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)
461 struct ib_send_wr *bad_wr;
462 struct mlx4_ib_demux_pv_ctx *tun_ctx;
463 struct mlx4_ib_demux_pv_qp *tun_qp;
464 struct mlx4_rcv_tunnel_mad *tun_mad;
465 struct ib_ah_attr attr;
467 struct ib_qp *src_qp = NULL;
468 unsigned tun_tx_ix = 0;
473 u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
475 if (dest_qpt > IB_QPT_GSI)
478 tun_ctx = dev->sriov.demux[port-1].tun[slave];
480 /* check if proxy qp created */
481 if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
485 tun_qp = &tun_ctx->qp[0];
487 tun_qp = &tun_ctx->qp[1];
489 /* compute P_Key index to put in tunnel header for slave */
492 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
496 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
499 tun_pkey_ix = pkey_ix;
501 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
503 dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
505 /* get tunnel tx data buf for slave */
508 /* create ah. Just need an empty one with the port num for the post send.
509 * The driver will set the force loopback bit in post_send */
510 memset(&attr, 0, sizeof attr);
511 attr.port_num = port;
513 memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
514 attr.ah_flags = IB_AH_GRH;
516 ah = ib_create_ah(tun_ctx->pd, &attr);
520 /* allocate tunnel tx buf after pass failure returns */
521 spin_lock(&tun_qp->tx_lock);
522 if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
523 (MLX4_NUM_TUNNEL_BUFS - 1))
526 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
527 spin_unlock(&tun_qp->tx_lock);
531 tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
532 if (tun_qp->tx_ring[tun_tx_ix].ah)
533 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
534 tun_qp->tx_ring[tun_tx_ix].ah = ah;
535 ib_dma_sync_single_for_cpu(&dev->ib_dev,
536 tun_qp->tx_ring[tun_tx_ix].buf.map,
537 sizeof (struct mlx4_rcv_tunnel_mad),
540 /* copy over to tunnel buffer */
542 memcpy(&tun_mad->grh, grh, sizeof *grh);
543 memcpy(&tun_mad->mad, mad, sizeof *mad);
545 /* adjust tunnel data */
546 tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
547 tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
548 tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
552 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
555 if (vlan != wc->vlan_id)
556 /* Packet vlan is not the VST-assigned vlan.
561 /* Remove the vlan tag before forwarding
562 * the packet to the VF.
569 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
570 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
571 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
573 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
574 tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
577 ib_dma_sync_single_for_device(&dev->ib_dev,
578 tun_qp->tx_ring[tun_tx_ix].buf.map,
579 sizeof (struct mlx4_rcv_tunnel_mad),
582 list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
583 list.length = sizeof (struct mlx4_rcv_tunnel_mad);
584 list.lkey = tun_ctx->pd->local_dma_lkey;
588 wr.remote_qkey = IB_QP_SET_QKEY;
589 wr.remote_qpn = dqpn;
591 wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
592 wr.wr.sg_list = &list;
594 wr.wr.opcode = IB_WR_SEND;
595 wr.wr.send_flags = IB_SEND_SIGNALED;
597 ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
601 spin_lock(&tun_qp->tx_lock);
602 tun_qp->tx_ix_tail++;
603 spin_unlock(&tun_qp->tx_lock);
604 tun_qp->tx_ring[tun_tx_ix].ah = NULL;
610 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
611 struct ib_wc *wc, struct ib_grh *grh,
614 struct mlx4_ib_dev *dev = to_mdev(ibdev);
620 if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
626 if (!(wc->wc_flags & IB_WC_GRH)) {
627 mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
630 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
631 mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
634 if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
635 mlx4_ib_warn(ibdev, "failed matching grh\n");
638 if (slave >= dev->dev->caps.sqp_demux) {
639 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
640 slave, dev->dev->caps.sqp_demux);
644 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
647 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
649 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
654 /* Initially assume that this mad is for us */
655 slave = mlx4_master_func_num(dev->dev);
657 /* See if the slave id is encoded in a response mad */
658 if (mad->mad_hdr.method & 0x80) {
659 slave_id = (u8 *) &mad->mad_hdr.tid;
661 if (slave != 255) /*255 indicates the dom0*/
662 *slave_id = 0; /* remap tid */
665 /* If a grh is present, we demux according to it */
666 if (wc->wc_flags & IB_WC_GRH) {
667 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
669 mlx4_ib_warn(ibdev, "failed matching grh\n");
673 /* Class-specific handling */
674 switch (mad->mad_hdr.mgmt_class) {
675 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
676 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
677 /* 255 indicates the dom0 */
678 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
679 if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
681 /* for a VF. drop unsolicited MADs */
682 if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
683 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
684 slave, mad->mad_hdr.mgmt_class,
685 mad->mad_hdr.method);
690 case IB_MGMT_CLASS_SUBN_ADM:
691 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
692 (struct ib_sa_mad *) mad))
695 case IB_MGMT_CLASS_CM:
696 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
699 case IB_MGMT_CLASS_DEVICE_MGMT:
700 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
704 /* Drop unsupported classes for slaves in tunnel mode */
705 if (slave != mlx4_master_func_num(dev->dev)) {
706 pr_debug("dropping unsupported ingress mad from class:%d "
707 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
711 /*make sure that no slave==255 was not handled yet.*/
712 if (slave >= dev->dev->caps.sqp_demux) {
713 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
714 slave, dev->dev->caps.sqp_demux);
718 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
720 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
725 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
726 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
727 const struct ib_mad *in_mad, struct ib_mad *out_mad)
729 u16 slid, prev_lid = 0;
731 struct ib_port_attr pattr;
733 if (in_wc && in_wc->qp->qp_num) {
734 pr_debug("received MAD: slid:%d sqpn:%d "
735 "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
736 in_wc->slid, in_wc->src_qp,
737 in_wc->dlid_path_bits,
740 in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
741 be16_to_cpu(in_mad->mad_hdr.attr_id));
742 if (in_wc->wc_flags & IB_WC_GRH) {
743 pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
744 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
745 be64_to_cpu(in_grh->sgid.global.interface_id));
746 pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
747 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
748 be64_to_cpu(in_grh->dgid.global.interface_id));
752 slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
754 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
755 forward_trap(to_mdev(ibdev), port_num, in_mad);
756 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
759 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
760 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
761 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
762 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET &&
763 in_mad->mad_hdr.method != IB_MGMT_METHOD_TRAP_REPRESS)
764 return IB_MAD_RESULT_SUCCESS;
767 * Don't process SMInfo queries -- the SMA can't handle them.
769 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
770 return IB_MAD_RESULT_SUCCESS;
771 } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
772 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1 ||
773 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2 ||
774 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
775 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
776 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET)
777 return IB_MAD_RESULT_SUCCESS;
779 return IB_MAD_RESULT_SUCCESS;
781 if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
782 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
783 in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
784 in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
785 !ib_query_port(ibdev, port_num, &pattr))
786 prev_lid = pattr.lid;
788 err = mlx4_MAD_IFC(to_mdev(ibdev),
789 (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
790 (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
791 MLX4_MAD_IFC_NET_VIEW,
792 port_num, in_wc, in_grh, in_mad, out_mad);
794 return IB_MAD_RESULT_FAILURE;
796 if (!out_mad->mad_hdr.status) {
797 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
798 smp_snoop(ibdev, port_num, in_mad, prev_lid);
799 /* slaves get node desc from FW */
800 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
801 node_desc_override(ibdev, out_mad);
804 /* set return bit in status of directed route responses */
805 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
806 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
808 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
809 /* no response for trap repress */
810 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
812 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
815 static void edit_counter(struct mlx4_counter *cnt,
816 struct ib_pma_portcounters *pma_cnt)
818 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
819 (be64_to_cpu(cnt->tx_bytes) >> 2));
820 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
821 (be64_to_cpu(cnt->rx_bytes) >> 2));
822 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
823 be64_to_cpu(cnt->tx_frames));
824 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
825 be64_to_cpu(cnt->rx_frames));
828 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
829 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
830 const struct ib_mad *in_mad, struct ib_mad *out_mad)
832 struct mlx4_counter counter_stats;
833 struct mlx4_ib_dev *dev = to_mdev(ibdev);
834 struct counter_index *tmp_counter;
835 int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
837 if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
840 memset(&counter_stats, 0, sizeof(counter_stats));
841 mutex_lock(&dev->counters_table[port_num - 1].mutex);
842 list_for_each_entry(tmp_counter,
843 &dev->counters_table[port_num - 1].counters_list,
845 err = mlx4_get_counter_stats(dev->dev,
849 err = IB_MAD_RESULT_FAILURE;
855 mutex_unlock(&dev->counters_table[port_num - 1].mutex);
857 memset(out_mad->data, 0, sizeof out_mad->data);
858 switch (counter_stats.counter_mode & 0xf) {
860 edit_counter(&counter_stats,
861 (void *)(out_mad->data + 40));
862 err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
865 err = IB_MAD_RESULT_FAILURE;
872 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
873 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
874 const struct ib_mad_hdr *in, size_t in_mad_size,
875 struct ib_mad_hdr *out, size_t *out_mad_size,
876 u16 *out_mad_pkey_index)
878 struct mlx4_ib_dev *dev = to_mdev(ibdev);
879 const struct ib_mad *in_mad = (const struct ib_mad *)in;
880 struct ib_mad *out_mad = (struct ib_mad *)out;
881 enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
883 if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
884 *out_mad_size != sizeof(*out_mad)))
885 return IB_MAD_RESULT_FAILURE;
887 /* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
888 * queries, should be called only by VFs and for that specific purpose
890 if (link == IB_LINK_LAYER_INFINIBAND) {
891 if (mlx4_is_slave(dev->dev) &&
892 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
893 in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS)
894 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
895 in_grh, in_mad, out_mad);
897 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
898 in_grh, in_mad, out_mad);
901 if (link == IB_LINK_LAYER_ETHERNET)
902 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
903 in_grh, in_mad, out_mad);
908 static void send_handler(struct ib_mad_agent *agent,
909 struct ib_mad_send_wc *mad_send_wc)
911 if (mad_send_wc->send_buf->context[0])
912 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
913 ib_free_send_mad(mad_send_wc->send_buf);
916 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
918 struct ib_mad_agent *agent;
921 enum rdma_link_layer ll;
923 for (p = 0; p < dev->num_ports; ++p) {
924 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
925 for (q = 0; q <= 1; ++q) {
926 if (ll == IB_LINK_LAYER_INFINIBAND) {
927 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
928 q ? IB_QPT_GSI : IB_QPT_SMI,
929 NULL, 0, send_handler,
932 ret = PTR_ERR(agent);
935 dev->send_agent[p][q] = agent;
937 dev->send_agent[p][q] = NULL;
944 for (p = 0; p < dev->num_ports; ++p)
945 for (q = 0; q <= 1; ++q)
946 if (dev->send_agent[p][q])
947 ib_unregister_mad_agent(dev->send_agent[p][q]);
952 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
954 struct ib_mad_agent *agent;
957 for (p = 0; p < dev->num_ports; ++p) {
958 for (q = 0; q <= 1; ++q) {
959 agent = dev->send_agent[p][q];
961 dev->send_agent[p][q] = NULL;
962 ib_unregister_mad_agent(agent);
967 ib_destroy_ah(dev->sm_ah[p]);
971 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
973 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
975 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
976 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
977 MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
980 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
982 /* re-configure the alias-guid and mcg's */
983 if (mlx4_is_master(dev->dev)) {
984 mlx4_ib_invalidate_all_guid_record(dev, port_num);
986 if (!dev->sriov.is_going_down) {
987 mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
988 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
989 MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
992 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
995 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
996 struct mlx4_eqe *eqe)
998 __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
999 GET_MASK_FROM_EQE(eqe));
1002 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1003 u32 guid_tbl_blk_num, u32 change_bitmap)
1005 struct ib_smp *in_mad = NULL;
1006 struct ib_smp *out_mad = NULL;
1009 if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1012 in_mad = kmalloc(sizeof *in_mad, GFP_KERNEL);
1013 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1014 if (!in_mad || !out_mad) {
1015 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1019 guid_tbl_blk_num *= 4;
1021 for (i = 0; i < 4; i++) {
1022 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1024 memset(in_mad, 0, sizeof *in_mad);
1025 memset(out_mad, 0, sizeof *out_mad);
1027 in_mad->base_version = 1;
1028 in_mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1029 in_mad->class_version = 1;
1030 in_mad->method = IB_MGMT_METHOD_GET;
1031 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
1032 in_mad->attr_mod = cpu_to_be32(guid_tbl_blk_num + i);
1034 if (mlx4_MAD_IFC(dev,
1035 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1036 port_num, NULL, NULL, in_mad, out_mad)) {
1037 mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1041 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1043 (u8 *)(&((struct ib_smp *)out_mad)->data));
1044 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1046 (u8 *)(&((struct ib_smp *)out_mad)->data));
1055 void handle_port_mgmt_change_event(struct work_struct *work)
1057 struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1058 struct mlx4_ib_dev *dev = ew->ib_dev;
1059 struct mlx4_eqe *eqe = &(ew->ib_eqe);
1060 u8 port = eqe->event.port_mgmt_change.port;
1065 switch (eqe->subtype) {
1066 case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1067 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1069 /* Update the SM ah - This should be done before handling
1070 the other changed attributes so that MADs can be sent to the SM */
1071 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1072 u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1073 u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1074 update_sm_ah(dev, port, lid, sl);
1077 /* Check if it is a lid change event */
1078 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1079 handle_lid_change_event(dev, port);
1081 /* Generate GUID changed event */
1082 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1083 if (mlx4_is_master(dev->dev)) {
1087 if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1088 err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1090 gid.global.subnet_prefix =
1091 eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1093 pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1096 pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1098 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1099 be64_to_cpu(gid.global.subnet_prefix));
1100 atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1101 be64_to_cpu(gid.global.subnet_prefix));
1104 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1105 /*if master, notify all slaves*/
1106 if (mlx4_is_master(dev->dev))
1107 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1108 MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1111 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1112 handle_client_rereg_event(dev, port);
1115 case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1116 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1117 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1118 propagate_pkey_ev(dev, port, eqe);
1120 case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1121 /* paravirtualized master's guid is guid 0 -- does not change */
1122 if (!mlx4_is_master(dev->dev))
1123 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1124 /*if master, notify relevant slaves*/
1125 else if (!dev->sriov.is_going_down) {
1126 tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1127 change_bitmap = GET_MASK_FROM_EQE(eqe);
1128 handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1132 pr_warn("Unsupported subtype 0x%x for "
1133 "Port Management Change event\n", eqe->subtype);
1139 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1140 enum ib_event_type type)
1142 struct ib_event event;
1144 event.device = &dev->ib_dev;
1145 event.element.port_num = port_num;
1148 ib_dispatch_event(&event);
1151 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1153 unsigned long flags;
1154 struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1155 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1156 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1157 if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1158 queue_work(ctx->wq, &ctx->work);
1159 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1162 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1163 struct mlx4_ib_demux_pv_qp *tun_qp,
1166 struct ib_sge sg_list;
1167 struct ib_recv_wr recv_wr, *bad_recv_wr;
1170 size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1171 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1173 sg_list.addr = tun_qp->ring[index].map;
1174 sg_list.length = size;
1175 sg_list.lkey = ctx->pd->local_dma_lkey;
1177 recv_wr.next = NULL;
1178 recv_wr.sg_list = &sg_list;
1179 recv_wr.num_sge = 1;
1180 recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1181 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1182 ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1183 size, DMA_FROM_DEVICE);
1184 return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1187 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1188 int slave, struct ib_sa_mad *sa_mad)
1192 /* dispatch to different sa handlers */
1193 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1194 case IB_SA_ATTR_MC_MEMBER_REC:
1195 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1203 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1205 int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1207 return (qpn >= proxy_start && qpn <= proxy_start + 1);
1211 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1212 enum ib_qp_type dest_qpt, u16 pkey_index,
1213 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1214 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1218 struct ib_send_wr *bad_wr;
1219 struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1220 struct mlx4_ib_demux_pv_qp *sqp;
1221 struct mlx4_mad_snd_buf *sqp_mad;
1223 struct ib_qp *send_qp = NULL;
1224 unsigned wire_tx_ix = 0;
1231 sqp_ctx = dev->sriov.sqps[port-1];
1233 /* check if proxy qp created */
1234 if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1237 if (dest_qpt == IB_QPT_SMI) {
1239 sqp = &sqp_ctx->qp[0];
1240 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1243 sqp = &sqp_ctx->qp[1];
1244 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1250 sgid_index = attr->grh.sgid_index;
1251 attr->grh.sgid_index = 0;
1252 ah = ib_create_ah(sqp_ctx->pd, attr);
1255 attr->grh.sgid_index = sgid_index;
1256 to_mah(ah)->av.ib.gid_index = sgid_index;
1257 /* get rid of force-loopback bit */
1258 to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1259 spin_lock(&sqp->tx_lock);
1260 if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1261 (MLX4_NUM_TUNNEL_BUFS - 1))
1264 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1265 spin_unlock(&sqp->tx_lock);
1269 sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1270 if (sqp->tx_ring[wire_tx_ix].ah)
1271 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1272 sqp->tx_ring[wire_tx_ix].ah = ah;
1273 ib_dma_sync_single_for_cpu(&dev->ib_dev,
1274 sqp->tx_ring[wire_tx_ix].buf.map,
1275 sizeof (struct mlx4_mad_snd_buf),
1278 memcpy(&sqp_mad->payload, mad, sizeof *mad);
1280 ib_dma_sync_single_for_device(&dev->ib_dev,
1281 sqp->tx_ring[wire_tx_ix].buf.map,
1282 sizeof (struct mlx4_mad_snd_buf),
1285 list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1286 list.length = sizeof (struct mlx4_mad_snd_buf);
1287 list.lkey = sqp_ctx->pd->local_dma_lkey;
1291 wr.pkey_index = wire_pkey_ix;
1292 wr.remote_qkey = qkey;
1293 wr.remote_qpn = remote_qpn;
1295 wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1296 wr.wr.sg_list = &list;
1298 wr.wr.opcode = IB_WR_SEND;
1299 wr.wr.send_flags = IB_SEND_SIGNALED;
1301 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1302 if (vlan_id < 0x1000)
1303 vlan_id |= (attr->sl & 7) << 13;
1304 to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1307 ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1311 spin_lock(&sqp->tx_lock);
1313 spin_unlock(&sqp->tx_lock);
1314 sqp->tx_ring[wire_tx_ix].ah = NULL;
1320 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1322 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1324 return mlx4_get_base_gid_ix(dev->dev, slave, port);
1327 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1328 struct ib_ah_attr *ah_attr)
1330 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1331 ah_attr->grh.sgid_index = slave;
1333 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1336 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1338 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1339 struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1340 int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1341 struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1342 struct mlx4_ib_ah ah;
1343 struct ib_ah_attr ah_attr;
1349 /* Get slave that sent this packet */
1350 if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1351 wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1352 (wc->src_qp & 0x1) != ctx->port - 1 ||
1354 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1357 slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1358 if (slave != ctx->slave) {
1359 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1360 "belongs to another slave\n", wc->src_qp);
1364 /* Map transaction ID */
1365 ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1366 sizeof (struct mlx4_tunnel_mad),
1368 switch (tunnel->mad.mad_hdr.method) {
1369 case IB_MGMT_METHOD_SET:
1370 case IB_MGMT_METHOD_GET:
1371 case IB_MGMT_METHOD_REPORT:
1372 case IB_SA_METHOD_GET_TABLE:
1373 case IB_SA_METHOD_DELETE:
1374 case IB_SA_METHOD_GET_MULTI:
1375 case IB_SA_METHOD_GET_TRACE_TBL:
1376 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1378 mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1379 "class:%d slave:%d\n", *slave_id,
1380 tunnel->mad.mad_hdr.mgmt_class, slave);
1388 /* Class-specific handling */
1389 switch (tunnel->mad.mad_hdr.mgmt_class) {
1390 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1391 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1392 if (slave != mlx4_master_func_num(dev->dev) &&
1393 !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1396 case IB_MGMT_CLASS_SUBN_ADM:
1397 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1398 (struct ib_sa_mad *) &tunnel->mad))
1401 case IB_MGMT_CLASS_CM:
1402 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1403 (struct ib_mad *) &tunnel->mad))
1406 case IB_MGMT_CLASS_DEVICE_MGMT:
1407 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1408 tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1412 /* Drop unsupported classes for slaves in tunnel mode */
1413 if (slave != mlx4_master_func_num(dev->dev)) {
1414 mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1415 "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1420 /* We are using standard ib_core services to send the mad, so generate a
1421 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1422 memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1423 ah.ibah.device = ctx->ib_dev;
1425 port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1426 port = mlx4_slave_convert_port(dev->dev, slave, port);
1429 ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1431 mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1432 if (ah_attr.ah_flags & IB_AH_GRH)
1433 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1435 memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1436 vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1437 /* if slave have default vlan use it */
1438 mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1439 &vlan_id, &ah_attr.sl);
1441 mlx4_ib_send_to_wire(dev, slave, ctx->port,
1442 is_proxy_qp0(dev, wc->src_qp, slave) ?
1443 IB_QPT_SMI : IB_QPT_GSI,
1444 be16_to_cpu(tunnel->hdr.pkey_index),
1445 be32_to_cpu(tunnel->hdr.remote_qpn),
1446 be32_to_cpu(tunnel->hdr.qkey),
1447 &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1450 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1451 enum ib_qp_type qp_type, int is_tun)
1454 struct mlx4_ib_demux_pv_qp *tun_qp;
1455 int rx_buf_size, tx_buf_size;
1457 if (qp_type > IB_QPT_GSI)
1460 tun_qp = &ctx->qp[qp_type];
1462 tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1467 tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1468 sizeof (struct mlx4_ib_tun_tx_buf),
1470 if (!tun_qp->tx_ring) {
1471 kfree(tun_qp->ring);
1472 tun_qp->ring = NULL;
1477 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1478 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1480 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1481 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1484 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1485 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1486 if (!tun_qp->ring[i].addr)
1488 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1489 tun_qp->ring[i].addr,
1492 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1493 kfree(tun_qp->ring[i].addr);
1498 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1499 tun_qp->tx_ring[i].buf.addr =
1500 kmalloc(tx_buf_size, GFP_KERNEL);
1501 if (!tun_qp->tx_ring[i].buf.addr)
1503 tun_qp->tx_ring[i].buf.map =
1504 ib_dma_map_single(ctx->ib_dev,
1505 tun_qp->tx_ring[i].buf.addr,
1508 if (ib_dma_mapping_error(ctx->ib_dev,
1509 tun_qp->tx_ring[i].buf.map)) {
1510 kfree(tun_qp->tx_ring[i].buf.addr);
1513 tun_qp->tx_ring[i].ah = NULL;
1515 spin_lock_init(&tun_qp->tx_lock);
1516 tun_qp->tx_ix_head = 0;
1517 tun_qp->tx_ix_tail = 0;
1518 tun_qp->proxy_qpt = qp_type;
1525 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1526 tx_buf_size, DMA_TO_DEVICE);
1527 kfree(tun_qp->tx_ring[i].buf.addr);
1529 kfree(tun_qp->tx_ring);
1530 tun_qp->tx_ring = NULL;
1531 i = MLX4_NUM_TUNNEL_BUFS;
1535 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1536 rx_buf_size, DMA_FROM_DEVICE);
1537 kfree(tun_qp->ring[i].addr);
1539 kfree(tun_qp->ring);
1540 tun_qp->ring = NULL;
1544 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1545 enum ib_qp_type qp_type, int is_tun)
1548 struct mlx4_ib_demux_pv_qp *tun_qp;
1549 int rx_buf_size, tx_buf_size;
1551 if (qp_type > IB_QPT_GSI)
1554 tun_qp = &ctx->qp[qp_type];
1556 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1557 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1559 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1560 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1564 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1565 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1566 rx_buf_size, DMA_FROM_DEVICE);
1567 kfree(tun_qp->ring[i].addr);
1570 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1571 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1572 tx_buf_size, DMA_TO_DEVICE);
1573 kfree(tun_qp->tx_ring[i].buf.addr);
1574 if (tun_qp->tx_ring[i].ah)
1575 ib_destroy_ah(tun_qp->tx_ring[i].ah);
1577 kfree(tun_qp->tx_ring);
1578 kfree(tun_qp->ring);
1581 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1583 struct mlx4_ib_demux_pv_ctx *ctx;
1584 struct mlx4_ib_demux_pv_qp *tun_qp;
1587 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1588 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1590 while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1591 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1592 if (wc.status == IB_WC_SUCCESS) {
1593 switch (wc.opcode) {
1595 mlx4_ib_multiplex_mad(ctx, &wc);
1596 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1598 (MLX4_NUM_TUNNEL_BUFS - 1));
1600 pr_err("Failed reposting tunnel "
1601 "buf:%lld\n", wc.wr_id);
1604 pr_debug("received tunnel send completion:"
1605 "wrid=0x%llx, status=0x%x\n",
1606 wc.wr_id, wc.status);
1607 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1608 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1609 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1611 spin_lock(&tun_qp->tx_lock);
1612 tun_qp->tx_ix_tail++;
1613 spin_unlock(&tun_qp->tx_lock);
1620 pr_debug("mlx4_ib: completion error in tunnel: %d."
1621 " status = %d, wrid = 0x%llx\n",
1622 ctx->slave, wc.status, wc.wr_id);
1623 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1624 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1625 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1626 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1628 spin_lock(&tun_qp->tx_lock);
1629 tun_qp->tx_ix_tail++;
1630 spin_unlock(&tun_qp->tx_lock);
1636 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1638 struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1640 /* It's worse than that! He's dead, Jim! */
1641 pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1642 event->event, sqp->port);
1645 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1646 enum ib_qp_type qp_type, int create_tun)
1649 struct mlx4_ib_demux_pv_qp *tun_qp;
1650 struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1651 struct ib_qp_attr attr;
1652 int qp_attr_mask_INIT;
1654 if (qp_type > IB_QPT_GSI)
1657 tun_qp = &ctx->qp[qp_type];
1659 memset(&qp_init_attr, 0, sizeof qp_init_attr);
1660 qp_init_attr.init_attr.send_cq = ctx->cq;
1661 qp_init_attr.init_attr.recv_cq = ctx->cq;
1662 qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1663 qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1664 qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1665 qp_init_attr.init_attr.cap.max_send_sge = 1;
1666 qp_init_attr.init_attr.cap.max_recv_sge = 1;
1668 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1669 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1670 qp_init_attr.port = ctx->port;
1671 qp_init_attr.slave = ctx->slave;
1672 qp_init_attr.proxy_qp_type = qp_type;
1673 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1674 IB_QP_QKEY | IB_QP_PORT;
1676 qp_init_attr.init_attr.qp_type = qp_type;
1677 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1678 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1680 qp_init_attr.init_attr.port_num = ctx->port;
1681 qp_init_attr.init_attr.qp_context = ctx;
1682 qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1683 tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1684 if (IS_ERR(tun_qp->qp)) {
1685 ret = PTR_ERR(tun_qp->qp);
1687 pr_err("Couldn't create %s QP (%d)\n",
1688 create_tun ? "tunnel" : "special", ret);
1692 memset(&attr, 0, sizeof attr);
1693 attr.qp_state = IB_QPS_INIT;
1696 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1697 ctx->port, IB_DEFAULT_PKEY_FULL,
1699 if (ret || !create_tun)
1701 to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1702 attr.qkey = IB_QP1_QKEY;
1703 attr.port_num = ctx->port;
1704 ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1706 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1707 create_tun ? "tunnel" : "special", ret);
1710 attr.qp_state = IB_QPS_RTR;
1711 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1713 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1714 create_tun ? "tunnel" : "special", ret);
1717 attr.qp_state = IB_QPS_RTS;
1719 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1721 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1722 create_tun ? "tunnel" : "special", ret);
1726 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1727 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1729 pr_err(" mlx4_ib_post_pv_buf error"
1730 " (err = %d, i = %d)\n", ret, i);
1737 ib_destroy_qp(tun_qp->qp);
1743 * IB MAD completion callback for real SQPs
1745 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1747 struct mlx4_ib_demux_pv_ctx *ctx;
1748 struct mlx4_ib_demux_pv_qp *sqp;
1753 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1754 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1756 while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1757 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1758 if (wc.status == IB_WC_SUCCESS) {
1759 switch (wc.opcode) {
1761 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1762 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1763 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1765 spin_lock(&sqp->tx_lock);
1767 spin_unlock(&sqp->tx_lock);
1770 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1771 (sqp->ring[wc.wr_id &
1772 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1773 grh = &(((struct mlx4_mad_rcv_buf *)
1774 (sqp->ring[wc.wr_id &
1775 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1776 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1777 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1778 (MLX4_NUM_TUNNEL_BUFS - 1)))
1779 pr_err("Failed reposting SQP "
1780 "buf:%lld\n", wc.wr_id);
1787 pr_debug("mlx4_ib: completion error in tunnel: %d."
1788 " status = %d, wrid = 0x%llx\n",
1789 ctx->slave, wc.status, wc.wr_id);
1790 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1791 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1792 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1793 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1795 spin_lock(&sqp->tx_lock);
1797 spin_unlock(&sqp->tx_lock);
1803 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1804 struct mlx4_ib_demux_pv_ctx **ret_ctx)
1806 struct mlx4_ib_demux_pv_ctx *ctx;
1809 ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1811 pr_err("failed allocating pv resource context "
1812 "for port %d, slave %d\n", port, slave);
1816 ctx->ib_dev = &dev->ib_dev;
1823 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1825 if (dev->sriov.demux[port - 1].tun[slave]) {
1826 kfree(dev->sriov.demux[port - 1].tun[slave]);
1827 dev->sriov.demux[port - 1].tun[slave] = NULL;
1831 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1832 int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1835 struct ib_cq_init_attr cq_attr = {};
1837 if (ctx->state != DEMUX_PV_STATE_DOWN)
1840 ctx->state = DEMUX_PV_STATE_STARTING;
1841 /* have QP0 only if link layer is IB */
1842 if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1843 IB_LINK_LAYER_INFINIBAND)
1847 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1849 pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1854 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1856 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1860 cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1864 cq_attr.cqe = cq_size;
1865 ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1866 NULL, ctx, &cq_attr);
1867 if (IS_ERR(ctx->cq)) {
1868 ret = PTR_ERR(ctx->cq);
1869 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1873 ctx->pd = ib_alloc_pd(ctx->ib_dev);
1874 if (IS_ERR(ctx->pd)) {
1875 ret = PTR_ERR(ctx->pd);
1876 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1881 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1883 pr_err("Couldn't create %s QP0 (%d)\n",
1884 create_tun ? "tunnel for" : "", ret);
1889 ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1891 pr_err("Couldn't create %s QP1 (%d)\n",
1892 create_tun ? "tunnel for" : "", ret);
1897 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1899 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1901 ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1903 ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1905 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1908 ctx->state = DEMUX_PV_STATE_ACTIVE;
1913 ib_destroy_qp(ctx->qp[1].qp);
1914 ctx->qp[1].qp = NULL;
1919 ib_destroy_qp(ctx->qp[0].qp);
1920 ctx->qp[0].qp = NULL;
1923 ib_dealloc_pd(ctx->pd);
1927 ib_destroy_cq(ctx->cq);
1931 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1935 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1937 ctx->state = DEMUX_PV_STATE_DOWN;
1941 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1942 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1946 if (ctx->state > DEMUX_PV_STATE_DOWN) {
1947 ctx->state = DEMUX_PV_STATE_DOWNING;
1949 flush_workqueue(ctx->wq);
1951 ib_destroy_qp(ctx->qp[0].qp);
1952 ctx->qp[0].qp = NULL;
1953 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1955 ib_destroy_qp(ctx->qp[1].qp);
1956 ctx->qp[1].qp = NULL;
1957 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1958 ib_dealloc_pd(ctx->pd);
1960 ib_destroy_cq(ctx->cq);
1962 ctx->state = DEMUX_PV_STATE_DOWN;
1966 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1967 int port, int do_init)
1972 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1973 /* for master, destroy real sqp resources */
1974 if (slave == mlx4_master_func_num(dev->dev))
1975 destroy_pv_resources(dev, slave, port,
1976 dev->sriov.sqps[port - 1], 1);
1977 /* destroy the tunnel qp resources */
1978 destroy_pv_resources(dev, slave, port,
1979 dev->sriov.demux[port - 1].tun[slave], 1);
1983 /* create the tunnel qp resources */
1984 ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1985 dev->sriov.demux[port - 1].tun[slave]);
1987 /* for master, create the real sqp resources */
1988 if (!ret && slave == mlx4_master_func_num(dev->dev))
1989 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1990 dev->sriov.sqps[port - 1]);
1994 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1996 struct mlx4_ib_demux_work *dmxw;
1998 dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1999 mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2005 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2006 struct mlx4_ib_demux_ctx *ctx,
2013 ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2014 sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2020 ctx->ib_dev = &dev->ib_dev;
2023 i < min(dev->dev->caps.sqp_demux,
2024 (u16)(dev->dev->persist->num_vfs + 1));
2026 struct mlx4_active_ports actv_ports =
2027 mlx4_get_active_ports(dev->dev, i);
2029 if (!test_bit(port - 1, actv_ports.ports))
2032 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2039 ret = mlx4_ib_mcg_port_init(ctx);
2041 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2045 snprintf(name, sizeof name, "mlx4_ibt%d", port);
2046 ctx->wq = create_singlethread_workqueue(name);
2048 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2053 snprintf(name, sizeof name, "mlx4_ibud%d", port);
2054 ctx->ud_wq = create_singlethread_workqueue(name);
2056 pr_err("Failed to create up/down WQ for port %d\n", port);
2064 destroy_workqueue(ctx->wq);
2068 mlx4_ib_mcg_port_cleanup(ctx, 1);
2070 for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2071 free_pv_object(dev, i, port);
2077 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2079 if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2080 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2081 flush_workqueue(sqp_ctx->wq);
2082 if (sqp_ctx->has_smi) {
2083 ib_destroy_qp(sqp_ctx->qp[0].qp);
2084 sqp_ctx->qp[0].qp = NULL;
2085 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2087 ib_destroy_qp(sqp_ctx->qp[1].qp);
2088 sqp_ctx->qp[1].qp = NULL;
2089 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2090 ib_dealloc_pd(sqp_ctx->pd);
2092 ib_destroy_cq(sqp_ctx->cq);
2094 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2098 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2102 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2103 mlx4_ib_mcg_port_cleanup(ctx, 1);
2104 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2107 if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2108 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2110 flush_workqueue(ctx->wq);
2111 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2112 destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2113 free_pv_object(dev, i, ctx->port);
2116 destroy_workqueue(ctx->ud_wq);
2117 destroy_workqueue(ctx->wq);
2121 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2125 if (!mlx4_is_master(dev->dev))
2127 /* initialize or tear down tunnel QPs for the master */
2128 for (i = 0; i < dev->dev->caps.num_ports; i++)
2129 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2133 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2138 if (!mlx4_is_mfunc(dev->dev))
2141 dev->sriov.is_going_down = 0;
2142 spin_lock_init(&dev->sriov.going_down_lock);
2143 mlx4_ib_cm_paravirt_init(dev);
2145 mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2147 if (mlx4_is_slave(dev->dev)) {
2148 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2152 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2153 if (i == mlx4_master_func_num(dev->dev))
2154 mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2156 mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2159 err = mlx4_ib_init_alias_guid_service(dev);
2161 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2164 err = mlx4_ib_device_register_sysfs(dev);
2166 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2170 mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2171 dev->dev->caps.sqp_demux);
2172 for (i = 0; i < dev->num_ports; i++) {
2174 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2177 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2178 atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2179 be64_to_cpu(gid.global.subnet_prefix));
2180 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2181 &dev->sriov.sqps[i]);
2184 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2188 mlx4_ib_master_tunnels(dev, 1);
2192 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2195 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2196 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2198 mlx4_ib_device_unregister_sysfs(dev);
2201 mlx4_ib_destroy_alias_guid_service(dev);
2204 mlx4_ib_cm_paravirt_clean(dev, -1);
2209 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2212 unsigned long flags;
2214 if (!mlx4_is_mfunc(dev->dev))
2217 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2218 dev->sriov.is_going_down = 1;
2219 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2220 if (mlx4_is_master(dev->dev)) {
2221 for (i = 0; i < dev->num_ports; i++) {
2222 flush_workqueue(dev->sriov.demux[i].ud_wq);
2223 mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2224 kfree(dev->sriov.sqps[i]);
2225 dev->sriov.sqps[i] = NULL;
2226 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2229 mlx4_ib_cm_paravirt_clean(dev, -1);
2230 mlx4_ib_destroy_alias_guid_service(dev);
2231 mlx4_ib_device_unregister_sysfs(dev);