt handler uses
103	 * so we would need to lock out receive interrupts while doing
104	 * local loopback.
105	 */
106	if (qp->ibqp.srq) {
107		srq = to_isrq(qp->ibqp.srq);
108		handler = srq->ibsrq.event_handler;
109		rq = &srq->rq;
110	} else {
111		srq = NULL;
112		handler = NULL;
113		rq = &qp->r_rq;
114	}
115
116	/*
117	 * Get the next work request entry to find where to put the data.
118	 * Note that it is safe to drop the lock after changing rq->tail
119	 * since ipath_post_receive() won't fill the empty slot.
120	 */
121	spin_lock_irqsave(&rq->lock, flags);
122	wq = rq->wq;
123	tail = wq->tail;
124	/* Validate tail before using it since it is user writable. */
125	if (tail >= rq->size)
126		tail = 0;
127	if (unlikely(tail == wq->head)) {
128		spin_unlock_irqrestore(&rq->lock, flags);
129		dev->n_pkt_drops++;
130		goto drop;
131	}
132	wqe = get_rwqe_ptr(rq, tail);
133	rsge.sg_list = qp->r_ud_sg_list;
134	if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
135		spin_unlock_irqrestore(&rq->lock, flags);
136		dev->n_pkt_drops++;
137		goto drop;
138	}
139	/* Silently drop packets which are too big. */
140	if (wc.byte_len > rlen) {
141		spin_unlock_irqrestore(&rq->lock, flags);
142		dev->n_pkt_drops++;
143		goto drop;
144	}
145	if (++tail >= rq->size)
146		tail = 0;
147	wq->tail = tail;
148	wc.wr_id = wqe->wr_id;
149	if (handler) {
150		u32 n;
151
152		/*
153		 * validate head pointer value and compute
154		 * the number of remaining WQEs.
155		 */
156		n = wq->head;
157		if (n >= rq->size)
158			n = 0;
159		if (n < tail)
160			n += rq->size - tail;
161		else
162			n -= tail;
163		if (n < srq->limit) {
164			struct ib_event ev;
165
166			srq->limit = 0;
167			spin_unlock_irqrestore(&rq->lock, flags);
168			ev.device = qp->ibqp.device;
169			ev.element.srq = qp->ibqp.srq;
170			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
171			handler(&ev, srq->ibsrq.srq_context);
172		} else
173			spin_unlock_irqrestore(&rq->lock, flags);
174	} else
175		spin_unlock_irqrestore(&rq->lock, flags);
176
177	ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
178	if (ah_attr->ah_flags & IB_AH_GRH) {
179		ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
180		wc.wc_flags |= IB_WC_GRH;
181	} else
182		ipath_skip_sge(&rsge, sizeof(struct ib_grh));
183	sge = swqe->sg_list;
184	while (length) {
185		u32 len = sge->length;
186
187		if (len > length)
188			len = length;
189		if (len > sge->sge_length)
190			len = sge->sge_length;
191		BUG_ON(len == 0);
192		ipath_copy_sge(&rsge, sge->vaddr, len);
193		sge->vaddr += len;
194		sge->length -= len;
195		sge->sge_length -= len;
196		if (sge->sge_length == 0) {
197			if (--swqe->wr.num_sge)
198				sge++;
199		} else if (sge->length == 0 && sge->mr != NULL) {
200			if (++sge->n >= IPATH_SEGSZ) {
201				if (++sge->m >= sge->mr->mapsz)
202					break;
203				sge->n = 0;
204			}
205			sge->vaddr =
206				sge->mr->map[sge->m]->segs[sge->n].vaddr;
207			sge->length =
208				sge->mr->map[sge->m]->segs[sge->n].length;
209		}
210		length -= len;
211	}
212	wc.status = IB_WC_SUCCESS;
213	wc.opcode = IB_WC_RECV;
214	wc.qp = &qp->ibqp;
215	wc.src_qp = sqp->ibqp.qp_num;
216	wc.pkey_index = 0;
217	wc.slid = dev->dd->ipath_lid |
218		(ah_attr->src_path_bits &
219		 ((1 << dev->dd->ipath_lmc) - 1));
220	wc.sl = ah_attr->sl;
221	wc.dlid_path_bits =
222		ah_attr->dlid & ((1 << dev->dd->ipath_lmc) - 1);
223	wc.port_num = 1;
224	/* Signal completion event if the solicited bit is set. */
225	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
226		       swqe->wr.send_flags & IB_SEND_SOLICITED);
227drop:
228	if (atomic_dec_and_test(&qp->refcount))
229		wake_up(&qp->wait);
230done:;
231}
232
233/**
234 * ipath_make_ud_req - construct a UD request packet
235 * @qp: the QP
236 *
237 * Return 1 if constructed; otherwise, return 0.
238 */
239int ipath_make_ud_req(struct ipath_qp *qp)
240{
241	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
242	struct ipath_other_headers *ohdr;
243	struct ib_ah_attr *ah_attr;
244	struct ipath_swqe *wqe;
245	unsigned long flags;
246	u32 nwords;
247	u32 extra_bytes;
248	u32 bth0;
249	u16 lrh0;
250	u16 lid;
251	int ret = 0;
252	int next_cur;
253
254	spin_lock_irqsave(&qp->s_lock, flags);
255
256	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
257		if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
258			goto bail;
259		/* We are in the error state, flush the work request. */
260		if (qp->s_last == qp->s_head)
261			goto bail;
262		/* If DMAs are in progress, we can't flush immediately. */
263		if (atomic_read(&qp->s_dma_busy)) {
264			qp->s_flags |= IPATH_S_WAIT_DMA;
265			goto bail;
266		}
267		wqe = get_swqe_ptr(qp, qp->s_last);
268		ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
269		goto done;
270	}
271
272	if (qp->s_cur == qp->s_head)
273		goto bail;
274
275	wqe = get_swqe_ptr(qp, qp->s_cur);
276	next_cur = qp->s_cur + 1;
277	if (next_cur >= qp->s_size)
278		next_cur = 0;
279
280	/* Construct the header. */
281	ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
282	if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
283		if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
284			dev->n_multicast_xmit++;
285		else
286			dev->n_unicast_xmit++;
287	} else {
288		dev->n_unicast_xmit++;
289		lid = ah_attr->dlid & ~((1 << dev->dd->ipath_lmc) - 1);
290		if (unlikely(lid == dev->dd->ipath_lid)) {
291			if (atomic_read(&qp->s_dma_busy)) {
292				qp->s_flags |= IPATH_S_WAIT_DMA;
293				goto bail;
294			}
295			qp->s_cur = next_cur;
296			spin_unlock_irqrestore(&qp->s_lock, flags);
297			ipath_ud_loopback(qp, wqe);
298			spin_lock_irqsave(&qp->s_lock, flags);
299			ipath_send_complete(qp, wqe, IB_WC_SUCCESS);
300			goto done;
301		}
302	}
303
304	qp->s_cur = next_cur;
305	extra_bytes = -wqe->length & 3;
306	nwords = (wqe->length + extra_bytes) >> 2;
307
308	/* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
309	qp->s_hdrwords = 7;
310	qp->s_cur_size = wqe->length;
311	qp->s_cur_sge = &qp->s_sge;
312	qp->s_dmult = ah_attr->static_rate;
313	qp->s_wqe = wqe;
314	qp->s_sge.sge = wqe->sg_list[0];
315	qp->s_sge.sg_list = wqe->sg_list + 1;
316	qp->s_sge.num_sge = wqe->wr.num_sge;
317
318	if (ah_attr->ah_flags & IB_AH_GRH) {
319		/* Header size in 32-bit words. */
320		qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
321						 &ah_attr->grh,
322						 qp->s_hdrwords, nwords);
323		lrh0 = IPATH_LRH_GRH;
324		ohdr = &qp->s_hdr.u.l.oth;
325		/*
326		 * Don't worry about sending to locally attached multicast
327		 * QPs.  It is unspecified by the spec. what happens.
328		 */
329	} else {
330		/* Header size in 32-bit words. */
331		lrh0 = IPATH_LRH_BTH;
332		ohdr = &qp->s_hdr.u.oth;
333	}
334	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
335		qp->s_hdrwords++;
336		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
337		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
338	} else
339		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
340	lrh0 |= ah_attr->sl << 4;
341	if (qp->ibqp.qp_type == IB_QPT_SMI)
342		lrh0 |= 0xF000;	/* Set VL (see ch. 13.5.3.1) */
343	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
344	qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);	/* DEST LID */
345	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
346					   SIZE_OF_CRC);
347	lid = dev->dd->ipath_lid;
348	if (lid) {
349		lid |= ah_attr->src_path_bits &
350			((1 << dev->dd->ipath_lmc) - 1);
351		qp->s_hdr.lrh[3] = cpu_to_be16(lid);
352	} else
353		qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
354	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
355		bth0 |= 1 << 23;
356	bth0 |= extra_bytes << 20;
357	bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
358		ipath_get_pkey(dev->dd, qp->s_pkey_index);
359	ohdr->bth[0] = cpu_to_be32(bth0);
360	/*
361	 * Use the multicast QP if the destination LID is a multicast LID.
362	 */
363	ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
364		ah_attr->dlid != IPATH_PERMISSIVE_LID ?
365		cpu_to_be32(IPATH_MULTICAST_QPN) :
366		cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
367	ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
368	/*
369	 * Qkeys with the high order bit set mean use the
370	 * qkey from the QP context instead of the WR (see 10.2.5).
371	 */
372	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
373					 qp->qkey : wqe->wr.wr.ud.remote_qkey);
374	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
375
376done:
377	ret = 1;
378	goto unlock;
379
380bail:
381	qp->s_flags &= ~IPATH_S_BUSY;
382unlock:
383	spin_unlock_irqrestore(&qp->s_lock, flags);
384	return ret;
385}
386
387/**
388 * ipath_ud_rcv - receive an incoming UD packet
389 * @dev: the device the packet came in on
390 * @hdr: the packet header
391 * @has_grh: true if the packet has a GRH
392 * @data: the packet data
393 * @tlen: the packet length
394 * @qp: the QP the packet came on
395 *
396 * This is called from ipath_qp_rcv() to process an incoming UD packet
397 * for the given QP.
398 * Called at interrupt level.
399 */
400void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
401		  int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
402{
403	struct ipath_other_headers *ohdr;
404	int opcode;
405	u32 hdrsize;
406	u32 pad;
407	struct ib_wc wc;
408	u32 qkey;
409	u32 src_qp;
410	u16 dlid;
411	int header_in_data;
412
413	/* Check for GRH */
414	if (!has_grh) {
415		ohdr = &hdr->u.oth;
416		hdrsize = 8 + 12 + 8;	/* LRH + BTH + DETH */
417		qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
418		src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
419		header_in_data = 0;
420	} else {
421		ohdr = &hdr->u.l.oth;
422		hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
423		/*
424		 * The header with GRH is 68 bytes and the core driver sets
425		 * the eager header buffer size to 56 bytes so the last 12
426		 * bytes of the IB header is in the data buffer.
427		 */
428		header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
429		if (header_in_data) {
430			qkey = be32_to_cpu(((__be32 *) data)[1]);
431			src_qp = be32_to_cpu(((__be32 *) data)[2]);
432			data += 12;
433		} else {
434			qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
435			src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
436		}
437	}
438	src_qp &= IPATH_QPN_MASK;
439
440	/*
441	 * Check that the permissive LID is only used on QP0
442	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
443	 */
444	if (qp->ibqp.qp_num) {
445		if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
446			     hdr->lrh[3] == IB_LID_PERMISSIVE)) {
447			dev->n_pkt_drops++;
448			goto bail;
449		}
450		if (unlikely(qkey != qp->qkey)) {
451			dev->qkey_violations++;
452			dev->n_pkt_drops++;
453			goto bail;
454		}
455	} else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
456		   hdr->lrh[3] == IB_LID_PERMISSIVE) {
457		struct ib_smp *smp = (struct ib_smp *) data;
458
459		if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
460			dev->n_pkt_drops++;
461			goto bail;
462		}
463	}
464
465	/*
466	 * The opcode is in the low byte when its in network order
467	 * (top byte when in host order).
468	 */
469	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
470	if (qp->ibqp.qp_num > 1 &&
471	    opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
472		if (header_in_data) {
473			wc.ex.imm_data = *(__be32 *) data;
474			data += sizeof(__be32);
475		} else
476			wc.ex.imm_data = ohdr->u.ud.imm_data;
477		wc.wc_flags = IB_WC_WITH_IMM;
478		hdrsize += sizeof(u32);
479	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
480		wc.ex.imm_data = 0;
481		wc.wc_flags = 0;
482	} else {
483		dev->n_pkt_drops++;
484		goto bail;
485	}
486
487	/* Get the number of bytes the message was padded by. */
488	pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
489	if (unlikely(tlen < (hdrsize + pad + 4))) {
490		/* Drop incomplete packets. */
491		dev->n_pkt_drops++;
492		goto bail;
493	}
494	tlen -= hdrsize + pad + 4;
495
496	/* Drop invalid MAD packets (see 13.5.3.1). */
497	if (unlikely((qp->ibqp.qp_num == 0 &&
498		      (tlen != 256 ||
499		       (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
500		     (qp->ibqp.qp_num == 1 &&
501		      (tlen != 256 ||
502		       (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
503		dev->n_pkt_drops++;
504		goto bail;
505	}
506
507	/*
508	 * A GRH is expected to preceed the data even if not
509	 * present on the wire.
510	 */
511	wc.byte_len = tlen + sizeof(struct ib_grh);
512
513	/*
514	 * Get the next work request entry to find where to put the data.
515	 */
516	if (qp->r_flags & IPATH_R_REUSE_SGE)
517		qp->r_flags &= ~IPATH_R_REUSE_SGE;
518	else if (!ipath_get_rwqe(qp, 0)) {
519		/*
520		 * Count VL15 packets dropped due to no receive buffer.
521		 * Otherwise, count them as buffer overruns since usually,
522		 * the HW will be able to receive packets even if there are
523		 * no QPs with posted receive buffers.
524		 */
525		if (qp->ibqp.qp_num == 0)
526			dev->n_vl15_dropped++;
527		else
528			dev->rcv_errors++;
529		goto bail;
530	}
531	/* Silently drop packets which are too big. */
532	if (wc.byte_len > qp->r_len) {
533		qp->r_flags |= IPATH_R_REUSE_SGE;
534		dev->n_pkt_drops++;
535		goto bail;
536	}
537	if (has_grh) {
538		ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
539			       sizeof(struct ib_grh));
540		wc.wc_flags |= IB_WC_GRH;
541	} else
542		ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
543	ipath_copy_sge(&qp->r_sge, data,
544		       wc.byte_len - sizeof(struct ib_grh));
545	if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
546		goto bail;
547	wc.wr_id = qp->r_wr_id;
548	wc.status = IB_WC_SUCCESS;
549	wc.opcode = IB_WC_RECV;
550	wc.vendor_err = 0;
551	wc.qp = &qp->ibqp;
552	wc.src_qp = src_qp;
553	wc.pkey_index = 0;
554	wc.slid = be16_to_cpu(hdr->lrh[3]);
555	wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
556	dlid = be16_to_cpu(hdr->lrh[1]);
557	/*
558	 * Save the LMC lower bits if the destination LID is a unicast LID.
559	 */
560	wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
561		dlid & ((1 << dev->dd->ipath_lmc) - 1);
562	wc.port_num = 1;
563	/* Signal completion event if the solicited bit is set. */
564	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
565		       (ohdr->bth[0] &
566			cpu_to_be32(1 << 23)) != 0);
567
568bail:;
569}
570