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>
40 #include <rdma/uverbs_ioctl.h>
42 #include <linux/sched.h>
43 #include <linux/slab.h>
44 #include <linux/stat.h>
46 #include <linux/export.h>
48 #include "mthca_dev.h"
49 #include "mthca_cmd.h"
50 #include <rdma/mthca-abi.h>
51 #include "mthca_memfree.h"
53 static void init_query_mad(struct ib_smp *mad)
55 mad->base_version = 1;
56 mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
57 mad->class_version = 1;
58 mad->method = IB_MGMT_METHOD_GET;
61 static int mthca_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
64 struct ib_smp *in_mad = NULL;
65 struct ib_smp *out_mad = NULL;
67 struct mthca_dev *mdev = to_mdev(ibdev);
69 if (uhw->inlen || uhw->outlen)
72 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
73 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
74 if (!in_mad || !out_mad)
77 memset(props, 0, sizeof *props);
79 props->fw_ver = mdev->fw_ver;
81 init_query_mad(in_mad);
82 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
84 err = mthca_MAD_IFC(mdev, 1, 1,
85 1, NULL, NULL, in_mad, out_mad);
89 props->device_cap_flags = mdev->device_cap_flags;
90 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
92 props->vendor_part_id = be16_to_cpup((__be16 *) (out_mad->data + 30));
93 props->hw_ver = be32_to_cpup((__be32 *) (out_mad->data + 32));
94 memcpy(&props->sys_image_guid, out_mad->data + 4, 8);
96 props->max_mr_size = ~0ull;
97 props->page_size_cap = mdev->limits.page_size_cap;
98 props->max_qp = mdev->limits.num_qps - mdev->limits.reserved_qps;
99 props->max_qp_wr = mdev->limits.max_wqes;
100 props->max_send_sge = mdev->limits.max_sg;
101 props->max_recv_sge = mdev->limits.max_sg;
102 props->max_sge_rd = mdev->limits.max_sg;
103 props->max_cq = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
104 props->max_cqe = mdev->limits.max_cqes;
105 props->max_mr = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
106 props->max_pd = mdev->limits.num_pds - mdev->limits.reserved_pds;
107 props->max_qp_rd_atom = 1 << mdev->qp_table.rdb_shift;
108 props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
109 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
110 props->max_srq = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
111 props->max_srq_wr = mdev->limits.max_srq_wqes;
112 props->max_srq_sge = mdev->limits.max_srq_sge;
113 props->local_ca_ack_delay = mdev->limits.local_ca_ack_delay;
114 props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
115 IB_ATOMIC_HCA : IB_ATOMIC_NONE;
116 props->max_pkeys = mdev->limits.pkey_table_len;
117 props->max_mcast_grp = mdev->limits.num_mgms + mdev->limits.num_amgms;
118 props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
119 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
120 props->max_mcast_grp;
122 * If Sinai memory key optimization is being used, then only
123 * the 8-bit key portion will change. For other HCAs, the
124 * unused index bits will also be used for FMR remapping.
126 if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
127 props->max_map_per_fmr = 255;
129 props->max_map_per_fmr =
130 (1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
139 static int mthca_query_port(struct ib_device *ibdev,
140 u8 port, struct ib_port_attr *props)
142 struct ib_smp *in_mad = NULL;
143 struct ib_smp *out_mad = NULL;
146 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
147 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
148 if (!in_mad || !out_mad)
151 /* props being zeroed by the caller, avoid zeroing it here */
153 init_query_mad(in_mad);
154 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
155 in_mad->attr_mod = cpu_to_be32(port);
157 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
158 port, NULL, NULL, in_mad, out_mad);
162 props->lid = be16_to_cpup((__be16 *) (out_mad->data + 16));
163 props->lmc = out_mad->data[34] & 0x7;
164 props->sm_lid = be16_to_cpup((__be16 *) (out_mad->data + 18));
165 props->sm_sl = out_mad->data[36] & 0xf;
166 props->state = out_mad->data[32] & 0xf;
167 props->phys_state = out_mad->data[33] >> 4;
168 props->port_cap_flags = be32_to_cpup((__be32 *) (out_mad->data + 20));
169 props->gid_tbl_len = to_mdev(ibdev)->limits.gid_table_len;
170 props->max_msg_sz = 0x80000000;
171 props->pkey_tbl_len = to_mdev(ibdev)->limits.pkey_table_len;
172 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46));
173 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48));
174 props->active_width = out_mad->data[31] & 0xf;
175 props->active_speed = out_mad->data[35] >> 4;
176 props->max_mtu = out_mad->data[41] & 0xf;
177 props->active_mtu = out_mad->data[36] >> 4;
178 props->subnet_timeout = out_mad->data[51] & 0x1f;
179 props->max_vl_num = out_mad->data[37] >> 4;
180 props->init_type_reply = out_mad->data[41] >> 4;
188 static int mthca_modify_device(struct ib_device *ibdev,
190 struct ib_device_modify *props)
192 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
195 if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
196 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
198 memcpy(ibdev->node_desc, props->node_desc,
199 IB_DEVICE_NODE_DESC_MAX);
200 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
206 static int mthca_modify_port(struct ib_device *ibdev,
207 u8 port, int port_modify_mask,
208 struct ib_port_modify *props)
210 struct mthca_set_ib_param set_ib;
211 struct ib_port_attr attr;
214 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
217 err = ib_query_port(ibdev, port, &attr);
221 set_ib.set_si_guid = 0;
222 set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
224 set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
225 ~props->clr_port_cap_mask;
227 err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
231 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
235 static int mthca_query_pkey(struct ib_device *ibdev,
236 u8 port, u16 index, u16 *pkey)
238 struct ib_smp *in_mad = NULL;
239 struct ib_smp *out_mad = NULL;
242 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
243 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
244 if (!in_mad || !out_mad)
247 init_query_mad(in_mad);
248 in_mad->attr_id = IB_SMP_ATTR_PKEY_TABLE;
249 in_mad->attr_mod = cpu_to_be32(index / 32);
251 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
252 port, NULL, NULL, in_mad, out_mad);
256 *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
264 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
265 int index, union ib_gid *gid)
267 struct ib_smp *in_mad = NULL;
268 struct ib_smp *out_mad = NULL;
271 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
272 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
273 if (!in_mad || !out_mad)
276 init_query_mad(in_mad);
277 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
278 in_mad->attr_mod = cpu_to_be32(port);
280 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
281 port, NULL, NULL, in_mad, out_mad);
285 memcpy(gid->raw, out_mad->data + 8, 8);
287 init_query_mad(in_mad);
288 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
289 in_mad->attr_mod = cpu_to_be32(index / 8);
291 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
292 port, NULL, NULL, in_mad, out_mad);
296 memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
304 static int mthca_alloc_ucontext(struct ib_ucontext *uctx,
305 struct ib_udata *udata)
307 struct ib_device *ibdev = uctx->device;
308 struct mthca_alloc_ucontext_resp uresp = {};
309 struct mthca_ucontext *context = to_mucontext(uctx);
312 if (!(to_mdev(ibdev)->active))
315 uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
316 if (mthca_is_memfree(to_mdev(ibdev)))
317 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
321 err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
325 context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
326 if (IS_ERR(context->db_tab)) {
327 err = PTR_ERR(context->db_tab);
328 mthca_uar_free(to_mdev(ibdev), &context->uar);
332 if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) {
333 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
334 mthca_uar_free(to_mdev(ibdev), &context->uar);
338 context->reg_mr_warned = 0;
343 static void 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);
350 static int mthca_mmap_uar(struct ib_ucontext *context,
351 struct vm_area_struct *vma)
353 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
356 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
358 if (io_remap_pfn_range(vma, vma->vm_start,
359 to_mucontext(context)->uar.pfn,
360 PAGE_SIZE, vma->vm_page_prot))
366 static int mthca_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
368 struct ib_device *ibdev = ibpd->device;
369 struct mthca_pd *pd = to_mpd(ibpd);
372 err = mthca_pd_alloc(to_mdev(ibdev), !udata, pd);
377 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
378 mthca_pd_free(to_mdev(ibdev), pd);
386 static void mthca_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
388 mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
391 static int mthca_ah_create(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr,
392 u32 flags, struct ib_udata *udata)
395 struct mthca_ah *ah = to_mah(ibah);
397 return mthca_create_ah(to_mdev(ibah->device), to_mpd(ibah->pd), ah_attr,
401 static void mthca_ah_destroy(struct ib_ah *ah, u32 flags)
403 mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
406 static int mthca_create_srq(struct ib_srq *ibsrq,
407 struct ib_srq_init_attr *init_attr,
408 struct ib_udata *udata)
410 struct mthca_create_srq ucmd;
411 struct mthca_ucontext *context = rdma_udata_to_drv_context(
412 udata, struct mthca_ucontext, ibucontext);
413 struct mthca_srq *srq = to_msrq(ibsrq);
416 if (init_attr->srq_type != IB_SRQT_BASIC)
420 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
423 err = mthca_map_user_db(to_mdev(ibsrq->device), &context->uar,
424 context->db_tab, ucmd.db_index,
430 srq->mr.ibmr.lkey = ucmd.lkey;
431 srq->db_index = ucmd.db_index;
434 err = mthca_alloc_srq(to_mdev(ibsrq->device), to_mpd(ibsrq->pd),
435 &init_attr->attr, srq, udata);
438 mthca_unmap_user_db(to_mdev(ibsrq->device), &context->uar,
439 context->db_tab, ucmd.db_index);
444 if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof(__u32))) {
445 mthca_free_srq(to_mdev(ibsrq->device), srq);
452 static void mthca_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
455 struct mthca_ucontext *context =
456 rdma_udata_to_drv_context(
458 struct mthca_ucontext,
461 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
462 context->db_tab, to_msrq(srq)->db_index);
465 mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
468 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
469 struct ib_qp_init_attr *init_attr,
470 struct ib_udata *udata)
472 struct mthca_ucontext *context = rdma_udata_to_drv_context(
473 udata, struct mthca_ucontext, ibucontext);
474 struct mthca_create_qp ucmd;
478 if (init_attr->create_flags)
479 return ERR_PTR(-EINVAL);
481 switch (init_attr->qp_type) {
486 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
488 return ERR_PTR(-ENOMEM);
491 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
493 return ERR_PTR(-EFAULT);
496 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
498 ucmd.sq_db_index, ucmd.sq_db_page);
504 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
506 ucmd.rq_db_index, ucmd.rq_db_page);
508 mthca_unmap_user_db(to_mdev(pd->device),
516 qp->mr.ibmr.lkey = ucmd.lkey;
517 qp->sq.db_index = ucmd.sq_db_index;
518 qp->rq.db_index = ucmd.rq_db_index;
521 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
522 to_mcq(init_attr->send_cq),
523 to_mcq(init_attr->recv_cq),
524 init_attr->qp_type, init_attr->sq_sig_type,
525 &init_attr->cap, qp, udata);
528 mthca_unmap_user_db(to_mdev(pd->device),
532 mthca_unmap_user_db(to_mdev(pd->device),
538 qp->ibqp.qp_num = qp->qpn;
544 /* Don't allow userspace to create special QPs */
546 return ERR_PTR(-EINVAL);
548 qp = kzalloc(sizeof(struct mthca_sqp), GFP_KERNEL);
550 return ERR_PTR(-ENOMEM);
552 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
554 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
555 to_mcq(init_attr->send_cq),
556 to_mcq(init_attr->recv_cq),
557 init_attr->sq_sig_type, &init_attr->cap,
558 qp->ibqp.qp_num, init_attr->port_num,
563 /* Don't support raw QPs */
564 return ERR_PTR(-ENOSYS);
572 init_attr->cap.max_send_wr = qp->sq.max;
573 init_attr->cap.max_recv_wr = qp->rq.max;
574 init_attr->cap.max_send_sge = qp->sq.max_gs;
575 init_attr->cap.max_recv_sge = qp->rq.max_gs;
576 init_attr->cap.max_inline_data = qp->max_inline_data;
581 static int mthca_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
584 struct mthca_ucontext *context =
585 rdma_udata_to_drv_context(
587 struct mthca_ucontext,
590 mthca_unmap_user_db(to_mdev(qp->device),
593 to_mqp(qp)->sq.db_index);
594 mthca_unmap_user_db(to_mdev(qp->device),
597 to_mqp(qp)->rq.db_index);
599 mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
604 static int mthca_create_cq(struct ib_cq *ibcq,
605 const struct ib_cq_init_attr *attr,
606 struct ib_udata *udata)
608 struct ib_device *ibdev = ibcq->device;
609 int entries = attr->cqe;
610 struct mthca_create_cq ucmd;
614 struct mthca_ucontext *context = rdma_udata_to_drv_context(
615 udata, struct mthca_ucontext, ibucontext);
620 if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
624 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
627 err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
628 context->db_tab, ucmd.set_db_index,
633 err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
634 context->db_tab, ucmd.arm_db_index,
643 cq->buf.mr.ibmr.lkey = ucmd.lkey;
644 cq->set_ci_db_index = ucmd.set_db_index;
645 cq->arm_db_index = ucmd.arm_db_index;
648 for (nent = 1; nent <= entries; nent <<= 1)
651 err = mthca_init_cq(to_mdev(ibdev), nent, context,
652 udata ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
657 if (udata && ib_copy_to_udata(udata, &cq->cqn, sizeof(__u32))) {
658 mthca_free_cq(to_mdev(ibdev), cq);
663 cq->resize_buf = NULL;
669 mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
670 context->db_tab, ucmd.arm_db_index);
674 mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
675 context->db_tab, ucmd.set_db_index);
680 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
685 spin_lock_irq(&cq->lock);
686 if (cq->resize_buf) {
691 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
692 if (!cq->resize_buf) {
697 cq->resize_buf->state = CQ_RESIZE_ALLOC;
702 spin_unlock_irq(&cq->lock);
707 ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
709 spin_lock_irq(&cq->lock);
710 kfree(cq->resize_buf);
711 cq->resize_buf = NULL;
712 spin_unlock_irq(&cq->lock);
716 cq->resize_buf->cqe = entries - 1;
718 spin_lock_irq(&cq->lock);
719 cq->resize_buf->state = CQ_RESIZE_READY;
720 spin_unlock_irq(&cq->lock);
725 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
727 struct mthca_dev *dev = to_mdev(ibcq->device);
728 struct mthca_cq *cq = to_mcq(ibcq);
729 struct mthca_resize_cq ucmd;
733 if (entries < 1 || entries > dev->limits.max_cqes)
736 mutex_lock(&cq->mutex);
738 entries = roundup_pow_of_two(entries + 1);
739 if (entries == ibcq->cqe + 1) {
745 ret = mthca_alloc_resize_buf(dev, cq, entries);
748 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
750 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
757 ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
760 if (cq->resize_buf) {
761 mthca_free_cq_buf(dev, &cq->resize_buf->buf,
762 cq->resize_buf->cqe);
763 kfree(cq->resize_buf);
764 spin_lock_irq(&cq->lock);
765 cq->resize_buf = NULL;
766 spin_unlock_irq(&cq->lock);
772 struct mthca_cq_buf tbuf;
775 spin_lock_irq(&cq->lock);
776 if (cq->resize_buf->state == CQ_RESIZE_READY) {
777 mthca_cq_resize_copy_cqes(cq);
780 cq->buf = cq->resize_buf->buf;
781 cq->ibcq.cqe = cq->resize_buf->cqe;
783 tbuf = cq->resize_buf->buf;
784 tcqe = cq->resize_buf->cqe;
787 kfree(cq->resize_buf);
788 cq->resize_buf = NULL;
789 spin_unlock_irq(&cq->lock);
791 mthca_free_cq_buf(dev, &tbuf, tcqe);
793 ibcq->cqe = entries - 1;
796 mutex_unlock(&cq->mutex);
801 static void mthca_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
804 struct mthca_ucontext *context =
805 rdma_udata_to_drv_context(
807 struct mthca_ucontext,
810 mthca_unmap_user_db(to_mdev(cq->device),
813 to_mcq(cq)->arm_db_index);
814 mthca_unmap_user_db(to_mdev(cq->device),
817 to_mcq(cq)->set_ci_db_index);
819 mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
822 static inline u32 convert_access(int acc)
824 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC : 0) |
825 (acc & IB_ACCESS_REMOTE_WRITE ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
826 (acc & IB_ACCESS_REMOTE_READ ? MTHCA_MPT_FLAG_REMOTE_READ : 0) |
827 (acc & IB_ACCESS_LOCAL_WRITE ? MTHCA_MPT_FLAG_LOCAL_WRITE : 0) |
828 MTHCA_MPT_FLAG_LOCAL_READ;
831 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
836 mr = kmalloc(sizeof *mr, GFP_KERNEL);
838 return ERR_PTR(-ENOMEM);
840 err = mthca_mr_alloc_notrans(to_mdev(pd->device),
842 convert_access(acc), mr);
854 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
855 u64 virt, int acc, struct ib_udata *udata)
857 struct mthca_dev *dev = to_mdev(pd->device);
858 struct sg_dma_page_iter sg_iter;
859 struct mthca_ucontext *context = rdma_udata_to_drv_context(
860 udata, struct mthca_ucontext, ibucontext);
862 struct mthca_reg_mr ucmd;
868 if (udata->inlen < sizeof ucmd) {
869 if (!context->reg_mr_warned) {
870 mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
872 mthca_warn(dev, " Update libmthca to fix this.\n");
874 ++context->reg_mr_warned;
876 } else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
877 return ERR_PTR(-EFAULT);
879 mr = kmalloc(sizeof *mr, GFP_KERNEL);
881 return ERR_PTR(-ENOMEM);
883 mr->umem = ib_umem_get(udata, start, length, acc);
884 if (IS_ERR(mr->umem)) {
885 err = PTR_ERR(mr->umem);
889 n = ib_umem_num_pages(mr->umem);
891 mr->mtt = mthca_alloc_mtt(dev, n);
892 if (IS_ERR(mr->mtt)) {
893 err = PTR_ERR(mr->mtt);
897 pages = (u64 *) __get_free_page(GFP_KERNEL);
905 write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
907 for_each_sg_dma_page(mr->umem->sg_head.sgl, &sg_iter, mr->umem->nmap, 0) {
908 pages[i++] = sg_page_iter_dma_address(&sg_iter);
911 * Be friendly to write_mtt and pass it chunks
912 * of appropriate size.
914 if (i == write_mtt_size) {
915 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
924 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
926 free_page((unsigned long) pages);
930 err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, PAGE_SHIFT, virt, length,
931 convert_access(acc), mr);
939 mthca_free_mtt(dev, mr->mtt);
942 ib_umem_release(mr->umem);
949 static int mthca_dereg_mr(struct ib_mr *mr, struct ib_udata *udata)
951 struct mthca_mr *mmr = to_mmr(mr);
953 mthca_free_mr(to_mdev(mr->device), mmr);
954 ib_umem_release(mmr->umem);
960 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
961 struct ib_fmr_attr *fmr_attr)
963 struct mthca_fmr *fmr;
966 fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
968 return ERR_PTR(-ENOMEM);
970 memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
971 err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
972 convert_access(mr_access_flags), fmr);
982 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
984 struct mthca_fmr *mfmr = to_mfmr(fmr);
987 err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
995 static int mthca_unmap_fmr(struct list_head *fmr_list)
999 struct mthca_dev *mdev = NULL;
1001 list_for_each_entry(fmr, fmr_list, list) {
1002 if (mdev && to_mdev(fmr->device) != mdev)
1004 mdev = to_mdev(fmr->device);
1010 if (mthca_is_memfree(mdev)) {
1011 list_for_each_entry(fmr, fmr_list, list)
1012 mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1016 list_for_each_entry(fmr, fmr_list, list)
1017 mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1019 err = mthca_SYNC_TPT(mdev);
1023 static ssize_t hw_rev_show(struct device *device,
1024 struct device_attribute *attr, char *buf)
1026 struct mthca_dev *dev =
1027 rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
1029 return sprintf(buf, "%x\n", dev->rev_id);
1031 static DEVICE_ATTR_RO(hw_rev);
1033 static ssize_t hca_type_show(struct device *device,
1034 struct device_attribute *attr, char *buf)
1036 struct mthca_dev *dev =
1037 rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
1039 switch (dev->pdev->device) {
1040 case PCI_DEVICE_ID_MELLANOX_TAVOR:
1041 return sprintf(buf, "MT23108\n");
1042 case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1043 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1044 case PCI_DEVICE_ID_MELLANOX_ARBEL:
1045 return sprintf(buf, "MT25208\n");
1046 case PCI_DEVICE_ID_MELLANOX_SINAI:
1047 case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1048 return sprintf(buf, "MT25204\n");
1050 return sprintf(buf, "unknown\n");
1053 static DEVICE_ATTR_RO(hca_type);
1055 static ssize_t board_id_show(struct device *device,
1056 struct device_attribute *attr, char *buf)
1058 struct mthca_dev *dev =
1059 rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
1061 return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1063 static DEVICE_ATTR_RO(board_id);
1065 static struct attribute *mthca_dev_attributes[] = {
1066 &dev_attr_hw_rev.attr,
1067 &dev_attr_hca_type.attr,
1068 &dev_attr_board_id.attr,
1072 static const struct attribute_group mthca_attr_group = {
1073 .attrs = mthca_dev_attributes,
1076 static int mthca_init_node_data(struct mthca_dev *dev)
1078 struct ib_smp *in_mad = NULL;
1079 struct ib_smp *out_mad = NULL;
1082 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
1083 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1084 if (!in_mad || !out_mad)
1087 init_query_mad(in_mad);
1088 in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1090 err = mthca_MAD_IFC(dev, 1, 1,
1091 1, NULL, NULL, in_mad, out_mad);
1095 memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX);
1097 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1099 err = mthca_MAD_IFC(dev, 1, 1,
1100 1, NULL, NULL, in_mad, out_mad);
1104 if (mthca_is_memfree(dev))
1105 dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1106 memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1114 static int mthca_port_immutable(struct ib_device *ibdev, u8 port_num,
1115 struct ib_port_immutable *immutable)
1117 struct ib_port_attr attr;
1120 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1122 err = ib_query_port(ibdev, port_num, &attr);
1126 immutable->pkey_tbl_len = attr.pkey_tbl_len;
1127 immutable->gid_tbl_len = attr.gid_tbl_len;
1128 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
1133 static void get_dev_fw_str(struct ib_device *device, char *str)
1135 struct mthca_dev *dev =
1136 container_of(device, struct mthca_dev, ib_dev);
1137 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d",
1138 (int) (dev->fw_ver >> 32),
1139 (int) (dev->fw_ver >> 16) & 0xffff,
1140 (int) dev->fw_ver & 0xffff);
1143 static const struct ib_device_ops mthca_dev_ops = {
1144 .owner = THIS_MODULE,
1145 .driver_id = RDMA_DRIVER_MTHCA,
1146 .uverbs_abi_ver = MTHCA_UVERBS_ABI_VERSION,
1147 .uverbs_no_driver_id_binding = 1,
1149 .alloc_pd = mthca_alloc_pd,
1150 .alloc_ucontext = mthca_alloc_ucontext,
1151 .attach_mcast = mthca_multicast_attach,
1152 .create_ah = mthca_ah_create,
1153 .create_cq = mthca_create_cq,
1154 .create_qp = mthca_create_qp,
1155 .dealloc_pd = mthca_dealloc_pd,
1156 .dealloc_ucontext = mthca_dealloc_ucontext,
1157 .dereg_mr = mthca_dereg_mr,
1158 .destroy_ah = mthca_ah_destroy,
1159 .destroy_cq = mthca_destroy_cq,
1160 .destroy_qp = mthca_destroy_qp,
1161 .detach_mcast = mthca_multicast_detach,
1162 .get_dev_fw_str = get_dev_fw_str,
1163 .get_dma_mr = mthca_get_dma_mr,
1164 .get_port_immutable = mthca_port_immutable,
1165 .mmap = mthca_mmap_uar,
1166 .modify_device = mthca_modify_device,
1167 .modify_port = mthca_modify_port,
1168 .modify_qp = mthca_modify_qp,
1169 .poll_cq = mthca_poll_cq,
1170 .process_mad = mthca_process_mad,
1171 .query_ah = mthca_ah_query,
1172 .query_device = mthca_query_device,
1173 .query_gid = mthca_query_gid,
1174 .query_pkey = mthca_query_pkey,
1175 .query_port = mthca_query_port,
1176 .query_qp = mthca_query_qp,
1177 .reg_user_mr = mthca_reg_user_mr,
1178 .resize_cq = mthca_resize_cq,
1180 INIT_RDMA_OBJ_SIZE(ib_ah, mthca_ah, ibah),
1181 INIT_RDMA_OBJ_SIZE(ib_cq, mthca_cq, ibcq),
1182 INIT_RDMA_OBJ_SIZE(ib_pd, mthca_pd, ibpd),
1183 INIT_RDMA_OBJ_SIZE(ib_ucontext, mthca_ucontext, ibucontext),
1186 static const struct ib_device_ops mthca_dev_arbel_srq_ops = {
1187 .create_srq = mthca_create_srq,
1188 .destroy_srq = mthca_destroy_srq,
1189 .modify_srq = mthca_modify_srq,
1190 .post_srq_recv = mthca_arbel_post_srq_recv,
1191 .query_srq = mthca_query_srq,
1193 INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq),
1196 static const struct ib_device_ops mthca_dev_tavor_srq_ops = {
1197 .create_srq = mthca_create_srq,
1198 .destroy_srq = mthca_destroy_srq,
1199 .modify_srq = mthca_modify_srq,
1200 .post_srq_recv = mthca_tavor_post_srq_recv,
1201 .query_srq = mthca_query_srq,
1203 INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq),
1206 static const struct ib_device_ops mthca_dev_arbel_fmr_ops = {
1207 .alloc_fmr = mthca_alloc_fmr,
1208 .dealloc_fmr = mthca_dealloc_fmr,
1209 .map_phys_fmr = mthca_arbel_map_phys_fmr,
1210 .unmap_fmr = mthca_unmap_fmr,
1213 static const struct ib_device_ops mthca_dev_tavor_fmr_ops = {
1214 .alloc_fmr = mthca_alloc_fmr,
1215 .dealloc_fmr = mthca_dealloc_fmr,
1216 .map_phys_fmr = mthca_tavor_map_phys_fmr,
1217 .unmap_fmr = mthca_unmap_fmr,
1220 static const struct ib_device_ops mthca_dev_arbel_ops = {
1221 .post_recv = mthca_arbel_post_receive,
1222 .post_send = mthca_arbel_post_send,
1223 .req_notify_cq = mthca_arbel_arm_cq,
1226 static const struct ib_device_ops mthca_dev_tavor_ops = {
1227 .post_recv = mthca_tavor_post_receive,
1228 .post_send = mthca_tavor_post_send,
1229 .req_notify_cq = mthca_tavor_arm_cq,
1232 int mthca_register_device(struct mthca_dev *dev)
1236 ret = mthca_init_node_data(dev);
1240 dev->ib_dev.uverbs_cmd_mask =
1241 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
1242 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
1243 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
1244 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
1245 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
1246 (1ull << IB_USER_VERBS_CMD_REG_MR) |
1247 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
1248 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1249 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
1250 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
1251 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
1252 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
1253 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
1254 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
1255 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
1256 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
1257 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1258 dev->ib_dev.node_type = RDMA_NODE_IB_CA;
1259 dev->ib_dev.phys_port_cnt = dev->limits.num_ports;
1260 dev->ib_dev.num_comp_vectors = 1;
1261 dev->ib_dev.dev.parent = &dev->pdev->dev;
1263 if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1264 dev->ib_dev.uverbs_cmd_mask |=
1265 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
1266 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
1267 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
1268 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1270 if (mthca_is_memfree(dev))
1271 ib_set_device_ops(&dev->ib_dev,
1272 &mthca_dev_arbel_srq_ops);
1274 ib_set_device_ops(&dev->ib_dev,
1275 &mthca_dev_tavor_srq_ops);
1278 if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1279 if (mthca_is_memfree(dev))
1280 ib_set_device_ops(&dev->ib_dev,
1281 &mthca_dev_arbel_fmr_ops);
1283 ib_set_device_ops(&dev->ib_dev,
1284 &mthca_dev_tavor_fmr_ops);
1287 ib_set_device_ops(&dev->ib_dev, &mthca_dev_ops);
1289 if (mthca_is_memfree(dev))
1290 ib_set_device_ops(&dev->ib_dev, &mthca_dev_arbel_ops);
1292 ib_set_device_ops(&dev->ib_dev, &mthca_dev_tavor_ops);
1294 mutex_init(&dev->cap_mask_mutex);
1296 rdma_set_device_sysfs_group(&dev->ib_dev, &mthca_attr_group);
1297 ret = ib_register_device(&dev->ib_dev, "mthca%d");
1301 mthca_start_catas_poll(dev);
1306 void mthca_unregister_device(struct mthca_dev *dev)
1308 mthca_stop_catas_poll(dev);
1309 ib_unregister_device(&dev->ib_dev);