2 * Copyright (c) 2012, 2013 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 <linux/err.h>
36 #include <linux/vmalloc.h>
37 #include <rdma/rdma_vt.h>
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/seq_file.h>
44 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
45 struct rvt_qpn_map *map, unsigned off)
47 return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
50 static inline unsigned find_next_offset(struct rvt_qpn_table *qpt,
51 struct rvt_qpn_map *map, unsigned off,
52 unsigned n, u16 qpt_mask)
56 if (((off & qpt_mask) >> 1) >= n)
57 off = (off | qpt_mask) + 2;
59 off = find_next_zero_bit(map->page, RVT_BITS_PER_PAGE, off);
65 * Convert the AETH credit code into the number of credits.
67 static u32 credit_table[31] = {
101 const struct rvt_operation_params qib_post_parms[RVT_OPERATION_MAX] = {
102 [IB_WR_RDMA_WRITE] = {
103 .length = sizeof(struct ib_rdma_wr),
104 .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
107 [IB_WR_RDMA_READ] = {
108 .length = sizeof(struct ib_rdma_wr),
109 .qpt_support = BIT(IB_QPT_RC),
110 .flags = RVT_OPERATION_ATOMIC,
113 [IB_WR_ATOMIC_CMP_AND_SWP] = {
114 .length = sizeof(struct ib_atomic_wr),
115 .qpt_support = BIT(IB_QPT_RC),
116 .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
119 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
120 .length = sizeof(struct ib_atomic_wr),
121 .qpt_support = BIT(IB_QPT_RC),
122 .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
125 [IB_WR_RDMA_WRITE_WITH_IMM] = {
126 .length = sizeof(struct ib_rdma_wr),
127 .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
131 .length = sizeof(struct ib_send_wr),
132 .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
133 BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
136 [IB_WR_SEND_WITH_IMM] = {
137 .length = sizeof(struct ib_send_wr),
138 .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
139 BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
144 static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map,
147 unsigned long page = get_zeroed_page(gfp);
150 * Free the page if someone raced with us installing it.
153 spin_lock(&qpt->lock);
157 map->page = (void *)page;
158 spin_unlock(&qpt->lock);
162 * Allocate the next available QPN or
163 * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
165 int qib_alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
166 enum ib_qp_type type, u8 port, gfp_t gfp)
168 u32 i, offset, max_scan, qpn;
169 struct rvt_qpn_map *map;
171 struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
172 struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
174 u16 qpt_mask = dd->qpn_mask;
176 if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
179 ret = type == IB_QPT_GSI;
180 n = 1 << (ret + 2 * (port - 1));
181 spin_lock(&qpt->lock);
186 spin_unlock(&qpt->lock);
191 if (qpn >= RVT_QPN_MAX)
193 if (qpt_mask && ((qpn & qpt_mask) >> 1) >= dd->n_krcv_queues)
194 qpn = (qpn | qpt_mask) + 2;
195 offset = qpn & RVT_BITS_PER_PAGE_MASK;
196 map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
197 max_scan = qpt->nmaps - !offset;
199 if (unlikely(!map->page)) {
200 get_map_page(qpt, map, gfp);
201 if (unlikely(!map->page))
205 if (!test_and_set_bit(offset, map->page)) {
210 offset = find_next_offset(qpt, map, offset,
211 dd->n_krcv_queues, qpt_mask);
212 qpn = mk_qpn(qpt, map, offset);
214 * This test differs from alloc_pidmap().
215 * If find_next_offset() does find a zero
216 * bit, we don't need to check for QPN
217 * wrapping around past our starting QPN.
218 * We just need to be sure we don't loop
221 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
223 * In order to keep the number of pages allocated to a
224 * minimum, we scan the all existing pages before increasing
225 * the size of the bitmap table.
227 if (++i > max_scan) {
228 if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
230 map = &qpt->map[qpt->nmaps++];
232 } else if (map < &qpt->map[qpt->nmaps]) {
239 qpn = mk_qpn(qpt, map, offset);
249 * qib_free_all_qps - check for QPs still in use
251 unsigned qib_free_all_qps(struct rvt_dev_info *rdi)
253 struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
254 struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
256 unsigned n, qp_inuse = 0;
258 for (n = 0; n < dd->num_pports; n++) {
259 struct qib_ibport *ibp = &dd->pport[n].ibport_data;
262 if (rcu_dereference(ibp->rvp.qp[0]))
264 if (rcu_dereference(ibp->rvp.qp[1]))
271 void qib_notify_qp_reset(struct rvt_qp *qp)
273 struct qib_qp_priv *priv = qp->priv;
275 atomic_set(&priv->s_dma_busy, 0);
278 void qib_notify_error_qp(struct rvt_qp *qp)
280 struct qib_qp_priv *priv = qp->priv;
281 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
283 spin_lock(&dev->rdi.pending_lock);
284 if (!list_empty(&priv->iowait) && !(qp->s_flags & RVT_S_BUSY)) {
285 qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
286 list_del_init(&priv->iowait);
288 spin_unlock(&dev->rdi.pending_lock);
290 if (!(qp->s_flags & RVT_S_BUSY)) {
293 rvt_put_mr(qp->s_rdma_mr);
294 qp->s_rdma_mr = NULL;
297 qib_put_txreq(priv->s_tx);
303 static int mtu_to_enum(u32 mtu)
309 enum_mtu = IB_MTU_4096;
312 enum_mtu = IB_MTU_2048;
315 enum_mtu = IB_MTU_1024;
318 enum_mtu = IB_MTU_512;
321 enum_mtu = IB_MTU_256;
324 enum_mtu = IB_MTU_2048;
329 int qib_get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
330 struct ib_qp_attr *attr)
332 int mtu, pmtu, pidx = qp->port_num - 1;
333 struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
334 struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
336 mtu = ib_mtu_enum_to_int(attr->path_mtu);
340 if (mtu > dd->pport[pidx].ibmtu)
341 pmtu = mtu_to_enum(dd->pport[pidx].ibmtu);
343 pmtu = attr->path_mtu;
347 int qib_mtu_to_path_mtu(u32 mtu)
349 return mtu_to_enum(mtu);
352 u32 qib_mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
354 return ib_mtu_enum_to_int(pmtu);
358 * qib_compute_aeth - compute the AETH (syndrome + MSN)
359 * @qp: the queue pair to compute the AETH for
363 __be32 qib_compute_aeth(struct rvt_qp *qp)
365 u32 aeth = qp->r_msn & QIB_MSN_MASK;
369 * Shared receive queues don't generate credits.
370 * Set the credit field to the invalid value.
372 aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
376 struct rvt_rwq *wq = qp->r_rq.wq;
380 /* sanity check pointers before trusting them */
382 if (head >= qp->r_rq.size)
385 if (tail >= qp->r_rq.size)
388 * Compute the number of credits available (RWQEs).
389 * XXX Not holding the r_rq.lock here so there is a small
390 * chance that the pair of reads are not atomic.
392 credits = head - tail;
393 if ((int)credits < 0)
394 credits += qp->r_rq.size;
396 * Binary search the credit table to find the code to
403 if (credit_table[x] == credits)
405 if (credit_table[x] > credits)
412 aeth |= x << QIB_AETH_CREDIT_SHIFT;
414 return cpu_to_be32(aeth);
417 void *qib_qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp, gfp_t gfp)
419 struct qib_qp_priv *priv;
421 priv = kzalloc(sizeof(*priv), gfp);
423 return ERR_PTR(-ENOMEM);
426 priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), gfp);
429 return ERR_PTR(-ENOMEM);
431 init_waitqueue_head(&priv->wait_dma);
432 INIT_WORK(&priv->s_work, _qib_do_send);
433 INIT_LIST_HEAD(&priv->iowait);
438 void qib_qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
440 struct qib_qp_priv *priv = qp->priv;
446 void qib_stop_send_queue(struct rvt_qp *qp)
448 struct qib_qp_priv *priv = qp->priv;
450 cancel_work_sync(&priv->s_work);
451 del_timer_sync(&qp->s_timer);
454 void qib_quiesce_qp(struct rvt_qp *qp)
456 struct qib_qp_priv *priv = qp->priv;
458 wait_event(priv->wait_dma, !atomic_read(&priv->s_dma_busy));
460 qib_put_txreq(priv->s_tx);
465 void qib_flush_qp_waiters(struct rvt_qp *qp)
467 struct qib_qp_priv *priv = qp->priv;
468 struct qib_ibdev *dev = to_idev(qp->ibqp.device);
470 spin_lock(&dev->rdi.pending_lock);
471 if (!list_empty(&priv->iowait))
472 list_del_init(&priv->iowait);
473 spin_unlock(&dev->rdi.pending_lock);
477 * qib_get_credit - flush the send work queue of a QP
478 * @qp: the qp who's send work queue to flush
479 * @aeth: the Acknowledge Extended Transport Header
481 * The QP s_lock should be held.
483 void qib_get_credit(struct rvt_qp *qp, u32 aeth)
485 u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
488 * If the credit is invalid, we can send
489 * as many packets as we like. Otherwise, we have to
490 * honor the credit field.
492 if (credit == QIB_AETH_CREDIT_INVAL) {
493 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
494 qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
495 if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
496 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
497 qib_schedule_send(qp);
500 } else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
501 /* Compute new LSN (i.e., MSN + credit) */
502 credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
503 if (qib_cmp24(credit, qp->s_lsn) > 0) {
505 if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
506 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
507 qib_schedule_send(qp);
514 * qib_check_send_wqe - validate wr/wqe
516 * @wqe - The built wqe
518 * validate wr/wqe. This is called
519 * prior to inserting the wqe into
520 * the ring but after the wqe has been
523 * Returns 1 to force direct progress, 0 otherwise, -EINVAL on failure
525 int qib_check_send_wqe(struct rvt_qp *qp,
526 struct rvt_swqe *wqe)
531 switch (qp->ibqp.qp_type) {
534 if (wqe->length > 0x80000000U)
540 ah = ibah_to_rvtah(wqe->ud_wr.ah);
541 if (wqe->length > (1 << ah->log_pmtu))
552 #ifdef CONFIG_DEBUG_FS
555 struct qib_ibdev *dev;
560 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
562 struct qib_qp_iter *iter;
564 iter = kzalloc(sizeof(*iter), GFP_KERNEL);
573 int qib_qp_iter_next(struct qib_qp_iter *iter)
575 struct qib_ibdev *dev = iter->dev;
578 struct rvt_qp *pqp = iter->qp;
581 for (; n < dev->rdi.qp_dev->qp_table_size; n++) {
583 qp = rcu_dereference(pqp->next);
585 qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]);
596 static const char * const qp_type_str[] = {
597 "SMI", "GSI", "RC", "UC", "UD",
600 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
602 struct rvt_swqe *wqe;
603 struct rvt_qp *qp = iter->qp;
604 struct qib_qp_priv *priv = qp->priv;
606 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
608 "N %d QP%u %s %u %u %u f=%x %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u) QP%u LID %x\n",
611 qp_type_str[qp->ibqp.qp_type],
616 atomic_read(&priv->s_dma_busy),
617 !list_empty(&priv->iowait),
622 qp->s_psn, qp->s_next_psn,
623 qp->s_sending_psn, qp->s_sending_hpsn,
624 qp->s_last, qp->s_acked, qp->s_cur,
625 qp->s_tail, qp->s_head, qp->s_size,
627 qp->remote_ah_attr.dlid);