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
37 #include <rdma/ib_smi.h>
38 #include <rdma/ib_umem.h>
39 #include <rdma/ib_user_verbs.h>
41 #include <linux/sched.h>
42 #include <linux/slab.h>
45 #include "mthca_dev.h"
46 #include "mthca_cmd.h"
47 #include "mthca_user.h"
48 #include "mthca_memfree.h"
50 static void init_query_mad(struct ib_smp *mad)
52 mad->base_version = 1;
53 mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
54 mad->class_version = 1;
55 mad->method = IB_MGMT_METHOD_GET;
58 static int mthca_query_device(struct ib_device *ibdev,
59 struct ib_device_attr *props)
61 struct ib_smp *in_mad = NULL;
62 struct ib_smp *out_mad = NULL;
64 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);
83 props->device_cap_flags = mdev->device_cap_flags;
84 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
86 props->vendor_part_id = be16_to_cpup((__be16 *) (out_mad->data + 30));
87 props->hw_ver = be32_to_cpup((__be32 *) (out_mad->data + 32));
88 memcpy(&props->sys_image_guid, out_mad->data + 4, 8);
90 props->max_mr_size = ~0ull;
91 props->page_size_cap = mdev->limits.page_size_cap;
92 props->max_qp = mdev->limits.num_qps - mdev->limits.reserved_qps;
93 props->max_qp_wr = mdev->limits.max_wqes;
94 props->max_sge = mdev->limits.max_sg;
95 props->max_cq = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
96 props->max_cqe = mdev->limits.max_cqes;
97 props->max_mr = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
98 props->max_pd = mdev->limits.num_pds - mdev->limits.reserved_pds;
99 props->max_qp_rd_atom = 1 << mdev->qp_table.rdb_shift;
100 props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
101 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
102 props->max_srq = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
103 props->max_srq_wr = mdev->limits.max_srq_wqes;
104 props->max_srq_sge = mdev->limits.max_srq_sge;
105 props->local_ca_ack_delay = mdev->limits.local_ca_ack_delay;
106 props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
107 IB_ATOMIC_HCA : IB_ATOMIC_NONE;
108 props->max_pkeys = mdev->limits.pkey_table_len;
109 props->max_mcast_grp = mdev->limits.num_mgms + mdev->limits.num_amgms;
110 props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
111 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
112 props->max_mcast_grp;
114 * If Sinai memory key optimization is being used, then only
115 * the 8-bit key portion will change. For other HCAs, the
116 * unused index bits will also be used for FMR remapping.
118 if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
119 props->max_map_per_fmr = 255;
121 props->max_map_per_fmr =
122 (1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
131 static int mthca_query_port(struct ib_device *ibdev,
132 u8 port, struct ib_port_attr *props)
134 struct ib_smp *in_mad = NULL;
135 struct ib_smp *out_mad = NULL;
138 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
139 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
140 if (!in_mad || !out_mad)
143 memset(props, 0, sizeof *props);
145 init_query_mad(in_mad);
146 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
147 in_mad->attr_mod = cpu_to_be32(port);
149 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
150 port, NULL, NULL, in_mad, out_mad);
154 props->lid = be16_to_cpup((__be16 *) (out_mad->data + 16));
155 props->lmc = out_mad->data[34] & 0x7;
156 props->sm_lid = be16_to_cpup((__be16 *) (out_mad->data + 18));
157 props->sm_sl = out_mad->data[36] & 0xf;
158 props->state = out_mad->data[32] & 0xf;
159 props->phys_state = out_mad->data[33] >> 4;
160 props->port_cap_flags = be32_to_cpup((__be32 *) (out_mad->data + 20));
161 props->gid_tbl_len = to_mdev(ibdev)->limits.gid_table_len;
162 props->max_msg_sz = 0x80000000;
163 props->pkey_tbl_len = to_mdev(ibdev)->limits.pkey_table_len;
164 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46));
165 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48));
166 props->active_width = out_mad->data[31] & 0xf;
167 props->active_speed = out_mad->data[35] >> 4;
168 props->max_mtu = out_mad->data[41] & 0xf;
169 props->active_mtu = out_mad->data[36] >> 4;
170 props->subnet_timeout = out_mad->data[51] & 0x1f;
171 props->max_vl_num = out_mad->data[37] >> 4;
172 props->init_type_reply = out_mad->data[41] >> 4;
180 static int mthca_modify_device(struct ib_device *ibdev,
182 struct ib_device_modify *props)
184 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
187 if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
188 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
190 memcpy(ibdev->node_desc, props->node_desc, 64);
191 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
197 static int mthca_modify_port(struct ib_device *ibdev,
198 u8 port, int port_modify_mask,
199 struct ib_port_modify *props)
201 struct mthca_set_ib_param set_ib;
202 struct ib_port_attr attr;
205 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
208 err = mthca_query_port(ibdev, port, &attr);
212 set_ib.set_si_guid = 0;
213 set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
215 set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
216 ~props->clr_port_cap_mask;
218 err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
222 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
226 static int mthca_query_pkey(struct ib_device *ibdev,
227 u8 port, u16 index, u16 *pkey)
229 struct ib_smp *in_mad = NULL;
230 struct ib_smp *out_mad = NULL;
233 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
234 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
235 if (!in_mad || !out_mad)
238 init_query_mad(in_mad);
239 in_mad->attr_id = IB_SMP_ATTR_PKEY_TABLE;
240 in_mad->attr_mod = cpu_to_be32(index / 32);
242 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
243 port, NULL, NULL, in_mad, out_mad);
247 *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
255 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
256 int index, union ib_gid *gid)
258 struct ib_smp *in_mad = NULL;
259 struct ib_smp *out_mad = NULL;
262 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
263 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
264 if (!in_mad || !out_mad)
267 init_query_mad(in_mad);
268 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
269 in_mad->attr_mod = cpu_to_be32(port);
271 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
272 port, NULL, NULL, in_mad, out_mad);
276 memcpy(gid->raw, out_mad->data + 8, 8);
278 init_query_mad(in_mad);
279 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
280 in_mad->attr_mod = cpu_to_be32(index / 8);
282 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
283 port, NULL, NULL, in_mad, out_mad);
287 memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
295 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
296 struct ib_udata *udata)
298 struct mthca_alloc_ucontext_resp uresp;
299 struct mthca_ucontext *context;
302 if (!(to_mdev(ibdev)->active))
303 return ERR_PTR(-EAGAIN);
305 memset(&uresp, 0, sizeof uresp);
307 uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
308 if (mthca_is_memfree(to_mdev(ibdev)))
309 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
313 context = kmalloc(sizeof *context, GFP_KERNEL);
315 return ERR_PTR(-ENOMEM);
317 err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
323 context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
324 if (IS_ERR(context->db_tab)) {
325 err = PTR_ERR(context->db_tab);
326 mthca_uar_free(to_mdev(ibdev), &context->uar);
331 if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
332 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
333 mthca_uar_free(to_mdev(ibdev), &context->uar);
335 return ERR_PTR(-EFAULT);
338 context->reg_mr_warned = 0;
340 return &context->ibucontext;
343 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
345 mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
346 to_mucontext(context)->db_tab);
347 mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
348 kfree(to_mucontext(context));
353 static int mthca_mmap_uar(struct ib_ucontext *context,
354 struct vm_area_struct *vma)
356 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
359 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
361 if (io_remap_pfn_range(vma, vma->vm_start,
362 to_mucontext(context)->uar.pfn,
363 PAGE_SIZE, vma->vm_page_prot))
369 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
370 struct ib_ucontext *context,
371 struct ib_udata *udata)
376 pd = kmalloc(sizeof *pd, GFP_KERNEL);
378 return ERR_PTR(-ENOMEM);
380 err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
387 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
388 mthca_pd_free(to_mdev(ibdev), pd);
390 return ERR_PTR(-EFAULT);
397 static int mthca_dealloc_pd(struct ib_pd *pd)
399 mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
405 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
406 struct ib_ah_attr *ah_attr)
411 ah = kmalloc(sizeof *ah, GFP_ATOMIC);
413 return ERR_PTR(-ENOMEM);
415 err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
424 static int mthca_ah_destroy(struct ib_ah *ah)
426 mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
432 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
433 struct ib_srq_init_attr *init_attr,
434 struct ib_udata *udata)
436 struct mthca_create_srq ucmd;
437 struct mthca_ucontext *context = NULL;
438 struct mthca_srq *srq;
441 srq = kmalloc(sizeof *srq, GFP_KERNEL);
443 return ERR_PTR(-ENOMEM);
446 context = to_mucontext(pd->uobject->context);
448 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
453 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
454 context->db_tab, ucmd.db_index,
460 srq->mr.ibmr.lkey = ucmd.lkey;
461 srq->db_index = ucmd.db_index;
464 err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
465 &init_attr->attr, srq);
467 if (err && pd->uobject)
468 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
469 context->db_tab, ucmd.db_index);
474 if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
475 mthca_free_srq(to_mdev(pd->device), srq);
488 static int mthca_destroy_srq(struct ib_srq *srq)
490 struct mthca_ucontext *context;
493 context = to_mucontext(srq->uobject->context);
495 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
496 context->db_tab, to_msrq(srq)->db_index);
499 mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
505 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
506 struct ib_qp_init_attr *init_attr,
507 struct ib_udata *udata)
509 struct mthca_create_qp ucmd;
513 if (init_attr->create_flags)
514 return ERR_PTR(-EINVAL);
516 switch (init_attr->qp_type) {
521 struct mthca_ucontext *context;
523 qp = kmalloc(sizeof *qp, GFP_KERNEL);
525 return ERR_PTR(-ENOMEM);
528 context = to_mucontext(pd->uobject->context);
530 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
532 return ERR_PTR(-EFAULT);
535 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
537 ucmd.sq_db_index, ucmd.sq_db_page);
543 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
545 ucmd.rq_db_index, ucmd.rq_db_page);
547 mthca_unmap_user_db(to_mdev(pd->device),
555 qp->mr.ibmr.lkey = ucmd.lkey;
556 qp->sq.db_index = ucmd.sq_db_index;
557 qp->rq.db_index = ucmd.rq_db_index;
560 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
561 to_mcq(init_attr->send_cq),
562 to_mcq(init_attr->recv_cq),
563 init_attr->qp_type, init_attr->sq_sig_type,
564 &init_attr->cap, qp);
566 if (err && pd->uobject) {
567 context = to_mucontext(pd->uobject->context);
569 mthca_unmap_user_db(to_mdev(pd->device),
573 mthca_unmap_user_db(to_mdev(pd->device),
579 qp->ibqp.qp_num = qp->qpn;
585 /* Don't allow userspace to create special QPs */
587 return ERR_PTR(-EINVAL);
589 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
591 return ERR_PTR(-ENOMEM);
593 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
595 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
596 to_mcq(init_attr->send_cq),
597 to_mcq(init_attr->recv_cq),
598 init_attr->sq_sig_type, &init_attr->cap,
599 qp->ibqp.qp_num, init_attr->port_num,
604 /* Don't support raw QPs */
605 return ERR_PTR(-ENOSYS);
613 init_attr->cap.max_send_wr = qp->sq.max;
614 init_attr->cap.max_recv_wr = qp->rq.max;
615 init_attr->cap.max_send_sge = qp->sq.max_gs;
616 init_attr->cap.max_recv_sge = qp->rq.max_gs;
617 init_attr->cap.max_inline_data = qp->max_inline_data;
622 static int mthca_destroy_qp(struct ib_qp *qp)
625 mthca_unmap_user_db(to_mdev(qp->device),
626 &to_mucontext(qp->uobject->context)->uar,
627 to_mucontext(qp->uobject->context)->db_tab,
628 to_mqp(qp)->sq.db_index);
629 mthca_unmap_user_db(to_mdev(qp->device),
630 &to_mucontext(qp->uobject->context)->uar,
631 to_mucontext(qp->uobject->context)->db_tab,
632 to_mqp(qp)->rq.db_index);
634 mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
639 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
641 struct ib_ucontext *context,
642 struct ib_udata *udata)
644 struct mthca_create_cq ucmd;
649 if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
650 return ERR_PTR(-EINVAL);
653 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
654 return ERR_PTR(-EFAULT);
656 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
657 to_mucontext(context)->db_tab,
658 ucmd.set_db_index, ucmd.set_db_page);
662 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
663 to_mucontext(context)->db_tab,
664 ucmd.arm_db_index, ucmd.arm_db_page);
669 cq = kmalloc(sizeof *cq, GFP_KERNEL);
676 cq->buf.mr.ibmr.lkey = ucmd.lkey;
677 cq->set_ci_db_index = ucmd.set_db_index;
678 cq->arm_db_index = ucmd.arm_db_index;
681 for (nent = 1; nent <= entries; nent <<= 1)
684 err = mthca_init_cq(to_mdev(ibdev), nent,
685 context ? to_mucontext(context) : NULL,
686 context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
691 if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
692 mthca_free_cq(to_mdev(ibdev), cq);
696 cq->resize_buf = NULL;
705 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
706 to_mucontext(context)->db_tab, ucmd.arm_db_index);
710 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
711 to_mucontext(context)->db_tab, ucmd.set_db_index);
716 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
721 spin_lock_irq(&cq->lock);
722 if (cq->resize_buf) {
727 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
728 if (!cq->resize_buf) {
733 cq->resize_buf->state = CQ_RESIZE_ALLOC;
738 spin_unlock_irq(&cq->lock);
743 ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
745 spin_lock_irq(&cq->lock);
746 kfree(cq->resize_buf);
747 cq->resize_buf = NULL;
748 spin_unlock_irq(&cq->lock);
752 cq->resize_buf->cqe = entries - 1;
754 spin_lock_irq(&cq->lock);
755 cq->resize_buf->state = CQ_RESIZE_READY;
756 spin_unlock_irq(&cq->lock);
761 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
763 struct mthca_dev *dev = to_mdev(ibcq->device);
764 struct mthca_cq *cq = to_mcq(ibcq);
765 struct mthca_resize_cq ucmd;
769 if (entries < 1 || entries > dev->limits.max_cqes)
772 mutex_lock(&cq->mutex);
774 entries = roundup_pow_of_two(entries + 1);
775 if (entries == ibcq->cqe + 1) {
781 ret = mthca_alloc_resize_buf(dev, cq, entries);
784 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
786 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
793 ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
796 if (cq->resize_buf) {
797 mthca_free_cq_buf(dev, &cq->resize_buf->buf,
798 cq->resize_buf->cqe);
799 kfree(cq->resize_buf);
800 spin_lock_irq(&cq->lock);
801 cq->resize_buf = NULL;
802 spin_unlock_irq(&cq->lock);
808 struct mthca_cq_buf tbuf;
811 spin_lock_irq(&cq->lock);
812 if (cq->resize_buf->state == CQ_RESIZE_READY) {
813 mthca_cq_resize_copy_cqes(cq);
816 cq->buf = cq->resize_buf->buf;
817 cq->ibcq.cqe = cq->resize_buf->cqe;
819 tbuf = cq->resize_buf->buf;
820 tcqe = cq->resize_buf->cqe;
823 kfree(cq->resize_buf);
824 cq->resize_buf = NULL;
825 spin_unlock_irq(&cq->lock);
827 mthca_free_cq_buf(dev, &tbuf, tcqe);
829 ibcq->cqe = entries - 1;
832 mutex_unlock(&cq->mutex);
837 static int mthca_destroy_cq(struct ib_cq *cq)
840 mthca_unmap_user_db(to_mdev(cq->device),
841 &to_mucontext(cq->uobject->context)->uar,
842 to_mucontext(cq->uobject->context)->db_tab,
843 to_mcq(cq)->arm_db_index);
844 mthca_unmap_user_db(to_mdev(cq->device),
845 &to_mucontext(cq->uobject->context)->uar,
846 to_mucontext(cq->uobject->context)->db_tab,
847 to_mcq(cq)->set_ci_db_index);
849 mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
855 static inline u32 convert_access(int acc)
857 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC : 0) |
858 (acc & IB_ACCESS_REMOTE_WRITE ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
859 (acc & IB_ACCESS_REMOTE_READ ? MTHCA_MPT_FLAG_REMOTE_READ : 0) |
860 (acc & IB_ACCESS_LOCAL_WRITE ? MTHCA_MPT_FLAG_LOCAL_WRITE : 0) |
861 MTHCA_MPT_FLAG_LOCAL_READ;
864 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
869 mr = kmalloc(sizeof *mr, GFP_KERNEL);
871 return ERR_PTR(-ENOMEM);
873 err = mthca_mr_alloc_notrans(to_mdev(pd->device),
875 convert_access(acc), mr);
887 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd *pd,
888 struct ib_phys_buf *buffer_list,
902 mask = buffer_list[0].addr ^ *iova_start;
904 for (i = 0; i < num_phys_buf; ++i) {
906 mask |= buffer_list[i].addr;
907 if (i != num_phys_buf - 1)
908 mask |= buffer_list[i].addr + buffer_list[i].size;
910 total_size += buffer_list[i].size;
913 if (mask & ~PAGE_MASK)
914 return ERR_PTR(-EINVAL);
916 shift = __ffs(mask | 1 << 31);
918 buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
919 buffer_list[0].addr &= ~0ull << shift;
921 mr = kmalloc(sizeof *mr, GFP_KERNEL);
923 return ERR_PTR(-ENOMEM);
926 for (i = 0; i < num_phys_buf; ++i)
927 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
932 page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
935 return ERR_PTR(-ENOMEM);
939 for (i = 0; i < num_phys_buf; ++i)
941 j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
943 page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
945 mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
946 "in PD %x; shift %d, npages %d.\n",
947 (unsigned long long) buffer_list[0].addr,
948 (unsigned long long) *iova_start,
952 err = mthca_mr_alloc_phys(to_mdev(pd->device),
954 page_list, shift, npages,
955 *iova_start, total_size,
956 convert_access(acc), mr);
970 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
971 u64 virt, int acc, struct ib_udata *udata)
973 struct mthca_dev *dev = to_mdev(pd->device);
974 struct ib_umem_chunk *chunk;
976 struct mthca_reg_mr ucmd;
983 if (udata->inlen - sizeof (struct ib_uverbs_cmd_hdr) < sizeof ucmd) {
984 if (!to_mucontext(pd->uobject->context)->reg_mr_warned) {
985 mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
987 mthca_warn(dev, " Update libmthca to fix this.\n");
989 ++to_mucontext(pd->uobject->context)->reg_mr_warned;
991 } else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
992 return ERR_PTR(-EFAULT);
994 mr = kmalloc(sizeof *mr, GFP_KERNEL);
996 return ERR_PTR(-ENOMEM);
998 mr->umem = ib_umem_get(pd->uobject->context, start, length, acc,
999 ucmd.mr_attrs & MTHCA_MR_DMASYNC);
1001 if (IS_ERR(mr->umem)) {
1002 err = PTR_ERR(mr->umem);
1006 shift = ffs(mr->umem->page_size) - 1;
1009 list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1012 mr->mtt = mthca_alloc_mtt(dev, n);
1013 if (IS_ERR(mr->mtt)) {
1014 err = PTR_ERR(mr->mtt);
1018 pages = (u64 *) __get_free_page(GFP_KERNEL);
1026 write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1028 list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1029 for (j = 0; j < chunk->nmap; ++j) {
1030 len = sg_dma_len(&chunk->page_list[j]) >> shift;
1031 for (k = 0; k < len; ++k) {
1032 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1033 mr->umem->page_size * k;
1035 * Be friendly to write_mtt and pass it chunks
1036 * of appropriate size.
1038 if (i == write_mtt_size) {
1039 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1049 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1051 free_page((unsigned long) pages);
1055 err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
1056 convert_access(acc), mr);
1064 mthca_free_mtt(dev, mr->mtt);
1067 ib_umem_release(mr->umem);
1071 return ERR_PTR(err);
1074 static int mthca_dereg_mr(struct ib_mr *mr)
1076 struct mthca_mr *mmr = to_mmr(mr);
1078 mthca_free_mr(to_mdev(mr->device), mmr);
1080 ib_umem_release(mmr->umem);
1086 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1087 struct ib_fmr_attr *fmr_attr)
1089 struct mthca_fmr *fmr;
1092 fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1094 return ERR_PTR(-ENOMEM);
1096 memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1097 err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1098 convert_access(mr_access_flags), fmr);
1102 return ERR_PTR(err);
1108 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1110 struct mthca_fmr *mfmr = to_mfmr(fmr);
1113 err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1121 static int mthca_unmap_fmr(struct list_head *fmr_list)
1125 struct mthca_dev *mdev = NULL;
1127 list_for_each_entry(fmr, fmr_list, list) {
1128 if (mdev && to_mdev(fmr->device) != mdev)
1130 mdev = to_mdev(fmr->device);
1136 if (mthca_is_memfree(mdev)) {
1137 list_for_each_entry(fmr, fmr_list, list)
1138 mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1142 list_for_each_entry(fmr, fmr_list, list)
1143 mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1145 err = mthca_SYNC_TPT(mdev);
1149 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1152 struct mthca_dev *dev =
1153 container_of(device, struct mthca_dev, ib_dev.dev);
1154 return sprintf(buf, "%x\n", dev->rev_id);
1157 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
1160 struct mthca_dev *dev =
1161 container_of(device, struct mthca_dev, ib_dev.dev);
1162 return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1163 (int) (dev->fw_ver >> 16) & 0xffff,
1164 (int) dev->fw_ver & 0xffff);
1167 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1170 struct mthca_dev *dev =
1171 container_of(device, struct mthca_dev, ib_dev.dev);
1172 switch (dev->pdev->device) {
1173 case PCI_DEVICE_ID_MELLANOX_TAVOR:
1174 return sprintf(buf, "MT23108\n");
1175 case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1176 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1177 case PCI_DEVICE_ID_MELLANOX_ARBEL:
1178 return sprintf(buf, "MT25208\n");
1179 case PCI_DEVICE_ID_MELLANOX_SINAI:
1180 case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1181 return sprintf(buf, "MT25204\n");
1183 return sprintf(buf, "unknown\n");
1187 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1190 struct mthca_dev *dev =
1191 container_of(device, struct mthca_dev, ib_dev.dev);
1192 return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1195 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1196 static DEVICE_ATTR(fw_ver, S_IRUGO, show_fw_ver, NULL);
1197 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1198 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
1200 static struct device_attribute *mthca_dev_attributes[] = {
1207 static int mthca_init_node_data(struct mthca_dev *dev)
1209 struct ib_smp *in_mad = NULL;
1210 struct ib_smp *out_mad = NULL;
1213 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
1214 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1215 if (!in_mad || !out_mad)
1218 init_query_mad(in_mad);
1219 in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1221 err = mthca_MAD_IFC(dev, 1, 1,
1222 1, NULL, NULL, in_mad, out_mad);
1226 memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1228 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1230 err = mthca_MAD_IFC(dev, 1, 1,
1231 1, NULL, NULL, in_mad, out_mad);
1235 if (mthca_is_memfree(dev))
1236 dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1237 memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1245 int mthca_register_device(struct mthca_dev *dev)
1250 ret = mthca_init_node_data(dev);
1254 strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1255 dev->ib_dev.owner = THIS_MODULE;
1257 dev->ib_dev.uverbs_abi_ver = MTHCA_UVERBS_ABI_VERSION;
1258 dev->ib_dev.uverbs_cmd_mask =
1259 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
1260 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
1261 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
1262 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
1263 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
1264 (1ull << IB_USER_VERBS_CMD_REG_MR) |
1265 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
1266 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1267 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
1268 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
1269 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
1270 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
1271 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
1272 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
1273 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
1274 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
1275 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1276 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
1277 dev->ib_dev.phys_port_cnt = dev->limits.num_ports;
1278 dev->ib_dev.num_comp_vectors = 1;
1279 dev->ib_dev.dma_device = &dev->pdev->dev;
1280 dev->ib_dev.query_device = mthca_query_device;
1281 dev->ib_dev.query_port = mthca_query_port;
1282 dev->ib_dev.modify_device = mthca_modify_device;
1283 dev->ib_dev.modify_port = mthca_modify_port;
1284 dev->ib_dev.query_pkey = mthca_query_pkey;
1285 dev->ib_dev.query_gid = mthca_query_gid;
1286 dev->ib_dev.alloc_ucontext = mthca_alloc_ucontext;
1287 dev->ib_dev.dealloc_ucontext = mthca_dealloc_ucontext;
1288 dev->ib_dev.mmap = mthca_mmap_uar;
1289 dev->ib_dev.alloc_pd = mthca_alloc_pd;
1290 dev->ib_dev.dealloc_pd = mthca_dealloc_pd;
1291 dev->ib_dev.create_ah = mthca_ah_create;
1292 dev->ib_dev.query_ah = mthca_ah_query;
1293 dev->ib_dev.destroy_ah = mthca_ah_destroy;
1295 if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1296 dev->ib_dev.create_srq = mthca_create_srq;
1297 dev->ib_dev.modify_srq = mthca_modify_srq;
1298 dev->ib_dev.query_srq = mthca_query_srq;
1299 dev->ib_dev.destroy_srq = mthca_destroy_srq;
1300 dev->ib_dev.uverbs_cmd_mask |=
1301 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
1302 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
1303 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
1304 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1306 if (mthca_is_memfree(dev))
1307 dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1309 dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1312 dev->ib_dev.create_qp = mthca_create_qp;
1313 dev->ib_dev.modify_qp = mthca_modify_qp;
1314 dev->ib_dev.query_qp = mthca_query_qp;
1315 dev->ib_dev.destroy_qp = mthca_destroy_qp;
1316 dev->ib_dev.create_cq = mthca_create_cq;
1317 dev->ib_dev.resize_cq = mthca_resize_cq;
1318 dev->ib_dev.destroy_cq = mthca_destroy_cq;
1319 dev->ib_dev.poll_cq = mthca_poll_cq;
1320 dev->ib_dev.get_dma_mr = mthca_get_dma_mr;
1321 dev->ib_dev.reg_phys_mr = mthca_reg_phys_mr;
1322 dev->ib_dev.reg_user_mr = mthca_reg_user_mr;
1323 dev->ib_dev.dereg_mr = mthca_dereg_mr;
1325 if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1326 dev->ib_dev.alloc_fmr = mthca_alloc_fmr;
1327 dev->ib_dev.unmap_fmr = mthca_unmap_fmr;
1328 dev->ib_dev.dealloc_fmr = mthca_dealloc_fmr;
1329 if (mthca_is_memfree(dev))
1330 dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1332 dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1335 dev->ib_dev.attach_mcast = mthca_multicast_attach;
1336 dev->ib_dev.detach_mcast = mthca_multicast_detach;
1337 dev->ib_dev.process_mad = mthca_process_mad;
1339 if (mthca_is_memfree(dev)) {
1340 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1341 dev->ib_dev.post_send = mthca_arbel_post_send;
1342 dev->ib_dev.post_recv = mthca_arbel_post_receive;
1344 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1345 dev->ib_dev.post_send = mthca_tavor_post_send;
1346 dev->ib_dev.post_recv = mthca_tavor_post_receive;
1349 mutex_init(&dev->cap_mask_mutex);
1351 ret = ib_register_device(&dev->ib_dev, NULL);
1355 for (i = 0; i < ARRAY_SIZE(mthca_dev_attributes); ++i) {
1356 ret = device_create_file(&dev->ib_dev.dev,
1357 mthca_dev_attributes[i]);
1359 ib_unregister_device(&dev->ib_dev);
1364 mthca_start_catas_poll(dev);
1369 void mthca_unregister_device(struct mthca_dev *dev)
1371 mthca_stop_catas_poll(dev);
1372 ib_unregister_device(&dev->ib_dev);