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 = rvt_get_sge_length(&sge, length);
149 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
150 (len != length && (len & (sizeof(u32) - 1)))) {
157 sge.sge_length -= len;
158 if (sge.sge_length == 0) {
161 } else if (sge.length == 0 && sge.mr->lkey) {
162 if (++sge.n >= RVT_SEGSZ) {
163 if (++sge.m >= sge.mr->mapsz)
168 sge.mr->map[sge.m]->segs[sge.n].vaddr;
170 sge.mr->map[sge.m]->segs[sge.n].length;
178 * Copy from the SGEs to the data buffer.
180 static void qib_copy_from_sge(void *data, struct rvt_sge_state *ss, u32 length)
182 struct rvt_sge *sge = &ss->sge;
185 u32 len = rvt_get_sge_length(sge, length);
187 memcpy(data, sge->vaddr, len);
190 sge->sge_length -= len;
191 if (sge->sge_length == 0) {
193 *sge = *ss->sg_list++;
194 } else if (sge->length == 0 && sge->mr->lkey) {
195 if (++sge->n >= RVT_SEGSZ) {
196 if (++sge->m >= sge->mr->mapsz)
201 sge->mr->map[sge->m]->segs[sge->n].vaddr;
203 sge->mr->map[sge->m]->segs[sge->n].length;
211 * qib_qp_rcv - processing an incoming packet on a QP
212 * @rcd: the context pointer
213 * @hdr: the packet header
214 * @has_grh: true if the packet has a GRH
215 * @data: the packet data
216 * @tlen: the packet length
217 * @qp: the QP the packet came on
219 * This is called from qib_ib_rcv() to process an incoming packet
221 * Called at interrupt level.
223 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
224 int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
226 struct qib_ibport *ibp = &rcd->ppd->ibport_data;
228 spin_lock(&qp->r_lock);
230 /* Check for valid receive state. */
231 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
232 ibp->rvp.n_pkt_drops++;
236 switch (qp->ibqp.qp_type) {
239 if (ib_qib_disable_sma)
243 qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp);
247 qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp);
251 qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp);
259 spin_unlock(&qp->r_lock);
263 * qib_ib_rcv - process an incoming packet
264 * @rcd: the context pointer
265 * @rhdr: the header of the packet
266 * @data: the packet payload
267 * @tlen: the packet length
269 * This is called from qib_kreceive() to process an incoming packet at
270 * interrupt level. Tlen is the length of the header + data + CRC in bytes.
272 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen)
274 struct qib_pportdata *ppd = rcd->ppd;
275 struct qib_ibport *ibp = &ppd->ibport_data;
276 struct ib_header *hdr = rhdr;
277 struct qib_devdata *dd = ppd->dd;
278 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
279 struct ib_other_headers *ohdr;
286 /* 24 == LRH+BTH+CRC */
287 if (unlikely(tlen < 24))
290 /* Check for a valid destination LID (see ch. 7.11.1). */
291 lid = be16_to_cpu(hdr->lrh[1]);
292 if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
293 lid &= ~((1 << ppd->lmc) - 1);
294 if (unlikely(lid != ppd->lid))
299 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
300 if (lnh == QIB_LRH_BTH)
302 else if (lnh == QIB_LRH_GRH) {
305 ohdr = &hdr->u.l.oth;
306 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
308 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
309 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
314 opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0x7f;
315 #ifdef CONFIG_DEBUG_FS
316 rcd->opstats->stats[opcode].n_bytes += tlen;
317 rcd->opstats->stats[opcode].n_packets++;
320 /* Get the destination QP number. */
321 qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
322 if (qp_num == QIB_MULTICAST_QPN) {
323 struct rvt_mcast *mcast;
324 struct rvt_mcast_qp *p;
326 if (lnh != QIB_LRH_GRH)
328 mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid, lid);
331 this_cpu_inc(ibp->pmastats->n_multicast_rcv);
333 list_for_each_entry_rcu(p, &mcast->qp_list, list)
334 qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp);
337 * Notify rvt_multicast_detach() if it is waiting for us
340 if (atomic_dec_return(&mcast->refcount) <= 1)
341 wake_up(&mcast->wait);
344 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
349 this_cpu_inc(ibp->pmastats->n_unicast_rcv);
350 qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp);
356 ibp->rvp.n_pkt_drops++;
360 * This is called from a timer to check for QPs
361 * which need kernel memory in order to send a packet.
363 static void mem_timer(struct timer_list *t)
365 struct qib_ibdev *dev = from_timer(dev, t, mem_timer);
366 struct list_head *list = &dev->memwait;
367 struct rvt_qp *qp = NULL;
368 struct qib_qp_priv *priv = NULL;
371 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
372 if (!list_empty(list)) {
373 priv = list_entry(list->next, struct qib_qp_priv, iowait);
375 list_del_init(&priv->iowait);
377 if (!list_empty(list))
378 mod_timer(&dev->mem_timer, jiffies + 1);
380 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
383 spin_lock_irqsave(&qp->s_lock, flags);
384 if (qp->s_flags & RVT_S_WAIT_KMEM) {
385 qp->s_flags &= ~RVT_S_WAIT_KMEM;
386 qib_schedule_send(qp);
388 spin_unlock_irqrestore(&qp->s_lock, flags);
393 #ifdef __LITTLE_ENDIAN
394 static inline u32 get_upper_bits(u32 data, u32 shift)
396 return data >> shift;
399 static inline u32 set_upper_bits(u32 data, u32 shift)
401 return data << shift;
404 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
406 data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
407 data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
411 static inline u32 get_upper_bits(u32 data, u32 shift)
413 return data << shift;
416 static inline u32 set_upper_bits(u32 data, u32 shift)
418 return data >> shift;
421 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
423 data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
424 data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
429 static void copy_io(u32 __iomem *piobuf, struct rvt_sge_state *ss,
430 u32 length, unsigned flush_wc)
437 u32 len = rvt_get_sge_length(&ss->sge, length);
440 /* If the source address is not aligned, try to align it. */
441 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
443 u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
445 u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
448 y = sizeof(u32) - off;
451 if (len + extra >= sizeof(u32)) {
452 data |= set_upper_bits(v, extra *
454 len = sizeof(u32) - extra;
459 __raw_writel(data, piobuf);
464 /* Clear unused upper bytes */
465 data |= clear_upper_bytes(v, len, extra);
473 /* Source address is aligned. */
474 u32 *addr = (u32 *) ss->sge.vaddr;
475 int shift = extra * BITS_PER_BYTE;
476 int ushift = 32 - shift;
479 while (l >= sizeof(u32)) {
482 data |= set_upper_bits(v, shift);
483 __raw_writel(data, piobuf);
484 data = get_upper_bits(v, ushift);
490 * We still have 'extra' number of bytes leftover.
495 if (l + extra >= sizeof(u32)) {
496 data |= set_upper_bits(v, shift);
497 len -= l + extra - sizeof(u32);
502 __raw_writel(data, piobuf);
507 /* Clear unused upper bytes */
508 data |= clear_upper_bytes(v, l, extra);
515 } else if (len == length) {
519 } else if (len == length) {
523 * Need to round up for the last dword in the
527 qib_pio_copy(piobuf, ss->sge.vaddr, w - 1);
529 last = ((u32 *) ss->sge.vaddr)[w - 1];
534 qib_pio_copy(piobuf, ss->sge.vaddr, w);
537 extra = len & (sizeof(u32) - 1);
539 u32 v = ((u32 *) ss->sge.vaddr)[w];
541 /* Clear unused upper bytes */
542 data = clear_upper_bytes(v, extra, 0);
545 rvt_update_sge(ss, len, false);
548 /* Update address before sending packet. */
549 rvt_update_sge(ss, length, false);
551 /* must flush early everything before trigger word */
553 __raw_writel(last, piobuf);
554 /* be sure trigger word is written */
557 __raw_writel(last, piobuf);
560 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev,
563 struct qib_qp_priv *priv = qp->priv;
564 struct qib_verbs_txreq *tx;
567 spin_lock_irqsave(&qp->s_lock, flags);
568 spin_lock(&dev->rdi.pending_lock);
570 if (!list_empty(&dev->txreq_free)) {
571 struct list_head *l = dev->txreq_free.next;
574 spin_unlock(&dev->rdi.pending_lock);
575 spin_unlock_irqrestore(&qp->s_lock, flags);
576 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
578 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK &&
579 list_empty(&priv->iowait)) {
581 qp->s_flags |= RVT_S_WAIT_TX;
582 list_add_tail(&priv->iowait, &dev->txwait);
584 qp->s_flags &= ~RVT_S_BUSY;
585 spin_unlock(&dev->rdi.pending_lock);
586 spin_unlock_irqrestore(&qp->s_lock, flags);
587 tx = ERR_PTR(-EBUSY);
592 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev,
595 struct qib_verbs_txreq *tx;
598 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
599 /* assume the list non empty */
600 if (likely(!list_empty(&dev->txreq_free))) {
601 struct list_head *l = dev->txreq_free.next;
604 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
605 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
607 /* call slow path to get the extra lock */
608 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
609 tx = __get_txreq(dev, qp);
614 void qib_put_txreq(struct qib_verbs_txreq *tx)
616 struct qib_ibdev *dev;
618 struct qib_qp_priv *priv;
622 dev = to_idev(qp->ibqp.device);
628 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) {
629 tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF;
630 dma_unmap_single(&dd_from_dev(dev)->pcidev->dev,
631 tx->txreq.addr, tx->hdr_dwords << 2,
633 kfree(tx->align_buf);
636 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
638 /* Put struct back on free list */
639 list_add(&tx->txreq.list, &dev->txreq_free);
641 if (!list_empty(&dev->txwait)) {
642 /* Wake up first QP wanting a free struct */
643 priv = list_entry(dev->txwait.next, struct qib_qp_priv,
646 list_del_init(&priv->iowait);
648 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
650 spin_lock_irqsave(&qp->s_lock, flags);
651 if (qp->s_flags & RVT_S_WAIT_TX) {
652 qp->s_flags &= ~RVT_S_WAIT_TX;
653 qib_schedule_send(qp);
655 spin_unlock_irqrestore(&qp->s_lock, flags);
659 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
663 * This is called when there are send DMA descriptors that might be
666 * This is called with ppd->sdma_lock held.
668 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail)
671 struct qib_qp_priv *qpp, *nqpp;
672 struct rvt_qp *qps[20];
673 struct qib_ibdev *dev;
677 dev = &ppd->dd->verbs_dev;
678 spin_lock(&dev->rdi.pending_lock);
680 /* Search wait list for first QP wanting DMA descriptors. */
681 list_for_each_entry_safe(qpp, nqpp, &dev->dmawait, iowait) {
683 if (qp->port_num != ppd->port)
685 if (n == ARRAY_SIZE(qps))
687 if (qpp->s_tx->txreq.sg_count > avail)
689 avail -= qpp->s_tx->txreq.sg_count;
690 list_del_init(&qpp->iowait);
695 spin_unlock(&dev->rdi.pending_lock);
697 for (i = 0; i < n; i++) {
699 spin_lock(&qp->s_lock);
700 if (qp->s_flags & RVT_S_WAIT_DMA_DESC) {
701 qp->s_flags &= ~RVT_S_WAIT_DMA_DESC;
702 qib_schedule_send(qp);
704 spin_unlock(&qp->s_lock);
710 * This is called with ppd->sdma_lock held.
712 static void sdma_complete(struct qib_sdma_txreq *cookie, int status)
714 struct qib_verbs_txreq *tx =
715 container_of(cookie, struct qib_verbs_txreq, txreq);
716 struct rvt_qp *qp = tx->qp;
717 struct qib_qp_priv *priv = qp->priv;
719 spin_lock(&qp->s_lock);
721 rvt_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
722 else if (qp->ibqp.qp_type == IB_QPT_RC) {
723 struct ib_header *hdr;
725 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF)
726 hdr = &tx->align_buf->hdr;
728 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
730 hdr = &dev->pio_hdrs[tx->hdr_inx].hdr;
732 qib_rc_send_complete(qp, hdr);
734 if (atomic_dec_and_test(&priv->s_dma_busy)) {
735 if (qp->state == IB_QPS_RESET)
736 wake_up(&priv->wait_dma);
737 else if (qp->s_flags & RVT_S_WAIT_DMA) {
738 qp->s_flags &= ~RVT_S_WAIT_DMA;
739 qib_schedule_send(qp);
742 spin_unlock(&qp->s_lock);
747 static int wait_kmem(struct qib_ibdev *dev, struct rvt_qp *qp)
749 struct qib_qp_priv *priv = qp->priv;
753 spin_lock_irqsave(&qp->s_lock, flags);
754 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
755 spin_lock(&dev->rdi.pending_lock);
756 if (list_empty(&priv->iowait)) {
757 if (list_empty(&dev->memwait))
758 mod_timer(&dev->mem_timer, jiffies + 1);
759 qp->s_flags |= RVT_S_WAIT_KMEM;
760 list_add_tail(&priv->iowait, &dev->memwait);
762 spin_unlock(&dev->rdi.pending_lock);
763 qp->s_flags &= ~RVT_S_BUSY;
766 spin_unlock_irqrestore(&qp->s_lock, flags);
771 static int qib_verbs_send_dma(struct rvt_qp *qp, struct ib_header *hdr,
772 u32 hdrwords, struct rvt_sge_state *ss, u32 len,
773 u32 plen, u32 dwords)
775 struct qib_qp_priv *priv = qp->priv;
776 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
777 struct qib_devdata *dd = dd_from_dev(dev);
778 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
779 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
780 struct qib_verbs_txreq *tx;
781 struct qib_pio_header *phdr;
789 /* resend previously constructed packet */
790 ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx);
794 tx = get_txreq(dev, qp);
798 control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
799 be16_to_cpu(hdr->lrh[0]) >> 12);
802 tx->mr = qp->s_rdma_mr;
804 qp->s_rdma_mr = NULL;
805 tx->txreq.callback = sdma_complete;
806 if (dd->flags & QIB_HAS_SDMA_TIMEOUT)
807 tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST;
809 tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ;
810 if (plen + 1 > dd->piosize2kmax_dwords)
811 tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF;
815 * Don't try to DMA if it takes more descriptors than
818 ndesc = qib_count_sge(ss, len);
819 if (ndesc >= ppd->sdma_descq_cnt)
824 phdr = &dev->pio_hdrs[tx->hdr_inx];
825 phdr->pbc[0] = cpu_to_le32(plen);
826 phdr->pbc[1] = cpu_to_le32(control);
827 memcpy(&phdr->hdr, hdr, hdrwords << 2);
828 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC;
829 tx->txreq.sg_count = ndesc;
830 tx->txreq.addr = dev->pio_hdrs_phys +
831 tx->hdr_inx * sizeof(struct qib_pio_header);
832 tx->hdr_dwords = hdrwords + 2; /* add PBC length */
833 ret = qib_sdma_verbs_send(ppd, ss, dwords, tx);
837 /* Allocate a buffer and copy the header and payload to it. */
838 tx->hdr_dwords = plen + 1;
839 phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC);
842 phdr->pbc[0] = cpu_to_le32(plen);
843 phdr->pbc[1] = cpu_to_le32(control);
844 memcpy(&phdr->hdr, hdr, hdrwords << 2);
845 qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len);
847 tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr,
848 tx->hdr_dwords << 2, DMA_TO_DEVICE);
849 if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr))
851 tx->align_buf = phdr;
852 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF;
853 tx->txreq.sg_count = 1;
854 ret = qib_sdma_verbs_send(ppd, NULL, 0, tx);
861 ret = wait_kmem(dev, qp);
863 ibp->rvp.n_unaligned++;
872 * If we are now in the error state, return zero to flush the
875 static int no_bufs_available(struct rvt_qp *qp)
877 struct qib_qp_priv *priv = qp->priv;
878 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
879 struct qib_devdata *dd;
884 * Note that as soon as want_buffer() is called and
885 * possibly before it returns, qib_ib_piobufavail()
886 * could be called. Therefore, put QP on the I/O wait list before
887 * enabling the PIO avail interrupt.
889 spin_lock_irqsave(&qp->s_lock, flags);
890 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
891 spin_lock(&dev->rdi.pending_lock);
892 if (list_empty(&priv->iowait)) {
894 qp->s_flags |= RVT_S_WAIT_PIO;
895 list_add_tail(&priv->iowait, &dev->piowait);
896 dd = dd_from_dev(dev);
897 dd->f_wantpiobuf_intr(dd, 1);
899 spin_unlock(&dev->rdi.pending_lock);
900 qp->s_flags &= ~RVT_S_BUSY;
903 spin_unlock_irqrestore(&qp->s_lock, flags);
907 static int qib_verbs_send_pio(struct rvt_qp *qp, struct ib_header *ibhdr,
908 u32 hdrwords, struct rvt_sge_state *ss, u32 len,
909 u32 plen, u32 dwords)
911 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
912 struct qib_pportdata *ppd = dd->pport + qp->port_num - 1;
913 u32 *hdr = (u32 *) ibhdr;
914 u32 __iomem *piobuf_orig;
922 control = dd->f_setpbc_control(ppd, plen, qp->s_srate,
923 be16_to_cpu(ibhdr->lrh[0]) >> 12);
924 pbc = ((u64) control << 32) | plen;
925 piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
926 if (unlikely(piobuf == NULL))
927 return no_bufs_available(qp);
931 * We have to flush after the PBC for correctness on some cpus
932 * or WC buffer can be written out of order.
935 piobuf_orig = piobuf;
938 flush_wc = dd->flags & QIB_PIO_FLUSH_WC;
941 * If there is just the header portion, must flush before
942 * writing last word of header for correctness, and after
943 * the last header word (trigger word).
947 qib_pio_copy(piobuf, hdr, hdrwords - 1);
949 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
952 qib_pio_copy(piobuf, hdr, hdrwords);
958 qib_pio_copy(piobuf, hdr, hdrwords);
961 /* The common case is aligned and contained in one segment. */
962 if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
963 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
964 u32 *addr = (u32 *) ss->sge.vaddr;
966 /* Update address before sending packet. */
967 rvt_update_sge(ss, len, false);
969 qib_pio_copy(piobuf, addr, dwords - 1);
970 /* must flush early everything before trigger word */
972 __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
973 /* be sure trigger word is written */
976 qib_pio_copy(piobuf, addr, dwords);
979 copy_io(piobuf, ss, len, flush_wc);
981 if (dd->flags & QIB_USE_SPCL_TRIG) {
982 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
985 __raw_writel(0xaebecede, piobuf_orig + spcl_off);
987 qib_sendbuf_done(dd, pbufn);
989 rvt_put_mr(qp->s_rdma_mr);
990 qp->s_rdma_mr = NULL;
993 spin_lock_irqsave(&qp->s_lock, flags);
994 rvt_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
995 spin_unlock_irqrestore(&qp->s_lock, flags);
996 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
997 spin_lock_irqsave(&qp->s_lock, flags);
998 qib_rc_send_complete(qp, ibhdr);
999 spin_unlock_irqrestore(&qp->s_lock, flags);
1005 * qib_verbs_send - send a packet
1006 * @qp: the QP to send on
1007 * @hdr: the packet header
1008 * @hdrwords: the number of 32-bit words in the header
1009 * @ss: the SGE to send
1010 * @len: the length of the packet in bytes
1012 * Return zero if packet is sent or queued OK.
1013 * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1015 int qib_verbs_send(struct rvt_qp *qp, struct ib_header *hdr,
1016 u32 hdrwords, struct rvt_sge_state *ss, u32 len)
1018 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1021 u32 dwords = (len + 3) >> 2;
1024 * Calculate the send buffer trigger address.
1025 * The +1 counts for the pbc control dword following the pbc length.
1027 plen = hdrwords + dwords + 1;
1030 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1031 * can defer SDMA restart until link goes ACTIVE without
1032 * worrying about just how we got there.
1034 if (qp->ibqp.qp_type == IB_QPT_SMI ||
1035 !(dd->flags & QIB_HAS_SEND_DMA))
1036 ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1039 ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1045 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
1046 u64 *rwords, u64 *spkts, u64 *rpkts,
1050 struct qib_devdata *dd = ppd->dd;
1052 if (!(dd->flags & QIB_PRESENT)) {
1053 /* no hardware, freeze, etc. */
1057 *swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND);
1058 *rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV);
1059 *spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND);
1060 *rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV);
1061 *xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL);
1070 * qib_get_counters - get various chip counters
1071 * @dd: the qlogic_ib device
1072 * @cntrs: counters are placed here
1074 * Return the counters needed by recv_pma_get_portcounters().
1076 int qib_get_counters(struct qib_pportdata *ppd,
1077 struct qib_verbs_counters *cntrs)
1081 if (!(ppd->dd->flags & QIB_PRESENT)) {
1082 /* no hardware, freeze, etc. */
1086 cntrs->symbol_error_counter =
1087 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR);
1088 cntrs->link_error_recovery_counter =
1089 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV);
1091 * The link downed counter counts when the other side downs the
1092 * connection. We add in the number of times we downed the link
1093 * due to local link integrity errors to compensate.
1095 cntrs->link_downed_counter =
1096 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN);
1097 cntrs->port_rcv_errors =
1098 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) +
1099 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) +
1100 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) +
1101 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) +
1102 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) +
1103 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) +
1104 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) +
1105 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) +
1106 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT);
1107 cntrs->port_rcv_errors +=
1108 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR);
1109 cntrs->port_rcv_errors +=
1110 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR);
1111 cntrs->port_rcv_remphys_errors =
1112 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP);
1113 cntrs->port_xmit_discards =
1114 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL);
1115 cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd,
1116 QIBPORTCNTR_WORDSEND);
1117 cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd,
1118 QIBPORTCNTR_WORDRCV);
1119 cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd,
1120 QIBPORTCNTR_PKTSEND);
1121 cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd,
1122 QIBPORTCNTR_PKTRCV);
1123 cntrs->local_link_integrity_errors =
1124 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI);
1125 cntrs->excessive_buffer_overrun_errors =
1126 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL);
1127 cntrs->vl15_dropped =
1128 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP);
1137 * qib_ib_piobufavail - callback when a PIO buffer is available
1138 * @dd: the device pointer
1140 * This is called from qib_intr() at interrupt level when a PIO buffer is
1141 * available after qib_verbs_send() returned an error that no buffers were
1142 * available. Disable the interrupt if there are no more QPs waiting.
1144 void qib_ib_piobufavail(struct qib_devdata *dd)
1146 struct qib_ibdev *dev = &dd->verbs_dev;
1147 struct list_head *list;
1148 struct rvt_qp *qps[5];
1150 unsigned long flags;
1152 struct qib_qp_priv *priv;
1154 list = &dev->piowait;
1158 * Note: checking that the piowait list is empty and clearing
1159 * the buffer available interrupt needs to be atomic or we
1160 * could end up with QPs on the wait list with the interrupt
1163 spin_lock_irqsave(&dev->rdi.pending_lock, flags);
1164 while (!list_empty(list)) {
1165 if (n == ARRAY_SIZE(qps))
1167 priv = list_entry(list->next, struct qib_qp_priv, iowait);
1169 list_del_init(&priv->iowait);
1173 dd->f_wantpiobuf_intr(dd, 0);
1175 spin_unlock_irqrestore(&dev->rdi.pending_lock, flags);
1177 for (i = 0; i < n; i++) {
1180 spin_lock_irqsave(&qp->s_lock, flags);
1181 if (qp->s_flags & RVT_S_WAIT_PIO) {
1182 qp->s_flags &= ~RVT_S_WAIT_PIO;
1183 qib_schedule_send(qp);
1185 spin_unlock_irqrestore(&qp->s_lock, flags);
1187 /* Notify qib_destroy_qp() if it is waiting. */
1192 static int qib_query_port(struct rvt_dev_info *rdi, u8 port_num,
1193 struct ib_port_attr *props)
1195 struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1196 struct qib_devdata *dd = dd_from_dev(ibdev);
1197 struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1201 /* props being zeroed by the caller, avoid zeroing it here */
1202 props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1203 props->lmc = ppd->lmc;
1204 props->state = dd->f_iblink_state(ppd->lastibcstat);
1205 props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat);
1206 props->gid_tbl_len = QIB_GUIDS_PER_PORT;
1207 props->active_width = ppd->link_width_active;
1208 /* See rate_show() */
1209 props->active_speed = ppd->link_speed_active;
1210 props->max_vl_num = qib_num_vls(ppd->vls_supported);
1212 props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096;
1213 switch (ppd->ibmtu) {
1232 props->active_mtu = mtu;
1237 static int qib_modify_device(struct ib_device *device,
1238 int device_modify_mask,
1239 struct ib_device_modify *device_modify)
1241 struct qib_devdata *dd = dd_from_ibdev(device);
1245 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1246 IB_DEVICE_MODIFY_NODE_DESC)) {
1251 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1252 memcpy(device->node_desc, device_modify->node_desc,
1253 IB_DEVICE_NODE_DESC_MAX);
1254 for (i = 0; i < dd->num_pports; i++) {
1255 struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1257 qib_node_desc_chg(ibp);
1261 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1262 ib_qib_sys_image_guid =
1263 cpu_to_be64(device_modify->sys_image_guid);
1264 for (i = 0; i < dd->num_pports; i++) {
1265 struct qib_ibport *ibp = &dd->pport[i].ibport_data;
1267 qib_sys_guid_chg(ibp);
1277 static int qib_shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1279 struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
1280 struct qib_devdata *dd = dd_from_dev(ibdev);
1281 struct qib_pportdata *ppd = &dd->pport[port_num - 1];
1283 qib_set_linkstate(ppd, QIB_IB_LINKDOWN);
1288 static int qib_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1289 int guid_index, __be64 *guid)
1291 struct qib_ibport *ibp = container_of(rvp, struct qib_ibport, rvp);
1292 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1294 if (guid_index == 0)
1296 else if (guid_index < QIB_GUIDS_PER_PORT)
1297 *guid = ibp->guids[guid_index - 1];
1304 int qib_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1306 if (rdma_ah_get_sl(ah_attr) > 15)
1309 if (rdma_ah_get_dlid(ah_attr) == 0)
1311 if (rdma_ah_get_dlid(ah_attr) >=
1312 be16_to_cpu(IB_MULTICAST_LID_BASE) &&
1313 rdma_ah_get_dlid(ah_attr) !=
1314 be16_to_cpu(IB_LID_PERMISSIVE) &&
1315 !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1321 static void qib_notify_new_ah(struct ib_device *ibdev,
1322 struct rdma_ah_attr *ah_attr,
1325 struct qib_ibport *ibp;
1326 struct qib_pportdata *ppd;
1329 * Do not trust reading anything from rvt_ah at this point as it is not
1330 * done being setup. We can however modify things which we need to set.
1333 ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1334 ppd = ppd_from_ibp(ibp);
1335 ah->vl = ibp->sl_to_vl[rdma_ah_get_sl(&ah->attr)];
1336 ah->log_pmtu = ilog2(ppd->ibmtu);
1339 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid)
1341 struct rdma_ah_attr attr;
1342 struct ib_ah *ah = ERR_PTR(-EINVAL);
1344 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1345 struct qib_devdata *dd = dd_from_ppd(ppd);
1346 u8 port_num = ppd->port;
1348 memset(&attr, 0, sizeof(attr));
1349 attr.type = rdma_ah_find_type(&dd->verbs_dev.rdi.ibdev, port_num);
1350 rdma_ah_set_dlid(&attr, dlid);
1351 rdma_ah_set_port_num(&attr, port_num);
1353 qp0 = rcu_dereference(ibp->rvp.qp[0]);
1355 ah = rdma_create_ah(qp0->ibqp.pd, &attr, 0);
1361 * qib_get_npkeys - return the size of the PKEY table for context 0
1362 * @dd: the qlogic_ib device
1364 unsigned qib_get_npkeys(struct qib_devdata *dd)
1366 return ARRAY_SIZE(dd->rcd[0]->pkeys);
1370 * Return the indexed PKEY from the port PKEY table.
1371 * No need to validate rcd[ctxt]; the port is setup if we are here.
1373 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index)
1375 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1376 struct qib_devdata *dd = ppd->dd;
1377 unsigned ctxt = ppd->hw_pidx;
1380 /* dd->rcd null if mini_init or some init failures */
1381 if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys))
1384 ret = dd->rcd[ctxt]->pkeys[index];
1389 static void init_ibport(struct qib_pportdata *ppd)
1391 struct qib_verbs_counters cntrs;
1392 struct qib_ibport *ibp = &ppd->ibport_data;
1394 spin_lock_init(&ibp->rvp.lock);
1395 /* Set the prefix to the default value (see ch. 4.1.1) */
1396 ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1397 ibp->rvp.sm_lid = be16_to_cpu(IB_LID_PERMISSIVE);
1398 ibp->rvp.port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP |
1399 IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP |
1400 IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP |
1401 IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP |
1402 IB_PORT_OTHER_LOCAL_CHANGES_SUP;
1403 if (ppd->dd->flags & QIB_HAS_LINK_LATENCY)
1404 ibp->rvp.port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
1405 ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1406 ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1407 ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1408 ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1409 ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1411 /* Snapshot current HW counters to "clear" them. */
1412 qib_get_counters(ppd, &cntrs);
1413 ibp->z_symbol_error_counter = cntrs.symbol_error_counter;
1414 ibp->z_link_error_recovery_counter =
1415 cntrs.link_error_recovery_counter;
1416 ibp->z_link_downed_counter = cntrs.link_downed_counter;
1417 ibp->z_port_rcv_errors = cntrs.port_rcv_errors;
1418 ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors;
1419 ibp->z_port_xmit_discards = cntrs.port_xmit_discards;
1420 ibp->z_port_xmit_data = cntrs.port_xmit_data;
1421 ibp->z_port_rcv_data = cntrs.port_rcv_data;
1422 ibp->z_port_xmit_packets = cntrs.port_xmit_packets;
1423 ibp->z_port_rcv_packets = cntrs.port_rcv_packets;
1424 ibp->z_local_link_integrity_errors =
1425 cntrs.local_link_integrity_errors;
1426 ibp->z_excessive_buffer_overrun_errors =
1427 cntrs.excessive_buffer_overrun_errors;
1428 ibp->z_vl15_dropped = cntrs.vl15_dropped;
1429 RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1430 RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1434 * qib_fill_device_attr - Fill in rvt dev info device attributes.
1435 * @dd: the device data structure
1437 static void qib_fill_device_attr(struct qib_devdata *dd)
1439 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1441 memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1443 rdi->dparms.props.max_pd = ib_qib_max_pds;
1444 rdi->dparms.props.max_ah = ib_qib_max_ahs;
1445 rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1446 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1447 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1448 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1449 rdi->dparms.props.page_size_cap = PAGE_SIZE;
1450 rdi->dparms.props.vendor_id =
1451 QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3;
1452 rdi->dparms.props.vendor_part_id = dd->deviceid;
1453 rdi->dparms.props.hw_ver = dd->minrev;
1454 rdi->dparms.props.sys_image_guid = ib_qib_sys_image_guid;
1455 rdi->dparms.props.max_mr_size = ~0ULL;
1456 rdi->dparms.props.max_qp = ib_qib_max_qps;
1457 rdi->dparms.props.max_qp_wr = ib_qib_max_qp_wrs;
1458 rdi->dparms.props.max_send_sge = ib_qib_max_sges;
1459 rdi->dparms.props.max_recv_sge = ib_qib_max_sges;
1460 rdi->dparms.props.max_sge_rd = ib_qib_max_sges;
1461 rdi->dparms.props.max_cq = ib_qib_max_cqs;
1462 rdi->dparms.props.max_cqe = ib_qib_max_cqes;
1463 rdi->dparms.props.max_ah = ib_qib_max_ahs;
1464 rdi->dparms.props.max_map_per_fmr = 32767;
1465 rdi->dparms.props.max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC;
1466 rdi->dparms.props.max_qp_init_rd_atom = 255;
1467 rdi->dparms.props.max_srq = ib_qib_max_srqs;
1468 rdi->dparms.props.max_srq_wr = ib_qib_max_srq_wrs;
1469 rdi->dparms.props.max_srq_sge = ib_qib_max_srq_sges;
1470 rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1471 rdi->dparms.props.max_pkeys = qib_get_npkeys(dd);
1472 rdi->dparms.props.max_mcast_grp = ib_qib_max_mcast_grps;
1473 rdi->dparms.props.max_mcast_qp_attach = ib_qib_max_mcast_qp_attached;
1474 rdi->dparms.props.max_total_mcast_qp_attach =
1475 rdi->dparms.props.max_mcast_qp_attach *
1476 rdi->dparms.props.max_mcast_grp;
1477 /* post send table */
1478 dd->verbs_dev.rdi.post_parms = qib_post_parms;
1480 /* opcode translation table */
1481 dd->verbs_dev.rdi.wc_opcode = ib_qib_wc_opcode;
1484 static const struct ib_device_ops qib_dev_ops = {
1485 .owner = THIS_MODULE,
1486 .driver_id = RDMA_DRIVER_QIB,
1488 .init_port = qib_create_port_files,
1489 .modify_device = qib_modify_device,
1490 .process_mad = qib_process_mad,
1494 * qib_register_ib_device - register our device with the infiniband core
1495 * @dd: the device data structure
1496 * Return the allocated qib_ibdev pointer or NULL on error.
1498 int qib_register_ib_device(struct qib_devdata *dd)
1500 struct qib_ibdev *dev = &dd->verbs_dev;
1501 struct ib_device *ibdev = &dev->rdi.ibdev;
1502 struct qib_pportdata *ppd = dd->pport;
1506 get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd));
1507 for (i = 0; i < dd->num_pports; i++)
1508 init_ibport(ppd + i);
1510 /* Only need to initialize non-zero fields. */
1511 timer_setup(&dev->mem_timer, mem_timer, 0);
1513 INIT_LIST_HEAD(&dev->piowait);
1514 INIT_LIST_HEAD(&dev->dmawait);
1515 INIT_LIST_HEAD(&dev->txwait);
1516 INIT_LIST_HEAD(&dev->memwait);
1517 INIT_LIST_HEAD(&dev->txreq_free);
1519 if (ppd->sdma_descq_cnt) {
1520 dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev,
1521 ppd->sdma_descq_cnt *
1522 sizeof(struct qib_pio_header),
1523 &dev->pio_hdrs_phys,
1525 if (!dev->pio_hdrs) {
1531 for (i = 0; i < ppd->sdma_descq_cnt; i++) {
1532 struct qib_verbs_txreq *tx;
1534 tx = kzalloc(sizeof(*tx), GFP_KERNEL);
1540 list_add(&tx->txreq.list, &dev->txreq_free);
1544 * The system image GUID is supposed to be the same for all
1545 * IB HCAs in a single system but since there can be other
1546 * device types in the system, we can't be sure this is unique.
1548 if (!ib_qib_sys_image_guid)
1549 ib_qib_sys_image_guid = ppd->guid;
1551 ibdev->node_guid = ppd->guid;
1552 ibdev->phys_port_cnt = dd->num_pports;
1553 ibdev->dev.parent = &dd->pcidev->dev;
1555 snprintf(ibdev->node_desc, sizeof(ibdev->node_desc),
1556 "Intel Infiniband HCA %s", init_utsname()->nodename);
1559 * Fill in rvt info object.
1561 dd->verbs_dev.rdi.driver_f.get_pci_dev = qib_get_pci_dev;
1562 dd->verbs_dev.rdi.driver_f.check_ah = qib_check_ah;
1563 dd->verbs_dev.rdi.driver_f.setup_wqe = qib_check_send_wqe;
1564 dd->verbs_dev.rdi.driver_f.notify_new_ah = qib_notify_new_ah;
1565 dd->verbs_dev.rdi.driver_f.alloc_qpn = qib_alloc_qpn;
1566 dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qib_qp_priv_alloc;
1567 dd->verbs_dev.rdi.driver_f.qp_priv_free = qib_qp_priv_free;
1568 dd->verbs_dev.rdi.driver_f.free_all_qps = qib_free_all_qps;
1569 dd->verbs_dev.rdi.driver_f.notify_qp_reset = qib_notify_qp_reset;
1570 dd->verbs_dev.rdi.driver_f.do_send = qib_do_send;
1571 dd->verbs_dev.rdi.driver_f.schedule_send = qib_schedule_send;
1572 dd->verbs_dev.rdi.driver_f.quiesce_qp = qib_quiesce_qp;
1573 dd->verbs_dev.rdi.driver_f.stop_send_queue = qib_stop_send_queue;
1574 dd->verbs_dev.rdi.driver_f.flush_qp_waiters = qib_flush_qp_waiters;
1575 dd->verbs_dev.rdi.driver_f.notify_error_qp = qib_notify_error_qp;
1576 dd->verbs_dev.rdi.driver_f.notify_restart_rc = qib_restart_rc;
1577 dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = qib_mtu_to_path_mtu;
1578 dd->verbs_dev.rdi.driver_f.mtu_from_qp = qib_mtu_from_qp;
1579 dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = qib_get_pmtu_from_attr;
1580 dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _qib_schedule_send;
1581 dd->verbs_dev.rdi.driver_f.query_port_state = qib_query_port;
1582 dd->verbs_dev.rdi.driver_f.shut_down_port = qib_shut_down_port;
1583 dd->verbs_dev.rdi.driver_f.cap_mask_chg = qib_cap_mask_chg;
1584 dd->verbs_dev.rdi.driver_f.notify_create_mad_agent =
1585 qib_notify_create_mad_agent;
1586 dd->verbs_dev.rdi.driver_f.notify_free_mad_agent =
1587 qib_notify_free_mad_agent;
1589 dd->verbs_dev.rdi.dparms.max_rdma_atomic = QIB_MAX_RDMA_ATOMIC;
1590 dd->verbs_dev.rdi.driver_f.get_guid_be = qib_get_guid_be;
1591 dd->verbs_dev.rdi.dparms.lkey_table_size = qib_lkey_table_size;
1592 dd->verbs_dev.rdi.dparms.qp_table_size = ib_qib_qp_table_size;
1593 dd->verbs_dev.rdi.dparms.qpn_start = 1;
1594 dd->verbs_dev.rdi.dparms.qpn_res_start = QIB_KD_QP;
1595 dd->verbs_dev.rdi.dparms.qpn_res_end = QIB_KD_QP; /* Reserve one QP */
1596 dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1597 dd->verbs_dev.rdi.dparms.qos_shift = 1;
1598 dd->verbs_dev.rdi.dparms.psn_mask = QIB_PSN_MASK;
1599 dd->verbs_dev.rdi.dparms.psn_shift = QIB_PSN_SHIFT;
1600 dd->verbs_dev.rdi.dparms.psn_modify_mask = QIB_PSN_MASK;
1601 dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1602 dd->verbs_dev.rdi.dparms.npkeys = qib_get_npkeys(dd);
1603 dd->verbs_dev.rdi.dparms.node = dd->assigned_node_id;
1604 dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1605 dd->verbs_dev.rdi.dparms.max_mad_size = IB_MGMT_MAD_SIZE;
1606 dd->verbs_dev.rdi.dparms.sge_copy_mode = RVT_SGE_COPY_MEMCPY;
1608 qib_fill_device_attr(dd);
1611 for (i = 0; i < dd->num_pports; i++, ppd++) {
1612 ctxt = ppd->hw_pidx;
1613 rvt_init_port(&dd->verbs_dev.rdi,
1614 &ppd->ibport_data.rvp,
1616 dd->rcd[ctxt]->pkeys);
1618 rdma_set_device_sysfs_group(&dd->verbs_dev.rdi.ibdev, &qib_attr_group);
1620 ib_set_device_ops(ibdev, &qib_dev_ops);
1621 ret = rvt_register_device(&dd->verbs_dev.rdi);
1628 while (!list_empty(&dev->txreq_free)) {
1629 struct list_head *l = dev->txreq_free.next;
1630 struct qib_verbs_txreq *tx;
1633 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1636 if (ppd->sdma_descq_cnt)
1637 dma_free_coherent(&dd->pcidev->dev,
1638 ppd->sdma_descq_cnt *
1639 sizeof(struct qib_pio_header),
1640 dev->pio_hdrs, dev->pio_hdrs_phys);
1642 qib_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1646 void qib_unregister_ib_device(struct qib_devdata *dd)
1648 struct qib_ibdev *dev = &dd->verbs_dev;
1650 qib_verbs_unregister_sysfs(dd);
1652 rvt_unregister_device(&dd->verbs_dev.rdi);
1654 if (!list_empty(&dev->piowait))
1655 qib_dev_err(dd, "piowait list not empty!\n");
1656 if (!list_empty(&dev->dmawait))
1657 qib_dev_err(dd, "dmawait list not empty!\n");
1658 if (!list_empty(&dev->txwait))
1659 qib_dev_err(dd, "txwait list not empty!\n");
1660 if (!list_empty(&dev->memwait))
1661 qib_dev_err(dd, "memwait list not empty!\n");
1663 del_timer_sync(&dev->mem_timer);
1664 while (!list_empty(&dev->txreq_free)) {
1665 struct list_head *l = dev->txreq_free.next;
1666 struct qib_verbs_txreq *tx;
1669 tx = list_entry(l, struct qib_verbs_txreq, txreq.list);
1672 if (dd->pport->sdma_descq_cnt)
1673 dma_free_coherent(&dd->pcidev->dev,
1674 dd->pport->sdma_descq_cnt *
1675 sizeof(struct qib_pio_header),
1676 dev->pio_hdrs, dev->pio_hdrs_phys);
1680 * _qib_schedule_send - schedule progress
1683 * This schedules progress w/o regard to the s_flags.
1685 * It is only used in post send, which doesn't hold
1688 bool _qib_schedule_send(struct rvt_qp *qp)
1690 struct qib_ibport *ibp =
1691 to_iport(qp->ibqp.device, qp->port_num);
1692 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1693 struct qib_qp_priv *priv = qp->priv;
1695 return queue_work(ppd->qib_wq, &priv->s_work);
1699 * qib_schedule_send - schedule progress
1702 * This schedules qp progress. The s_lock
1705 bool qib_schedule_send(struct rvt_qp *qp)
1707 if (qib_send_ok(qp))
1708 return _qib_schedule_send(qp);