m68k/defconfig: Enable automounting of devtmpfs at /dev
[linux/fpc-iii.git] / drivers / infiniband / hw / mlx4 / mad.c
blob82a7dd87089b66efa0e925116b4d72053c341a3f
1 /*
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
12 * conditions are met:
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
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
30 * SOFTWARE.
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>
43 #include "mlx4_ib.h"
45 enum {
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 {
68 struct ib_grh grh;
69 u8 payload[256];
70 } __packed;
72 struct mlx4_mad_snd_buf {
73 u8 payload[256];
74 } __packed;
76 struct mlx4_tunnel_mad {
77 struct ib_grh grh;
78 struct mlx4_ib_tunnel_header hdr;
79 struct ib_mad mad;
80 } __packed;
82 struct mlx4_rcv_tunnel_mad {
83 struct mlx4_rcv_tunnel_hdr hdr;
84 struct ib_grh grh;
85 struct ib_mad mad;
86 } __packed;
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, struct ib_wc *in_wc, struct ib_grh *in_grh,
107 void *in_mad, void *response_mad)
109 struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
110 void *inbox;
111 int err;
112 u32 in_modifier = port;
113 u8 op_modifier = 0;
115 inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
116 if (IS_ERR(inmailbox))
117 return PTR_ERR(inmailbox);
118 inbox = inmailbox->buf;
120 outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
121 if (IS_ERR(outmailbox)) {
122 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
123 return PTR_ERR(outmailbox);
126 memcpy(inbox, in_mad, 256);
129 * Key check traps can't be generated unless we have in_wc to
130 * tell us where to send the trap.
132 if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
133 op_modifier |= 0x1;
134 if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
135 op_modifier |= 0x2;
136 if (mlx4_is_mfunc(dev->dev) &&
137 (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
138 op_modifier |= 0x8;
140 if (in_wc) {
141 struct {
142 __be32 my_qpn;
143 u32 reserved1;
144 __be32 rqpn;
145 u8 sl;
146 u8 g_path;
147 u16 reserved2[2];
148 __be16 pkey;
149 u32 reserved3[11];
150 u8 grh[40];
151 } *ext_info;
153 memset(inbox + 256, 0, 256);
154 ext_info = inbox + 256;
156 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
157 ext_info->rqpn = cpu_to_be32(in_wc->src_qp);
158 ext_info->sl = in_wc->sl << 4;
159 ext_info->g_path = in_wc->dlid_path_bits |
160 (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
161 ext_info->pkey = cpu_to_be16(in_wc->pkey_index);
163 if (in_grh)
164 memcpy(ext_info->grh, in_grh, 40);
166 op_modifier |= 0x4;
168 in_modifier |= in_wc->slid << 16;
171 err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
172 mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
173 MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
174 (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
176 if (!err)
177 memcpy(response_mad, outmailbox->buf, 256);
179 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
180 mlx4_free_cmd_mailbox(dev->dev, outmailbox);
182 return err;
185 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
187 struct ib_ah *new_ah;
188 struct ib_ah_attr ah_attr;
189 unsigned long flags;
191 if (!dev->send_agent[port_num - 1][0])
192 return;
194 memset(&ah_attr, 0, sizeof ah_attr);
195 ah_attr.dlid = lid;
196 ah_attr.sl = sl;
197 ah_attr.port_num = port_num;
199 new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
200 &ah_attr);
201 if (IS_ERR(new_ah))
202 return;
204 spin_lock_irqsave(&dev->sm_lock, flags);
205 if (dev->sm_ah[port_num - 1])
206 ib_destroy_ah(dev->sm_ah[port_num - 1]);
207 dev->sm_ah[port_num - 1] = new_ah;
208 spin_unlock_irqrestore(&dev->sm_lock, flags);
212 * Snoop SM MADs for port info, GUID info, and P_Key table sets, so we can
213 * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
215 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
216 u16 prev_lid)
218 struct ib_port_info *pinfo;
219 u16 lid;
220 __be16 *base;
221 u32 bn, pkey_change_bitmap;
222 int i;
225 struct mlx4_ib_dev *dev = to_mdev(ibdev);
226 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
227 mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
228 mad->mad_hdr.method == IB_MGMT_METHOD_SET)
229 switch (mad->mad_hdr.attr_id) {
230 case IB_SMP_ATTR_PORT_INFO:
231 pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
232 lid = be16_to_cpu(pinfo->lid);
234 update_sm_ah(dev, port_num,
235 be16_to_cpu(pinfo->sm_lid),
236 pinfo->neighbormtu_mastersmsl & 0xf);
238 if (pinfo->clientrereg_resv_subnetto & 0x80)
239 handle_client_rereg_event(dev, port_num);
241 if (prev_lid != lid)
242 handle_lid_change_event(dev, port_num);
243 break;
245 case IB_SMP_ATTR_PKEY_TABLE:
246 if (!mlx4_is_mfunc(dev->dev)) {
247 mlx4_ib_dispatch_event(dev, port_num,
248 IB_EVENT_PKEY_CHANGE);
249 break;
252 /* at this point, we are running in the master.
253 * Slaves do not receive SMPs.
255 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
256 base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
257 pkey_change_bitmap = 0;
258 for (i = 0; i < 32; i++) {
259 pr_debug("PKEY[%d] = x%x\n",
260 i + bn*32, be16_to_cpu(base[i]));
261 if (be16_to_cpu(base[i]) !=
262 dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
263 pkey_change_bitmap |= (1 << i);
264 dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
265 be16_to_cpu(base[i]);
268 pr_debug("PKEY Change event: port=%d, "
269 "block=0x%x, change_bitmap=0x%x\n",
270 port_num, bn, pkey_change_bitmap);
272 if (pkey_change_bitmap) {
273 mlx4_ib_dispatch_event(dev, port_num,
274 IB_EVENT_PKEY_CHANGE);
275 if (!dev->sriov.is_going_down)
276 __propagate_pkey_ev(dev, port_num, bn,
277 pkey_change_bitmap);
279 break;
281 case IB_SMP_ATTR_GUID_INFO:
282 /* paravirtualized master's guid is guid 0 -- does not change */
283 if (!mlx4_is_master(dev->dev))
284 mlx4_ib_dispatch_event(dev, port_num,
285 IB_EVENT_GID_CHANGE);
286 /*if master, notify relevant slaves*/
287 if (mlx4_is_master(dev->dev) &&
288 !dev->sriov.is_going_down) {
289 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
290 mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
291 (u8 *)(&((struct ib_smp *)mad)->data));
292 mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
293 (u8 *)(&((struct ib_smp *)mad)->data));
295 break;
297 default:
298 break;
302 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
303 int block, u32 change_bitmap)
305 int i, ix, slave, err;
306 int have_event = 0;
308 for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
309 if (slave == mlx4_master_func_num(dev->dev))
310 continue;
311 if (!mlx4_is_slave_active(dev->dev, slave))
312 continue;
314 have_event = 0;
315 for (i = 0; i < 32; i++) {
316 if (!(change_bitmap & (1 << i)))
317 continue;
318 for (ix = 0;
319 ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
320 if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
321 [ix] == i + 32 * block) {
322 err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
323 pr_debug("propagate_pkey_ev: slave %d,"
324 " port %d, ix %d (%d)\n",
325 slave, port_num, ix, err);
326 have_event = 1;
327 break;
330 if (have_event)
331 break;
336 static void node_desc_override(struct ib_device *dev,
337 struct ib_mad *mad)
339 unsigned long flags;
341 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
342 mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
343 mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
344 mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
345 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
346 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
347 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
351 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
353 int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
354 struct ib_mad_send_buf *send_buf;
355 struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
356 int ret;
357 unsigned long flags;
359 if (agent) {
360 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
361 IB_MGMT_MAD_DATA, GFP_ATOMIC);
362 if (IS_ERR(send_buf))
363 return;
365 * We rely here on the fact that MLX QPs don't use the
366 * address handle after the send is posted (this is
367 * wrong following the IB spec strictly, but we know
368 * it's OK for our devices).
370 spin_lock_irqsave(&dev->sm_lock, flags);
371 memcpy(send_buf->mad, mad, sizeof *mad);
372 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
373 ret = ib_post_send_mad(send_buf, NULL);
374 else
375 ret = -EINVAL;
376 spin_unlock_irqrestore(&dev->sm_lock, flags);
378 if (ret)
379 ib_free_send_mad(send_buf);
383 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
384 struct ib_sa_mad *sa_mad)
386 int ret = 0;
388 /* dispatch to different sa handlers */
389 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
390 case IB_SA_ATTR_MC_MEMBER_REC:
391 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
392 break;
393 default:
394 break;
396 return ret;
399 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
401 struct mlx4_ib_dev *dev = to_mdev(ibdev);
402 int i;
404 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
405 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
406 return i;
408 return -1;
412 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
413 u8 port, u16 pkey, u16 *ix)
415 int i, ret;
416 u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
417 u16 slot_pkey;
419 if (slave == mlx4_master_func_num(dev->dev))
420 return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
422 unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
424 for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
425 if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
426 continue;
428 pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
430 ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
431 if (ret)
432 continue;
433 if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
434 if (slot_pkey & 0x8000) {
435 *ix = (u16) pkey_ix;
436 return 0;
437 } else {
438 /* take first partial pkey index found */
439 if (partial_ix == 0xFF)
440 partial_ix = pkey_ix;
445 if (partial_ix < 0xFF) {
446 *ix = (u16) partial_ix;
447 return 0;
450 return -EINVAL;
453 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
454 enum ib_qp_type dest_qpt, struct ib_wc *wc,
455 struct ib_grh *grh, struct ib_mad *mad)
457 struct ib_sge list;
458 struct ib_send_wr wr, *bad_wr;
459 struct mlx4_ib_demux_pv_ctx *tun_ctx;
460 struct mlx4_ib_demux_pv_qp *tun_qp;
461 struct mlx4_rcv_tunnel_mad *tun_mad;
462 struct ib_ah_attr attr;
463 struct ib_ah *ah;
464 struct ib_qp *src_qp = NULL;
465 unsigned tun_tx_ix = 0;
466 int dqpn;
467 int ret = 0;
468 u16 tun_pkey_ix;
469 u16 cached_pkey;
470 u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
472 if (dest_qpt > IB_QPT_GSI)
473 return -EINVAL;
475 tun_ctx = dev->sriov.demux[port-1].tun[slave];
477 /* check if proxy qp created */
478 if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
479 return -EAGAIN;
481 if (!dest_qpt)
482 tun_qp = &tun_ctx->qp[0];
483 else
484 tun_qp = &tun_ctx->qp[1];
486 /* compute P_Key index to put in tunnel header for slave */
487 if (dest_qpt) {
488 u16 pkey_ix;
489 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
490 if (ret)
491 return -EINVAL;
493 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
494 if (ret)
495 return -EINVAL;
496 tun_pkey_ix = pkey_ix;
497 } else
498 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
500 dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
502 /* get tunnel tx data buf for slave */
503 src_qp = tun_qp->qp;
505 /* create ah. Just need an empty one with the port num for the post send.
506 * The driver will set the force loopback bit in post_send */
507 memset(&attr, 0, sizeof attr);
508 attr.port_num = port;
509 if (is_eth) {
510 memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
511 attr.ah_flags = IB_AH_GRH;
513 ah = ib_create_ah(tun_ctx->pd, &attr);
514 if (IS_ERR(ah))
515 return -ENOMEM;
517 /* allocate tunnel tx buf after pass failure returns */
518 spin_lock(&tun_qp->tx_lock);
519 if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
520 (MLX4_NUM_TUNNEL_BUFS - 1))
521 ret = -EAGAIN;
522 else
523 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
524 spin_unlock(&tun_qp->tx_lock);
525 if (ret)
526 goto out;
528 tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
529 if (tun_qp->tx_ring[tun_tx_ix].ah)
530 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
531 tun_qp->tx_ring[tun_tx_ix].ah = ah;
532 ib_dma_sync_single_for_cpu(&dev->ib_dev,
533 tun_qp->tx_ring[tun_tx_ix].buf.map,
534 sizeof (struct mlx4_rcv_tunnel_mad),
535 DMA_TO_DEVICE);
537 /* copy over to tunnel buffer */
538 if (grh)
539 memcpy(&tun_mad->grh, grh, sizeof *grh);
540 memcpy(&tun_mad->mad, mad, sizeof *mad);
542 /* adjust tunnel data */
543 tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
544 tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
545 tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
547 if (is_eth) {
548 u16 vlan = 0;
549 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
550 NULL)) {
551 /* VST mode */
552 if (vlan != wc->vlan_id)
553 /* Packet vlan is not the VST-assigned vlan.
554 * Drop the packet.
556 goto out;
557 else
558 /* Remove the vlan tag before forwarding
559 * the packet to the VF.
561 vlan = 0xffff;
562 } else {
563 vlan = wc->vlan_id;
566 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
567 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
568 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
569 } else {
570 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
571 tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
574 ib_dma_sync_single_for_device(&dev->ib_dev,
575 tun_qp->tx_ring[tun_tx_ix].buf.map,
576 sizeof (struct mlx4_rcv_tunnel_mad),
577 DMA_TO_DEVICE);
579 list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
580 list.length = sizeof (struct mlx4_rcv_tunnel_mad);
581 list.lkey = tun_ctx->mr->lkey;
583 wr.wr.ud.ah = ah;
584 wr.wr.ud.port_num = port;
585 wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
586 wr.wr.ud.remote_qpn = dqpn;
587 wr.next = NULL;
588 wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
589 wr.sg_list = &list;
590 wr.num_sge = 1;
591 wr.opcode = IB_WR_SEND;
592 wr.send_flags = IB_SEND_SIGNALED;
594 ret = ib_post_send(src_qp, &wr, &bad_wr);
595 out:
596 if (ret)
597 ib_destroy_ah(ah);
598 return ret;
601 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
602 struct ib_wc *wc, struct ib_grh *grh,
603 struct ib_mad *mad)
605 struct mlx4_ib_dev *dev = to_mdev(ibdev);
606 int err;
607 int slave;
608 u8 *slave_id;
609 int is_eth = 0;
611 if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
612 is_eth = 0;
613 else
614 is_eth = 1;
616 if (is_eth) {
617 if (!(wc->wc_flags & IB_WC_GRH)) {
618 mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
619 return -EINVAL;
621 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
622 mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
623 return -EINVAL;
625 if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
626 mlx4_ib_warn(ibdev, "failed matching grh\n");
627 return -ENOENT;
629 if (slave >= dev->dev->caps.sqp_demux) {
630 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
631 slave, dev->dev->caps.sqp_demux);
632 return -ENOENT;
635 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
636 return 0;
638 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
639 if (err)
640 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
641 slave, err);
642 return 0;
645 /* Initially assume that this mad is for us */
646 slave = mlx4_master_func_num(dev->dev);
648 /* See if the slave id is encoded in a response mad */
649 if (mad->mad_hdr.method & 0x80) {
650 slave_id = (u8 *) &mad->mad_hdr.tid;
651 slave = *slave_id;
652 if (slave != 255) /*255 indicates the dom0*/
653 *slave_id = 0; /* remap tid */
656 /* If a grh is present, we demux according to it */
657 if (wc->wc_flags & IB_WC_GRH) {
658 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
659 if (slave < 0) {
660 mlx4_ib_warn(ibdev, "failed matching grh\n");
661 return -ENOENT;
664 /* Class-specific handling */
665 switch (mad->mad_hdr.mgmt_class) {
666 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
667 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
668 /* 255 indicates the dom0 */
669 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
670 if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
671 return -EPERM;
672 /* for a VF. drop unsolicited MADs */
673 if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
674 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
675 slave, mad->mad_hdr.mgmt_class,
676 mad->mad_hdr.method);
677 return -EINVAL;
680 break;
681 case IB_MGMT_CLASS_SUBN_ADM:
682 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
683 (struct ib_sa_mad *) mad))
684 return 0;
685 break;
686 case IB_MGMT_CLASS_CM:
687 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
688 return 0;
689 break;
690 case IB_MGMT_CLASS_DEVICE_MGMT:
691 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
692 return 0;
693 break;
694 default:
695 /* Drop unsupported classes for slaves in tunnel mode */
696 if (slave != mlx4_master_func_num(dev->dev)) {
697 pr_debug("dropping unsupported ingress mad from class:%d "
698 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
699 return 0;
702 /*make sure that no slave==255 was not handled yet.*/
703 if (slave >= dev->dev->caps.sqp_demux) {
704 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
705 slave, dev->dev->caps.sqp_demux);
706 return -ENOENT;
709 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
710 if (err)
711 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
712 slave, err);
713 return 0;
716 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
717 struct ib_wc *in_wc, struct ib_grh *in_grh,
718 struct ib_mad *in_mad, struct ib_mad *out_mad)
720 u16 slid, prev_lid = 0;
721 int err;
722 struct ib_port_attr pattr;
724 if (in_wc && in_wc->qp->qp_num) {
725 pr_debug("received MAD: slid:%d sqpn:%d "
726 "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
727 in_wc->slid, in_wc->src_qp,
728 in_wc->dlid_path_bits,
729 in_wc->qp->qp_num,
730 in_wc->wc_flags,
731 in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
732 be16_to_cpu(in_mad->mad_hdr.attr_id));
733 if (in_wc->wc_flags & IB_WC_GRH) {
734 pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
735 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
736 be64_to_cpu(in_grh->sgid.global.interface_id));
737 pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
738 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
739 be64_to_cpu(in_grh->dgid.global.interface_id));
743 slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
745 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
746 forward_trap(to_mdev(ibdev), port_num, in_mad);
747 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
750 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
751 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
752 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
753 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET &&
754 in_mad->mad_hdr.method != IB_MGMT_METHOD_TRAP_REPRESS)
755 return IB_MAD_RESULT_SUCCESS;
758 * Don't process SMInfo queries -- the SMA can't handle them.
760 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
761 return IB_MAD_RESULT_SUCCESS;
762 } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
763 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1 ||
764 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2 ||
765 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
766 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
767 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET)
768 return IB_MAD_RESULT_SUCCESS;
769 } else
770 return IB_MAD_RESULT_SUCCESS;
772 if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
773 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
774 in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
775 in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
776 !ib_query_port(ibdev, port_num, &pattr))
777 prev_lid = pattr.lid;
779 err = mlx4_MAD_IFC(to_mdev(ibdev),
780 (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
781 (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
782 MLX4_MAD_IFC_NET_VIEW,
783 port_num, in_wc, in_grh, in_mad, out_mad);
784 if (err)
785 return IB_MAD_RESULT_FAILURE;
787 if (!out_mad->mad_hdr.status) {
788 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
789 smp_snoop(ibdev, port_num, in_mad, prev_lid);
790 /* slaves get node desc from FW */
791 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
792 node_desc_override(ibdev, out_mad);
795 /* set return bit in status of directed route responses */
796 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
797 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
799 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
800 /* no response for trap repress */
801 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
803 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
806 static void edit_counter(struct mlx4_counter *cnt,
807 struct ib_pma_portcounters *pma_cnt)
809 pma_cnt->port_xmit_data = cpu_to_be32((be64_to_cpu(cnt->tx_bytes)>>2));
810 pma_cnt->port_rcv_data = cpu_to_be32((be64_to_cpu(cnt->rx_bytes)>>2));
811 pma_cnt->port_xmit_packets = cpu_to_be32(be64_to_cpu(cnt->tx_frames));
812 pma_cnt->port_rcv_packets = cpu_to_be32(be64_to_cpu(cnt->rx_frames));
815 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
816 struct ib_wc *in_wc, struct ib_grh *in_grh,
817 struct ib_mad *in_mad, struct ib_mad *out_mad)
819 struct mlx4_cmd_mailbox *mailbox;
820 struct mlx4_ib_dev *dev = to_mdev(ibdev);
821 int err;
822 u32 inmod = dev->counters[port_num - 1] & 0xffff;
823 u8 mode;
825 if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
826 return -EINVAL;
828 mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
829 if (IS_ERR(mailbox))
830 return IB_MAD_RESULT_FAILURE;
832 err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
833 MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
834 MLX4_CMD_WRAPPED);
835 if (err)
836 err = IB_MAD_RESULT_FAILURE;
837 else {
838 memset(out_mad->data, 0, sizeof out_mad->data);
839 mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
840 switch (mode & 0xf) {
841 case 0:
842 edit_counter(mailbox->buf,
843 (void *)(out_mad->data + 40));
844 err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
845 break;
846 default:
847 err = IB_MAD_RESULT_FAILURE;
851 mlx4_free_cmd_mailbox(dev->dev, mailbox);
853 return err;
856 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
857 struct ib_wc *in_wc, struct ib_grh *in_grh,
858 struct ib_mad *in_mad, struct ib_mad *out_mad)
860 switch (rdma_port_get_link_layer(ibdev, port_num)) {
861 case IB_LINK_LAYER_INFINIBAND:
862 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
863 in_grh, in_mad, out_mad);
864 case IB_LINK_LAYER_ETHERNET:
865 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
866 in_grh, in_mad, out_mad);
867 default:
868 return -EINVAL;
872 static void send_handler(struct ib_mad_agent *agent,
873 struct ib_mad_send_wc *mad_send_wc)
875 if (mad_send_wc->send_buf->context[0])
876 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
877 ib_free_send_mad(mad_send_wc->send_buf);
880 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
882 struct ib_mad_agent *agent;
883 int p, q;
884 int ret;
885 enum rdma_link_layer ll;
887 for (p = 0; p < dev->num_ports; ++p) {
888 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
889 for (q = 0; q <= 1; ++q) {
890 if (ll == IB_LINK_LAYER_INFINIBAND) {
891 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
892 q ? IB_QPT_GSI : IB_QPT_SMI,
893 NULL, 0, send_handler,
894 NULL, NULL, 0);
895 if (IS_ERR(agent)) {
896 ret = PTR_ERR(agent);
897 goto err;
899 dev->send_agent[p][q] = agent;
900 } else
901 dev->send_agent[p][q] = NULL;
905 return 0;
907 err:
908 for (p = 0; p < dev->num_ports; ++p)
909 for (q = 0; q <= 1; ++q)
910 if (dev->send_agent[p][q])
911 ib_unregister_mad_agent(dev->send_agent[p][q]);
913 return ret;
916 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
918 struct ib_mad_agent *agent;
919 int p, q;
921 for (p = 0; p < dev->num_ports; ++p) {
922 for (q = 0; q <= 1; ++q) {
923 agent = dev->send_agent[p][q];
924 if (agent) {
925 dev->send_agent[p][q] = NULL;
926 ib_unregister_mad_agent(agent);
930 if (dev->sm_ah[p])
931 ib_destroy_ah(dev->sm_ah[p]);
935 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
937 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
939 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
940 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
941 MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
944 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
946 /* re-configure the alias-guid and mcg's */
947 if (mlx4_is_master(dev->dev)) {
948 mlx4_ib_invalidate_all_guid_record(dev, port_num);
950 if (!dev->sriov.is_going_down) {
951 mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
952 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
953 MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
956 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
959 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
960 struct mlx4_eqe *eqe)
962 __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
963 GET_MASK_FROM_EQE(eqe));
966 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
967 u32 guid_tbl_blk_num, u32 change_bitmap)
969 struct ib_smp *in_mad = NULL;
970 struct ib_smp *out_mad = NULL;
971 u16 i;
973 if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
974 return;
976 in_mad = kmalloc(sizeof *in_mad, GFP_KERNEL);
977 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
978 if (!in_mad || !out_mad) {
979 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
980 goto out;
983 guid_tbl_blk_num *= 4;
985 for (i = 0; i < 4; i++) {
986 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
987 continue;
988 memset(in_mad, 0, sizeof *in_mad);
989 memset(out_mad, 0, sizeof *out_mad);
991 in_mad->base_version = 1;
992 in_mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
993 in_mad->class_version = 1;
994 in_mad->method = IB_MGMT_METHOD_GET;
995 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
996 in_mad->attr_mod = cpu_to_be32(guid_tbl_blk_num + i);
998 if (mlx4_MAD_IFC(dev,
999 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1000 port_num, NULL, NULL, in_mad, out_mad)) {
1001 mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1002 goto out;
1005 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1006 port_num,
1007 (u8 *)(&((struct ib_smp *)out_mad)->data));
1008 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1009 port_num,
1010 (u8 *)(&((struct ib_smp *)out_mad)->data));
1013 out:
1014 kfree(in_mad);
1015 kfree(out_mad);
1016 return;
1019 void handle_port_mgmt_change_event(struct work_struct *work)
1021 struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1022 struct mlx4_ib_dev *dev = ew->ib_dev;
1023 struct mlx4_eqe *eqe = &(ew->ib_eqe);
1024 u8 port = eqe->event.port_mgmt_change.port;
1025 u32 changed_attr;
1026 u32 tbl_block;
1027 u32 change_bitmap;
1029 switch (eqe->subtype) {
1030 case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1031 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1033 /* Update the SM ah - This should be done before handling
1034 the other changed attributes so that MADs can be sent to the SM */
1035 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1036 u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1037 u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1038 update_sm_ah(dev, port, lid, sl);
1041 /* Check if it is a lid change event */
1042 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1043 handle_lid_change_event(dev, port);
1045 /* Generate GUID changed event */
1046 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1047 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1048 /*if master, notify all slaves*/
1049 if (mlx4_is_master(dev->dev))
1050 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1051 MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1054 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1055 handle_client_rereg_event(dev, port);
1056 break;
1058 case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1059 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1060 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1061 propagate_pkey_ev(dev, port, eqe);
1062 break;
1063 case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1064 /* paravirtualized master's guid is guid 0 -- does not change */
1065 if (!mlx4_is_master(dev->dev))
1066 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1067 /*if master, notify relevant slaves*/
1068 else if (!dev->sriov.is_going_down) {
1069 tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1070 change_bitmap = GET_MASK_FROM_EQE(eqe);
1071 handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1073 break;
1074 default:
1075 pr_warn("Unsupported subtype 0x%x for "
1076 "Port Management Change event\n", eqe->subtype);
1079 kfree(ew);
1082 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1083 enum ib_event_type type)
1085 struct ib_event event;
1087 event.device = &dev->ib_dev;
1088 event.element.port_num = port_num;
1089 event.event = type;
1091 ib_dispatch_event(&event);
1094 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1096 unsigned long flags;
1097 struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1098 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1099 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1100 if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1101 queue_work(ctx->wq, &ctx->work);
1102 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1105 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1106 struct mlx4_ib_demux_pv_qp *tun_qp,
1107 int index)
1109 struct ib_sge sg_list;
1110 struct ib_recv_wr recv_wr, *bad_recv_wr;
1111 int size;
1113 size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1114 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1116 sg_list.addr = tun_qp->ring[index].map;
1117 sg_list.length = size;
1118 sg_list.lkey = ctx->mr->lkey;
1120 recv_wr.next = NULL;
1121 recv_wr.sg_list = &sg_list;
1122 recv_wr.num_sge = 1;
1123 recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1124 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1125 ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1126 size, DMA_FROM_DEVICE);
1127 return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1130 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1131 int slave, struct ib_sa_mad *sa_mad)
1133 int ret = 0;
1135 /* dispatch to different sa handlers */
1136 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1137 case IB_SA_ATTR_MC_MEMBER_REC:
1138 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1139 break;
1140 default:
1141 break;
1143 return ret;
1146 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1148 int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1150 return (qpn >= proxy_start && qpn <= proxy_start + 1);
1154 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1155 enum ib_qp_type dest_qpt, u16 pkey_index,
1156 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1157 u8 *s_mac, struct ib_mad *mad)
1159 struct ib_sge list;
1160 struct ib_send_wr wr, *bad_wr;
1161 struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1162 struct mlx4_ib_demux_pv_qp *sqp;
1163 struct mlx4_mad_snd_buf *sqp_mad;
1164 struct ib_ah *ah;
1165 struct ib_qp *send_qp = NULL;
1166 unsigned wire_tx_ix = 0;
1167 int ret = 0;
1168 u16 wire_pkey_ix;
1169 int src_qpnum;
1170 u8 sgid_index;
1173 sqp_ctx = dev->sriov.sqps[port-1];
1175 /* check if proxy qp created */
1176 if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1177 return -EAGAIN;
1179 if (dest_qpt == IB_QPT_SMI) {
1180 src_qpnum = 0;
1181 sqp = &sqp_ctx->qp[0];
1182 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1183 } else {
1184 src_qpnum = 1;
1185 sqp = &sqp_ctx->qp[1];
1186 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1189 send_qp = sqp->qp;
1191 /* create ah */
1192 sgid_index = attr->grh.sgid_index;
1193 attr->grh.sgid_index = 0;
1194 ah = ib_create_ah(sqp_ctx->pd, attr);
1195 if (IS_ERR(ah))
1196 return -ENOMEM;
1197 attr->grh.sgid_index = sgid_index;
1198 to_mah(ah)->av.ib.gid_index = sgid_index;
1199 /* get rid of force-loopback bit */
1200 to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1201 spin_lock(&sqp->tx_lock);
1202 if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1203 (MLX4_NUM_TUNNEL_BUFS - 1))
1204 ret = -EAGAIN;
1205 else
1206 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1207 spin_unlock(&sqp->tx_lock);
1208 if (ret)
1209 goto out;
1211 sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1212 if (sqp->tx_ring[wire_tx_ix].ah)
1213 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1214 sqp->tx_ring[wire_tx_ix].ah = ah;
1215 ib_dma_sync_single_for_cpu(&dev->ib_dev,
1216 sqp->tx_ring[wire_tx_ix].buf.map,
1217 sizeof (struct mlx4_mad_snd_buf),
1218 DMA_TO_DEVICE);
1220 memcpy(&sqp_mad->payload, mad, sizeof *mad);
1222 ib_dma_sync_single_for_device(&dev->ib_dev,
1223 sqp->tx_ring[wire_tx_ix].buf.map,
1224 sizeof (struct mlx4_mad_snd_buf),
1225 DMA_TO_DEVICE);
1227 list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1228 list.length = sizeof (struct mlx4_mad_snd_buf);
1229 list.lkey = sqp_ctx->mr->lkey;
1231 wr.wr.ud.ah = ah;
1232 wr.wr.ud.port_num = port;
1233 wr.wr.ud.pkey_index = wire_pkey_ix;
1234 wr.wr.ud.remote_qkey = qkey;
1235 wr.wr.ud.remote_qpn = remote_qpn;
1236 wr.next = NULL;
1237 wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1238 wr.sg_list = &list;
1239 wr.num_sge = 1;
1240 wr.opcode = IB_WR_SEND;
1241 wr.send_flags = IB_SEND_SIGNALED;
1242 if (s_mac)
1243 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1246 ret = ib_post_send(send_qp, &wr, &bad_wr);
1247 out:
1248 if (ret)
1249 ib_destroy_ah(ah);
1250 return ret;
1253 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1255 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1256 return slave;
1257 return mlx4_get_base_gid_ix(dev->dev, slave, port);
1260 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1261 struct ib_ah_attr *ah_attr)
1263 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1264 ah_attr->grh.sgid_index = slave;
1265 else
1266 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1269 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1271 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1272 struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1273 int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1274 struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1275 struct mlx4_ib_ah ah;
1276 struct ib_ah_attr ah_attr;
1277 u8 *slave_id;
1278 int slave;
1279 int port;
1281 /* Get slave that sent this packet */
1282 if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1283 wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1284 (wc->src_qp & 0x1) != ctx->port - 1 ||
1285 wc->src_qp & 0x4) {
1286 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1287 return;
1289 slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1290 if (slave != ctx->slave) {
1291 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1292 "belongs to another slave\n", wc->src_qp);
1293 return;
1296 /* Map transaction ID */
1297 ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1298 sizeof (struct mlx4_tunnel_mad),
1299 DMA_FROM_DEVICE);
1300 switch (tunnel->mad.mad_hdr.method) {
1301 case IB_MGMT_METHOD_SET:
1302 case IB_MGMT_METHOD_GET:
1303 case IB_MGMT_METHOD_REPORT:
1304 case IB_SA_METHOD_GET_TABLE:
1305 case IB_SA_METHOD_DELETE:
1306 case IB_SA_METHOD_GET_MULTI:
1307 case IB_SA_METHOD_GET_TRACE_TBL:
1308 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1309 if (*slave_id) {
1310 mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1311 "class:%d slave:%d\n", *slave_id,
1312 tunnel->mad.mad_hdr.mgmt_class, slave);
1313 return;
1314 } else
1315 *slave_id = slave;
1316 default:
1317 /* nothing */;
1320 /* Class-specific handling */
1321 switch (tunnel->mad.mad_hdr.mgmt_class) {
1322 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1323 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1324 if (slave != mlx4_master_func_num(dev->dev) &&
1325 !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1326 return;
1327 break;
1328 case IB_MGMT_CLASS_SUBN_ADM:
1329 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1330 (struct ib_sa_mad *) &tunnel->mad))
1331 return;
1332 break;
1333 case IB_MGMT_CLASS_CM:
1334 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1335 (struct ib_mad *) &tunnel->mad))
1336 return;
1337 break;
1338 case IB_MGMT_CLASS_DEVICE_MGMT:
1339 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1340 tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1341 return;
1342 break;
1343 default:
1344 /* Drop unsupported classes for slaves in tunnel mode */
1345 if (slave != mlx4_master_func_num(dev->dev)) {
1346 mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1347 "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1348 return;
1352 /* We are using standard ib_core services to send the mad, so generate a
1353 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1354 memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1355 ah.ibah.device = ctx->ib_dev;
1356 mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1357 if (ah_attr.ah_flags & IB_AH_GRH)
1358 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1360 port = mlx4_slave_convert_port(dev->dev, slave, ah_attr.port_num);
1361 if (port < 0)
1362 return;
1363 ah_attr.port_num = port;
1364 memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1365 ah_attr.vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1366 /* if slave have default vlan use it */
1367 mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1368 &ah_attr.vlan_id, &ah_attr.sl);
1370 mlx4_ib_send_to_wire(dev, slave, ctx->port,
1371 is_proxy_qp0(dev, wc->src_qp, slave) ?
1372 IB_QPT_SMI : IB_QPT_GSI,
1373 be16_to_cpu(tunnel->hdr.pkey_index),
1374 be32_to_cpu(tunnel->hdr.remote_qpn),
1375 be32_to_cpu(tunnel->hdr.qkey),
1376 &ah_attr, wc->smac, &tunnel->mad);
1379 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1380 enum ib_qp_type qp_type, int is_tun)
1382 int i;
1383 struct mlx4_ib_demux_pv_qp *tun_qp;
1384 int rx_buf_size, tx_buf_size;
1386 if (qp_type > IB_QPT_GSI)
1387 return -EINVAL;
1389 tun_qp = &ctx->qp[qp_type];
1391 tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1392 GFP_KERNEL);
1393 if (!tun_qp->ring)
1394 return -ENOMEM;
1396 tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1397 sizeof (struct mlx4_ib_tun_tx_buf),
1398 GFP_KERNEL);
1399 if (!tun_qp->tx_ring) {
1400 kfree(tun_qp->ring);
1401 tun_qp->ring = NULL;
1402 return -ENOMEM;
1405 if (is_tun) {
1406 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1407 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1408 } else {
1409 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1410 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1413 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1414 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1415 if (!tun_qp->ring[i].addr)
1416 goto err;
1417 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1418 tun_qp->ring[i].addr,
1419 rx_buf_size,
1420 DMA_FROM_DEVICE);
1423 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1424 tun_qp->tx_ring[i].buf.addr =
1425 kmalloc(tx_buf_size, GFP_KERNEL);
1426 if (!tun_qp->tx_ring[i].buf.addr)
1427 goto tx_err;
1428 tun_qp->tx_ring[i].buf.map =
1429 ib_dma_map_single(ctx->ib_dev,
1430 tun_qp->tx_ring[i].buf.addr,
1431 tx_buf_size,
1432 DMA_TO_DEVICE);
1433 tun_qp->tx_ring[i].ah = NULL;
1435 spin_lock_init(&tun_qp->tx_lock);
1436 tun_qp->tx_ix_head = 0;
1437 tun_qp->tx_ix_tail = 0;
1438 tun_qp->proxy_qpt = qp_type;
1440 return 0;
1442 tx_err:
1443 while (i > 0) {
1444 --i;
1445 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1446 tx_buf_size, DMA_TO_DEVICE);
1447 kfree(tun_qp->tx_ring[i].buf.addr);
1449 kfree(tun_qp->tx_ring);
1450 tun_qp->tx_ring = NULL;
1451 i = MLX4_NUM_TUNNEL_BUFS;
1452 err:
1453 while (i > 0) {
1454 --i;
1455 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1456 rx_buf_size, DMA_FROM_DEVICE);
1457 kfree(tun_qp->ring[i].addr);
1459 kfree(tun_qp->ring);
1460 tun_qp->ring = NULL;
1461 return -ENOMEM;
1464 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1465 enum ib_qp_type qp_type, int is_tun)
1467 int i;
1468 struct mlx4_ib_demux_pv_qp *tun_qp;
1469 int rx_buf_size, tx_buf_size;
1471 if (qp_type > IB_QPT_GSI)
1472 return;
1474 tun_qp = &ctx->qp[qp_type];
1475 if (is_tun) {
1476 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1477 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1478 } else {
1479 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1480 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1484 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1485 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1486 rx_buf_size, DMA_FROM_DEVICE);
1487 kfree(tun_qp->ring[i].addr);
1490 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1491 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1492 tx_buf_size, DMA_TO_DEVICE);
1493 kfree(tun_qp->tx_ring[i].buf.addr);
1494 if (tun_qp->tx_ring[i].ah)
1495 ib_destroy_ah(tun_qp->tx_ring[i].ah);
1497 kfree(tun_qp->tx_ring);
1498 kfree(tun_qp->ring);
1501 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1503 struct mlx4_ib_demux_pv_ctx *ctx;
1504 struct mlx4_ib_demux_pv_qp *tun_qp;
1505 struct ib_wc wc;
1506 int ret;
1507 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1508 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1510 while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1511 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1512 if (wc.status == IB_WC_SUCCESS) {
1513 switch (wc.opcode) {
1514 case IB_WC_RECV:
1515 mlx4_ib_multiplex_mad(ctx, &wc);
1516 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1517 wc.wr_id &
1518 (MLX4_NUM_TUNNEL_BUFS - 1));
1519 if (ret)
1520 pr_err("Failed reposting tunnel "
1521 "buf:%lld\n", wc.wr_id);
1522 break;
1523 case IB_WC_SEND:
1524 pr_debug("received tunnel send completion:"
1525 "wrid=0x%llx, status=0x%x\n",
1526 wc.wr_id, wc.status);
1527 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1528 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1529 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1530 = NULL;
1531 spin_lock(&tun_qp->tx_lock);
1532 tun_qp->tx_ix_tail++;
1533 spin_unlock(&tun_qp->tx_lock);
1535 break;
1536 default:
1537 break;
1539 } else {
1540 pr_debug("mlx4_ib: completion error in tunnel: %d."
1541 " status = %d, wrid = 0x%llx\n",
1542 ctx->slave, wc.status, wc.wr_id);
1543 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1544 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1545 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1546 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1547 = NULL;
1548 spin_lock(&tun_qp->tx_lock);
1549 tun_qp->tx_ix_tail++;
1550 spin_unlock(&tun_qp->tx_lock);
1556 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1558 struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1560 /* It's worse than that! He's dead, Jim! */
1561 pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1562 event->event, sqp->port);
1565 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1566 enum ib_qp_type qp_type, int create_tun)
1568 int i, ret;
1569 struct mlx4_ib_demux_pv_qp *tun_qp;
1570 struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1571 struct ib_qp_attr attr;
1572 int qp_attr_mask_INIT;
1574 if (qp_type > IB_QPT_GSI)
1575 return -EINVAL;
1577 tun_qp = &ctx->qp[qp_type];
1579 memset(&qp_init_attr, 0, sizeof qp_init_attr);
1580 qp_init_attr.init_attr.send_cq = ctx->cq;
1581 qp_init_attr.init_attr.recv_cq = ctx->cq;
1582 qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1583 qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1584 qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1585 qp_init_attr.init_attr.cap.max_send_sge = 1;
1586 qp_init_attr.init_attr.cap.max_recv_sge = 1;
1587 if (create_tun) {
1588 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1589 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1590 qp_init_attr.port = ctx->port;
1591 qp_init_attr.slave = ctx->slave;
1592 qp_init_attr.proxy_qp_type = qp_type;
1593 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1594 IB_QP_QKEY | IB_QP_PORT;
1595 } else {
1596 qp_init_attr.init_attr.qp_type = qp_type;
1597 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1598 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1600 qp_init_attr.init_attr.port_num = ctx->port;
1601 qp_init_attr.init_attr.qp_context = ctx;
1602 qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1603 tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1604 if (IS_ERR(tun_qp->qp)) {
1605 ret = PTR_ERR(tun_qp->qp);
1606 tun_qp->qp = NULL;
1607 pr_err("Couldn't create %s QP (%d)\n",
1608 create_tun ? "tunnel" : "special", ret);
1609 return ret;
1612 memset(&attr, 0, sizeof attr);
1613 attr.qp_state = IB_QPS_INIT;
1614 ret = 0;
1615 if (create_tun)
1616 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1617 ctx->port, IB_DEFAULT_PKEY_FULL,
1618 &attr.pkey_index);
1619 if (ret || !create_tun)
1620 attr.pkey_index =
1621 to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1622 attr.qkey = IB_QP1_QKEY;
1623 attr.port_num = ctx->port;
1624 ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1625 if (ret) {
1626 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1627 create_tun ? "tunnel" : "special", ret);
1628 goto err_qp;
1630 attr.qp_state = IB_QPS_RTR;
1631 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1632 if (ret) {
1633 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1634 create_tun ? "tunnel" : "special", ret);
1635 goto err_qp;
1637 attr.qp_state = IB_QPS_RTS;
1638 attr.sq_psn = 0;
1639 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1640 if (ret) {
1641 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1642 create_tun ? "tunnel" : "special", ret);
1643 goto err_qp;
1646 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1647 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1648 if (ret) {
1649 pr_err(" mlx4_ib_post_pv_buf error"
1650 " (err = %d, i = %d)\n", ret, i);
1651 goto err_qp;
1654 return 0;
1656 err_qp:
1657 ib_destroy_qp(tun_qp->qp);
1658 tun_qp->qp = NULL;
1659 return ret;
1663 * IB MAD completion callback for real SQPs
1665 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1667 struct mlx4_ib_demux_pv_ctx *ctx;
1668 struct mlx4_ib_demux_pv_qp *sqp;
1669 struct ib_wc wc;
1670 struct ib_grh *grh;
1671 struct ib_mad *mad;
1673 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1674 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1676 while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1677 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1678 if (wc.status == IB_WC_SUCCESS) {
1679 switch (wc.opcode) {
1680 case IB_WC_SEND:
1681 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1682 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1683 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1684 = NULL;
1685 spin_lock(&sqp->tx_lock);
1686 sqp->tx_ix_tail++;
1687 spin_unlock(&sqp->tx_lock);
1688 break;
1689 case IB_WC_RECV:
1690 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1691 (sqp->ring[wc.wr_id &
1692 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1693 grh = &(((struct mlx4_mad_rcv_buf *)
1694 (sqp->ring[wc.wr_id &
1695 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1696 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1697 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1698 (MLX4_NUM_TUNNEL_BUFS - 1)))
1699 pr_err("Failed reposting SQP "
1700 "buf:%lld\n", wc.wr_id);
1701 break;
1702 default:
1703 BUG_ON(1);
1704 break;
1706 } else {
1707 pr_debug("mlx4_ib: completion error in tunnel: %d."
1708 " status = %d, wrid = 0x%llx\n",
1709 ctx->slave, wc.status, wc.wr_id);
1710 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1711 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1712 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1713 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1714 = NULL;
1715 spin_lock(&sqp->tx_lock);
1716 sqp->tx_ix_tail++;
1717 spin_unlock(&sqp->tx_lock);
1723 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1724 struct mlx4_ib_demux_pv_ctx **ret_ctx)
1726 struct mlx4_ib_demux_pv_ctx *ctx;
1728 *ret_ctx = NULL;
1729 ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1730 if (!ctx) {
1731 pr_err("failed allocating pv resource context "
1732 "for port %d, slave %d\n", port, slave);
1733 return -ENOMEM;
1736 ctx->ib_dev = &dev->ib_dev;
1737 ctx->port = port;
1738 ctx->slave = slave;
1739 *ret_ctx = ctx;
1740 return 0;
1743 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1745 if (dev->sriov.demux[port - 1].tun[slave]) {
1746 kfree(dev->sriov.demux[port - 1].tun[slave]);
1747 dev->sriov.demux[port - 1].tun[slave] = NULL;
1751 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1752 int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1754 int ret, cq_size;
1756 if (ctx->state != DEMUX_PV_STATE_DOWN)
1757 return -EEXIST;
1759 ctx->state = DEMUX_PV_STATE_STARTING;
1760 /* have QP0 only if link layer is IB */
1761 if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1762 IB_LINK_LAYER_INFINIBAND)
1763 ctx->has_smi = 1;
1765 if (ctx->has_smi) {
1766 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1767 if (ret) {
1768 pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1769 goto err_out;
1773 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1774 if (ret) {
1775 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1776 goto err_out_qp0;
1779 cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1780 if (ctx->has_smi)
1781 cq_size *= 2;
1783 ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1784 NULL, ctx, cq_size, 0);
1785 if (IS_ERR(ctx->cq)) {
1786 ret = PTR_ERR(ctx->cq);
1787 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1788 goto err_buf;
1791 ctx->pd = ib_alloc_pd(ctx->ib_dev);
1792 if (IS_ERR(ctx->pd)) {
1793 ret = PTR_ERR(ctx->pd);
1794 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1795 goto err_cq;
1798 ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1799 if (IS_ERR(ctx->mr)) {
1800 ret = PTR_ERR(ctx->mr);
1801 pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1802 goto err_pd;
1805 if (ctx->has_smi) {
1806 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1807 if (ret) {
1808 pr_err("Couldn't create %s QP0 (%d)\n",
1809 create_tun ? "tunnel for" : "", ret);
1810 goto err_mr;
1814 ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1815 if (ret) {
1816 pr_err("Couldn't create %s QP1 (%d)\n",
1817 create_tun ? "tunnel for" : "", ret);
1818 goto err_qp0;
1821 if (create_tun)
1822 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1823 else
1824 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1826 ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1828 ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1829 if (ret) {
1830 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1831 goto err_wq;
1833 ctx->state = DEMUX_PV_STATE_ACTIVE;
1834 return 0;
1836 err_wq:
1837 ctx->wq = NULL;
1838 ib_destroy_qp(ctx->qp[1].qp);
1839 ctx->qp[1].qp = NULL;
1842 err_qp0:
1843 if (ctx->has_smi)
1844 ib_destroy_qp(ctx->qp[0].qp);
1845 ctx->qp[0].qp = NULL;
1847 err_mr:
1848 ib_dereg_mr(ctx->mr);
1849 ctx->mr = NULL;
1851 err_pd:
1852 ib_dealloc_pd(ctx->pd);
1853 ctx->pd = NULL;
1855 err_cq:
1856 ib_destroy_cq(ctx->cq);
1857 ctx->cq = NULL;
1859 err_buf:
1860 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1862 err_out_qp0:
1863 if (ctx->has_smi)
1864 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1865 err_out:
1866 ctx->state = DEMUX_PV_STATE_DOWN;
1867 return ret;
1870 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1871 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1873 if (!ctx)
1874 return;
1875 if (ctx->state > DEMUX_PV_STATE_DOWN) {
1876 ctx->state = DEMUX_PV_STATE_DOWNING;
1877 if (flush)
1878 flush_workqueue(ctx->wq);
1879 if (ctx->has_smi) {
1880 ib_destroy_qp(ctx->qp[0].qp);
1881 ctx->qp[0].qp = NULL;
1882 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1884 ib_destroy_qp(ctx->qp[1].qp);
1885 ctx->qp[1].qp = NULL;
1886 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1887 ib_dereg_mr(ctx->mr);
1888 ctx->mr = NULL;
1889 ib_dealloc_pd(ctx->pd);
1890 ctx->pd = NULL;
1891 ib_destroy_cq(ctx->cq);
1892 ctx->cq = NULL;
1893 ctx->state = DEMUX_PV_STATE_DOWN;
1897 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1898 int port, int do_init)
1900 int ret = 0;
1902 if (!do_init) {
1903 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1904 /* for master, destroy real sqp resources */
1905 if (slave == mlx4_master_func_num(dev->dev))
1906 destroy_pv_resources(dev, slave, port,
1907 dev->sriov.sqps[port - 1], 1);
1908 /* destroy the tunnel qp resources */
1909 destroy_pv_resources(dev, slave, port,
1910 dev->sriov.demux[port - 1].tun[slave], 1);
1911 return 0;
1914 /* create the tunnel qp resources */
1915 ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1916 dev->sriov.demux[port - 1].tun[slave]);
1918 /* for master, create the real sqp resources */
1919 if (!ret && slave == mlx4_master_func_num(dev->dev))
1920 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1921 dev->sriov.sqps[port - 1]);
1922 return ret;
1925 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1927 struct mlx4_ib_demux_work *dmxw;
1929 dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1930 mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1931 dmxw->do_init);
1932 kfree(dmxw);
1933 return;
1936 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1937 struct mlx4_ib_demux_ctx *ctx,
1938 int port)
1940 char name[12];
1941 int ret = 0;
1942 int i;
1944 ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1945 sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1946 if (!ctx->tun)
1947 return -ENOMEM;
1949 ctx->dev = dev;
1950 ctx->port = port;
1951 ctx->ib_dev = &dev->ib_dev;
1953 for (i = 0;
1954 i < min(dev->dev->caps.sqp_demux, (u16)(dev->dev->num_vfs + 1));
1955 i++) {
1956 struct mlx4_active_ports actv_ports =
1957 mlx4_get_active_ports(dev->dev, i);
1959 if (!test_bit(port - 1, actv_ports.ports))
1960 continue;
1962 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1963 if (ret) {
1964 ret = -ENOMEM;
1965 goto err_mcg;
1969 ret = mlx4_ib_mcg_port_init(ctx);
1970 if (ret) {
1971 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1972 goto err_mcg;
1975 snprintf(name, sizeof name, "mlx4_ibt%d", port);
1976 ctx->wq = create_singlethread_workqueue(name);
1977 if (!ctx->wq) {
1978 pr_err("Failed to create tunnelling WQ for port %d\n", port);
1979 ret = -ENOMEM;
1980 goto err_wq;
1983 snprintf(name, sizeof name, "mlx4_ibud%d", port);
1984 ctx->ud_wq = create_singlethread_workqueue(name);
1985 if (!ctx->ud_wq) {
1986 pr_err("Failed to create up/down WQ for port %d\n", port);
1987 ret = -ENOMEM;
1988 goto err_udwq;
1991 return 0;
1993 err_udwq:
1994 destroy_workqueue(ctx->wq);
1995 ctx->wq = NULL;
1997 err_wq:
1998 mlx4_ib_mcg_port_cleanup(ctx, 1);
1999 err_mcg:
2000 for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2001 free_pv_object(dev, i, port);
2002 kfree(ctx->tun);
2003 ctx->tun = NULL;
2004 return ret;
2007 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2009 if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2010 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2011 flush_workqueue(sqp_ctx->wq);
2012 if (sqp_ctx->has_smi) {
2013 ib_destroy_qp(sqp_ctx->qp[0].qp);
2014 sqp_ctx->qp[0].qp = NULL;
2015 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2017 ib_destroy_qp(sqp_ctx->qp[1].qp);
2018 sqp_ctx->qp[1].qp = NULL;
2019 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2020 ib_dereg_mr(sqp_ctx->mr);
2021 sqp_ctx->mr = NULL;
2022 ib_dealloc_pd(sqp_ctx->pd);
2023 sqp_ctx->pd = NULL;
2024 ib_destroy_cq(sqp_ctx->cq);
2025 sqp_ctx->cq = NULL;
2026 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2030 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2032 int i;
2033 if (ctx) {
2034 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2035 mlx4_ib_mcg_port_cleanup(ctx, 1);
2036 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2037 if (!ctx->tun[i])
2038 continue;
2039 if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2040 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2042 flush_workqueue(ctx->wq);
2043 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2044 destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2045 free_pv_object(dev, i, ctx->port);
2047 kfree(ctx->tun);
2048 destroy_workqueue(ctx->ud_wq);
2049 destroy_workqueue(ctx->wq);
2053 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2055 int i;
2057 if (!mlx4_is_master(dev->dev))
2058 return;
2059 /* initialize or tear down tunnel QPs for the master */
2060 for (i = 0; i < dev->dev->caps.num_ports; i++)
2061 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2062 return;
2065 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2067 int i = 0;
2068 int err;
2070 if (!mlx4_is_mfunc(dev->dev))
2071 return 0;
2073 dev->sriov.is_going_down = 0;
2074 spin_lock_init(&dev->sriov.going_down_lock);
2075 mlx4_ib_cm_paravirt_init(dev);
2077 mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2079 if (mlx4_is_slave(dev->dev)) {
2080 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2081 return 0;
2084 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2085 if (i == mlx4_master_func_num(dev->dev))
2086 mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2087 else
2088 mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2091 err = mlx4_ib_init_alias_guid_service(dev);
2092 if (err) {
2093 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2094 goto paravirt_err;
2096 err = mlx4_ib_device_register_sysfs(dev);
2097 if (err) {
2098 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2099 goto sysfs_err;
2102 mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2103 dev->dev->caps.sqp_demux);
2104 for (i = 0; i < dev->num_ports; i++) {
2105 union ib_gid gid;
2106 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2107 if (err)
2108 goto demux_err;
2109 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2110 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2111 &dev->sriov.sqps[i]);
2112 if (err)
2113 goto demux_err;
2114 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2115 if (err)
2116 goto free_pv;
2118 mlx4_ib_master_tunnels(dev, 1);
2119 return 0;
2121 free_pv:
2122 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2123 demux_err:
2124 while (--i >= 0) {
2125 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2126 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2128 mlx4_ib_device_unregister_sysfs(dev);
2130 sysfs_err:
2131 mlx4_ib_destroy_alias_guid_service(dev);
2133 paravirt_err:
2134 mlx4_ib_cm_paravirt_clean(dev, -1);
2136 return err;
2139 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2141 int i;
2142 unsigned long flags;
2144 if (!mlx4_is_mfunc(dev->dev))
2145 return;
2147 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2148 dev->sriov.is_going_down = 1;
2149 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2150 if (mlx4_is_master(dev->dev)) {
2151 for (i = 0; i < dev->num_ports; i++) {
2152 flush_workqueue(dev->sriov.demux[i].ud_wq);
2153 mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2154 kfree(dev->sriov.sqps[i]);
2155 dev->sriov.sqps[i] = NULL;
2156 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2159 mlx4_ib_cm_paravirt_clean(dev, -1);
2160 mlx4_ib_destroy_alias_guid_service(dev);
2161 mlx4_ib_device_unregister_sysfs(dev);