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.
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:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
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.
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
35 #include <rdma/ib_mad.h>
36 #include <rdma/ib_user_verbs.h>
38 #include <linux/module.h>
39 #include <linux/utsname.h>
40 #include <linux/rculist.h>
42 #include <linux/random.h>
43 #include <linux/vmalloc.h>
44 #include <rdma/rdma_vt.h>
47 #include "qib_common.h"
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");
53 static unsigned int qib_lkey_table_size = 16;
54 module_param_named(lkey_table_size, qib_lkey_table_size, uint,
56 MODULE_PARM_DESC(lkey_table_size,
57 "LKEY table size in bits (2^n, 1 <= n <= 23)");
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");
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");
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");
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");
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");
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");
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");
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");
94 unsigned int ib_qib_max_mcast_qp_attached = 16;
95 module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached,
97 MODULE_PARM_DESC(max_mcast_qp_attached,
98 "Maximum number of attached QPs to support");
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");
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");
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");
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");
117 * Translate ib_wr_opcode into ib_wc_opcode.
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
132 __be64 ib_qib_sys_image_guid;
135 * Count the number of DMA descriptors needed to send length bytes of data.
136 * Don't modify the qib_sge_state to get the count.
137 * Return zero if any of the segments is not aligned.
139 static u32 qib_count_sge(struct rvt_sge_state *ss, u32 length)
141 struct rvt_sge *sg_list = ss->sg_list;
142 struct rvt_sge sge = ss->sge;
143 u8 num_sge = ss->num_sge;
144 u32 ndesc = 1; /* count the header */
147 u32 len = sge.length;
151 if (len > sge.sge_length)
152 len = sge.sge_length;
153 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
154 (len != length && (len & (sizeof(u32) - 1)))) {
161 sge.sge_length -= len;
162 if (sge.sge_length == 0) {
165 } else if (sge.length == 0 && sge.mr->lkey) {
166 if (++sge.n >= RVT_SEGSZ) {
167 if (++sge.m >= sge.mr->mapsz)
172 sge.mr->map[sge.m]->segs[sge.n].vaddr;
174 sge.mr->map[sge.m]->segs[sge.n].length;
182 * Copy from the SGEs to the data buffer.
184 static void qib_copy_from_sge(void *data, struct rvt_sge_state *ss, u32 length)
186 struct rvt_sge *sge = &ss->sge;
189 u32 len = sge->length;
193 if (len > sge->sge_length)
194 len = sge->sge_length;
195 memcpy(data, sge->vaddr, len);
198 sge->sge_length -= len;
199 if (sge->sge_length == 0) {
201 *sge = *ss->sg_list++;
202 } else if (sge->length == 0 && sge->mr->lkey) {
203 if (++sge->n >= RVT_SEGSZ) {
204 if (++sge->m >= sge->mr->mapsz)
209 sge->mr->map[sge->m]->segs[sge->n].vaddr;
211 sge->mr->map[sge->m]->segs[sge->n].length;
219 * qib_qp_rcv - processing an incoming packet on a QP
220 * @rcd: the context pointer
221 * @hdr: the packet header
222 * @has_grh: true if the packet has a GRH
223 * @data: the packet data
224 * @tlen: the packet length
225 * @qp: the QP the packet came on
227 * This is called from qib_ib_rcv() to process an incoming packet
229 * Called at interrupt level.
231 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
232 int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
234 struct qib_ibport *ibp = &rcd->ppd->ibport_data;
236 spin_lock(&qp->r_lock);
238 /* Check for valid receive state. */
239 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
240 ibp->rvp.n_pkt_drops++;
244 switch (qp->ibqp.qp_type) {
247 if (ib_qib_disable_sma)
251 qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
255 qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
259 qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
267 spin_unlock(&qp->r_lock);
271 * qib_ib_rcv - process an incoming packet
272 * @rcd: the context pointer
273 * @rhdr: the header of the packet
274 * @data: the packet payload
275 * @tlen: the packet length
277 * This is called from qib_kreceive() to process an incoming packet at
278 * interrupt level. Tlen is the length of the header + data + CRC in bytes.
280 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
282 struct qib_pportdata *ppd = rcd->ppd;
283 struct qib_ibport *ibp = &ppd->ibport_data;
284 struct ib_header *hdr = rhdr;
285 struct qib_devdata *dd = ppd->dd;
286 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
287 struct ib_other_headers *ohdr;
294 /* 24 == LRH+BTH+CRC */
295 if (unlikely(tlen < 24))
298 /* Check for a valid destination LID (see ch. 7.11.1). */
299 lid = be16_to_cpu(hdr->lrh[1]);
300 if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
301 lid &= ~((1 << ppd->lmc) - 1);
302 if (unlikely(lid != ppd->lid))
307 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
308 if (lnh == QIB_LRH_BTH)
310 else if (lnh == QIB_LRH_GRH) {
313 ohdr = &hdr->u.l.oth;
314 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
316 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
317 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
322 opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
323 #ifdef CONFIG_DEBUG_FS
324 rcd->opstats->stats[opcode].n_bytes += tlen;
325 rcd->opstats->stats[opcode].n_packets++;
328 /* Get the destination QP number. */
329 qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
330 if (qp_num == QIB_MULTICAST_QPN) {
331 struct rvt_mcast *mcast;
332 struct rvt_mcast_qp *p;
334 if (lnh != QIB_LRH_GRH)
336 mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid, lid);
339 this_cpu_inc(ibp->pmastats->n_multicast_rcv);
340 list_for_each_entry_rcu(p, &mcast->qp_list, list)
341 qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
343 * Notify rvt_multicast_detach() if it is waiting for us
346 if (atomic_dec_return(&mcast->refcount) <= 1)
347 wake_up(&mcast->wait);
350 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
355 this_cpu_inc(ibp->pmastats->n_unicast_rcv);
356 qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
362 ibp->rvp.n_pkt_drops++;
366 * This is called from a timer to check for QPs
367 * which need kernel memory in order to send a packet.
369 static void mem_timer(struct timer_list *t)
371 struct qib_ibdev *dev = from_timer(dev, t, mem_timer);
372 struct list_head *list = &dev->memwait;
373 struct rvt_qp *qp = NULL;
374 struct qib_qp_priv *priv = NULL;
377 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
378 if (!list_empty(list)) {
379 priv = list_entry(list->next, struct qib_qp_priv, iowait);
381 list_del_init(&priv->iowait);
383 if (!list_empty(list))
384 mod_timer(&dev->mem_timer, jiffies + 1);
386 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
389 spin_lock_irqsave(&qp->s_lock, flags);
390 if (qp->s_flags & RVT_S_WAIT_KMEM) {
391 qp->s_flags &= ~RVT_S_WAIT_KMEM;
392 qib_schedule_send(qp);
394 spin_unlock_irqrestore(&qp->s_lock, flags);
399 #ifdef __LITTLE_ENDIAN
400 static inline u32 get_upper_bits(u32 data, u32 shift)
402 return data >> shift;
405 static inline u32 set_upper_bits(u32 data, u32 shift)
407 return data << shift;
410 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
412 data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
413 data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
417 static inline u32 get_upper_bits(u32 data, u32 shift)
419 return data << shift;
422 static inline u32 set_upper_bits(u32 data, u32 shift)
424 return data >> shift;
427 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
429 data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
430 data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
435 static void copy_io(u32 __iomem *piobuf, struct rvt_sge_state *ss,
436 u32 length, unsigned flush_wc)
443 u32 len = ss->sge.length;
448 if (len > ss->sge.sge_length)
449 len = ss->sge.sge_length;
450 /* If the source address is not aligned, try to align it. */
451 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
453 u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
455 u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
458 y = sizeof(u32) - off;
461 if (len + extra >= sizeof(u32)) {
462 data |= set_upper_bits(v, extra *
464 len = sizeof(u32) - extra;
469 __raw_writel(data, piobuf);
474 /* Clear unused upper bytes */
475 data |= clear_upper_bytes(v, len, extra);
483 /* Source address is aligned. */
484 u32 *addr = (u32 *) ss->sge.vaddr;
485 int shift = extra * BITS_PER_BYTE;
486 int ushift = 32 - shift;
489 while (l >= sizeof(u32)) {
492 data |= set_upper_bits(v, shift);
493 __raw_writel(data, piobuf);
494 data = get_upper_bits(v, ushift);
500 * We still have 'extra' number of bytes leftover.
505 if (l + extra >= sizeof(u32)) {
506 data |= set_upper_bits(v, shift);
507 len -= l + extra - sizeof(u32);
512 __raw_writel(data, piobuf);
517 /* Clear unused upper bytes */
518 data |= clear_upper_bytes(v, l, extra);
525 } else if (len == length) {
529 } else if (len == length) {
533 * Need to round up for the last dword in the
537 qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
539 last = ((u32 *) ss->sge.vaddr)[w - 1];
544 qib_pio_copy(piobuf, ss->sge.vaddr, w);
547 extra = len & (sizeof(u32) - 1);
549 u32 v = ((u32 *) ss->sge.vaddr)[w];
551 /* Clear unused upper bytes */
552 data = clear_upper_bytes(v, extra, 0);
555 rvt_update_sge(ss, len, false);
558 /* Update address before sending packet. */
559 rvt_update_sge(ss, length, false);
561 /* must flush early everything before trigger word */
563 __raw_writel(last, piobuf);
564 /* be sure trigger word is written */
567 __raw_writel(last, piobuf);
570 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
573 struct qib_qp_priv *priv = qp->priv;
574 struct qib_verbs_txreq *tx;
577 spin_lock_irqsave(&qp->s_lock, flags);
578 spin_lock(&dev->rdi.pending_lock);
580 if (!list_empty(&dev->txreq_free)) {
581 struct list_head *l = dev->txreq_free.next;
584 spin_unlock(&dev->rdi.pending_lock);
585 spin_unlock_irqrestore(&qp->s_lock, flags);
586 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
588 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK &&
589 list_empty(&priv->iowait)) {
591 qp->s_flags |= RVT_S_WAIT_TX;
592 list_add_tail(&priv->iowait, &dev->txwait);
594 qp->s_flags &= ~RVT_S_BUSY;
595 spin_unlock(&dev->rdi.pending_lock);
596 spin_unlock_irqrestore(&qp->s_lock, flags);
597 tx = ERR_PTR(-EBUSY);
602 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
605 struct qib_verbs_txreq *tx;
608 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
609 /* assume the list non empty */
610 if (likely(!list_empty(&dev->txreq_free))) {
611 struct list_head *l = dev->txreq_free.next;
614 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
615 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
617 /* call slow path to get the extra lock */
618 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
619 tx = __get_txreq(dev, qp);
624 void qib_put_txreq(struct qib_verbs_txreq *tx)
626 struct qib_ibdev *dev;
628 struct qib_qp_priv *priv;
632 dev = to_idev(qp->ibqp.device);
638 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
639 tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
640 dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
641 tx->txreq.addr, tx->hdr_dwords << 2,
643 kfree(tx->align_buf);
646 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
648 /* Put struct back on free list */
649 list_add(&tx->txreq.list, &dev->txreq_free);
651 if (!list_empty(&dev->txwait)) {
652 /* Wake up first QP wanting a free struct */
653 priv = list_entry(dev->txwait.next, struct qib_qp_priv,
656 list_del_init(&priv->iowait);
658 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
660 spin_lock_irqsave(&qp->s_lock, flags);
661 if (qp->s_flags & RVT_S_WAIT_TX) {
662 qp->s_flags &= ~RVT_S_WAIT_TX;
663 qib_schedule_send(qp);
665 spin_unlock_irqrestore(&qp->s_lock, flags);
669 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
673 * This is called when there are send DMA descriptors that might be
676 * This is called with ppd->sdma_lock held.
678 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
681 struct qib_qp_priv *qpp, *nqpp;
682 struct rvt_qp *qps[20];
683 struct qib_ibdev *dev;
687 dev = &ppd->dd->verbs_dev;
688 spin_lock(&dev->rdi.pending_lock);
690 /* Search wait list for first QP wanting DMA descriptors. */
691 list_for_each_entry_safe(qpp, nqpp, &dev->dmawait, iowait) {
693 if (qp->port_num != ppd->port)
695 if (n == ARRAY_SIZE(qps))
697 if (qpp->s_tx->txreq.sg_count > avail)
699 avail -= qpp->s_tx->txreq.sg_count;
700 list_del_init(&qpp->iowait);
705 spin_unlock(&dev->rdi.pending_lock);
707 for (i = 0; i < n; i++) {
709 spin_lock(&qp->s_lock);
710 if (qp->s_flags & RVT_S_WAIT_DMA_DESC) {
711 qp->s_flags &= ~RVT_S_WAIT_DMA_DESC;
712 qib_schedule_send(qp);
714 spin_unlock(&qp->s_lock);
720 * This is called with ppd->sdma_lock held.
722 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
724 struct qib_verbs_txreq *tx =
725 container_of(cookie, struct qib_verbs_txreq, txreq);
726 struct rvt_qp *qp = tx->qp;
727 struct qib_qp_priv *priv = qp->priv;
729 spin_lock(&qp->s_lock);
731 rvt_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
732 else if (qp->ibqp.qp_type == IB_QPT_RC) {
733 struct ib_header *hdr;
735 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
736 hdr = &tx->align_buf->hdr;
738 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
740 hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
742 qib_rc_send_complete(qp, hdr);
744 if (atomic_dec_and_test(&priv->s_dma_busy)) {
745 if (qp->state == IB_QPS_RESET)
746 wake_up(&priv->wait_dma);
747 else if (qp->s_flags & RVT_S_WAIT_DMA) {
748 qp->s_flags &= ~RVT_S_WAIT_DMA;
749 qib_schedule_send(qp);
752 spin_unlock(&qp->s_lock);
757 static int wait_kmem(struct qib_ibdev *dev, struct rvt_qp *qp)
759 struct qib_qp_priv *priv = qp->priv;
763 spin_lock_irqsave(&qp->s_lock, flags);
764 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
765 spin_lock(&dev->rdi.pending_lock);
766 if (list_empty(&priv->iowait)) {
767 if (list_empty(&dev->memwait))
768 mod_timer(&dev->mem_timer, jiffies + 1);
769 qp->s_flags |= RVT_S_WAIT_KMEM;
770 list_add_tail(&priv->iowait, &dev->memwait);
772 spin_unlock(&dev->rdi.pending_lock);
773 qp->s_flags &= ~RVT_S_BUSY;
776 spin_unlock_irqrestore(&qp->s_lock, flags);
781 static int qib_verbs_send_dma(struct rvt_qp *qp, struct ib_header *hdr,
782 u32 hdrwords, struct rvt_sge_state *ss, u32 len,
783 u32 plen, u32 dwords)
785 struct qib_qp_priv *priv = qp->priv;
786 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
787 struct qib_devdata *dd = dd_from_dev(dev);
788 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
789 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
790 struct qib_verbs_txreq *tx;
791 struct qib_pio_header *phdr;
799 /* resend previously constructed packet */
800 ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
804 tx = get_txreq(dev, qp);
808 control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
809 be16_to_cpu(hdr->lrh[0]) >> 12);
812 tx->mr = qp->s_rdma_mr;
814 qp->s_rdma_mr = NULL;
815 tx->txreq.callback = sdma_complete;
816 if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
817 tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
819 tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
820 if (plen + 1 > dd->piosize2kmax_dwords)
821 tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
825 * Don't try to DMA if it takes more descriptors than
828 ndesc = qib_count_sge(ss, len);
829 if (ndesc >= ppd->sdma_descq_cnt)
834 phdr = &dev->pio_hdrs[tx->hdr_inx];
835 phdr->pbc[0] = cpu_to_le32(plen);
836 phdr->pbc[1] = cpu_to_le32(control);
837 memcpy(&phdr->hdr, hdr, hdrwords << 2);
838 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
839 tx->txreq.sg_count = ndesc;
840 tx->txreq.addr = dev->pio_hdrs_phys +
841 tx->hdr_inx * sizeof(struct qib_pio_header);
842 tx->hdr_dwords = hdrwords + 2; /* add PBC length */
843 ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
847 /* Allocate a buffer and copy the header and payload to it. */
848 tx->hdr_dwords = plen + 1;
849 phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
852 phdr->pbc[0] = cpu_to_le32(plen);
853 phdr->pbc[1] = cpu_to_le32(control);
854 memcpy(&phdr->hdr, hdr, hdrwords << 2);
855 qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
857 tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
858 tx->hdr_dwords << 2, DMA_TO_DEVICE);
859 if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
861 tx->align_buf = phdr;
862 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
863 tx->txreq.sg_count = 1;
864 ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
871 ret = wait_kmem(dev, qp);
873 ibp->rvp.n_unaligned++;
882 * If we are now in the error state, return zero to flush the
885 static int no_bufs_available(struct rvt_qp *qp)
887 struct qib_qp_priv *priv = qp->priv;
888 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
889 struct qib_devdata *dd;
894 * Note that as soon as want_buffer() is called and
895 * possibly before it returns, qib_ib_piobufavail()
896 * could be called. Therefore, put QP on the I/O wait list before
897 * enabling the PIO avail interrupt.
899 spin_lock_irqsave(&qp->s_lock, flags);
900 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
901 spin_lock(&dev->rdi.pending_lock);
902 if (list_empty(&priv->iowait)) {
904 qp->s_flags |= RVT_S_WAIT_PIO;
905 list_add_tail(&priv->iowait, &dev->piowait);
906 dd = dd_from_dev(dev);
907 dd->f_wantpiobuf_intr(dd, 1);
909 spin_unlock(&dev->rdi.pending_lock);
910 qp->s_flags &= ~RVT_S_BUSY;
913 spin_unlock_irqrestore(&qp->s_lock, flags);
917 static int qib_verbs_send_pio(struct rvt_qp *qp, struct ib_header *ibhdr,
918 u32 hdrwords, struct rvt_sge_state *ss, u32 len,
919 u32 plen, u32 dwords)
921 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
922 struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
923 u32 *hdr = (u32 *) ibhdr;
924 u32 __iomem *piobuf_orig;
932 control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
933 be16_to_cpu(ibhdr->lrh[0]) >> 12);
934 pbc = ((u64) control << 32) | plen;
935 piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
936 if (unlikely(piobuf == NULL))
937 return no_bufs_available(qp);
941 * We have to flush after the PBC for correctness on some cpus
942 * or WC buffer can be written out of order.
945 piobuf_orig = piobuf;
948 flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
951 * If there is just the header portion, must flush before
952 * writing last word of header for correctness, and after
953 * the last header word (trigger word).
957 qib_pio_copy(piobuf, hdr, hdrwords - 1);
959 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
962 qib_pio_copy(piobuf, hdr, hdrwords);
968 qib_pio_copy(piobuf, hdr, hdrwords);
971 /* The common case is aligned and contained in one segment. */
972 if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
973 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
974 u32 *addr = (u32 *) ss->sge.vaddr;
976 /* Update address before sending packet. */
977 rvt_update_sge(ss, len, false);
979 qib_pio_copy(piobuf, addr, dwords - 1);
980 /* must flush early everything before trigger word */
982 __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
983 /* be sure trigger word is written */
986 qib_pio_copy(piobuf, addr, dwords);
989 copy_io(piobuf, ss, len, flush_wc);
991 if (dd->flags & QIB_USE_SPCL_TRIG) {
992 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
995 __raw_writel(0xaebecede, piobuf_orig + spcl_off);
997 qib_sendbuf_done(dd, pbufn);
999 rvt_put_mr(qp->s_rdma_mr);
1000 qp->s_rdma_mr = NULL;
1003 spin_lock_irqsave(&qp->s_lock, flags);
1004 rvt_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1005 spin_unlock_irqrestore(&qp->s_lock, flags);
1006 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1007 spin_lock_irqsave(&qp->s_lock, flags);
1008 qib_rc_send_complete(qp, ibhdr);
1009 spin_unlock_irqrestore(&qp->s_lock, flags);
1015 * qib_verbs_send - send a packet
1016 * @qp: the QP to send on
1017 * @hdr: the packet header
1018 * @hdrwords: the number of 32-bit words in the header
1019 * @ss: the SGE to send
1020 * @len: the length of the packet in bytes
1022 * Return zero if packet is sent or queued OK.
1023 * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1025 int qib_verbs_send(struct rvt_qp *qp, struct ib_header *hdr,
1026 u32 hdrwords, struct rvt_sge_state *ss, u32 len)
1028 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1031 u32 dwords = (len + 3) >> 2;
1034 * Calculate the send buffer trigger address.
1035 * The +1 counts for the pbc control dword following the pbc length.
1037 plen = hdrwords + dwords + 1;
1040 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1041 * can defer SDMA restart until link goes ACTIVE without
1042 * worrying about just how we got there.
1044 if (qp->ibqp.qp_type == IB_QPT_SMI ||
1045 !(dd->flags & QIB_HAS_SEND_DMA))
1046 ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1049 ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1055 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1056 u64 *rwords, u64 *spkts, u64 *rpkts,
1060 struct qib_devdata *dd = ppd->dd;
1062 if (!(dd->flags & QIB_PRESENT)) {
1063 /* no hardware, freeze, etc. */
1067 *swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1068 *rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1069 *spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1070 *rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1071 *xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1080 * qib_get_counters - get various chip counters
1081 * @dd: the qlogic_ib device
1082 * @cntrs: counters are placed here
1084 * Return the counters needed by recv_pma_get_portcounters().
1086 int qib_get_counters(struct qib_pportdata *ppd,
1087 struct qib_verbs_counters *cntrs)
1091 if (!(ppd->dd->flags & QIB_PRESENT)) {
1092 /* no hardware, freeze, etc. */
1096 cntrs->symbol_error_counter =
1097 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1098 cntrs->link_error_recovery_counter =
1099 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1101 * The link downed counter counts when the other side downs the
1102 * connection. We add in the number of times we downed the link
1103 * due to local link integrity errors to compensate.
1105 cntrs->link_downed_counter =
1106 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1107 cntrs->port_rcv_errors =
1108 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1109 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1110 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1111 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1112 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1113 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1114 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1115 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1116 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1117 cntrs->port_rcv_errors +=
1118 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1119 cntrs->port_rcv_errors +=
1120 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1121 cntrs->port_rcv_remphys_errors =
1122 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1123 cntrs->port_xmit_discards =
1124 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1125 cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1126 QIBPORTCNTR_WORDSEND);
1127 cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1128 QIBPORTCNTR_WORDRCV);
1129 cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1130 QIBPORTCNTR_PKTSEND);
1131 cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1132 QIBPORTCNTR_PKTRCV);
1133 cntrs->local_link_integrity_errors =
1134 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1135 cntrs->excessive_buffer_overrun_errors =
1136 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1137 cntrs->vl15_dropped =
1138 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1147 * qib_ib_piobufavail - callback when a PIO buffer is available
1148 * @dd: the device pointer
1150 * This is called from qib_intr() at interrupt level when a PIO buffer is
1151 * available after qib_verbs_send() returned an error that no buffers were
1152 * available. Disable the interrupt if there are no more QPs waiting.
1154 void qib_ib_piobufavail(struct qib_devdata *dd)
1156 struct qib_ibdev *dev = &dd->verbs_dev;
1157 struct list_head *list;
1158 struct rvt_qp *qps[5];
1160 unsigned long flags;
1162 struct qib_qp_priv *priv;
1164 list = &dev->piowait;
1168 * Note: checking that the piowait list is empty and clearing
1169 * the buffer available interrupt needs to be atomic or we
1170 * could end up with QPs on the wait list with the interrupt
1173 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
1174 while (!list_empty(list)) {
1175 if (n == ARRAY_SIZE(qps))
1177 priv = list_entry(list->next, struct qib_qp_priv, iowait);
1179 list_del_init(&priv->iowait);
1183 dd->f_wantpiobuf_intr(dd, 0);
1185 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
1187 for (i = 0; i < n; i++) {
1190 spin_lock_irqsave(&qp->s_lock, flags);
1191 if (qp->s_flags & RVT_S_WAIT_PIO) {
1192 qp->s_flags &= ~RVT_S_WAIT_PIO;
1193 qib_schedule_send(qp);
1195 spin_unlock_irqrestore(&qp->s_lock, flags);
1197 /* Notify qib_destroy_qp() if it is waiting. */
1202 static int qib_query_port(struct rvt_dev_info *rdi, u8 port_num,
1203 struct ib_port_attr *props)
1205 struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1206 struct qib_devdata *dd = dd_from_dev(ibdev);
1207 struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1211 /* props being zeroed by the caller, avoid zeroing it here */
1212 props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1213 props->lmc = ppd->lmc;
1214 props->state = dd->f_iblink_state(ppd->lastibcstat);
1215 props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1216 props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1217 props->active_width = ppd->link_width_active;
1218 /* See rate_show() */
1219 props->active_speed = ppd->link_speed_active;
1220 props->max_vl_num = qib_num_vls(ppd->vls_supported);
1222 props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1223 switch (ppd->ibmtu) {
1242 props->active_mtu = mtu;
1247 static int qib_modify_device(struct ib_device *device,
1248 int device_modify_mask,
1249 struct ib_device_modify *device_modify)
1251 struct qib_devdata *dd = dd_from_ibdev(device);
1255 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1256 IB_DEVICE_MODIFY_NODE_DESC)) {
1261 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1262 memcpy(device->node_desc, device_modify->node_desc,
1263 IB_DEVICE_NODE_DESC_MAX);
1264 for (i = 0; i < dd->num_pports; i++) {
1265 struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1267 qib_node_desc_chg(ibp);
1271 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1272 ib_qib_sys_image_guid =
1273 cpu_to_be64(device_modify->sys_image_guid);
1274 for (i = 0; i < dd->num_pports; i++) {
1275 struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1277 qib_sys_guid_chg(ibp);
1287 static int qib_shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1289 struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1290 struct qib_devdata *dd = dd_from_dev(ibdev);
1291 struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1293 qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1298 static int qib_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1299 int guid_index, __be64 *guid)
1301 struct qib_ibport *ibp = container_of(rvp, struct qib_ibport, rvp);
1302 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1304 if (guid_index == 0)
1306 else if (guid_index < QIB_GUIDS_PER_PORT)
1307 *guid = ibp->guids[guid_index - 1];
1314 int qib_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1316 if (rdma_ah_get_sl(ah_attr) > 15)
1319 if (rdma_ah_get_dlid(ah_attr) == 0)
1321 if (rdma_ah_get_dlid(ah_attr) >=
1322 be16_to_cpu(IB_MULTICAST_LID_BASE) &&
1323 rdma_ah_get_dlid(ah_attr) !=
1324 be16_to_cpu(IB_LID_PERMISSIVE) &&
1325 !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1331 static void qib_notify_new_ah(struct ib_device *ibdev,
1332 struct rdma_ah_attr *ah_attr,
1335 struct qib_ibport *ibp;
1336 struct qib_pportdata *ppd;
1339 * Do not trust reading anything from rvt_ah at this point as it is not
1340 * done being setup. We can however modify things which we need to set.
1343 ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1344 ppd = ppd_from_ibp(ibp);
1345 ah->vl = ibp->sl_to_vl[rdma_ah_get_sl(&ah->attr)];
1346 ah->log_pmtu = ilog2(ppd->ibmtu);
1349 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1351 struct rdma_ah_attr attr;
1352 struct ib_ah *ah = ERR_PTR(-EINVAL);
1354 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1355 struct qib_devdata *dd = dd_from_ppd(ppd);
1356 u8 port_num = ppd->port;
1358 memset(&attr, 0, sizeof(attr));
1359 attr.type = rdma_ah_find_type(&dd->verbs_dev.rdi.ibdev, port_num);
1360 rdma_ah_set_dlid(&attr, dlid);
1361 rdma_ah_set_port_num(&attr, port_num);
1363 qp0 = rcu_dereference(ibp->rvp.qp[0]);
1365 ah = rdma_create_ah(qp0->ibqp.pd, &attr);
1371 * qib_get_npkeys - return the size of the PKEY table for context 0
1372 * @dd: the qlogic_ib device
1374 unsigned qib_get_npkeys(struct qib_devdata *dd)
1376 return ARRAY_SIZE(dd->rcd[0]->pkeys);
1380 * Return the indexed PKEY from the port PKEY table.
1381 * No need to validate rcd[ctxt]; the port is setup if we are here.
1383 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1385 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1386 struct qib_devdata *dd = ppd->dd;
1387 unsigned ctxt = ppd->hw_pidx;
1390 /* dd->rcd null if mini_init or some init failures */
1391 if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1394 ret = dd->rcd[ctxt]->pkeys[index];
1399 static void init_ibport(struct qib_pportdata *ppd)
1401 struct qib_verbs_counters cntrs;
1402 struct qib_ibport *ibp = &ppd->ibport_data;
1404 spin_lock_init(&ibp->rvp.lock);
1405 /* Set the prefix to the default value (see ch. 4.1.1) */
1406 ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1407 ibp->rvp.sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
1408 ibp->rvp.port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
1409 IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
1410 IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
1411 IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
1412 IB_PORT_OTHER_LOCAL_CHANGES_SUP;
1413 if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
1414 ibp->rvp.port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
1415 ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1416 ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1417 ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1418 ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1419 ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1421 /* Snapshot current HW counters to "clear" them. */
1422 qib_get_counters(ppd, &cntrs);
1423 ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
1424 ibp->z_link_error_recovery_counter =
1425 cntrs.link_error_recovery_counter;
1426 ibp->z_link_downed_counter = cntrs.link_downed_counter;
1427 ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
1428 ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
1429 ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
1430 ibp->z_port_xmit_data = cntrs.port_xmit_data;
1431 ibp->z_port_rcv_data = cntrs.port_rcv_data;
1432 ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
1433 ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
1434 ibp->z_local_link_integrity_errors =
1435 cntrs.local_link_integrity_errors;
1436 ibp->z_excessive_buffer_overrun_errors =
1437 cntrs.excessive_buffer_overrun_errors;
1438 ibp->z_vl15_dropped = cntrs.vl15_dropped;
1439 RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1440 RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1444 * qib_fill_device_attr - Fill in rvt dev info device attributes.
1445 * @dd: the device data structure
1447 static void qib_fill_device_attr(struct qib_devdata *dd)
1449 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1451 memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1453 rdi->dparms.props.max_pd = ib_qib_max_pds;
1454 rdi->dparms.props.max_ah = ib_qib_max_ahs;
1455 rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1456 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1457 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1458 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1459 rdi->dparms.props.page_size_cap = PAGE_SIZE;
1460 rdi->dparms.props.vendor_id =
1461 QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1462 rdi->dparms.props.vendor_part_id = dd->deviceid;
1463 rdi->dparms.props.hw_ver = dd->minrev;
1464 rdi->dparms.props.sys_image_guid = ib_qib_sys_image_guid;
1465 rdi->dparms.props.max_mr_size = ~0ULL;
1466 rdi->dparms.props.max_qp = ib_qib_max_qps;
1467 rdi->dparms.props.max_qp_wr = ib_qib_max_qp_wrs;
1468 rdi->dparms.props.max_send_sge = ib_qib_max_sges;
1469 rdi->dparms.props.max_recv_sge = ib_qib_max_sges;
1470 rdi->dparms.props.max_sge_rd = ib_qib_max_sges;
1471 rdi->dparms.props.max_cq = ib_qib_max_cqs;
1472 rdi->dparms.props.max_cqe = ib_qib_max_cqes;
1473 rdi->dparms.props.max_ah = ib_qib_max_ahs;
1474 rdi->dparms.props.max_mr = rdi->lkey_table.max;
1475 rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1476 rdi->dparms.props.max_map_per_fmr = 32767;
1477 rdi->dparms.props.max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1478 rdi->dparms.props.max_qp_init_rd_atom = 255;
1479 rdi->dparms.props.max_srq = ib_qib_max_srqs;
1480 rdi->dparms.props.max_srq_wr = ib_qib_max_srq_wrs;
1481 rdi->dparms.props.max_srq_sge = ib_qib_max_srq_sges;
1482 rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1483 rdi->dparms.props.max_pkeys = qib_get_npkeys(dd);
1484 rdi->dparms.props.max_mcast_grp = ib_qib_max_mcast_grps;
1485 rdi->dparms.props.max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1486 rdi->dparms.props.max_total_mcast_qp_attach =
1487 rdi->dparms.props.max_mcast_qp_attach *
1488 rdi->dparms.props.max_mcast_grp;
1489 /* post send table */
1490 dd->verbs_dev.rdi.post_parms = qib_post_parms;
1492 /* opcode translation table */
1493 dd->verbs_dev.rdi.wc_opcode = ib_qib_wc_opcode;
1497 * qib_register_ib_device - register our device with the infiniband core
1498 * @dd: the device data structure
1499 * Return the allocated qib_ibdev pointer or NULL on error.
1501 int qib_register_ib_device(struct qib_devdata *dd)
1503 struct qib_ibdev *dev = &dd->verbs_dev;
1504 struct ib_device *ibdev = &dev->rdi.ibdev;
1505 struct qib_pportdata *ppd = dd->pport;
1509 get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
1510 for (i = 0; i < dd->num_pports; i++)
1511 init_ibport(ppd + i);
1513 /* Only need to initialize non-zero fields. */
1514 timer_setup(&dev->mem_timer, mem_timer, 0);
1516 INIT_LIST_HEAD(&dev->piowait);
1517 INIT_LIST_HEAD(&dev->dmawait);
1518 INIT_LIST_HEAD(&dev->txwait);
1519 INIT_LIST_HEAD(&dev->memwait);
1520 INIT_LIST_HEAD(&dev->txreq_free);
1522 if (ppd->sdma_descq_cnt) {
1523 dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
1524 ppd->sdma_descq_cnt *
1525 sizeof(struct qib_pio_header),
1526 &dev->pio_hdrs_phys,
1528 if (!dev->pio_hdrs) {
1534 for (i = 0; i < ppd->sdma_descq_cnt; i++) {
1535 struct qib_verbs_txreq *tx;
1537 tx = kzalloc(sizeof(*tx), GFP_KERNEL);
1543 list_add(&tx->txreq.list, &dev->txreq_free);
1547 * The system image GUID is supposed to be the same for all
1548 * IB HCAs in a single system but since there can be other
1549 * device types in the system, we can't be sure this is unique.
1551 if (!ib_qib_sys_image_guid)
1552 ib_qib_sys_image_guid = ppd->guid;
1554 ibdev->owner = THIS_MODULE;
1555 ibdev->node_guid = ppd->guid;
1556 ibdev->phys_port_cnt = dd->num_pports;
1557 ibdev->dev.parent = &dd->pcidev->dev;
1558 ibdev->modify_device = qib_modify_device;
1559 ibdev->process_mad = qib_process_mad;
1561 snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
1562 "Intel Infiniband HCA %s", init_utsname()->nodename);
1565 * Fill in rvt info object.
1567 dd->verbs_dev.rdi.driver_f.port_callback = qib_create_port_files;
1568 dd->verbs_dev.rdi.driver_f.get_pci_dev = qib_get_pci_dev;
1569 dd->verbs_dev.rdi.driver_f.check_ah = qib_check_ah;
1570 dd->verbs_dev.rdi.driver_f.setup_wqe = qib_check_send_wqe;
1571 dd->verbs_dev.rdi.driver_f.notify_new_ah = qib_notify_new_ah;
1572 dd->verbs_dev.rdi.driver_f.alloc_qpn = qib_alloc_qpn;
1573 dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qib_qp_priv_alloc;
1574 dd->verbs_dev.rdi.driver_f.qp_priv_free = qib_qp_priv_free;
1575 dd->verbs_dev.rdi.driver_f.free_all_qps = qib_free_all_qps;
1576 dd->verbs_dev.rdi.driver_f.notify_qp_reset = qib_notify_qp_reset;
1577 dd->verbs_dev.rdi.driver_f.do_send = qib_do_send;
1578 dd->verbs_dev.rdi.driver_f.schedule_send = qib_schedule_send;
1579 dd->verbs_dev.rdi.driver_f.quiesce_qp = qib_quiesce_qp;
1580 dd->verbs_dev.rdi.driver_f.stop_send_queue = qib_stop_send_queue;
1581 dd->verbs_dev.rdi.driver_f.flush_qp_waiters = qib_flush_qp_waiters;
1582 dd->verbs_dev.rdi.driver_f.notify_error_qp = qib_notify_error_qp;
1583 dd->verbs_dev.rdi.driver_f.notify_restart_rc = qib_restart_rc;
1584 dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = qib_mtu_to_path_mtu;
1585 dd->verbs_dev.rdi.driver_f.mtu_from_qp = qib_mtu_from_qp;
1586 dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = qib_get_pmtu_from_attr;
1587 dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _qib_schedule_send;
1588 dd->verbs_dev.rdi.driver_f.query_port_state = qib_query_port;
1589 dd->verbs_dev.rdi.driver_f.shut_down_port = qib_shut_down_port;
1590 dd->verbs_dev.rdi.driver_f.cap_mask_chg = qib_cap_mask_chg;
1591 dd->verbs_dev.rdi.driver_f.notify_create_mad_agent =
1592 qib_notify_create_mad_agent;
1593 dd->verbs_dev.rdi.driver_f.notify_free_mad_agent =
1594 qib_notify_free_mad_agent;
1596 dd->verbs_dev.rdi.dparms.max_rdma_atomic = QIB_MAX_RDMA_ATOMIC;
1597 dd->verbs_dev.rdi.driver_f.get_guid_be = qib_get_guid_be;
1598 dd->verbs_dev.rdi.dparms.lkey_table_size = qib_lkey_table_size;
1599 dd->verbs_dev.rdi.dparms.qp_table_size = ib_qib_qp_table_size;
1600 dd->verbs_dev.rdi.dparms.qpn_start = 1;
1601 dd->verbs_dev.rdi.dparms.qpn_res_start = QIB_KD_QP;
1602 dd->verbs_dev.rdi.dparms.qpn_res_end = QIB_KD_QP; /* Reserve one QP */
1603 dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1604 dd->verbs_dev.rdi.dparms.qos_shift = 1;
1605 dd->verbs_dev.rdi.dparms.psn_mask = QIB_PSN_MASK;
1606 dd->verbs_dev.rdi.dparms.psn_shift = QIB_PSN_SHIFT;
1607 dd->verbs_dev.rdi.dparms.psn_modify_mask = QIB_PSN_MASK;
1608 dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1609 dd->verbs_dev.rdi.dparms.npkeys = qib_get_npkeys(dd);
1610 dd->verbs_dev.rdi.dparms.node = dd->assigned_node_id;
1611 dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1612 dd->verbs_dev.rdi.dparms.max_mad_size = IB_MGMT_MAD_SIZE;
1613 dd->verbs_dev.rdi.dparms.sge_copy_mode = RVT_SGE_COPY_MEMCPY;
1615 qib_fill_device_attr(dd);
1618 for (i = 0; i < dd->num_pports; i++, ppd++) {
1619 ctxt = ppd->hw_pidx;
1620 rvt_init_port(&dd->verbs_dev.rdi,
1621 &ppd->ibport_data.rvp,
1623 dd->rcd[ctxt]->pkeys);
1625 rdma_set_device_sysfs_group(&dd->verbs_dev.rdi.ibdev, &qib_attr_group);
1627 ret = rvt_register_device(&dd->verbs_dev.rdi, RDMA_DRIVER_QIB);
1634 while (!list_empty(&dev->txreq_free)) {
1635 struct list_head *l = dev->txreq_free.next;
1636 struct qib_verbs_txreq *tx;
1639 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1642 if (ppd->sdma_descq_cnt)
1643 dma_free_coherent(&dd->pcidev->dev,
1644 ppd->sdma_descq_cnt *
1645 sizeof(struct qib_pio_header),
1646 dev->pio_hdrs, dev->pio_hdrs_phys);
1648 qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1652 void qib_unregister_ib_device(struct qib_devdata *dd)
1654 struct qib_ibdev *dev = &dd->verbs_dev;
1656 qib_verbs_unregister_sysfs(dd);
1658 rvt_unregister_device(&dd->verbs_dev.rdi);
1660 if (!list_empty(&dev->piowait))
1661 qib_dev_err(dd, "piowait list not empty!\n");
1662 if (!list_empty(&dev->dmawait))
1663 qib_dev_err(dd, "dmawait list not empty!\n");
1664 if (!list_empty(&dev->txwait))
1665 qib_dev_err(dd, "txwait list not empty!\n");
1666 if (!list_empty(&dev->memwait))
1667 qib_dev_err(dd, "memwait list not empty!\n");
1669 del_timer_sync(&dev->mem_timer);
1670 while (!list_empty(&dev->txreq_free)) {
1671 struct list_head *l = dev->txreq_free.next;
1672 struct qib_verbs_txreq *tx;
1675 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1678 if (dd->pport->sdma_descq_cnt)
1679 dma_free_coherent(&dd->pcidev->dev,
1680 dd->pport->sdma_descq_cnt *
1681 sizeof(struct qib_pio_header),
1682 dev->pio_hdrs, dev->pio_hdrs_phys);
1686 * _qib_schedule_send - schedule progress
1689 * This schedules progress w/o regard to the s_flags.
1691 * It is only used in post send, which doesn't hold
1694 bool _qib_schedule_send(struct rvt_qp *qp)
1696 struct qib_ibport *ibp =
1697 to_iport(qp->ibqp.device, qp->port_num);
1698 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1699 struct qib_qp_priv *priv = qp->priv;
1701 return queue_work(ppd->qib_wq, &priv->s_work);
1705 * qib_schedule_send - schedule progress
1708 * This schedules qp progress. The s_lock
1711 bool qib_schedule_send(struct rvt_qp *qp)
1713 if (qib_send_ok(qp))
1714 return _qib_schedule_send(qp);