2 * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
34 #include <rdma/ib_mad.h>
35 #include <rdma/ib_user_verbs.h>
37 #include <linux/utsname.h>
39 #include "ipath_kernel.h"
40 #include "ipath_verbs.h"
41 #include "ipath_common.h"
43 static unsigned int ib_ipath_qp_table_size = 251;
44 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
45 MODULE_PARM_DESC(qp_table_size, "QP table size");
47 unsigned int ib_ipath_lkey_table_size = 12;
48 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
50 MODULE_PARM_DESC(lkey_table_size,
51 "LKEY table size in bits (2^n, 1 <= n <= 23)");
53 static unsigned int ib_ipath_max_pds = 0xFFFF;
54 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
55 MODULE_PARM_DESC(max_pds,
56 "Maximum number of protection domains to support");
58 static unsigned int ib_ipath_max_ahs = 0xFFFF;
59 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
60 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
62 unsigned int ib_ipath_max_cqes = 0x2FFFF;
63 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
64 MODULE_PARM_DESC(max_cqes,
65 "Maximum number of completion queue entries to support");
67 unsigned int ib_ipath_max_cqs = 0x1FFFF;
68 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
69 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
71 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
72 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
74 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
76 unsigned int ib_ipath_max_qps = 16384;
77 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
78 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
80 unsigned int ib_ipath_max_sges = 0x60;
81 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
82 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
84 unsigned int ib_ipath_max_mcast_grps = 16384;
85 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
87 MODULE_PARM_DESC(max_mcast_grps,
88 "Maximum number of multicast groups to support");
90 unsigned int ib_ipath_max_mcast_qp_attached = 16;
91 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
92 uint, S_IWUSR | S_IRUGO);
93 MODULE_PARM_DESC(max_mcast_qp_attached,
94 "Maximum number of attached QPs to support");
96 unsigned int ib_ipath_max_srqs = 1024;
97 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
98 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
100 unsigned int ib_ipath_max_srq_sges = 128;
101 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
102 uint, S_IWUSR | S_IRUGO);
103 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
105 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
106 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
107 uint, S_IWUSR | S_IRUGO);
108 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
110 static unsigned int ib_ipath_disable_sma;
111 module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
112 MODULE_PARM_DESC(ib_ipath_disable_sma, "Disable the SMA");
114 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
116 [IB_QPS_INIT] = IPATH_POST_RECV_OK,
117 [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
118 [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
119 IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
120 [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
122 [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
126 struct ipath_ucontext {
127 struct ib_ucontext ibucontext;
130 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
133 return container_of(ibucontext, struct ipath_ucontext, ibucontext);
137 * Translate ib_wr_opcode into ib_wc_opcode.
139 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
140 [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
141 [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
142 [IB_WR_SEND] = IB_WC_SEND,
143 [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
144 [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
145 [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
146 [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
152 static __be64 sys_image_guid;
155 * ipath_copy_sge - copy data to SGE memory
157 * @data: the data to copy
158 * @length: the length of the data
160 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
162 struct ipath_sge *sge = &ss->sge;
165 u32 len = sge->length;
169 if (len > sge->sge_length)
170 len = sge->sge_length;
172 memcpy(sge->vaddr, data, len);
175 sge->sge_length -= len;
176 if (sge->sge_length == 0) {
178 *sge = *ss->sg_list++;
179 } else if (sge->length == 0 && sge->mr != NULL) {
180 if (++sge->n >= IPATH_SEGSZ) {
181 if (++sge->m >= sge->mr->mapsz)
186 sge->mr->map[sge->m]->segs[sge->n].vaddr;
188 sge->mr->map[sge->m]->segs[sge->n].length;
196 * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
198 * @length: the number of bytes to skip
200 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
202 struct ipath_sge *sge = &ss->sge;
205 u32 len = sge->length;
209 if (len > sge->sge_length)
210 len = sge->sge_length;
214 sge->sge_length -= len;
215 if (sge->sge_length == 0) {
217 *sge = *ss->sg_list++;
218 } else if (sge->length == 0 && sge->mr != NULL) {
219 if (++sge->n >= IPATH_SEGSZ) {
220 if (++sge->m >= sge->mr->mapsz)
225 sge->mr->map[sge->m]->segs[sge->n].vaddr;
227 sge->mr->map[sge->m]->segs[sge->n].length;
234 * ipath_post_one_send - post one RC, UC, or UD send work request
235 * @qp: the QP to post on
236 * @wr: the work request to send
238 static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
240 struct ipath_swqe *wqe;
248 spin_lock_irqsave(&qp->s_lock, flags);
250 /* Check that state is OK to post send. */
251 if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK))
254 /* IB spec says that num_sge == 0 is OK. */
255 if (wr->num_sge > qp->s_max_sge)
259 * Don't allow RDMA reads or atomic operations on UC or
260 * undefined operations.
261 * Make sure buffer is large enough to hold the result for atomics.
263 if (qp->ibqp.qp_type == IB_QPT_UC) {
264 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
266 } else if (qp->ibqp.qp_type == IB_QPT_UD) {
267 /* Check UD opcode */
268 if (wr->opcode != IB_WR_SEND &&
269 wr->opcode != IB_WR_SEND_WITH_IMM)
271 /* Check UD destination address PD */
272 if (qp->ibqp.pd != wr->wr.ud.ah->pd)
274 } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
276 else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
278 wr->sg_list[0].length < sizeof(u64) ||
279 wr->sg_list[0].addr & (sizeof(u64) - 1)))
281 else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
284 next = qp->s_head + 1;
285 if (next >= qp->s_size)
287 if (next == qp->s_last)
290 wqe = get_swqe_ptr(qp, qp->s_head);
292 wqe->ssn = qp->s_ssn++;
295 acc = wr->opcode >= IB_WR_RDMA_READ ?
296 IB_ACCESS_LOCAL_WRITE : 0;
297 for (i = 0, j = 0; i < wr->num_sge; i++) {
298 u32 length = wr->sg_list[i].length;
303 ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
304 &wr->sg_list[i], acc);
307 wqe->length += length;
312 if (qp->ibqp.qp_type == IB_QPT_UC ||
313 qp->ibqp.qp_type == IB_QPT_RC) {
314 if (wqe->length > 0x80000000U)
316 } else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
326 spin_unlock_irqrestore(&qp->s_lock, flags);
331 * ipath_post_send - post a send on a QP
332 * @ibqp: the QP to post the send on
333 * @wr: the list of work requests to post
334 * @bad_wr: the first bad WR is put here
336 * This may be called from interrupt context.
338 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
339 struct ib_send_wr **bad_wr)
341 struct ipath_qp *qp = to_iqp(ibqp);
344 for (; wr; wr = wr->next) {
345 err = ipath_post_one_send(qp, wr);
352 /* Try to do the send work in the caller's context. */
353 ipath_do_send((unsigned long) qp);
360 * ipath_post_receive - post a receive on a QP
361 * @ibqp: the QP to post the receive on
362 * @wr: the WR to post
363 * @bad_wr: the first bad WR is put here
365 * This may be called from interrupt context.
367 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
368 struct ib_recv_wr **bad_wr)
370 struct ipath_qp *qp = to_iqp(ibqp);
371 struct ipath_rwq *wq = qp->r_rq.wq;
375 /* Check that state is OK to post receive. */
376 if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
382 for (; wr; wr = wr->next) {
383 struct ipath_rwqe *wqe;
387 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
393 spin_lock_irqsave(&qp->r_rq.lock, flags);
395 if (next >= qp->r_rq.size)
397 if (next == wq->tail) {
398 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
404 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
405 wqe->wr_id = wr->wr_id;
406 wqe->num_sge = wr->num_sge;
407 for (i = 0; i < wr->num_sge; i++)
408 wqe->sg_list[i] = wr->sg_list[i];
409 /* Make sure queue entry is written before the head index. */
412 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
421 * ipath_qp_rcv - processing an incoming packet on a QP
422 * @dev: the device the packet came on
423 * @hdr: the packet header
424 * @has_grh: true if the packet has a GRH
425 * @data: the packet data
426 * @tlen: the packet length
427 * @qp: the QP the packet came on
429 * This is called from ipath_ib_rcv() to process an incoming packet
431 * Called at interrupt level.
433 static void ipath_qp_rcv(struct ipath_ibdev *dev,
434 struct ipath_ib_header *hdr, int has_grh,
435 void *data, u32 tlen, struct ipath_qp *qp)
437 /* Check for valid receive state. */
438 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
443 switch (qp->ibqp.qp_type) {
446 if (ib_ipath_disable_sma)
450 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
454 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
458 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
467 * ipath_ib_rcv - process an incoming packet
468 * @arg: the device pointer
469 * @rhdr: the header of the packet
470 * @data: the packet data
471 * @tlen: the packet length
473 * This is called from ipath_kreceive() to process an incoming packet at
474 * interrupt level. Tlen is the length of the header + data + CRC in bytes.
476 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
479 struct ipath_ib_header *hdr = rhdr;
480 struct ipath_other_headers *ohdr;
487 if (unlikely(dev == NULL))
490 if (unlikely(tlen < 24)) { /* LRH+BTH+CRC */
495 /* Check for a valid destination LID (see ch. 7.11.1). */
496 lid = be16_to_cpu(hdr->lrh[1]);
497 if (lid < IPATH_MULTICAST_LID_BASE) {
498 lid &= ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
499 if (unlikely(lid != dev->dd->ipath_lid)) {
506 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
507 if (lnh == IPATH_LRH_BTH)
509 else if (lnh == IPATH_LRH_GRH)
510 ohdr = &hdr->u.l.oth;
516 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
517 dev->opstats[opcode].n_bytes += tlen;
518 dev->opstats[opcode].n_packets++;
520 /* Get the destination QP number. */
521 qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
522 if (qp_num == IPATH_MULTICAST_QPN) {
523 struct ipath_mcast *mcast;
524 struct ipath_mcast_qp *p;
526 if (lnh != IPATH_LRH_GRH) {
530 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
535 dev->n_multicast_rcv++;
536 list_for_each_entry_rcu(p, &mcast->qp_list, list)
537 ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
539 * Notify ipath_multicast_detach() if it is waiting for us
542 if (atomic_dec_return(&mcast->refcount) <= 1)
543 wake_up(&mcast->wait);
545 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
547 dev->n_unicast_rcv++;
548 ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
551 * Notify ipath_destroy_qp() if it is waiting
554 if (atomic_dec_and_test(&qp->refcount))
564 * ipath_ib_timer - verbs timer
565 * @arg: the device pointer
567 * This is called from ipath_do_rcv_timer() at interrupt level to check for
568 * QPs which need retransmits and to collect performance numbers.
570 static void ipath_ib_timer(struct ipath_ibdev *dev)
572 struct ipath_qp *resend = NULL;
573 struct list_head *last;
580 spin_lock_irqsave(&dev->pending_lock, flags);
581 /* Start filling the next pending queue. */
582 if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
583 dev->pending_index = 0;
584 /* Save any requests still in the new queue, they have timed out. */
585 last = &dev->pending[dev->pending_index];
586 while (!list_empty(last)) {
587 qp = list_entry(last->next, struct ipath_qp, timerwait);
588 list_del_init(&qp->timerwait);
589 qp->timer_next = resend;
591 atomic_inc(&qp->refcount);
593 last = &dev->rnrwait;
594 if (!list_empty(last)) {
595 qp = list_entry(last->next, struct ipath_qp, timerwait);
596 if (--qp->s_rnr_timeout == 0) {
598 list_del_init(&qp->timerwait);
599 tasklet_hi_schedule(&qp->s_task);
600 if (list_empty(last))
602 qp = list_entry(last->next, struct ipath_qp,
604 } while (qp->s_rnr_timeout == 0);
608 * We should only be in the started state if pma_sample_start != 0
610 if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
611 --dev->pma_sample_start == 0) {
612 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
613 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
617 &dev->ipath_xmit_wait);
619 if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
620 if (dev->pma_sample_interval == 0) {
621 u64 ta, tb, tc, td, te;
623 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
624 ipath_snapshot_counters(dev->dd, &ta, &tb,
627 dev->ipath_sword = ta - dev->ipath_sword;
628 dev->ipath_rword = tb - dev->ipath_rword;
629 dev->ipath_spkts = tc - dev->ipath_spkts;
630 dev->ipath_rpkts = td - dev->ipath_rpkts;
631 dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
634 dev->pma_sample_interval--;
636 spin_unlock_irqrestore(&dev->pending_lock, flags);
638 /* XXX What if timer fires again while this is running? */
639 for (qp = resend; qp != NULL; qp = qp->timer_next) {
642 spin_lock_irqsave(&qp->s_lock, flags);
643 if (qp->s_last != qp->s_tail && qp->state == IB_QPS_RTS) {
645 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
647 spin_unlock_irqrestore(&qp->s_lock, flags);
649 /* Notify ipath_destroy_qp() if it is waiting. */
650 if (atomic_dec_and_test(&qp->refcount))
655 static void update_sge(struct ipath_sge_state *ss, u32 length)
657 struct ipath_sge *sge = &ss->sge;
659 sge->vaddr += length;
660 sge->length -= length;
661 sge->sge_length -= length;
662 if (sge->sge_length == 0) {
664 *sge = *ss->sg_list++;
665 } else if (sge->length == 0 && sge->mr != NULL) {
666 if (++sge->n >= IPATH_SEGSZ) {
667 if (++sge->m >= sge->mr->mapsz)
671 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
672 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
676 #ifdef __LITTLE_ENDIAN
677 static inline u32 get_upper_bits(u32 data, u32 shift)
679 return data >> shift;
682 static inline u32 set_upper_bits(u32 data, u32 shift)
684 return data << shift;
687 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
689 data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
690 data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
694 static inline u32 get_upper_bits(u32 data, u32 shift)
696 return data << shift;
699 static inline u32 set_upper_bits(u32 data, u32 shift)
701 return data >> shift;
704 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
706 data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
707 data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
712 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
713 u32 length, unsigned flush_wc)
720 u32 len = ss->sge.length;
725 if (len > ss->sge.sge_length)
726 len = ss->sge.sge_length;
728 /* If the source address is not aligned, try to align it. */
729 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
731 u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
733 u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
736 y = sizeof(u32) - off;
739 if (len + extra >= sizeof(u32)) {
740 data |= set_upper_bits(v, extra *
742 len = sizeof(u32) - extra;
747 __raw_writel(data, piobuf);
752 /* Clear unused upper bytes */
753 data |= clear_upper_bytes(v, len, extra);
761 /* Source address is aligned. */
762 u32 *addr = (u32 *) ss->sge.vaddr;
763 int shift = extra * BITS_PER_BYTE;
764 int ushift = 32 - shift;
767 while (l >= sizeof(u32)) {
770 data |= set_upper_bits(v, shift);
771 __raw_writel(data, piobuf);
772 data = get_upper_bits(v, ushift);
778 * We still have 'extra' number of bytes leftover.
783 if (l + extra >= sizeof(u32)) {
784 data |= set_upper_bits(v, shift);
785 len -= l + extra - sizeof(u32);
790 __raw_writel(data, piobuf);
795 /* Clear unused upper bytes */
796 data |= clear_upper_bytes(v, l,
804 } else if (len == length) {
808 } else if (len == length) {
812 * Need to round up for the last dword in the
816 __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
818 last = ((u32 *) ss->sge.vaddr)[w - 1];
823 __iowrite32_copy(piobuf, ss->sge.vaddr, w);
826 extra = len & (sizeof(u32) - 1);
828 u32 v = ((u32 *) ss->sge.vaddr)[w];
830 /* Clear unused upper bytes */
831 data = clear_upper_bytes(v, extra, 0);
837 /* Update address before sending packet. */
838 update_sge(ss, length);
840 /* must flush early everything before trigger word */
842 __raw_writel(last, piobuf);
843 /* be sure trigger word is written */
846 __raw_writel(last, piobuf);
849 static int ipath_verbs_send_pio(struct ipath_qp *qp, u32 *hdr, u32 hdrwords,
850 struct ipath_sge_state *ss, u32 len,
851 u32 plen, u32 dwords)
853 struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
858 piobuf = ipath_getpiobuf(dd, NULL);
859 if (unlikely(piobuf == NULL)) {
865 * Write len to control qword, no flags.
866 * We have to flush after the PBC for correctness on some cpus
867 * or WC buffer can be written out of order.
869 writeq(plen, piobuf);
872 flush_wc = dd->ipath_flags & IPATH_PIO_FLUSH_WC;
875 * If there is just the header portion, must flush before
876 * writing last word of header for correctness, and after
877 * the last header word (trigger word).
881 __iowrite32_copy(piobuf, hdr, hdrwords - 1);
883 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
886 __iowrite32_copy(piobuf, hdr, hdrwords);
892 __iowrite32_copy(piobuf, hdr, hdrwords);
895 /* The common case is aligned and contained in one segment. */
896 if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
897 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
898 u32 *addr = (u32 *) ss->sge.vaddr;
900 /* Update address before sending packet. */
903 __iowrite32_copy(piobuf, addr, dwords - 1);
904 /* must flush early everything before trigger word */
906 __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
907 /* be sure trigger word is written */
910 __iowrite32_copy(piobuf, addr, dwords);
913 copy_io(piobuf, ss, len, flush_wc);
916 ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
923 * ipath_verbs_send - send a packet
924 * @qp: the QP to send on
925 * @hdr: the packet header
926 * @hdrwords: the number of words in the header
927 * @ss: the SGE to send
928 * @len: the length of the packet in bytes
930 int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
931 u32 hdrwords, struct ipath_sge_state *ss, u32 len)
935 u32 dwords = (len + 3) >> 2;
937 /* +1 is for the qword padding of pbc */
938 plen = hdrwords + dwords + 1;
940 ret = ipath_verbs_send_pio(qp, (u32 *) hdr, hdrwords,
941 ss, len, plen, dwords);
946 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
947 u64 *rwords, u64 *spkts, u64 *rpkts,
952 if (!(dd->ipath_flags & IPATH_INITTED)) {
953 /* no hardware, freeze, etc. */
957 *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
958 *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
959 *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
960 *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
961 *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
970 * ipath_get_counters - get various chip counters
971 * @dd: the infinipath device
972 * @cntrs: counters are placed here
974 * Return the counters needed by recv_pma_get_portcounters().
976 int ipath_get_counters(struct ipath_devdata *dd,
977 struct ipath_verbs_counters *cntrs)
979 struct ipath_cregs const *crp = dd->ipath_cregs;
982 if (!(dd->ipath_flags & IPATH_INITTED)) {
983 /* no hardware, freeze, etc. */
987 cntrs->symbol_error_counter =
988 ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
989 cntrs->link_error_recovery_counter =
990 ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
992 * The link downed counter counts when the other side downs the
993 * connection. We add in the number of times we downed the link
994 * due to local link integrity errors to compensate.
996 cntrs->link_downed_counter =
997 ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
998 cntrs->port_rcv_errors =
999 ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
1000 ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
1001 ipath_snap_cntr(dd, crp->cr_portovflcnt) +
1002 ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
1003 ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
1004 ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
1005 ipath_snap_cntr(dd, crp->cr_erricrccnt) +
1006 ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
1007 ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1008 ipath_snap_cntr(dd, crp->cr_badformatcnt) +
1009 dd->ipath_rxfc_unsupvl_errs;
1010 cntrs->port_rcv_remphys_errors =
1011 ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
1012 cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
1013 cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
1014 cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
1015 cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
1016 cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
1017 cntrs->local_link_integrity_errors =
1018 (dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
1019 dd->ipath_lli_errs : dd->ipath_lli_errors;
1020 cntrs->excessive_buffer_overrun_errors = dd->ipath_overrun_thresh_errs;
1029 * ipath_ib_piobufavail - callback when a PIO buffer is available
1030 * @arg: the device pointer
1032 * This is called from ipath_intr() at interrupt level when a PIO buffer is
1033 * available after ipath_verbs_send() returned an error that no buffers were
1034 * available. Return 1 if we consumed all the PIO buffers and we still have
1035 * QPs waiting for buffers (for now, just do a tasklet_hi_schedule and
1038 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
1040 struct ipath_qp *qp;
1041 unsigned long flags;
1046 spin_lock_irqsave(&dev->pending_lock, flags);
1047 while (!list_empty(&dev->piowait)) {
1048 qp = list_entry(dev->piowait.next, struct ipath_qp,
1050 list_del_init(&qp->piowait);
1051 clear_bit(IPATH_S_BUSY, &qp->s_busy);
1052 tasklet_hi_schedule(&qp->s_task);
1054 spin_unlock_irqrestore(&dev->pending_lock, flags);
1060 static int ipath_query_device(struct ib_device *ibdev,
1061 struct ib_device_attr *props)
1063 struct ipath_ibdev *dev = to_idev(ibdev);
1065 memset(props, 0, sizeof(*props));
1067 props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1068 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1069 IB_DEVICE_SYS_IMAGE_GUID;
1070 props->page_size_cap = PAGE_SIZE;
1071 props->vendor_id = dev->dd->ipath_vendorid;
1072 props->vendor_part_id = dev->dd->ipath_deviceid;
1073 props->hw_ver = dev->dd->ipath_pcirev;
1075 props->sys_image_guid = dev->sys_image_guid;
1077 props->max_mr_size = ~0ull;
1078 props->max_qp = ib_ipath_max_qps;
1079 props->max_qp_wr = ib_ipath_max_qp_wrs;
1080 props->max_sge = ib_ipath_max_sges;
1081 props->max_cq = ib_ipath_max_cqs;
1082 props->max_ah = ib_ipath_max_ahs;
1083 props->max_cqe = ib_ipath_max_cqes;
1084 props->max_mr = dev->lk_table.max;
1085 props->max_fmr = dev->lk_table.max;
1086 props->max_map_per_fmr = 32767;
1087 props->max_pd = ib_ipath_max_pds;
1088 props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
1089 props->max_qp_init_rd_atom = 255;
1090 /* props->max_res_rd_atom */
1091 props->max_srq = ib_ipath_max_srqs;
1092 props->max_srq_wr = ib_ipath_max_srq_wrs;
1093 props->max_srq_sge = ib_ipath_max_srq_sges;
1094 /* props->local_ca_ack_delay */
1095 props->atomic_cap = IB_ATOMIC_GLOB;
1096 props->max_pkeys = ipath_get_npkeys(dev->dd);
1097 props->max_mcast_grp = ib_ipath_max_mcast_grps;
1098 props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1099 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1100 props->max_mcast_grp;
1105 const u8 ipath_cvt_physportstate[16] = {
1106 [INFINIPATH_IBCS_LT_STATE_DISABLED] = 3,
1107 [INFINIPATH_IBCS_LT_STATE_LINKUP] = 5,
1108 [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = 2,
1109 [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = 2,
1110 [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = 1,
1111 [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = 1,
1112 [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] = 4,
1113 [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] = 4,
1114 [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] = 4,
1115 [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = 4,
1116 [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] = 6,
1117 [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] = 6,
1118 [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] = 6,
1121 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1123 return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1126 static int ipath_query_port(struct ib_device *ibdev,
1127 u8 port, struct ib_port_attr *props)
1129 struct ipath_ibdev *dev = to_idev(ibdev);
1130 struct ipath_devdata *dd = dev->dd;
1132 u16 lid = dd->ipath_lid;
1135 memset(props, 0, sizeof(*props));
1136 props->lid = lid ? lid : __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1137 props->lmc = dev->mkeyprot_resv_lmc & 7;
1138 props->sm_lid = dev->sm_lid;
1139 props->sm_sl = dev->sm_sl;
1140 ibcstat = dd->ipath_lastibcstat;
1141 props->state = ((ibcstat >> 4) & 0x3) + 1;
1142 /* See phys_state_show() */
1143 props->phys_state = ipath_cvt_physportstate[
1144 dd->ipath_lastibcstat & 0xf];
1145 props->port_cap_flags = dev->port_cap_flags;
1146 props->gid_tbl_len = 1;
1147 props->max_msg_sz = 0x80000000;
1148 props->pkey_tbl_len = ipath_get_npkeys(dd);
1149 props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1150 dev->z_pkey_violations;
1151 props->qkey_viol_cntr = dev->qkey_violations;
1152 props->active_width = IB_WIDTH_4X;
1153 /* See rate_show() */
1154 props->active_speed = 1; /* Regular 10Mbs speed. */
1155 props->max_vl_num = 1; /* VLCap = VL0 */
1156 props->init_type_reply = 0;
1159 * Note: the chip supports a maximum MTU of 4096, but the driver
1160 * hasn't implemented this feature yet, so set the maximum value
1163 props->max_mtu = IB_MTU_2048;
1164 switch (dd->ipath_ibmtu) {
1183 props->active_mtu = mtu;
1184 props->subnet_timeout = dev->subnet_timeout;
1189 static int ipath_modify_device(struct ib_device *device,
1190 int device_modify_mask,
1191 struct ib_device_modify *device_modify)
1195 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1196 IB_DEVICE_MODIFY_NODE_DESC)) {
1201 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1202 memcpy(device->node_desc, device_modify->node_desc, 64);
1204 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1205 to_idev(device)->sys_image_guid =
1206 cpu_to_be64(device_modify->sys_image_guid);
1214 static int ipath_modify_port(struct ib_device *ibdev,
1215 u8 port, int port_modify_mask,
1216 struct ib_port_modify *props)
1218 struct ipath_ibdev *dev = to_idev(ibdev);
1220 dev->port_cap_flags |= props->set_port_cap_mask;
1221 dev->port_cap_flags &= ~props->clr_port_cap_mask;
1222 if (port_modify_mask & IB_PORT_SHUTDOWN)
1223 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1224 if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1225 dev->qkey_violations = 0;
1229 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1230 int index, union ib_gid *gid)
1232 struct ipath_ibdev *dev = to_idev(ibdev);
1239 gid->global.subnet_prefix = dev->gid_prefix;
1240 gid->global.interface_id = dev->dd->ipath_guid;
1248 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1249 struct ib_ucontext *context,
1250 struct ib_udata *udata)
1252 struct ipath_ibdev *dev = to_idev(ibdev);
1253 struct ipath_pd *pd;
1257 * This is actually totally arbitrary. Some correctness tests
1258 * assume there's a maximum number of PDs that can be allocated.
1259 * We don't actually have this limit, but we fail the test if
1260 * we allow allocations of more than we report for this value.
1263 pd = kmalloc(sizeof *pd, GFP_KERNEL);
1265 ret = ERR_PTR(-ENOMEM);
1269 spin_lock(&dev->n_pds_lock);
1270 if (dev->n_pds_allocated == ib_ipath_max_pds) {
1271 spin_unlock(&dev->n_pds_lock);
1273 ret = ERR_PTR(-ENOMEM);
1277 dev->n_pds_allocated++;
1278 spin_unlock(&dev->n_pds_lock);
1280 /* ib_alloc_pd() will initialize pd->ibpd. */
1281 pd->user = udata != NULL;
1289 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1291 struct ipath_pd *pd = to_ipd(ibpd);
1292 struct ipath_ibdev *dev = to_idev(ibpd->device);
1294 spin_lock(&dev->n_pds_lock);
1295 dev->n_pds_allocated--;
1296 spin_unlock(&dev->n_pds_lock);
1304 * ipath_create_ah - create an address handle
1305 * @pd: the protection domain
1306 * @ah_attr: the attributes of the AH
1308 * This may be called from interrupt context.
1310 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1311 struct ib_ah_attr *ah_attr)
1313 struct ipath_ah *ah;
1315 struct ipath_ibdev *dev = to_idev(pd->device);
1316 unsigned long flags;
1318 /* A multicast address requires a GRH (see ch. 8.4.1). */
1319 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1320 ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1321 !(ah_attr->ah_flags & IB_AH_GRH)) {
1322 ret = ERR_PTR(-EINVAL);
1326 if (ah_attr->dlid == 0) {
1327 ret = ERR_PTR(-EINVAL);
1331 if (ah_attr->port_num < 1 ||
1332 ah_attr->port_num > pd->device->phys_port_cnt) {
1333 ret = ERR_PTR(-EINVAL);
1337 ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1339 ret = ERR_PTR(-ENOMEM);
1343 spin_lock_irqsave(&dev->n_ahs_lock, flags);
1344 if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1345 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1347 ret = ERR_PTR(-ENOMEM);
1351 dev->n_ahs_allocated++;
1352 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1354 /* ib_create_ah() will initialize ah->ibah. */
1355 ah->attr = *ah_attr;
1364 * ipath_destroy_ah - destroy an address handle
1365 * @ibah: the AH to destroy
1367 * This may be called from interrupt context.
1369 static int ipath_destroy_ah(struct ib_ah *ibah)
1371 struct ipath_ibdev *dev = to_idev(ibah->device);
1372 struct ipath_ah *ah = to_iah(ibah);
1373 unsigned long flags;
1375 spin_lock_irqsave(&dev->n_ahs_lock, flags);
1376 dev->n_ahs_allocated--;
1377 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1384 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1386 struct ipath_ah *ah = to_iah(ibah);
1388 *ah_attr = ah->attr;
1394 * ipath_get_npkeys - return the size of the PKEY table for port 0
1395 * @dd: the infinipath device
1397 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1399 return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1403 * ipath_get_pkey - return the indexed PKEY from the port 0 PKEY table
1404 * @dd: the infinipath device
1405 * @index: the PKEY index
1407 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1411 if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1414 ret = dd->ipath_pd[0]->port_pkeys[index];
1419 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1422 struct ipath_ibdev *dev = to_idev(ibdev);
1425 if (index >= ipath_get_npkeys(dev->dd)) {
1430 *pkey = ipath_get_pkey(dev->dd, index);
1438 * ipath_alloc_ucontext - allocate a ucontest
1439 * @ibdev: the infiniband device
1440 * @udata: not used by the InfiniPath driver
1443 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1444 struct ib_udata *udata)
1446 struct ipath_ucontext *context;
1447 struct ib_ucontext *ret;
1449 context = kmalloc(sizeof *context, GFP_KERNEL);
1451 ret = ERR_PTR(-ENOMEM);
1455 ret = &context->ibucontext;
1461 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1463 kfree(to_iucontext(context));
1467 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1469 static void __verbs_timer(unsigned long arg)
1471 struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1473 /* Handle verbs layer timeouts. */
1474 ipath_ib_timer(dd->verbs_dev);
1476 mod_timer(&dd->verbs_timer, jiffies + 1);
1479 static int enable_timer(struct ipath_devdata *dd)
1482 * Early chips had a design flaw where the chip and kernel idea
1483 * of the tail register don't always agree, and therefore we won't
1484 * get an interrupt on the next packet received.
1485 * If the board supports per packet receive interrupts, use it.
1486 * Otherwise, the timer function periodically checks for packets
1487 * to cover this case.
1488 * Either way, the timer is needed for verbs layer related
1491 if (dd->ipath_flags & IPATH_GPIO_INTR) {
1492 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1493 0x2074076542310ULL);
1494 /* Enable GPIO bit 2 interrupt */
1495 dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1496 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1497 dd->ipath_gpio_mask);
1500 init_timer(&dd->verbs_timer);
1501 dd->verbs_timer.function = __verbs_timer;
1502 dd->verbs_timer.data = (unsigned long)dd;
1503 dd->verbs_timer.expires = jiffies + 1;
1504 add_timer(&dd->verbs_timer);
1509 static int disable_timer(struct ipath_devdata *dd)
1511 /* Disable GPIO bit 2 interrupt */
1512 if (dd->ipath_flags & IPATH_GPIO_INTR) {
1513 /* Disable GPIO bit 2 interrupt */
1514 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1515 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1516 dd->ipath_gpio_mask);
1518 * We might want to undo changes to debugportselect,
1523 del_timer_sync(&dd->verbs_timer);
1529 * ipath_register_ib_device - register our device with the infiniband core
1530 * @dd: the device data structure
1531 * Return the allocated ipath_ibdev pointer or NULL on error.
1533 int ipath_register_ib_device(struct ipath_devdata *dd)
1535 struct ipath_verbs_counters cntrs;
1536 struct ipath_ibdev *idev;
1537 struct ib_device *dev;
1540 idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1548 /* Only need to initialize non-zero fields. */
1549 spin_lock_init(&idev->n_pds_lock);
1550 spin_lock_init(&idev->n_ahs_lock);
1551 spin_lock_init(&idev->n_cqs_lock);
1552 spin_lock_init(&idev->n_qps_lock);
1553 spin_lock_init(&idev->n_srqs_lock);
1554 spin_lock_init(&idev->n_mcast_grps_lock);
1556 spin_lock_init(&idev->qp_table.lock);
1557 spin_lock_init(&idev->lk_table.lock);
1558 idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1559 /* Set the prefix to the default value (see ch. 4.1.1) */
1560 idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
1562 ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
1567 * The top ib_ipath_lkey_table_size bits are used to index the
1568 * table. The lower 8 bits can be owned by the user (copied from
1569 * the LKEY). The remaining bits act as a generation number or tag.
1571 idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
1572 idev->lk_table.table = kzalloc(idev->lk_table.max *
1573 sizeof(*idev->lk_table.table),
1575 if (idev->lk_table.table == NULL) {
1579 INIT_LIST_HEAD(&idev->pending_mmaps);
1580 spin_lock_init(&idev->pending_lock);
1581 idev->mmap_offset = PAGE_SIZE;
1582 spin_lock_init(&idev->mmap_offset_lock);
1583 INIT_LIST_HEAD(&idev->pending[0]);
1584 INIT_LIST_HEAD(&idev->pending[1]);
1585 INIT_LIST_HEAD(&idev->pending[2]);
1586 INIT_LIST_HEAD(&idev->piowait);
1587 INIT_LIST_HEAD(&idev->rnrwait);
1588 idev->pending_index = 0;
1589 idev->port_cap_flags =
1590 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
1591 idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1592 idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1593 idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1594 idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1595 idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1596 idev->link_width_enabled = 3; /* 1x or 4x */
1598 /* Snapshot current HW counters to "clear" them. */
1599 ipath_get_counters(dd, &cntrs);
1600 idev->z_symbol_error_counter = cntrs.symbol_error_counter;
1601 idev->z_link_error_recovery_counter =
1602 cntrs.link_error_recovery_counter;
1603 idev->z_link_downed_counter = cntrs.link_downed_counter;
1604 idev->z_port_rcv_errors = cntrs.port_rcv_errors;
1605 idev->z_port_rcv_remphys_errors =
1606 cntrs.port_rcv_remphys_errors;
1607 idev->z_port_xmit_discards = cntrs.port_xmit_discards;
1608 idev->z_port_xmit_data = cntrs.port_xmit_data;
1609 idev->z_port_rcv_data = cntrs.port_rcv_data;
1610 idev->z_port_xmit_packets = cntrs.port_xmit_packets;
1611 idev->z_port_rcv_packets = cntrs.port_rcv_packets;
1612 idev->z_local_link_integrity_errors =
1613 cntrs.local_link_integrity_errors;
1614 idev->z_excessive_buffer_overrun_errors =
1615 cntrs.excessive_buffer_overrun_errors;
1618 * The system image GUID is supposed to be the same for all
1619 * IB HCAs in a single system but since there can be other
1620 * device types in the system, we can't be sure this is unique.
1622 if (!sys_image_guid)
1623 sys_image_guid = dd->ipath_guid;
1624 idev->sys_image_guid = sys_image_guid;
1625 idev->ib_unit = dd->ipath_unit;
1628 strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
1629 dev->owner = THIS_MODULE;
1630 dev->node_guid = dd->ipath_guid;
1631 dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
1632 dev->uverbs_cmd_mask =
1633 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
1634 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
1635 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
1636 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
1637 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
1638 (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
1639 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) |
1640 (1ull << IB_USER_VERBS_CMD_QUERY_AH) |
1641 (1ull << IB_USER_VERBS_CMD_REG_MR) |
1642 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
1643 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1644 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
1645 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
1646 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
1647 (1ull << IB_USER_VERBS_CMD_POLL_CQ) |
1648 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) |
1649 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
1650 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
1651 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
1652 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
1653 (1ull << IB_USER_VERBS_CMD_POST_SEND) |
1654 (1ull << IB_USER_VERBS_CMD_POST_RECV) |
1655 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
1656 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
1657 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
1658 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
1659 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
1660 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
1661 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
1662 dev->node_type = RDMA_NODE_IB_CA;
1663 dev->phys_port_cnt = 1;
1664 dev->num_comp_vectors = 1;
1665 dev->dma_device = &dd->pcidev->dev;
1666 dev->query_device = ipath_query_device;
1667 dev->modify_device = ipath_modify_device;
1668 dev->query_port = ipath_query_port;
1669 dev->modify_port = ipath_modify_port;
1670 dev->query_pkey = ipath_query_pkey;
1671 dev->query_gid = ipath_query_gid;
1672 dev->alloc_ucontext = ipath_alloc_ucontext;
1673 dev->dealloc_ucontext = ipath_dealloc_ucontext;
1674 dev->alloc_pd = ipath_alloc_pd;
1675 dev->dealloc_pd = ipath_dealloc_pd;
1676 dev->create_ah = ipath_create_ah;
1677 dev->destroy_ah = ipath_destroy_ah;
1678 dev->query_ah = ipath_query_ah;
1679 dev->create_srq = ipath_create_srq;
1680 dev->modify_srq = ipath_modify_srq;
1681 dev->query_srq = ipath_query_srq;
1682 dev->destroy_srq = ipath_destroy_srq;
1683 dev->create_qp = ipath_create_qp;
1684 dev->modify_qp = ipath_modify_qp;
1685 dev->query_qp = ipath_query_qp;
1686 dev->destroy_qp = ipath_destroy_qp;
1687 dev->post_send = ipath_post_send;
1688 dev->post_recv = ipath_post_receive;
1689 dev->post_srq_recv = ipath_post_srq_receive;
1690 dev->create_cq = ipath_create_cq;
1691 dev->destroy_cq = ipath_destroy_cq;
1692 dev->resize_cq = ipath_resize_cq;
1693 dev->poll_cq = ipath_poll_cq;
1694 dev->req_notify_cq = ipath_req_notify_cq;
1695 dev->get_dma_mr = ipath_get_dma_mr;
1696 dev->reg_phys_mr = ipath_reg_phys_mr;
1697 dev->reg_user_mr = ipath_reg_user_mr;
1698 dev->dereg_mr = ipath_dereg_mr;
1699 dev->alloc_fmr = ipath_alloc_fmr;
1700 dev->map_phys_fmr = ipath_map_phys_fmr;
1701 dev->unmap_fmr = ipath_unmap_fmr;
1702 dev->dealloc_fmr = ipath_dealloc_fmr;
1703 dev->attach_mcast = ipath_multicast_attach;
1704 dev->detach_mcast = ipath_multicast_detach;
1705 dev->process_mad = ipath_process_mad;
1706 dev->mmap = ipath_mmap;
1707 dev->dma_ops = &ipath_dma_mapping_ops;
1709 snprintf(dev->node_desc, sizeof(dev->node_desc),
1710 IPATH_IDSTR " %s", init_utsname()->nodename);
1712 ret = ib_register_device(dev);
1716 if (ipath_verbs_register_sysfs(dev))
1724 ib_unregister_device(dev);
1726 kfree(idev->lk_table.table);
1728 kfree(idev->qp_table.table);
1730 ib_dealloc_device(dev);
1731 ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1735 dd->verbs_dev = idev;
1739 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
1741 struct ib_device *ibdev = &dev->ibdev;
1743 disable_timer(dev->dd);
1745 ib_unregister_device(ibdev);
1747 if (!list_empty(&dev->pending[0]) ||
1748 !list_empty(&dev->pending[1]) ||
1749 !list_empty(&dev->pending[2]))
1750 ipath_dev_err(dev->dd, "pending list not empty!\n");
1751 if (!list_empty(&dev->piowait))
1752 ipath_dev_err(dev->dd, "piowait list not empty!\n");
1753 if (!list_empty(&dev->rnrwait))
1754 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
1755 if (!ipath_mcast_tree_empty())
1756 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
1758 * Note that ipath_unregister_ib_device() can be called before all
1759 * the QPs are destroyed!
1761 ipath_free_all_qps(&dev->qp_table);
1762 kfree(dev->qp_table.table);
1763 kfree(dev->lk_table.table);
1764 ib_dealloc_device(ibdev);
1767 static ssize_t show_rev(struct class_device *cdev, char *buf)
1769 struct ipath_ibdev *dev =
1770 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1772 return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
1775 static ssize_t show_hca(struct class_device *cdev, char *buf)
1777 struct ipath_ibdev *dev =
1778 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1781 ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
1791 static ssize_t show_stats(struct class_device *cdev, char *buf)
1793 struct ipath_ibdev *dev =
1794 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1813 dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
1814 dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
1815 dev->n_other_naks, dev->n_timeouts,
1816 dev->n_rdma_dup_busy, dev->n_rc_stalls, dev->n_piowait,
1817 dev->n_no_piobuf, dev->n_pkt_drops, dev->n_wqe_errs);
1818 for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
1819 const struct ipath_opcode_stats *si = &dev->opstats[i];
1821 if (!si->n_packets && !si->n_bytes)
1823 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
1824 (unsigned long long) si->n_packets,
1825 (unsigned long long) si->n_bytes);
1830 static CLASS_DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1831 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1832 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
1833 static CLASS_DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
1835 static struct class_device_attribute *ipath_class_attributes[] = {
1836 &class_device_attr_hw_rev,
1837 &class_device_attr_hca_type,
1838 &class_device_attr_board_id,
1839 &class_device_attr_stats
1842 static int ipath_verbs_register_sysfs(struct ib_device *dev)
1847 for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
1848 if (class_device_create_file(&dev->class_dev,
1849 ipath_class_attributes[i])) {