2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <rdma/ib_mad.h>
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_sa.h>
36 #include <rdma/ib_cache.h>
38 #include <linux/random.h>
39 #include <linux/mlx4/cmd.h>
40 #include <linux/gfp.h>
41 #include <rdma/ib_pma.h>
46 MLX4_IB_VENDOR_CLASS1 = 0x9,
47 MLX4_IB_VENDOR_CLASS2 = 0xa
50 #define MLX4_TUN_SEND_WRID_SHIFT 34
51 #define MLX4_TUN_QPN_SHIFT 32
52 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
53 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
55 #define MLX4_TUN_IS_RECV(a) (((a) >> MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
56 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
58 /* Port mgmt change event handling */
60 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
61 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
62 #define NUM_IDX_IN_PKEY_TBL_BLK 32
63 #define GUID_TBL_ENTRY_SIZE 8 /* size in bytes */
64 #define GUID_TBL_BLK_NUM_ENTRIES 8
65 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
67 struct mlx4_mad_rcv_buf {
72 struct mlx4_mad_snd_buf {
76 struct mlx4_tunnel_mad {
78 struct mlx4_ib_tunnel_header hdr;
82 struct mlx4_rcv_tunnel_mad {
83 struct mlx4_rcv_tunnel_hdr hdr;
88 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
89 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
90 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
91 int block, u32 change_bitmap);
93 __be64 mlx4_ib_gen_node_guid(void)
95 #define NODE_GUID_HI ((u64) (((u64)IB_OPENIB_OUI) << 40))
96 return cpu_to_be64(NODE_GUID_HI | prandom_u32());
99 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
101 return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
102 cpu_to_be64(0xff00000000000000LL);
105 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
106 int port, const struct ib_wc *in_wc,
107 const struct ib_grh *in_grh,
108 const void *in_mad, void *response_mad)
110 struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
113 u32 in_modifier = port;
116 inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
117 if (IS_ERR(inmailbox))
118 return PTR_ERR(inmailbox);
119 inbox = inmailbox->buf;
121 outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
122 if (IS_ERR(outmailbox)) {
123 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
124 return PTR_ERR(outmailbox);
127 memcpy(inbox, in_mad, 256);
130 * Key check traps can't be generated unless we have in_wc to
131 * tell us where to send the trap.
133 if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
135 if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
137 if (mlx4_is_mfunc(dev->dev) &&
138 (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
154 memset(inbox + 256, 0, 256);
155 ext_info = inbox + 256;
157 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
158 ext_info->rqpn = cpu_to_be32(in_wc->src_qp);
159 ext_info->sl = in_wc->sl << 4;
160 ext_info->g_path = in_wc->dlid_path_bits |
161 (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
162 ext_info->pkey = cpu_to_be16(in_wc->pkey_index);
165 memcpy(ext_info->grh, in_grh, 40);
169 in_modifier |= in_wc->slid << 16;
172 err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
173 mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
174 MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
175 (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
178 memcpy(response_mad, outmailbox->buf, 256);
180 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
181 mlx4_free_cmd_mailbox(dev->dev, outmailbox);
186 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
188 struct ib_ah *new_ah;
189 struct ib_ah_attr ah_attr;
192 if (!dev->send_agent[port_num - 1][0])
195 memset(&ah_attr, 0, sizeof ah_attr);
198 ah_attr.port_num = port_num;
200 new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
205 spin_lock_irqsave(&dev->sm_lock, flags);
206 if (dev->sm_ah[port_num - 1])
207 ib_destroy_ah(dev->sm_ah[port_num - 1]);
208 dev->sm_ah[port_num - 1] = new_ah;
209 spin_unlock_irqrestore(&dev->sm_lock, flags);
213 * Snoop SM MADs for port info, GUID info, and P_Key table sets, so we can
214 * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
216 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
219 struct ib_port_info *pinfo;
222 u32 bn, pkey_change_bitmap;
226 struct mlx4_ib_dev *dev = to_mdev(ibdev);
227 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
228 mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
229 mad->mad_hdr.method == IB_MGMT_METHOD_SET)
230 switch (mad->mad_hdr.attr_id) {
231 case IB_SMP_ATTR_PORT_INFO:
232 pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
233 lid = be16_to_cpu(pinfo->lid);
235 update_sm_ah(dev, port_num,
236 be16_to_cpu(pinfo->sm_lid),
237 pinfo->neighbormtu_mastersmsl & 0xf);
239 if (pinfo->clientrereg_resv_subnetto & 0x80)
240 handle_client_rereg_event(dev, port_num);
243 handle_lid_change_event(dev, port_num);
246 case IB_SMP_ATTR_PKEY_TABLE:
247 if (!mlx4_is_mfunc(dev->dev)) {
248 mlx4_ib_dispatch_event(dev, port_num,
249 IB_EVENT_PKEY_CHANGE);
253 /* at this point, we are running in the master.
254 * Slaves do not receive SMPs.
256 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
257 base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
258 pkey_change_bitmap = 0;
259 for (i = 0; i < 32; i++) {
260 pr_debug("PKEY[%d] = x%x\n",
261 i + bn*32, be16_to_cpu(base[i]));
262 if (be16_to_cpu(base[i]) !=
263 dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
264 pkey_change_bitmap |= (1 << i);
265 dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
266 be16_to_cpu(base[i]);
269 pr_debug("PKEY Change event: port=%d, "
270 "block=0x%x, change_bitmap=0x%x\n",
271 port_num, bn, pkey_change_bitmap);
273 if (pkey_change_bitmap) {
274 mlx4_ib_dispatch_event(dev, port_num,
275 IB_EVENT_PKEY_CHANGE);
276 if (!dev->sriov.is_going_down)
277 __propagate_pkey_ev(dev, port_num, bn,
282 case IB_SMP_ATTR_GUID_INFO:
283 /* paravirtualized master's guid is guid 0 -- does not change */
284 if (!mlx4_is_master(dev->dev))
285 mlx4_ib_dispatch_event(dev, port_num,
286 IB_EVENT_GID_CHANGE);
287 /*if master, notify relevant slaves*/
288 if (mlx4_is_master(dev->dev) &&
289 !dev->sriov.is_going_down) {
290 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
291 mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
292 (u8 *)(&((struct ib_smp *)mad)->data));
293 mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
294 (u8 *)(&((struct ib_smp *)mad)->data));
303 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
304 int block, u32 change_bitmap)
306 int i, ix, slave, err;
309 for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
310 if (slave == mlx4_master_func_num(dev->dev))
312 if (!mlx4_is_slave_active(dev->dev, slave))
316 for (i = 0; i < 32; i++) {
317 if (!(change_bitmap & (1 << i)))
320 ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
321 if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
322 [ix] == i + 32 * block) {
323 err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
324 pr_debug("propagate_pkey_ev: slave %d,"
325 " port %d, ix %d (%d)\n",
326 slave, port_num, ix, err);
337 static void node_desc_override(struct ib_device *dev,
342 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
343 mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
344 mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
345 mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
346 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
347 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
348 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
352 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
354 int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
355 struct ib_mad_send_buf *send_buf;
356 struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
361 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
362 IB_MGMT_MAD_DATA, GFP_ATOMIC,
363 IB_MGMT_BASE_VERSION);
364 if (IS_ERR(send_buf))
367 * We rely here on the fact that MLX QPs don't use the
368 * address handle after the send is posted (this is
369 * wrong following the IB spec strictly, but we know
370 * it's OK for our devices).
372 spin_lock_irqsave(&dev->sm_lock, flags);
373 memcpy(send_buf->mad, mad, sizeof *mad);
374 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
375 ret = ib_post_send_mad(send_buf, NULL);
378 spin_unlock_irqrestore(&dev->sm_lock, flags);
381 ib_free_send_mad(send_buf);
385 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
386 struct ib_sa_mad *sa_mad)
390 /* dispatch to different sa handlers */
391 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
392 case IB_SA_ATTR_MC_MEMBER_REC:
393 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
401 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
403 struct mlx4_ib_dev *dev = to_mdev(ibdev);
406 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
407 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
414 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
415 u8 port, u16 pkey, u16 *ix)
418 u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
421 if (slave == mlx4_master_func_num(dev->dev))
422 return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
424 unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
426 for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
427 if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
430 pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
432 ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
435 if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
436 if (slot_pkey & 0x8000) {
440 /* take first partial pkey index found */
441 if (partial_ix == 0xFF)
442 partial_ix = pkey_ix;
447 if (partial_ix < 0xFF) {
448 *ix = (u16) partial_ix;
455 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
456 enum ib_qp_type dest_qpt, struct ib_wc *wc,
457 struct ib_grh *grh, struct ib_mad *mad)
460 struct ib_send_wr wr, *bad_wr;
461 struct mlx4_ib_demux_pv_ctx *tun_ctx;
462 struct mlx4_ib_demux_pv_qp *tun_qp;
463 struct mlx4_rcv_tunnel_mad *tun_mad;
464 struct ib_ah_attr attr;
466 struct ib_qp *src_qp = NULL;
467 unsigned tun_tx_ix = 0;
472 u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
474 if (dest_qpt > IB_QPT_GSI)
477 tun_ctx = dev->sriov.demux[port-1].tun[slave];
479 /* check if proxy qp created */
480 if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
484 tun_qp = &tun_ctx->qp[0];
486 tun_qp = &tun_ctx->qp[1];
488 /* compute P_Key index to put in tunnel header for slave */
491 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
495 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
498 tun_pkey_ix = pkey_ix;
500 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
502 dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
504 /* get tunnel tx data buf for slave */
507 /* create ah. Just need an empty one with the port num for the post send.
508 * The driver will set the force loopback bit in post_send */
509 memset(&attr, 0, sizeof attr);
510 attr.port_num = port;
512 memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
513 attr.ah_flags = IB_AH_GRH;
515 ah = ib_create_ah(tun_ctx->pd, &attr);
519 /* allocate tunnel tx buf after pass failure returns */
520 spin_lock(&tun_qp->tx_lock);
521 if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
522 (MLX4_NUM_TUNNEL_BUFS - 1))
525 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
526 spin_unlock(&tun_qp->tx_lock);
530 tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
531 if (tun_qp->tx_ring[tun_tx_ix].ah)
532 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
533 tun_qp->tx_ring[tun_tx_ix].ah = ah;
534 ib_dma_sync_single_for_cpu(&dev->ib_dev,
535 tun_qp->tx_ring[tun_tx_ix].buf.map,
536 sizeof (struct mlx4_rcv_tunnel_mad),
539 /* copy over to tunnel buffer */
541 memcpy(&tun_mad->grh, grh, sizeof *grh);
542 memcpy(&tun_mad->mad, mad, sizeof *mad);
544 /* adjust tunnel data */
545 tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
546 tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
547 tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
551 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
554 if (vlan != wc->vlan_id)
555 /* Packet vlan is not the VST-assigned vlan.
560 /* Remove the vlan tag before forwarding
561 * the packet to the VF.
568 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
569 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
570 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
572 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
573 tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
576 ib_dma_sync_single_for_device(&dev->ib_dev,
577 tun_qp->tx_ring[tun_tx_ix].buf.map,
578 sizeof (struct mlx4_rcv_tunnel_mad),
581 list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
582 list.length = sizeof (struct mlx4_rcv_tunnel_mad);
583 list.lkey = tun_ctx->mr->lkey;
586 wr.wr.ud.port_num = port;
587 wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
588 wr.wr.ud.remote_qpn = dqpn;
590 wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
593 wr.opcode = IB_WR_SEND;
594 wr.send_flags = IB_SEND_SIGNALED;
596 ret = ib_post_send(src_qp, &wr, &bad_wr);
603 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
604 struct ib_wc *wc, struct ib_grh *grh,
607 struct mlx4_ib_dev *dev = to_mdev(ibdev);
613 if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
619 if (!(wc->wc_flags & IB_WC_GRH)) {
620 mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
623 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
624 mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
627 if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
628 mlx4_ib_warn(ibdev, "failed matching grh\n");
631 if (slave >= dev->dev->caps.sqp_demux) {
632 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
633 slave, dev->dev->caps.sqp_demux);
637 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
640 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
642 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
647 /* Initially assume that this mad is for us */
648 slave = mlx4_master_func_num(dev->dev);
650 /* See if the slave id is encoded in a response mad */
651 if (mad->mad_hdr.method & 0x80) {
652 slave_id = (u8 *) &mad->mad_hdr.tid;
654 if (slave != 255) /*255 indicates the dom0*/
655 *slave_id = 0; /* remap tid */
658 /* If a grh is present, we demux according to it */
659 if (wc->wc_flags & IB_WC_GRH) {
660 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
662 mlx4_ib_warn(ibdev, "failed matching grh\n");
666 /* Class-specific handling */
667 switch (mad->mad_hdr.mgmt_class) {
668 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
669 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
670 /* 255 indicates the dom0 */
671 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
672 if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
674 /* for a VF. drop unsolicited MADs */
675 if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
676 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
677 slave, mad->mad_hdr.mgmt_class,
678 mad->mad_hdr.method);
683 case IB_MGMT_CLASS_SUBN_ADM:
684 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
685 (struct ib_sa_mad *) mad))
688 case IB_MGMT_CLASS_CM:
689 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
692 case IB_MGMT_CLASS_DEVICE_MGMT:
693 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
697 /* Drop unsupported classes for slaves in tunnel mode */
698 if (slave != mlx4_master_func_num(dev->dev)) {
699 pr_debug("dropping unsupported ingress mad from class:%d "
700 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
704 /*make sure that no slave==255 was not handled yet.*/
705 if (slave >= dev->dev->caps.sqp_demux) {
706 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
707 slave, dev->dev->caps.sqp_demux);
711 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
713 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
718 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
719 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
720 const struct ib_mad *in_mad, struct ib_mad *out_mad)
722 u16 slid, prev_lid = 0;
724 struct ib_port_attr pattr;
726 if (in_wc && in_wc->qp->qp_num) {
727 pr_debug("received MAD: slid:%d sqpn:%d "
728 "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
729 in_wc->slid, in_wc->src_qp,
730 in_wc->dlid_path_bits,
733 in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
734 be16_to_cpu(in_mad->mad_hdr.attr_id));
735 if (in_wc->wc_flags & IB_WC_GRH) {
736 pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
737 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
738 be64_to_cpu(in_grh->sgid.global.interface_id));
739 pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
740 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
741 be64_to_cpu(in_grh->dgid.global.interface_id));
745 slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
747 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
748 forward_trap(to_mdev(ibdev), port_num, in_mad);
749 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
752 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
753 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
754 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
755 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET &&
756 in_mad->mad_hdr.method != IB_MGMT_METHOD_TRAP_REPRESS)
757 return IB_MAD_RESULT_SUCCESS;
760 * Don't process SMInfo queries -- the SMA can't handle them.
762 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
763 return IB_MAD_RESULT_SUCCESS;
764 } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
765 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1 ||
766 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2 ||
767 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
768 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
769 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET)
770 return IB_MAD_RESULT_SUCCESS;
772 return IB_MAD_RESULT_SUCCESS;
774 if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
775 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
776 in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
777 in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
778 !ib_query_port(ibdev, port_num, &pattr))
779 prev_lid = pattr.lid;
781 err = mlx4_MAD_IFC(to_mdev(ibdev),
782 (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
783 (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
784 MLX4_MAD_IFC_NET_VIEW,
785 port_num, in_wc, in_grh, in_mad, out_mad);
787 return IB_MAD_RESULT_FAILURE;
789 if (!out_mad->mad_hdr.status) {
790 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
791 smp_snoop(ibdev, port_num, in_mad, prev_lid);
792 /* slaves get node desc from FW */
793 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
794 node_desc_override(ibdev, out_mad);
797 /* set return bit in status of directed route responses */
798 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
799 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
801 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
802 /* no response for trap repress */
803 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
805 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
808 static void edit_counter(struct mlx4_counter *cnt,
809 struct ib_pma_portcounters *pma_cnt)
811 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
812 (be64_to_cpu(cnt->tx_bytes) >> 2));
813 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
814 (be64_to_cpu(cnt->rx_bytes) >> 2));
815 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
816 be64_to_cpu(cnt->tx_frames));
817 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
818 be64_to_cpu(cnt->rx_frames));
821 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
822 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
823 const struct ib_mad *in_mad, struct ib_mad *out_mad)
825 struct mlx4_counter counter_stats;
826 struct mlx4_ib_dev *dev = to_mdev(ibdev);
829 if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
832 memset(&counter_stats, 0, sizeof(counter_stats));
833 err = mlx4_get_counter_stats(dev->dev,
834 dev->counters[port_num - 1].index,
837 err = IB_MAD_RESULT_FAILURE;
839 memset(out_mad->data, 0, sizeof out_mad->data);
840 switch (counter_stats.counter_mode & 0xf) {
842 edit_counter(&counter_stats,
843 (void *)(out_mad->data + 40));
844 err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
847 err = IB_MAD_RESULT_FAILURE;
854 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
855 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
856 const struct ib_mad_hdr *in, size_t in_mad_size,
857 struct ib_mad_hdr *out, size_t *out_mad_size,
858 u16 *out_mad_pkey_index)
860 struct mlx4_ib_dev *dev = to_mdev(ibdev);
861 const struct ib_mad *in_mad = (const struct ib_mad *)in;
862 struct ib_mad *out_mad = (struct ib_mad *)out;
864 if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
865 *out_mad_size != sizeof(*out_mad)))
866 return IB_MAD_RESULT_FAILURE;
868 switch (rdma_port_get_link_layer(ibdev, port_num)) {
869 case IB_LINK_LAYER_INFINIBAND:
870 if (!mlx4_is_slave(dev->dev))
871 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
872 in_grh, in_mad, out_mad);
873 case IB_LINK_LAYER_ETHERNET:
874 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
875 in_grh, in_mad, out_mad);
881 static void send_handler(struct ib_mad_agent *agent,
882 struct ib_mad_send_wc *mad_send_wc)
884 if (mad_send_wc->send_buf->context[0])
885 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
886 ib_free_send_mad(mad_send_wc->send_buf);
889 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
891 struct ib_mad_agent *agent;
894 enum rdma_link_layer ll;
896 for (p = 0; p < dev->num_ports; ++p) {
897 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
898 for (q = 0; q <= 1; ++q) {
899 if (ll == IB_LINK_LAYER_INFINIBAND) {
900 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
901 q ? IB_QPT_GSI : IB_QPT_SMI,
902 NULL, 0, send_handler,
905 ret = PTR_ERR(agent);
908 dev->send_agent[p][q] = agent;
910 dev->send_agent[p][q] = NULL;
917 for (p = 0; p < dev->num_ports; ++p)
918 for (q = 0; q <= 1; ++q)
919 if (dev->send_agent[p][q])
920 ib_unregister_mad_agent(dev->send_agent[p][q]);
925 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
927 struct ib_mad_agent *agent;
930 for (p = 0; p < dev->num_ports; ++p) {
931 for (q = 0; q <= 1; ++q) {
932 agent = dev->send_agent[p][q];
934 dev->send_agent[p][q] = NULL;
935 ib_unregister_mad_agent(agent);
940 ib_destroy_ah(dev->sm_ah[p]);
944 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
946 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
948 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
949 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
950 MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
953 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
955 /* re-configure the alias-guid and mcg's */
956 if (mlx4_is_master(dev->dev)) {
957 mlx4_ib_invalidate_all_guid_record(dev, port_num);
959 if (!dev->sriov.is_going_down) {
960 mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
961 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
962 MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
965 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
968 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
969 struct mlx4_eqe *eqe)
971 __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
972 GET_MASK_FROM_EQE(eqe));
975 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
976 u32 guid_tbl_blk_num, u32 change_bitmap)
978 struct ib_smp *in_mad = NULL;
979 struct ib_smp *out_mad = NULL;
982 if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
985 in_mad = kmalloc(sizeof *in_mad, GFP_KERNEL);
986 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
987 if (!in_mad || !out_mad) {
988 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
992 guid_tbl_blk_num *= 4;
994 for (i = 0; i < 4; i++) {
995 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
997 memset(in_mad, 0, sizeof *in_mad);
998 memset(out_mad, 0, sizeof *out_mad);
1000 in_mad->base_version = 1;
1001 in_mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1002 in_mad->class_version = 1;
1003 in_mad->method = IB_MGMT_METHOD_GET;
1004 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
1005 in_mad->attr_mod = cpu_to_be32(guid_tbl_blk_num + i);
1007 if (mlx4_MAD_IFC(dev,
1008 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1009 port_num, NULL, NULL, in_mad, out_mad)) {
1010 mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1014 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1016 (u8 *)(&((struct ib_smp *)out_mad)->data));
1017 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1019 (u8 *)(&((struct ib_smp *)out_mad)->data));
1028 void handle_port_mgmt_change_event(struct work_struct *work)
1030 struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1031 struct mlx4_ib_dev *dev = ew->ib_dev;
1032 struct mlx4_eqe *eqe = &(ew->ib_eqe);
1033 u8 port = eqe->event.port_mgmt_change.port;
1038 switch (eqe->subtype) {
1039 case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1040 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1042 /* Update the SM ah - This should be done before handling
1043 the other changed attributes so that MADs can be sent to the SM */
1044 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1045 u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1046 u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1047 update_sm_ah(dev, port, lid, sl);
1050 /* Check if it is a lid change event */
1051 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1052 handle_lid_change_event(dev, port);
1054 /* Generate GUID changed event */
1055 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1056 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1057 /*if master, notify all slaves*/
1058 if (mlx4_is_master(dev->dev))
1059 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1060 MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1063 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1064 handle_client_rereg_event(dev, port);
1067 case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1068 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1069 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1070 propagate_pkey_ev(dev, port, eqe);
1072 case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1073 /* paravirtualized master's guid is guid 0 -- does not change */
1074 if (!mlx4_is_master(dev->dev))
1075 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1076 /*if master, notify relevant slaves*/
1077 else if (!dev->sriov.is_going_down) {
1078 tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1079 change_bitmap = GET_MASK_FROM_EQE(eqe);
1080 handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1084 pr_warn("Unsupported subtype 0x%x for "
1085 "Port Management Change event\n", eqe->subtype);
1091 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1092 enum ib_event_type type)
1094 struct ib_event event;
1096 event.device = &dev->ib_dev;
1097 event.element.port_num = port_num;
1100 ib_dispatch_event(&event);
1103 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1105 unsigned long flags;
1106 struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1107 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1108 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1109 if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1110 queue_work(ctx->wq, &ctx->work);
1111 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1114 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1115 struct mlx4_ib_demux_pv_qp *tun_qp,
1118 struct ib_sge sg_list;
1119 struct ib_recv_wr recv_wr, *bad_recv_wr;
1122 size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1123 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1125 sg_list.addr = tun_qp->ring[index].map;
1126 sg_list.length = size;
1127 sg_list.lkey = ctx->mr->lkey;
1129 recv_wr.next = NULL;
1130 recv_wr.sg_list = &sg_list;
1131 recv_wr.num_sge = 1;
1132 recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1133 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1134 ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1135 size, DMA_FROM_DEVICE);
1136 return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1139 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1140 int slave, struct ib_sa_mad *sa_mad)
1144 /* dispatch to different sa handlers */
1145 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1146 case IB_SA_ATTR_MC_MEMBER_REC:
1147 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1155 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1157 int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1159 return (qpn >= proxy_start && qpn <= proxy_start + 1);
1163 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1164 enum ib_qp_type dest_qpt, u16 pkey_index,
1165 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1166 u8 *s_mac, struct ib_mad *mad)
1169 struct ib_send_wr wr, *bad_wr;
1170 struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1171 struct mlx4_ib_demux_pv_qp *sqp;
1172 struct mlx4_mad_snd_buf *sqp_mad;
1174 struct ib_qp *send_qp = NULL;
1175 unsigned wire_tx_ix = 0;
1182 sqp_ctx = dev->sriov.sqps[port-1];
1184 /* check if proxy qp created */
1185 if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1188 if (dest_qpt == IB_QPT_SMI) {
1190 sqp = &sqp_ctx->qp[0];
1191 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1194 sqp = &sqp_ctx->qp[1];
1195 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1201 sgid_index = attr->grh.sgid_index;
1202 attr->grh.sgid_index = 0;
1203 ah = ib_create_ah(sqp_ctx->pd, attr);
1206 attr->grh.sgid_index = sgid_index;
1207 to_mah(ah)->av.ib.gid_index = sgid_index;
1208 /* get rid of force-loopback bit */
1209 to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1210 spin_lock(&sqp->tx_lock);
1211 if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1212 (MLX4_NUM_TUNNEL_BUFS - 1))
1215 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1216 spin_unlock(&sqp->tx_lock);
1220 sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1221 if (sqp->tx_ring[wire_tx_ix].ah)
1222 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1223 sqp->tx_ring[wire_tx_ix].ah = ah;
1224 ib_dma_sync_single_for_cpu(&dev->ib_dev,
1225 sqp->tx_ring[wire_tx_ix].buf.map,
1226 sizeof (struct mlx4_mad_snd_buf),
1229 memcpy(&sqp_mad->payload, mad, sizeof *mad);
1231 ib_dma_sync_single_for_device(&dev->ib_dev,
1232 sqp->tx_ring[wire_tx_ix].buf.map,
1233 sizeof (struct mlx4_mad_snd_buf),
1236 list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1237 list.length = sizeof (struct mlx4_mad_snd_buf);
1238 list.lkey = sqp_ctx->mr->lkey;
1241 wr.wr.ud.port_num = port;
1242 wr.wr.ud.pkey_index = wire_pkey_ix;
1243 wr.wr.ud.remote_qkey = qkey;
1244 wr.wr.ud.remote_qpn = remote_qpn;
1246 wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1249 wr.opcode = IB_WR_SEND;
1250 wr.send_flags = IB_SEND_SIGNALED;
1252 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1255 ret = ib_post_send(send_qp, &wr, &bad_wr);
1262 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1264 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1266 return mlx4_get_base_gid_ix(dev->dev, slave, port);
1269 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1270 struct ib_ah_attr *ah_attr)
1272 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1273 ah_attr->grh.sgid_index = slave;
1275 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1278 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1280 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1281 struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1282 int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1283 struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1284 struct mlx4_ib_ah ah;
1285 struct ib_ah_attr ah_attr;
1290 /* Get slave that sent this packet */
1291 if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1292 wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1293 (wc->src_qp & 0x1) != ctx->port - 1 ||
1295 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1298 slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1299 if (slave != ctx->slave) {
1300 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1301 "belongs to another slave\n", wc->src_qp);
1305 /* Map transaction ID */
1306 ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1307 sizeof (struct mlx4_tunnel_mad),
1309 switch (tunnel->mad.mad_hdr.method) {
1310 case IB_MGMT_METHOD_SET:
1311 case IB_MGMT_METHOD_GET:
1312 case IB_MGMT_METHOD_REPORT:
1313 case IB_SA_METHOD_GET_TABLE:
1314 case IB_SA_METHOD_DELETE:
1315 case IB_SA_METHOD_GET_MULTI:
1316 case IB_SA_METHOD_GET_TRACE_TBL:
1317 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1319 mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1320 "class:%d slave:%d\n", *slave_id,
1321 tunnel->mad.mad_hdr.mgmt_class, slave);
1329 /* Class-specific handling */
1330 switch (tunnel->mad.mad_hdr.mgmt_class) {
1331 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1332 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1333 if (slave != mlx4_master_func_num(dev->dev) &&
1334 !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1337 case IB_MGMT_CLASS_SUBN_ADM:
1338 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1339 (struct ib_sa_mad *) &tunnel->mad))
1342 case IB_MGMT_CLASS_CM:
1343 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1344 (struct ib_mad *) &tunnel->mad))
1347 case IB_MGMT_CLASS_DEVICE_MGMT:
1348 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1349 tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1353 /* Drop unsupported classes for slaves in tunnel mode */
1354 if (slave != mlx4_master_func_num(dev->dev)) {
1355 mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1356 "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1361 /* We are using standard ib_core services to send the mad, so generate a
1362 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1363 memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1364 ah.ibah.device = ctx->ib_dev;
1366 port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1367 port = mlx4_slave_convert_port(dev->dev, slave, port);
1370 ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1372 mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1373 if (ah_attr.ah_flags & IB_AH_GRH)
1374 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1376 memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1377 ah_attr.vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1378 /* if slave have default vlan use it */
1379 mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1380 &ah_attr.vlan_id, &ah_attr.sl);
1382 mlx4_ib_send_to_wire(dev, slave, ctx->port,
1383 is_proxy_qp0(dev, wc->src_qp, slave) ?
1384 IB_QPT_SMI : IB_QPT_GSI,
1385 be16_to_cpu(tunnel->hdr.pkey_index),
1386 be32_to_cpu(tunnel->hdr.remote_qpn),
1387 be32_to_cpu(tunnel->hdr.qkey),
1388 &ah_attr, wc->smac, &tunnel->mad);
1391 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1392 enum ib_qp_type qp_type, int is_tun)
1395 struct mlx4_ib_demux_pv_qp *tun_qp;
1396 int rx_buf_size, tx_buf_size;
1398 if (qp_type > IB_QPT_GSI)
1401 tun_qp = &ctx->qp[qp_type];
1403 tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1408 tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1409 sizeof (struct mlx4_ib_tun_tx_buf),
1411 if (!tun_qp->tx_ring) {
1412 kfree(tun_qp->ring);
1413 tun_qp->ring = NULL;
1418 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1419 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1421 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1422 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1425 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1426 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1427 if (!tun_qp->ring[i].addr)
1429 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1430 tun_qp->ring[i].addr,
1433 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1434 kfree(tun_qp->ring[i].addr);
1439 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1440 tun_qp->tx_ring[i].buf.addr =
1441 kmalloc(tx_buf_size, GFP_KERNEL);
1442 if (!tun_qp->tx_ring[i].buf.addr)
1444 tun_qp->tx_ring[i].buf.map =
1445 ib_dma_map_single(ctx->ib_dev,
1446 tun_qp->tx_ring[i].buf.addr,
1449 if (ib_dma_mapping_error(ctx->ib_dev,
1450 tun_qp->tx_ring[i].buf.map)) {
1451 kfree(tun_qp->tx_ring[i].buf.addr);
1454 tun_qp->tx_ring[i].ah = NULL;
1456 spin_lock_init(&tun_qp->tx_lock);
1457 tun_qp->tx_ix_head = 0;
1458 tun_qp->tx_ix_tail = 0;
1459 tun_qp->proxy_qpt = qp_type;
1466 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1467 tx_buf_size, DMA_TO_DEVICE);
1468 kfree(tun_qp->tx_ring[i].buf.addr);
1470 kfree(tun_qp->tx_ring);
1471 tun_qp->tx_ring = NULL;
1472 i = MLX4_NUM_TUNNEL_BUFS;
1476 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1477 rx_buf_size, DMA_FROM_DEVICE);
1478 kfree(tun_qp->ring[i].addr);
1480 kfree(tun_qp->ring);
1481 tun_qp->ring = NULL;
1485 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1486 enum ib_qp_type qp_type, int is_tun)
1489 struct mlx4_ib_demux_pv_qp *tun_qp;
1490 int rx_buf_size, tx_buf_size;
1492 if (qp_type > IB_QPT_GSI)
1495 tun_qp = &ctx->qp[qp_type];
1497 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1498 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1500 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1501 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1505 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1506 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1507 rx_buf_size, DMA_FROM_DEVICE);
1508 kfree(tun_qp->ring[i].addr);
1511 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1512 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1513 tx_buf_size, DMA_TO_DEVICE);
1514 kfree(tun_qp->tx_ring[i].buf.addr);
1515 if (tun_qp->tx_ring[i].ah)
1516 ib_destroy_ah(tun_qp->tx_ring[i].ah);
1518 kfree(tun_qp->tx_ring);
1519 kfree(tun_qp->ring);
1522 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1524 struct mlx4_ib_demux_pv_ctx *ctx;
1525 struct mlx4_ib_demux_pv_qp *tun_qp;
1528 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1529 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1531 while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1532 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1533 if (wc.status == IB_WC_SUCCESS) {
1534 switch (wc.opcode) {
1536 mlx4_ib_multiplex_mad(ctx, &wc);
1537 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1539 (MLX4_NUM_TUNNEL_BUFS - 1));
1541 pr_err("Failed reposting tunnel "
1542 "buf:%lld\n", wc.wr_id);
1545 pr_debug("received tunnel send completion:"
1546 "wrid=0x%llx, status=0x%x\n",
1547 wc.wr_id, wc.status);
1548 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1549 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1550 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1552 spin_lock(&tun_qp->tx_lock);
1553 tun_qp->tx_ix_tail++;
1554 spin_unlock(&tun_qp->tx_lock);
1561 pr_debug("mlx4_ib: completion error in tunnel: %d."
1562 " status = %d, wrid = 0x%llx\n",
1563 ctx->slave, wc.status, wc.wr_id);
1564 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1565 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1566 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1567 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1569 spin_lock(&tun_qp->tx_lock);
1570 tun_qp->tx_ix_tail++;
1571 spin_unlock(&tun_qp->tx_lock);
1577 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1579 struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1581 /* It's worse than that! He's dead, Jim! */
1582 pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1583 event->event, sqp->port);
1586 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1587 enum ib_qp_type qp_type, int create_tun)
1590 struct mlx4_ib_demux_pv_qp *tun_qp;
1591 struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1592 struct ib_qp_attr attr;
1593 int qp_attr_mask_INIT;
1595 if (qp_type > IB_QPT_GSI)
1598 tun_qp = &ctx->qp[qp_type];
1600 memset(&qp_init_attr, 0, sizeof qp_init_attr);
1601 qp_init_attr.init_attr.send_cq = ctx->cq;
1602 qp_init_attr.init_attr.recv_cq = ctx->cq;
1603 qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1604 qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1605 qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1606 qp_init_attr.init_attr.cap.max_send_sge = 1;
1607 qp_init_attr.init_attr.cap.max_recv_sge = 1;
1609 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1610 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1611 qp_init_attr.port = ctx->port;
1612 qp_init_attr.slave = ctx->slave;
1613 qp_init_attr.proxy_qp_type = qp_type;
1614 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1615 IB_QP_QKEY | IB_QP_PORT;
1617 qp_init_attr.init_attr.qp_type = qp_type;
1618 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1619 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1621 qp_init_attr.init_attr.port_num = ctx->port;
1622 qp_init_attr.init_attr.qp_context = ctx;
1623 qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1624 tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1625 if (IS_ERR(tun_qp->qp)) {
1626 ret = PTR_ERR(tun_qp->qp);
1628 pr_err("Couldn't create %s QP (%d)\n",
1629 create_tun ? "tunnel" : "special", ret);
1633 memset(&attr, 0, sizeof attr);
1634 attr.qp_state = IB_QPS_INIT;
1637 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1638 ctx->port, IB_DEFAULT_PKEY_FULL,
1640 if (ret || !create_tun)
1642 to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1643 attr.qkey = IB_QP1_QKEY;
1644 attr.port_num = ctx->port;
1645 ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1647 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1648 create_tun ? "tunnel" : "special", ret);
1651 attr.qp_state = IB_QPS_RTR;
1652 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1654 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1655 create_tun ? "tunnel" : "special", ret);
1658 attr.qp_state = IB_QPS_RTS;
1660 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1662 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1663 create_tun ? "tunnel" : "special", ret);
1667 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1668 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1670 pr_err(" mlx4_ib_post_pv_buf error"
1671 " (err = %d, i = %d)\n", ret, i);
1678 ib_destroy_qp(tun_qp->qp);
1684 * IB MAD completion callback for real SQPs
1686 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1688 struct mlx4_ib_demux_pv_ctx *ctx;
1689 struct mlx4_ib_demux_pv_qp *sqp;
1694 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1695 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1697 while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1698 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1699 if (wc.status == IB_WC_SUCCESS) {
1700 switch (wc.opcode) {
1702 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1703 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1704 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1706 spin_lock(&sqp->tx_lock);
1708 spin_unlock(&sqp->tx_lock);
1711 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1712 (sqp->ring[wc.wr_id &
1713 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1714 grh = &(((struct mlx4_mad_rcv_buf *)
1715 (sqp->ring[wc.wr_id &
1716 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1717 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1718 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1719 (MLX4_NUM_TUNNEL_BUFS - 1)))
1720 pr_err("Failed reposting SQP "
1721 "buf:%lld\n", wc.wr_id);
1728 pr_debug("mlx4_ib: completion error in tunnel: %d."
1729 " status = %d, wrid = 0x%llx\n",
1730 ctx->slave, wc.status, wc.wr_id);
1731 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1732 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1733 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1734 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1736 spin_lock(&sqp->tx_lock);
1738 spin_unlock(&sqp->tx_lock);
1744 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1745 struct mlx4_ib_demux_pv_ctx **ret_ctx)
1747 struct mlx4_ib_demux_pv_ctx *ctx;
1750 ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1752 pr_err("failed allocating pv resource context "
1753 "for port %d, slave %d\n", port, slave);
1757 ctx->ib_dev = &dev->ib_dev;
1764 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1766 if (dev->sriov.demux[port - 1].tun[slave]) {
1767 kfree(dev->sriov.demux[port - 1].tun[slave]);
1768 dev->sriov.demux[port - 1].tun[slave] = NULL;
1772 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1773 int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1776 struct ib_cq_init_attr cq_attr = {};
1778 if (ctx->state != DEMUX_PV_STATE_DOWN)
1781 ctx->state = DEMUX_PV_STATE_STARTING;
1782 /* have QP0 only if link layer is IB */
1783 if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1784 IB_LINK_LAYER_INFINIBAND)
1788 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1790 pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1795 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1797 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1801 cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1805 cq_attr.cqe = cq_size;
1806 ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1807 NULL, ctx, &cq_attr);
1808 if (IS_ERR(ctx->cq)) {
1809 ret = PTR_ERR(ctx->cq);
1810 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1814 ctx->pd = ib_alloc_pd(ctx->ib_dev);
1815 if (IS_ERR(ctx->pd)) {
1816 ret = PTR_ERR(ctx->pd);
1817 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1821 ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1822 if (IS_ERR(ctx->mr)) {
1823 ret = PTR_ERR(ctx->mr);
1824 pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1829 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1831 pr_err("Couldn't create %s QP0 (%d)\n",
1832 create_tun ? "tunnel for" : "", ret);
1837 ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1839 pr_err("Couldn't create %s QP1 (%d)\n",
1840 create_tun ? "tunnel for" : "", ret);
1845 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1847 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1849 ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1851 ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1853 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1856 ctx->state = DEMUX_PV_STATE_ACTIVE;
1861 ib_destroy_qp(ctx->qp[1].qp);
1862 ctx->qp[1].qp = NULL;
1867 ib_destroy_qp(ctx->qp[0].qp);
1868 ctx->qp[0].qp = NULL;
1871 ib_dereg_mr(ctx->mr);
1875 ib_dealloc_pd(ctx->pd);
1879 ib_destroy_cq(ctx->cq);
1883 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1887 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1889 ctx->state = DEMUX_PV_STATE_DOWN;
1893 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1894 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1898 if (ctx->state > DEMUX_PV_STATE_DOWN) {
1899 ctx->state = DEMUX_PV_STATE_DOWNING;
1901 flush_workqueue(ctx->wq);
1903 ib_destroy_qp(ctx->qp[0].qp);
1904 ctx->qp[0].qp = NULL;
1905 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1907 ib_destroy_qp(ctx->qp[1].qp);
1908 ctx->qp[1].qp = NULL;
1909 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1910 ib_dereg_mr(ctx->mr);
1912 ib_dealloc_pd(ctx->pd);
1914 ib_destroy_cq(ctx->cq);
1916 ctx->state = DEMUX_PV_STATE_DOWN;
1920 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1921 int port, int do_init)
1926 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1927 /* for master, destroy real sqp resources */
1928 if (slave == mlx4_master_func_num(dev->dev))
1929 destroy_pv_resources(dev, slave, port,
1930 dev->sriov.sqps[port - 1], 1);
1931 /* destroy the tunnel qp resources */
1932 destroy_pv_resources(dev, slave, port,
1933 dev->sriov.demux[port - 1].tun[slave], 1);
1937 /* create the tunnel qp resources */
1938 ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1939 dev->sriov.demux[port - 1].tun[slave]);
1941 /* for master, create the real sqp resources */
1942 if (!ret && slave == mlx4_master_func_num(dev->dev))
1943 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1944 dev->sriov.sqps[port - 1]);
1948 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1950 struct mlx4_ib_demux_work *dmxw;
1952 dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1953 mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1959 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1960 struct mlx4_ib_demux_ctx *ctx,
1967 ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1968 sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1974 ctx->ib_dev = &dev->ib_dev;
1977 i < min(dev->dev->caps.sqp_demux,
1978 (u16)(dev->dev->persist->num_vfs + 1));
1980 struct mlx4_active_ports actv_ports =
1981 mlx4_get_active_ports(dev->dev, i);
1983 if (!test_bit(port - 1, actv_ports.ports))
1986 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1993 ret = mlx4_ib_mcg_port_init(ctx);
1995 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1999 snprintf(name, sizeof name, "mlx4_ibt%d", port);
2000 ctx->wq = create_singlethread_workqueue(name);
2002 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2007 snprintf(name, sizeof name, "mlx4_ibud%d", port);
2008 ctx->ud_wq = create_singlethread_workqueue(name);
2010 pr_err("Failed to create up/down WQ for port %d\n", port);
2018 destroy_workqueue(ctx->wq);
2022 mlx4_ib_mcg_port_cleanup(ctx, 1);
2024 for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2025 free_pv_object(dev, i, port);
2031 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2033 if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2034 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2035 flush_workqueue(sqp_ctx->wq);
2036 if (sqp_ctx->has_smi) {
2037 ib_destroy_qp(sqp_ctx->qp[0].qp);
2038 sqp_ctx->qp[0].qp = NULL;
2039 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2041 ib_destroy_qp(sqp_ctx->qp[1].qp);
2042 sqp_ctx->qp[1].qp = NULL;
2043 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2044 ib_dereg_mr(sqp_ctx->mr);
2046 ib_dealloc_pd(sqp_ctx->pd);
2048 ib_destroy_cq(sqp_ctx->cq);
2050 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2054 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2058 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2059 mlx4_ib_mcg_port_cleanup(ctx, 1);
2060 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2063 if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2064 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2066 flush_workqueue(ctx->wq);
2067 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2068 destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2069 free_pv_object(dev, i, ctx->port);
2072 destroy_workqueue(ctx->ud_wq);
2073 destroy_workqueue(ctx->wq);
2077 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2081 if (!mlx4_is_master(dev->dev))
2083 /* initialize or tear down tunnel QPs for the master */
2084 for (i = 0; i < dev->dev->caps.num_ports; i++)
2085 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2089 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2094 if (!mlx4_is_mfunc(dev->dev))
2097 dev->sriov.is_going_down = 0;
2098 spin_lock_init(&dev->sriov.going_down_lock);
2099 mlx4_ib_cm_paravirt_init(dev);
2101 mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2103 if (mlx4_is_slave(dev->dev)) {
2104 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2108 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2109 if (i == mlx4_master_func_num(dev->dev))
2110 mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2112 mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2115 err = mlx4_ib_init_alias_guid_service(dev);
2117 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2120 err = mlx4_ib_device_register_sysfs(dev);
2122 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2126 mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2127 dev->dev->caps.sqp_demux);
2128 for (i = 0; i < dev->num_ports; i++) {
2130 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2133 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2134 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2135 &dev->sriov.sqps[i]);
2138 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2142 mlx4_ib_master_tunnels(dev, 1);
2146 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2149 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2150 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2152 mlx4_ib_device_unregister_sysfs(dev);
2155 mlx4_ib_destroy_alias_guid_service(dev);
2158 mlx4_ib_cm_paravirt_clean(dev, -1);
2163 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2166 unsigned long flags;
2168 if (!mlx4_is_mfunc(dev->dev))
2171 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2172 dev->sriov.is_going_down = 1;
2173 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2174 if (mlx4_is_master(dev->dev)) {
2175 for (i = 0; i < dev->num_ports; i++) {
2176 flush_workqueue(dev->sriov.demux[i].ud_wq);
2177 mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2178 kfree(dev->sriov.sqps[i]);
2179 dev->sriov.sqps[i] = NULL;
2180 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2183 mlx4_ib_cm_paravirt_clean(dev, -1);
2184 mlx4_ib_destroy_alias_guid_service(dev);
2185 mlx4_ib_device_unregister_sysfs(dev);