2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
5 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
8 * This software is available to you under a choice of one of two
9 * licenses. You may choose to be licensed under the terms of the GNU
10 * General Public License (GPL) Version 2, available from the file
11 * COPYING in the main directory of this source tree, or the
12 * OpenIB.org BSD license below:
14 * Redistribution and use in source and binary forms, with or
15 * without modification, are permitted provided that the following
18 * - Redistributions of source code must retain the above
19 * copyright notice, this list of conditions and the following
22 * - Redistributions in binary form must reproduce the above
23 * copyright notice, this list of conditions and the following
24 * disclaimer in the documentation and/or other materials
25 * provided with the distribution.
27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 * $Id: mthca_provider.c 4859 2006-01-09 21:55:10Z roland $
39 #include <rdma/ib_smi.h>
40 #include <rdma/ib_user_verbs.h>
43 #include "mthca_dev.h"
44 #include "mthca_cmd.h"
45 #include "mthca_user.h"
46 #include "mthca_memfree.h"
48 static void init_query_mad(struct ib_smp *mad)
50 mad->base_version = 1;
51 mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
52 mad->class_version = 1;
53 mad->method = IB_MGMT_METHOD_GET;
56 static int mthca_query_device(struct ib_device *ibdev,
57 struct ib_device_attr *props)
59 struct ib_smp *in_mad = NULL;
60 struct ib_smp *out_mad = NULL;
62 struct mthca_dev* mdev = to_mdev(ibdev);
66 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
67 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
68 if (!in_mad || !out_mad)
71 memset(props, 0, sizeof *props);
73 props->fw_ver = mdev->fw_ver;
75 init_query_mad(in_mad);
76 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
78 err = mthca_MAD_IFC(mdev, 1, 1,
79 1, NULL, NULL, in_mad, out_mad,
88 props->device_cap_flags = mdev->device_cap_flags;
89 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
91 props->vendor_part_id = be16_to_cpup((__be16 *) (out_mad->data + 30));
92 props->hw_ver = be32_to_cpup((__be32 *) (out_mad->data + 32));
93 memcpy(&props->sys_image_guid, out_mad->data + 4, 8);
95 props->max_mr_size = ~0ull;
96 props->page_size_cap = mdev->limits.page_size_cap;
97 props->max_qp = mdev->limits.num_qps - mdev->limits.reserved_qps;
98 props->max_qp_wr = mdev->limits.max_wqes;
99 props->max_sge = mdev->limits.max_sg;
100 props->max_cq = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
101 props->max_cqe = mdev->limits.max_cqes;
102 props->max_mr = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
103 props->max_pd = mdev->limits.num_pds - mdev->limits.reserved_pds;
104 props->max_qp_rd_atom = 1 << mdev->qp_table.rdb_shift;
105 props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
106 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
107 props->max_srq = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
108 props->max_srq_wr = mdev->limits.max_srq_wqes;
109 props->max_srq_sge = mdev->limits.max_srq_sge;
110 props->local_ca_ack_delay = mdev->limits.local_ca_ack_delay;
111 props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
112 IB_ATOMIC_HCA : IB_ATOMIC_NONE;
113 props->max_pkeys = mdev->limits.pkey_table_len;
114 props->max_mcast_grp = mdev->limits.num_mgms + mdev->limits.num_amgms;
115 props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
116 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
117 props->max_mcast_grp;
119 * If Sinai memory key optimization is being used, then only
120 * the 8-bit key portion will change. For other HCAs, the
121 * unused index bits will also be used for FMR remapping.
123 if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
124 props->max_map_per_fmr = 255;
126 props->max_map_per_fmr =
127 (1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
136 static int mthca_query_port(struct ib_device *ibdev,
137 u8 port, struct ib_port_attr *props)
139 struct ib_smp *in_mad = NULL;
140 struct ib_smp *out_mad = NULL;
144 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
145 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
146 if (!in_mad || !out_mad)
149 memset(props, 0, sizeof *props);
151 init_query_mad(in_mad);
152 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
153 in_mad->attr_mod = cpu_to_be32(port);
155 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
156 port, NULL, NULL, in_mad, out_mad,
165 props->lid = be16_to_cpup((__be16 *) (out_mad->data + 16));
166 props->lmc = out_mad->data[34] & 0x7;
167 props->sm_lid = be16_to_cpup((__be16 *) (out_mad->data + 18));
168 props->sm_sl = out_mad->data[36] & 0xf;
169 props->state = out_mad->data[32] & 0xf;
170 props->phys_state = out_mad->data[33] >> 4;
171 props->port_cap_flags = be32_to_cpup((__be32 *) (out_mad->data + 20));
172 props->gid_tbl_len = to_mdev(ibdev)->limits.gid_table_len;
173 props->max_msg_sz = 0x80000000;
174 props->pkey_tbl_len = to_mdev(ibdev)->limits.pkey_table_len;
175 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46));
176 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48));
177 props->active_width = out_mad->data[31] & 0xf;
178 props->active_speed = out_mad->data[35] >> 4;
179 props->max_mtu = out_mad->data[41] & 0xf;
180 props->active_mtu = out_mad->data[36] >> 4;
181 props->subnet_timeout = out_mad->data[51] & 0x1f;
182 props->max_vl_num = out_mad->data[37] >> 4;
183 props->init_type_reply = out_mad->data[41] >> 4;
191 static int mthca_modify_device(struct ib_device *ibdev,
193 struct ib_device_modify *props)
195 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
198 if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
199 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
201 memcpy(ibdev->node_desc, props->node_desc, 64);
202 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
208 static int mthca_modify_port(struct ib_device *ibdev,
209 u8 port, int port_modify_mask,
210 struct ib_port_modify *props)
212 struct mthca_set_ib_param set_ib;
213 struct ib_port_attr attr;
217 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
220 err = mthca_query_port(ibdev, port, &attr);
224 set_ib.set_si_guid = 0;
225 set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
227 set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
228 ~props->clr_port_cap_mask;
230 err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
239 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
243 static int mthca_query_pkey(struct ib_device *ibdev,
244 u8 port, u16 index, u16 *pkey)
246 struct ib_smp *in_mad = NULL;
247 struct ib_smp *out_mad = NULL;
251 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
252 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
253 if (!in_mad || !out_mad)
256 init_query_mad(in_mad);
257 in_mad->attr_id = IB_SMP_ATTR_PKEY_TABLE;
258 in_mad->attr_mod = cpu_to_be32(index / 32);
260 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
261 port, NULL, NULL, in_mad, out_mad,
270 *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
278 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
279 int index, union ib_gid *gid)
281 struct ib_smp *in_mad = NULL;
282 struct ib_smp *out_mad = NULL;
286 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
287 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
288 if (!in_mad || !out_mad)
291 init_query_mad(in_mad);
292 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
293 in_mad->attr_mod = cpu_to_be32(port);
295 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
296 port, NULL, NULL, in_mad, out_mad,
305 memcpy(gid->raw, out_mad->data + 8, 8);
307 init_query_mad(in_mad);
308 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
309 in_mad->attr_mod = cpu_to_be32(index / 8);
311 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
312 port, NULL, NULL, in_mad, out_mad,
321 memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
329 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
330 struct ib_udata *udata)
332 struct mthca_alloc_ucontext_resp uresp;
333 struct mthca_ucontext *context;
336 memset(&uresp, 0, sizeof uresp);
338 uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
339 if (mthca_is_memfree(to_mdev(ibdev)))
340 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
344 context = kmalloc(sizeof *context, GFP_KERNEL);
346 return ERR_PTR(-ENOMEM);
348 err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
354 context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
355 if (IS_ERR(context->db_tab)) {
356 err = PTR_ERR(context->db_tab);
357 mthca_uar_free(to_mdev(ibdev), &context->uar);
362 if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
363 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
364 mthca_uar_free(to_mdev(ibdev), &context->uar);
366 return ERR_PTR(-EFAULT);
369 return &context->ibucontext;
372 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
374 mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
375 to_mucontext(context)->db_tab);
376 mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
377 kfree(to_mucontext(context));
382 static int mthca_mmap_uar(struct ib_ucontext *context,
383 struct vm_area_struct *vma)
385 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
388 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
390 if (io_remap_pfn_range(vma, vma->vm_start,
391 to_mucontext(context)->uar.pfn,
392 PAGE_SIZE, vma->vm_page_prot))
398 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
399 struct ib_ucontext *context,
400 struct ib_udata *udata)
405 pd = kmalloc(sizeof *pd, GFP_KERNEL);
407 return ERR_PTR(-ENOMEM);
409 err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
416 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
417 mthca_pd_free(to_mdev(ibdev), pd);
419 return ERR_PTR(-EFAULT);
426 static int mthca_dealloc_pd(struct ib_pd *pd)
428 mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
434 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
435 struct ib_ah_attr *ah_attr)
440 ah = kmalloc(sizeof *ah, GFP_ATOMIC);
442 return ERR_PTR(-ENOMEM);
444 err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
453 static int mthca_ah_destroy(struct ib_ah *ah)
455 mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
461 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
462 struct ib_srq_init_attr *init_attr,
463 struct ib_udata *udata)
465 struct mthca_create_srq ucmd;
466 struct mthca_ucontext *context = NULL;
467 struct mthca_srq *srq;
470 srq = kmalloc(sizeof *srq, GFP_KERNEL);
472 return ERR_PTR(-ENOMEM);
475 context = to_mucontext(pd->uobject->context);
477 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
482 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
483 context->db_tab, ucmd.db_index,
489 srq->mr.ibmr.lkey = ucmd.lkey;
490 srq->db_index = ucmd.db_index;
493 err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
494 &init_attr->attr, srq);
496 if (err && pd->uobject)
497 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
498 context->db_tab, ucmd.db_index);
503 if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
504 mthca_free_srq(to_mdev(pd->device), srq);
517 static int mthca_destroy_srq(struct ib_srq *srq)
519 struct mthca_ucontext *context;
522 context = to_mucontext(srq->uobject->context);
524 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
525 context->db_tab, to_msrq(srq)->db_index);
528 mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
534 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
535 struct ib_qp_init_attr *init_attr,
536 struct ib_udata *udata)
538 struct mthca_create_qp ucmd;
542 switch (init_attr->qp_type) {
547 struct mthca_ucontext *context;
549 qp = kmalloc(sizeof *qp, GFP_KERNEL);
551 return ERR_PTR(-ENOMEM);
554 context = to_mucontext(pd->uobject->context);
556 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
558 return ERR_PTR(-EFAULT);
561 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
563 ucmd.sq_db_index, ucmd.sq_db_page);
569 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
571 ucmd.rq_db_index, ucmd.rq_db_page);
573 mthca_unmap_user_db(to_mdev(pd->device),
581 qp->mr.ibmr.lkey = ucmd.lkey;
582 qp->sq.db_index = ucmd.sq_db_index;
583 qp->rq.db_index = ucmd.rq_db_index;
586 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
587 to_mcq(init_attr->send_cq),
588 to_mcq(init_attr->recv_cq),
589 init_attr->qp_type, init_attr->sq_sig_type,
590 &init_attr->cap, qp);
592 if (err && pd->uobject) {
593 context = to_mucontext(pd->uobject->context);
595 mthca_unmap_user_db(to_mdev(pd->device),
599 mthca_unmap_user_db(to_mdev(pd->device),
605 qp->ibqp.qp_num = qp->qpn;
611 /* Don't allow userspace to create special QPs */
613 return ERR_PTR(-EINVAL);
615 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
617 return ERR_PTR(-ENOMEM);
619 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
621 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
622 to_mcq(init_attr->send_cq),
623 to_mcq(init_attr->recv_cq),
624 init_attr->sq_sig_type, &init_attr->cap,
625 qp->ibqp.qp_num, init_attr->port_num,
630 /* Don't support raw QPs */
631 return ERR_PTR(-ENOSYS);
639 init_attr->cap.max_send_wr = qp->sq.max;
640 init_attr->cap.max_recv_wr = qp->rq.max;
641 init_attr->cap.max_send_sge = qp->sq.max_gs;
642 init_attr->cap.max_recv_sge = qp->rq.max_gs;
643 init_attr->cap.max_inline_data = qp->max_inline_data;
648 static int mthca_destroy_qp(struct ib_qp *qp)
651 mthca_unmap_user_db(to_mdev(qp->device),
652 &to_mucontext(qp->uobject->context)->uar,
653 to_mucontext(qp->uobject->context)->db_tab,
654 to_mqp(qp)->sq.db_index);
655 mthca_unmap_user_db(to_mdev(qp->device),
656 &to_mucontext(qp->uobject->context)->uar,
657 to_mucontext(qp->uobject->context)->db_tab,
658 to_mqp(qp)->rq.db_index);
660 mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
665 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
667 struct ib_ucontext *context,
668 struct ib_udata *udata)
670 struct mthca_create_cq ucmd;
675 if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
676 return ERR_PTR(-EINVAL);
679 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
680 return ERR_PTR(-EFAULT);
682 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
683 to_mucontext(context)->db_tab,
684 ucmd.set_db_index, ucmd.set_db_page);
688 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
689 to_mucontext(context)->db_tab,
690 ucmd.arm_db_index, ucmd.arm_db_page);
695 cq = kmalloc(sizeof *cq, GFP_KERNEL);
702 cq->buf.mr.ibmr.lkey = ucmd.lkey;
703 cq->set_ci_db_index = ucmd.set_db_index;
704 cq->arm_db_index = ucmd.arm_db_index;
707 for (nent = 1; nent <= entries; nent <<= 1)
710 err = mthca_init_cq(to_mdev(ibdev), nent,
711 context ? to_mucontext(context) : NULL,
712 context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
717 if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
718 mthca_free_cq(to_mdev(ibdev), cq);
722 cq->resize_buf = NULL;
731 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
732 to_mucontext(context)->db_tab, ucmd.arm_db_index);
736 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
737 to_mucontext(context)->db_tab, ucmd.set_db_index);
742 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
747 spin_lock_irq(&cq->lock);
748 if (cq->resize_buf) {
753 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
754 if (!cq->resize_buf) {
759 cq->resize_buf->state = CQ_RESIZE_ALLOC;
764 spin_unlock_irq(&cq->lock);
769 ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
771 spin_lock_irq(&cq->lock);
772 kfree(cq->resize_buf);
773 cq->resize_buf = NULL;
774 spin_unlock_irq(&cq->lock);
778 cq->resize_buf->cqe = entries - 1;
780 spin_lock_irq(&cq->lock);
781 cq->resize_buf->state = CQ_RESIZE_READY;
782 spin_unlock_irq(&cq->lock);
787 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
789 struct mthca_dev *dev = to_mdev(ibcq->device);
790 struct mthca_cq *cq = to_mcq(ibcq);
791 struct mthca_resize_cq ucmd;
796 if (entries < 1 || entries > dev->limits.max_cqes)
799 mutex_lock(&cq->mutex);
801 entries = roundup_pow_of_two(entries + 1);
802 if (entries == ibcq->cqe + 1) {
808 ret = mthca_alloc_resize_buf(dev, cq, entries);
811 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
813 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
820 ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries), &status);
825 if (cq->resize_buf) {
826 mthca_free_cq_buf(dev, &cq->resize_buf->buf,
827 cq->resize_buf->cqe);
828 kfree(cq->resize_buf);
829 spin_lock_irq(&cq->lock);
830 cq->resize_buf = NULL;
831 spin_unlock_irq(&cq->lock);
837 struct mthca_cq_buf tbuf;
840 spin_lock_irq(&cq->lock);
841 if (cq->resize_buf->state == CQ_RESIZE_READY) {
842 mthca_cq_resize_copy_cqes(cq);
845 cq->buf = cq->resize_buf->buf;
846 cq->ibcq.cqe = cq->resize_buf->cqe;
848 tbuf = cq->resize_buf->buf;
849 tcqe = cq->resize_buf->cqe;
852 kfree(cq->resize_buf);
853 cq->resize_buf = NULL;
854 spin_unlock_irq(&cq->lock);
856 mthca_free_cq_buf(dev, &tbuf, tcqe);
858 ibcq->cqe = entries - 1;
861 mutex_unlock(&cq->mutex);
866 static int mthca_destroy_cq(struct ib_cq *cq)
869 mthca_unmap_user_db(to_mdev(cq->device),
870 &to_mucontext(cq->uobject->context)->uar,
871 to_mucontext(cq->uobject->context)->db_tab,
872 to_mcq(cq)->arm_db_index);
873 mthca_unmap_user_db(to_mdev(cq->device),
874 &to_mucontext(cq->uobject->context)->uar,
875 to_mucontext(cq->uobject->context)->db_tab,
876 to_mcq(cq)->set_ci_db_index);
878 mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
884 static inline u32 convert_access(int acc)
886 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC : 0) |
887 (acc & IB_ACCESS_REMOTE_WRITE ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
888 (acc & IB_ACCESS_REMOTE_READ ? MTHCA_MPT_FLAG_REMOTE_READ : 0) |
889 (acc & IB_ACCESS_LOCAL_WRITE ? MTHCA_MPT_FLAG_LOCAL_WRITE : 0) |
890 MTHCA_MPT_FLAG_LOCAL_READ;
893 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
898 mr = kmalloc(sizeof *mr, GFP_KERNEL);
900 return ERR_PTR(-ENOMEM);
902 err = mthca_mr_alloc_notrans(to_mdev(pd->device),
904 convert_access(acc), mr);
914 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd *pd,
915 struct ib_phys_buf *buffer_list,
929 /* First check that we have enough alignment */
930 if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
931 return ERR_PTR(-EINVAL);
935 for (i = 0; i < num_phys_buf; ++i) {
937 mask |= buffer_list[i].addr;
938 if (i != num_phys_buf - 1)
939 mask |= buffer_list[i].addr + buffer_list[i].size;
941 total_size += buffer_list[i].size;
944 if (mask & ~PAGE_MASK)
945 return ERR_PTR(-EINVAL);
947 /* Find largest page shift we can use to cover buffers */
948 for (shift = PAGE_SHIFT; shift < 31; ++shift)
949 if (num_phys_buf > 1) {
950 if ((1ULL << shift) & mask)
954 buffer_list[0].size +
955 (buffer_list[0].addr & ((1ULL << shift) - 1)))
959 buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
960 buffer_list[0].addr &= ~0ull << shift;
962 mr = kmalloc(sizeof *mr, GFP_KERNEL);
964 return ERR_PTR(-ENOMEM);
967 for (i = 0; i < num_phys_buf; ++i)
968 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
973 page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
976 return ERR_PTR(-ENOMEM);
980 for (i = 0; i < num_phys_buf; ++i)
982 j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
984 page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
986 mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
987 "in PD %x; shift %d, npages %d.\n",
988 (unsigned long long) buffer_list[0].addr,
989 (unsigned long long) *iova_start,
993 err = mthca_mr_alloc_phys(to_mdev(pd->device),
995 page_list, shift, npages,
996 *iova_start, total_size,
997 convert_access(acc), mr);
1002 return ERR_PTR(err);
1009 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
1010 int acc, struct ib_udata *udata)
1012 struct mthca_dev *dev = to_mdev(pd->device);
1013 struct ib_umem_chunk *chunk;
1014 struct mthca_mr *mr;
1021 shift = ffs(region->page_size) - 1;
1023 mr = kmalloc(sizeof *mr, GFP_KERNEL);
1025 return ERR_PTR(-ENOMEM);
1028 list_for_each_entry(chunk, ®ion->chunk_list, list)
1031 mr->mtt = mthca_alloc_mtt(dev, n);
1032 if (IS_ERR(mr->mtt)) {
1033 err = PTR_ERR(mr->mtt);
1037 pages = (u64 *) __get_free_page(GFP_KERNEL);
1045 write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1047 list_for_each_entry(chunk, ®ion->chunk_list, list)
1048 for (j = 0; j < chunk->nmap; ++j) {
1049 len = sg_dma_len(&chunk->page_list[j]) >> shift;
1050 for (k = 0; k < len; ++k) {
1051 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1052 region->page_size * k;
1054 * Be friendly to write_mtt and pass it chunks
1055 * of appropriate size.
1057 if (i == write_mtt_size) {
1058 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1068 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1070 free_page((unsigned long) pages);
1074 err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
1075 region->length, convert_access(acc), mr);
1083 mthca_free_mtt(dev, mr->mtt);
1087 return ERR_PTR(err);
1090 static int mthca_dereg_mr(struct ib_mr *mr)
1092 struct mthca_mr *mmr = to_mmr(mr);
1093 mthca_free_mr(to_mdev(mr->device), mmr);
1098 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1099 struct ib_fmr_attr *fmr_attr)
1101 struct mthca_fmr *fmr;
1104 fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1106 return ERR_PTR(-ENOMEM);
1108 memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1109 err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1110 convert_access(mr_access_flags), fmr);
1114 return ERR_PTR(err);
1120 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1122 struct mthca_fmr *mfmr = to_mfmr(fmr);
1125 err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1133 static int mthca_unmap_fmr(struct list_head *fmr_list)
1138 struct mthca_dev *mdev = NULL;
1140 list_for_each_entry(fmr, fmr_list, list) {
1141 if (mdev && to_mdev(fmr->device) != mdev)
1143 mdev = to_mdev(fmr->device);
1149 if (mthca_is_memfree(mdev)) {
1150 list_for_each_entry(fmr, fmr_list, list)
1151 mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1155 list_for_each_entry(fmr, fmr_list, list)
1156 mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1158 err = mthca_SYNC_TPT(mdev, &status);
1166 static ssize_t show_rev(struct class_device *cdev, char *buf)
1168 struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1169 return sprintf(buf, "%x\n", dev->rev_id);
1172 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1174 struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1175 return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1176 (int) (dev->fw_ver >> 16) & 0xffff,
1177 (int) dev->fw_ver & 0xffff);
1180 static ssize_t show_hca(struct class_device *cdev, char *buf)
1182 struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1183 switch (dev->pdev->device) {
1184 case PCI_DEVICE_ID_MELLANOX_TAVOR:
1185 return sprintf(buf, "MT23108\n");
1186 case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1187 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1188 case PCI_DEVICE_ID_MELLANOX_ARBEL:
1189 return sprintf(buf, "MT25208\n");
1190 case PCI_DEVICE_ID_MELLANOX_SINAI:
1191 case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1192 return sprintf(buf, "MT25204\n");
1194 return sprintf(buf, "unknown\n");
1198 static ssize_t show_board(struct class_device *cdev, char *buf)
1200 struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1201 return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1204 static CLASS_DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1205 static CLASS_DEVICE_ATTR(fw_ver, S_IRUGO, show_fw_ver, NULL);
1206 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1207 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
1209 static struct class_device_attribute *mthca_class_attributes[] = {
1210 &class_device_attr_hw_rev,
1211 &class_device_attr_fw_ver,
1212 &class_device_attr_hca_type,
1213 &class_device_attr_board_id
1216 static int mthca_init_node_data(struct mthca_dev *dev)
1218 struct ib_smp *in_mad = NULL;
1219 struct ib_smp *out_mad = NULL;
1223 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
1224 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1225 if (!in_mad || !out_mad)
1228 init_query_mad(in_mad);
1229 in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1231 err = mthca_MAD_IFC(dev, 1, 1,
1232 1, NULL, NULL, in_mad, out_mad,
1241 memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1243 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1245 err = mthca_MAD_IFC(dev, 1, 1,
1246 1, NULL, NULL, in_mad, out_mad,
1255 memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1263 int mthca_register_device(struct mthca_dev *dev)
1268 ret = mthca_init_node_data(dev);
1272 strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1273 dev->ib_dev.owner = THIS_MODULE;
1275 dev->ib_dev.uverbs_abi_ver = MTHCA_UVERBS_ABI_VERSION;
1276 dev->ib_dev.uverbs_cmd_mask =
1277 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
1278 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
1279 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
1280 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
1281 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
1282 (1ull << IB_USER_VERBS_CMD_REG_MR) |
1283 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
1284 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1285 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
1286 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
1287 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
1288 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
1289 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
1290 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
1291 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
1292 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
1293 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1294 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
1295 dev->ib_dev.phys_port_cnt = dev->limits.num_ports;
1296 dev->ib_dev.num_comp_vectors = 1;
1297 dev->ib_dev.dma_device = &dev->pdev->dev;
1298 dev->ib_dev.query_device = mthca_query_device;
1299 dev->ib_dev.query_port = mthca_query_port;
1300 dev->ib_dev.modify_device = mthca_modify_device;
1301 dev->ib_dev.modify_port = mthca_modify_port;
1302 dev->ib_dev.query_pkey = mthca_query_pkey;
1303 dev->ib_dev.query_gid = mthca_query_gid;
1304 dev->ib_dev.alloc_ucontext = mthca_alloc_ucontext;
1305 dev->ib_dev.dealloc_ucontext = mthca_dealloc_ucontext;
1306 dev->ib_dev.mmap = mthca_mmap_uar;
1307 dev->ib_dev.alloc_pd = mthca_alloc_pd;
1308 dev->ib_dev.dealloc_pd = mthca_dealloc_pd;
1309 dev->ib_dev.create_ah = mthca_ah_create;
1310 dev->ib_dev.query_ah = mthca_ah_query;
1311 dev->ib_dev.destroy_ah = mthca_ah_destroy;
1313 if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1314 dev->ib_dev.create_srq = mthca_create_srq;
1315 dev->ib_dev.modify_srq = mthca_modify_srq;
1316 dev->ib_dev.query_srq = mthca_query_srq;
1317 dev->ib_dev.destroy_srq = mthca_destroy_srq;
1318 dev->ib_dev.uverbs_cmd_mask |=
1319 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
1320 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
1321 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
1322 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1324 if (mthca_is_memfree(dev))
1325 dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1327 dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1330 dev->ib_dev.create_qp = mthca_create_qp;
1331 dev->ib_dev.modify_qp = mthca_modify_qp;
1332 dev->ib_dev.query_qp = mthca_query_qp;
1333 dev->ib_dev.destroy_qp = mthca_destroy_qp;
1334 dev->ib_dev.create_cq = mthca_create_cq;
1335 dev->ib_dev.resize_cq = mthca_resize_cq;
1336 dev->ib_dev.destroy_cq = mthca_destroy_cq;
1337 dev->ib_dev.poll_cq = mthca_poll_cq;
1338 dev->ib_dev.get_dma_mr = mthca_get_dma_mr;
1339 dev->ib_dev.reg_phys_mr = mthca_reg_phys_mr;
1340 dev->ib_dev.reg_user_mr = mthca_reg_user_mr;
1341 dev->ib_dev.dereg_mr = mthca_dereg_mr;
1343 if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1344 dev->ib_dev.alloc_fmr = mthca_alloc_fmr;
1345 dev->ib_dev.unmap_fmr = mthca_unmap_fmr;
1346 dev->ib_dev.dealloc_fmr = mthca_dealloc_fmr;
1347 if (mthca_is_memfree(dev))
1348 dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1350 dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1353 dev->ib_dev.attach_mcast = mthca_multicast_attach;
1354 dev->ib_dev.detach_mcast = mthca_multicast_detach;
1355 dev->ib_dev.process_mad = mthca_process_mad;
1357 if (mthca_is_memfree(dev)) {
1358 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1359 dev->ib_dev.post_send = mthca_arbel_post_send;
1360 dev->ib_dev.post_recv = mthca_arbel_post_receive;
1362 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1363 dev->ib_dev.post_send = mthca_tavor_post_send;
1364 dev->ib_dev.post_recv = mthca_tavor_post_receive;
1367 mutex_init(&dev->cap_mask_mutex);
1369 ret = ib_register_device(&dev->ib_dev);
1373 for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1374 ret = class_device_create_file(&dev->ib_dev.class_dev,
1375 mthca_class_attributes[i]);
1377 ib_unregister_device(&dev->ib_dev);
1382 mthca_start_catas_poll(dev);
1387 void mthca_unregister_device(struct mthca_dev *dev)
1389 mthca_stop_catas_poll(dev);
1390 ib_unregister_device(&dev->ib_dev);