1 /*
2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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.
31 */
32
33 #include <rdma/ib_mad.h>
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_sa.h>
36 #include <rdma/ib_cache.h>
37
38 #include <linux/random.h>
39 #include <linux/mlx4/cmd.h>
40 #include <linux/gfp.h>
41 #include <rdma/ib_pma.h>
42 #include <linux/ip.h>
43 #include <net/ipv6.h>
44
45 #include <linux/mlx4/driver.h>
46 #include "mlx4_ib.h"
47
48 enum {
49 MLX4_IB_VENDOR_CLASS1 = 0x9,
50 MLX4_IB_VENDOR_CLASS2 = 0xa
51 };
52
53 #define MLX4_TUN_SEND_WRID_SHIFT 34
54 #define MLX4_TUN_QPN_SHIFT 32
55 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
56 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
57
58 #define MLX4_TUN_IS_RECV(a) (((a) >> MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
59 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
60
61 /* Port mgmt change event handling */
62
63 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
64 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
65 #define NUM_IDX_IN_PKEY_TBL_BLK 32
66 #define GUID_TBL_ENTRY_SIZE 8 /* size in bytes */
67 #define GUID_TBL_BLK_NUM_ENTRIES 8
68 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
69
70 struct mlx4_mad_rcv_buf {
71 struct ib_grh grh;
72 u8 payload[256];
73 } __packed;
74
75 struct mlx4_mad_snd_buf {
76 u8 payload[256];
77 } __packed;
78
79 struct mlx4_tunnel_mad {
80 struct ib_grh grh;
81 struct mlx4_ib_tunnel_header hdr;
82 struct ib_mad mad;
83 } __packed;
84
85 struct mlx4_rcv_tunnel_mad {
86 struct mlx4_rcv_tunnel_hdr hdr;
87 struct ib_grh grh;
88 struct ib_mad mad;
89 } __packed;
90
91 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
92 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
93 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
94 int block, u32 change_bitmap);
95
mlx4_ib_gen_node_guid(void)96 __be64 mlx4_ib_gen_node_guid(void)
97 {
98 #define NODE_GUID_HI ((u64) (((u64)IB_OPENIB_OUI) << 40))
99 return cpu_to_be64(NODE_GUID_HI | prandom_u32());
100 }
101
mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx * ctx)102 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
103 {
104 return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
105 cpu_to_be64(0xff00000000000000LL);
106 }
107
mlx4_MAD_IFC(struct mlx4_ib_dev * dev,int mad_ifc_flags,int port,const struct ib_wc * in_wc,const struct ib_grh * in_grh,const void * in_mad,void * response_mad)108 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
109 int port, const struct ib_wc *in_wc,
110 const struct ib_grh *in_grh,
111 const void *in_mad, void *response_mad)
112 {
113 struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
114 void *inbox;
115 int err;
116 u32 in_modifier = port;
117 u8 op_modifier = 0;
118
119 inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
120 if (IS_ERR(inmailbox))
121 return PTR_ERR(inmailbox);
122 inbox = inmailbox->buf;
123
124 outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
125 if (IS_ERR(outmailbox)) {
126 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
127 return PTR_ERR(outmailbox);
128 }
129
130 memcpy(inbox, in_mad, 256);
131
132 /*
133 * Key check traps can't be generated unless we have in_wc to
134 * tell us where to send the trap.
135 */
136 if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
137 op_modifier |= 0x1;
138 if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
139 op_modifier |= 0x2;
140 if (mlx4_is_mfunc(dev->dev) &&
141 (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
142 op_modifier |= 0x8;
143
144 if (in_wc) {
145 struct {
146 __be32 my_qpn;
147 u32 reserved1;
148 __be32 rqpn;
149 u8 sl;
150 u8 g_path;
151 u16 reserved2[2];
152 __be16 pkey;
153 u32 reserved3[11];
154 u8 grh[40];
155 } *ext_info;
156
157 memset(inbox + 256, 0, 256);
158 ext_info = inbox + 256;
159
160 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
161 ext_info->rqpn = cpu_to_be32(in_wc->src_qp);
162 ext_info->sl = in_wc->sl << 4;
163 ext_info->g_path = in_wc->dlid_path_bits |
164 (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
165 ext_info->pkey = cpu_to_be16(in_wc->pkey_index);
166
167 if (in_grh)
168 memcpy(ext_info->grh, in_grh, 40);
169
170 op_modifier |= 0x4;
171
172 in_modifier |= ib_lid_cpu16(in_wc->slid) << 16;
173 }
174
175 err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
176 mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
177 MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
178 (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
179
180 if (!err)
181 memcpy(response_mad, outmailbox->buf, 256);
182
183 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
184 mlx4_free_cmd_mailbox(dev->dev, outmailbox);
185
186 return err;
187 }
188
update_sm_ah(struct mlx4_ib_dev * dev,u8 port_num,u16 lid,u8 sl)189 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
190 {
191 struct ib_ah *new_ah;
192 struct rdma_ah_attr ah_attr;
193 unsigned long flags;
194
195 if (!dev->send_agent[port_num - 1][0])
196 return;
197
198 memset(&ah_attr, 0, sizeof ah_attr);
199 ah_attr.type = rdma_ah_find_type(&dev->ib_dev, port_num);
200 rdma_ah_set_dlid(&ah_attr, lid);
201 rdma_ah_set_sl(&ah_attr, sl);
202 rdma_ah_set_port_num(&ah_attr, port_num);
203
204 new_ah = rdma_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
205 &ah_attr);
206 if (IS_ERR(new_ah))
207 return;
208
209 spin_lock_irqsave(&dev->sm_lock, flags);
210 if (dev->sm_ah[port_num - 1])
211 rdma_destroy_ah(dev->sm_ah[port_num - 1]);
212 dev->sm_ah[port_num - 1] = new_ah;
213 spin_unlock_irqrestore(&dev->sm_lock, flags);
214 }
215
216 /*
217 * Snoop SM MADs for port info, GUID info, and P_Key table sets, so we can
218 * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
219 */
smp_snoop(struct ib_device * ibdev,u8 port_num,const struct ib_mad * mad,u16 prev_lid)220 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
221 u16 prev_lid)
222 {
223 struct ib_port_info *pinfo;
224 u16 lid;
225 __be16 *base;
226 u32 bn, pkey_change_bitmap;
227 int i;
228
229
230 struct mlx4_ib_dev *dev = to_mdev(ibdev);
231 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
232 mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
233 mad->mad_hdr.method == IB_MGMT_METHOD_SET)
234 switch (mad->mad_hdr.attr_id) {
235 case IB_SMP_ATTR_PORT_INFO:
236 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
237 return;
238 pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
239 lid = be16_to_cpu(pinfo->lid);
240
241 update_sm_ah(dev, port_num,
242 be16_to_cpu(pinfo->sm_lid),
243 pinfo->neighbormtu_mastersmsl & 0xf);
244
245 if (pinfo->clientrereg_resv_subnetto & 0x80)
246 handle_client_rereg_event(dev, port_num);
247
248 if (prev_lid != lid)
249 handle_lid_change_event(dev, port_num);
250 break;
251
252 case IB_SMP_ATTR_PKEY_TABLE:
253 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
254 return;
255 if (!mlx4_is_mfunc(dev->dev)) {
256 mlx4_ib_dispatch_event(dev, port_num,
257 IB_EVENT_PKEY_CHANGE);
258 break;
259 }
260
261 /* at this point, we are running in the master.
262 * Slaves do not receive SMPs.
263 */
264 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
265 base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
266 pkey_change_bitmap = 0;
267 for (i = 0; i < 32; i++) {
268 pr_debug("PKEY[%d] = x%x\n",
269 i + bn*32, be16_to_cpu(base[i]));
270 if (be16_to_cpu(base[i]) !=
271 dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
272 pkey_change_bitmap |= (1 << i);
273 dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
274 be16_to_cpu(base[i]);
275 }
276 }
277 pr_debug("PKEY Change event: port=%d, "
278 "block=0x%x, change_bitmap=0x%x\n",
279 port_num, bn, pkey_change_bitmap);
280
281 if (pkey_change_bitmap) {
282 mlx4_ib_dispatch_event(dev, port_num,
283 IB_EVENT_PKEY_CHANGE);
284 if (!dev->sriov.is_going_down)
285 __propagate_pkey_ev(dev, port_num, bn,
286 pkey_change_bitmap);
287 }
288 break;
289
290 case IB_SMP_ATTR_GUID_INFO:
291 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
292 return;
293 /* paravirtualized master's guid is guid 0 -- does not change */
294 if (!mlx4_is_master(dev->dev))
295 mlx4_ib_dispatch_event(dev, port_num,
296 IB_EVENT_GID_CHANGE);
297 /*if master, notify relevant slaves*/
298 if (mlx4_is_master(dev->dev) &&
299 !dev->sriov.is_going_down) {
300 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
301 mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
302 (u8 *)(&((struct ib_smp *)mad)->data));
303 mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
304 (u8 *)(&((struct ib_smp *)mad)->data));
305 }
306 break;
307
308 case IB_SMP_ATTR_SL_TO_VL_TABLE:
309 /* cache sl to vl mapping changes for use in
310 * filling QP1 LRH VL field when sending packets
311 */
312 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV &&
313 dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)
314 return;
315 if (!mlx4_is_slave(dev->dev)) {
316 union sl2vl_tbl_to_u64 sl2vl64;
317 int jj;
318
319 for (jj = 0; jj < 8; jj++) {
320 sl2vl64.sl8[jj] = ((struct ib_smp *)mad)->data[jj];
321 pr_debug("port %u, sl2vl[%d] = %02x\n",
322 port_num, jj, sl2vl64.sl8[jj]);
323 }
324 atomic64_set(&dev->sl2vl[port_num - 1], sl2vl64.sl64);
325 }
326 break;
327
328 default:
329 break;
330 }
331 }
332
__propagate_pkey_ev(struct mlx4_ib_dev * dev,int port_num,int block,u32 change_bitmap)333 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
334 int block, u32 change_bitmap)
335 {
336 int i, ix, slave, err;
337 int have_event = 0;
338
339 for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
340 if (slave == mlx4_master_func_num(dev->dev))
341 continue;
342 if (!mlx4_is_slave_active(dev->dev, slave))
343 continue;
344
345 have_event = 0;
346 for (i = 0; i < 32; i++) {
347 if (!(change_bitmap & (1 << i)))
348 continue;
349 for (ix = 0;
350 ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
351 if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
352 [ix] == i + 32 * block) {
353 err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
354 pr_debug("propagate_pkey_ev: slave %d,"
355 " port %d, ix %d (%d)\n",
356 slave, port_num, ix, err);
357 have_event = 1;
358 break;
359 }
360 }
361 if (have_event)
362 break;
363 }
364 }
365 }
366
node_desc_override(struct ib_device * dev,struct ib_mad * mad)367 static void node_desc_override(struct ib_device *dev,
368 struct ib_mad *mad)
369 {
370 unsigned long flags;
371
372 if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
373 mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
374 mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
375 mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
376 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
377 memcpy(((struct ib_smp *) mad)->data, dev->node_desc,
378 IB_DEVICE_NODE_DESC_MAX);
379 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
380 }
381 }
382
forward_trap(struct mlx4_ib_dev * dev,u8 port_num,const struct ib_mad * mad)383 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
384 {
385 int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
386 struct ib_mad_send_buf *send_buf;
387 struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
388 int ret;
389 unsigned long flags;
390
391 if (agent) {
392 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
393 IB_MGMT_MAD_DATA, GFP_ATOMIC,
394 IB_MGMT_BASE_VERSION);
395 if (IS_ERR(send_buf))
396 return;
397 /*
398 * We rely here on the fact that MLX QPs don't use the
399 * address handle after the send is posted (this is
400 * wrong following the IB spec strictly, but we know
401 * it's OK for our devices).
402 */
403 spin_lock_irqsave(&dev->sm_lock, flags);
404 memcpy(send_buf->mad, mad, sizeof *mad);
405 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
406 ret = ib_post_send_mad(send_buf, NULL);
407 else
408 ret = -EINVAL;
409 spin_unlock_irqrestore(&dev->sm_lock, flags);
410
411 if (ret)
412 ib_free_send_mad(send_buf);
413 }
414 }
415
mlx4_ib_demux_sa_handler(struct ib_device * ibdev,int port,int slave,struct ib_sa_mad * sa_mad)416 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
417 struct ib_sa_mad *sa_mad)
418 {
419 int ret = 0;
420
421 /* dispatch to different sa handlers */
422 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
423 case IB_SA_ATTR_MC_MEMBER_REC:
424 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
425 break;
426 default:
427 break;
428 }
429 return ret;
430 }
431
mlx4_ib_find_real_gid(struct ib_device * ibdev,u8 port,__be64 guid)432 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
433 {
434 struct mlx4_ib_dev *dev = to_mdev(ibdev);
435 int i;
436
437 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
438 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
439 return i;
440 }
441 return -1;
442 }
443
444
find_slave_port_pkey_ix(struct mlx4_ib_dev * dev,int slave,u8 port,u16 pkey,u16 * ix)445 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
446 u8 port, u16 pkey, u16 *ix)
447 {
448 int i, ret;
449 u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
450 u16 slot_pkey;
451
452 if (slave == mlx4_master_func_num(dev->dev))
453 return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
454
455 unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
456
457 for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
458 if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
459 continue;
460
461 pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
462
463 ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
464 if (ret)
465 continue;
466 if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
467 if (slot_pkey & 0x8000) {
468 *ix = (u16) pkey_ix;
469 return 0;
470 } else {
471 /* take first partial pkey index found */
472 if (partial_ix == 0xFF)
473 partial_ix = pkey_ix;
474 }
475 }
476 }
477
478 if (partial_ix < 0xFF) {
479 *ix = (u16) partial_ix;
480 return 0;
481 }
482
483 return -EINVAL;
484 }
485
get_gids_from_l3_hdr(struct ib_grh * grh,union ib_gid * sgid,union ib_gid * dgid)486 static int get_gids_from_l3_hdr(struct ib_grh *grh, union ib_gid *sgid,
487 union ib_gid *dgid)
488 {
489 int version = ib_get_rdma_header_version((const union rdma_network_hdr *)grh);
490 enum rdma_network_type net_type;
491
492 if (version == 4)
493 net_type = RDMA_NETWORK_IPV4;
494 else if (version == 6)
495 net_type = RDMA_NETWORK_IPV6;
496 else
497 return -EINVAL;
498
499 return ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
500 sgid, dgid);
501 }
502
mlx4_ib_send_to_slave(struct mlx4_ib_dev * dev,int slave,u8 port,enum ib_qp_type dest_qpt,struct ib_wc * wc,struct ib_grh * grh,struct ib_mad * mad)503 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
504 enum ib_qp_type dest_qpt, struct ib_wc *wc,
505 struct ib_grh *grh, struct ib_mad *mad)
506 {
507 struct ib_sge list;
508 struct ib_ud_wr wr;
509 const struct ib_send_wr *bad_wr;
510 struct mlx4_ib_demux_pv_ctx *tun_ctx;
511 struct mlx4_ib_demux_pv_qp *tun_qp;
512 struct mlx4_rcv_tunnel_mad *tun_mad;
513 struct rdma_ah_attr attr;
514 struct ib_ah *ah;
515 struct ib_qp *src_qp = NULL;
516 unsigned tun_tx_ix = 0;
517 int dqpn;
518 int ret = 0;
519 u16 tun_pkey_ix;
520 u16 cached_pkey;
521 u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
522
523 if (dest_qpt > IB_QPT_GSI)
524 return -EINVAL;
525
526 tun_ctx = dev->sriov.demux[port-1].tun[slave];
527
528 /* check if proxy qp created */
529 if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
530 return -EAGAIN;
531
532 if (!dest_qpt)
533 tun_qp = &tun_ctx->qp[0];
534 else
535 tun_qp = &tun_ctx->qp[1];
536
537 /* compute P_Key index to put in tunnel header for slave */
538 if (dest_qpt) {
539 u16 pkey_ix;
540 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
541 if (ret)
542 return -EINVAL;
543
544 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
545 if (ret)
546 return -EINVAL;
547 tun_pkey_ix = pkey_ix;
548 } else
549 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
550
551 dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
552
553 /* get tunnel tx data buf for slave */
554 src_qp = tun_qp->qp;
555
556 /* create ah. Just need an empty one with the port num for the post send.
557 * The driver will set the force loopback bit in post_send */
558 memset(&attr, 0, sizeof attr);
559 attr.type = rdma_ah_find_type(&dev->ib_dev, port);
560
561 rdma_ah_set_port_num(&attr, port);
562 if (is_eth) {
563 union ib_gid sgid;
564 union ib_gid dgid;
565
566 if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
567 return -EINVAL;
568 rdma_ah_set_grh(&attr, &dgid, 0, 0, 0, 0);
569 }
570 ah = rdma_create_ah(tun_ctx->pd, &attr);
571 if (IS_ERR(ah))
572 return -ENOMEM;
573
574 /* allocate tunnel tx buf after pass failure returns */
575 spin_lock(&tun_qp->tx_lock);
576 if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
577 (MLX4_NUM_TUNNEL_BUFS - 1))
578 ret = -EAGAIN;
579 else
580 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
581 spin_unlock(&tun_qp->tx_lock);
582 if (ret)
583 goto end;
584
585 tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
586 if (tun_qp->tx_ring[tun_tx_ix].ah)
587 rdma_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
588 tun_qp->tx_ring[tun_tx_ix].ah = ah;
589 ib_dma_sync_single_for_cpu(&dev->ib_dev,
590 tun_qp->tx_ring[tun_tx_ix].buf.map,
591 sizeof (struct mlx4_rcv_tunnel_mad),
592 DMA_TO_DEVICE);
593
594 /* copy over to tunnel buffer */
595 if (grh)
596 memcpy(&tun_mad->grh, grh, sizeof *grh);
597 memcpy(&tun_mad->mad, mad, sizeof *mad);
598
599 /* adjust tunnel data */
600 tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
601 tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
602 tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
603
604 if (is_eth) {
605 u16 vlan = 0;
606 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
607 NULL)) {
608 /* VST mode */
609 if (vlan != wc->vlan_id)
610 /* Packet vlan is not the VST-assigned vlan.
611 * Drop the packet.
612 */
613 goto out;
614 else
615 /* Remove the vlan tag before forwarding
616 * the packet to the VF.
617 */
618 vlan = 0xffff;
619 } else {
620 vlan = wc->vlan_id;
621 }
622
623 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
624 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
625 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
626 } else {
627 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
628 tun_mad->hdr.slid_mac_47_32 = ib_lid_be16(wc->slid);
629 }
630
631 ib_dma_sync_single_for_device(&dev->ib_dev,
632 tun_qp->tx_ring[tun_tx_ix].buf.map,
633 sizeof (struct mlx4_rcv_tunnel_mad),
634 DMA_TO_DEVICE);
635
636 list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
637 list.length = sizeof (struct mlx4_rcv_tunnel_mad);
638 list.lkey = tun_ctx->pd->local_dma_lkey;
639
640 wr.ah = ah;
641 wr.port_num = port;
642 wr.remote_qkey = IB_QP_SET_QKEY;
643 wr.remote_qpn = dqpn;
644 wr.wr.next = NULL;
645 wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
646 wr.wr.sg_list = &list;
647 wr.wr.num_sge = 1;
648 wr.wr.opcode = IB_WR_SEND;
649 wr.wr.send_flags = IB_SEND_SIGNALED;
650
651 ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
652 if (!ret)
653 return 0;
654 out:
655 spin_lock(&tun_qp->tx_lock);
656 tun_qp->tx_ix_tail++;
657 spin_unlock(&tun_qp->tx_lock);
658 tun_qp->tx_ring[tun_tx_ix].ah = NULL;
659 end:
660 rdma_destroy_ah(ah);
661 return ret;
662 }
663
mlx4_ib_demux_mad(struct ib_device * ibdev,u8 port,struct ib_wc * wc,struct ib_grh * grh,struct ib_mad * mad)664 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
665 struct ib_wc *wc, struct ib_grh *grh,
666 struct ib_mad *mad)
667 {
668 struct mlx4_ib_dev *dev = to_mdev(ibdev);
669 int err, other_port;
670 int slave = -1;
671 u8 *slave_id;
672 int is_eth = 0;
673
674 if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
675 is_eth = 0;
676 else
677 is_eth = 1;
678
679 if (is_eth) {
680 union ib_gid dgid;
681 union ib_gid sgid;
682
683 if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
684 return -EINVAL;
685 if (!(wc->wc_flags & IB_WC_GRH)) {
686 mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
687 return -EINVAL;
688 }
689 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
690 mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
691 return -EINVAL;
692 }
693 err = mlx4_get_slave_from_roce_gid(dev->dev, port, dgid.raw, &slave);
694 if (err && mlx4_is_mf_bonded(dev->dev)) {
695 other_port = (port == 1) ? 2 : 1;
696 err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, dgid.raw, &slave);
697 if (!err) {
698 port = other_port;
699 pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
700 slave, grh->dgid.raw, port, other_port);
701 }
702 }
703 if (err) {
704 mlx4_ib_warn(ibdev, "failed matching grh\n");
705 return -ENOENT;
706 }
707 if (slave >= dev->dev->caps.sqp_demux) {
708 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
709 slave, dev->dev->caps.sqp_demux);
710 return -ENOENT;
711 }
712
713 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
714 return 0;
715
716 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
717 if (err)
718 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
719 slave, err);
720 return 0;
721 }
722
723 /* Initially assume that this mad is for us */
724 slave = mlx4_master_func_num(dev->dev);
725
726 /* See if the slave id is encoded in a response mad */
727 if (mad->mad_hdr.method & 0x80) {
728 slave_id = (u8 *) &mad->mad_hdr.tid;
729 slave = *slave_id;
730 if (slave != 255) /*255 indicates the dom0*/
731 *slave_id = 0; /* remap tid */
732 }
733
734 /* If a grh is present, we demux according to it */
735 if (wc->wc_flags & IB_WC_GRH) {
736 if (grh->dgid.global.interface_id ==
737 cpu_to_be64(IB_SA_WELL_KNOWN_GUID) &&
738 grh->dgid.global.subnet_prefix == cpu_to_be64(
739 atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix))) {
740 slave = 0;
741 } else {
742 slave = mlx4_ib_find_real_gid(ibdev, port,
743 grh->dgid.global.interface_id);
744 if (slave < 0) {
745 mlx4_ib_warn(ibdev, "failed matching grh\n");
746 return -ENOENT;
747 }
748 }
749 }
750 /* Class-specific handling */
751 switch (mad->mad_hdr.mgmt_class) {
752 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
753 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
754 /* 255 indicates the dom0 */
755 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
756 if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
757 return -EPERM;
758 /* for a VF. drop unsolicited MADs */
759 if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
760 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
761 slave, mad->mad_hdr.mgmt_class,
762 mad->mad_hdr.method);
763 return -EINVAL;
764 }
765 }
766 break;
767 case IB_MGMT_CLASS_SUBN_ADM:
768 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
769 (struct ib_sa_mad *) mad))
770 return 0;
771 break;
772 case IB_MGMT_CLASS_CM:
773 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
774 return 0;
775 break;
776 case IB_MGMT_CLASS_DEVICE_MGMT:
777 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
778 return 0;
779 break;
780 default:
781 /* Drop unsupported classes for slaves in tunnel mode */
782 if (slave != mlx4_master_func_num(dev->dev)) {
783 pr_debug("dropping unsupported ingress mad from class:%d "
784 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
785 return 0;
786 }
787 }
788 /*make sure that no slave==255 was not handled yet.*/
789 if (slave >= dev->dev->caps.sqp_demux) {
790 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
791 slave, dev->dev->caps.sqp_demux);
792 return -ENOENT;
793 }
794
795 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
796 if (err)
797 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
798 slave, err);
799 return 0;
800 }
801
ib_process_mad(struct ib_device * ibdev,int mad_flags,u8 port_num,const struct ib_wc * in_wc,const struct ib_grh * in_grh,const struct ib_mad * in_mad,struct ib_mad * out_mad)802 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
803 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
804 const struct ib_mad *in_mad, struct ib_mad *out_mad)
805 {
806 u16 slid, prev_lid = 0;
807 int err;
808 struct ib_port_attr pattr;
809
810 if (in_wc && in_wc->qp->qp_num) {
811 pr_debug("received MAD: slid:%d sqpn:%d "
812 "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
813 in_wc->slid, in_wc->src_qp,
814 in_wc->dlid_path_bits,
815 in_wc->qp->qp_num,
816 in_wc->wc_flags,
817 in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
818 be16_to_cpu(in_mad->mad_hdr.attr_id));
819 if (in_wc->wc_flags & IB_WC_GRH) {
820 pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
821 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
822 be64_to_cpu(in_grh->sgid.global.interface_id));
823 pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
824 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
825 be64_to_cpu(in_grh->dgid.global.interface_id));
826 }
827 }
828
829 slid = in_wc ? ib_lid_cpu16(in_wc->slid) : be16_to_cpu(IB_LID_PERMISSIVE);
830
831 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
832 forward_trap(to_mdev(ibdev), port_num, in_mad);
833 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
834 }
835
836 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
837 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
838 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
839 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET &&
840 in_mad->mad_hdr.method != IB_MGMT_METHOD_TRAP_REPRESS)
841 return IB_MAD_RESULT_SUCCESS;
842
843 /*
844 * Don't process SMInfo queries -- the SMA can't handle them.
845 */
846 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
847 return IB_MAD_RESULT_SUCCESS;
848 } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
849 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1 ||
850 in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2 ||
851 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
852 if (in_mad->mad_hdr.method != IB_MGMT_METHOD_GET &&
853 in_mad->mad_hdr.method != IB_MGMT_METHOD_SET)
854 return IB_MAD_RESULT_SUCCESS;
855 } else
856 return IB_MAD_RESULT_SUCCESS;
857
858 if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
859 in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
860 in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
861 in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
862 !ib_query_port(ibdev, port_num, &pattr))
863 prev_lid = ib_lid_cpu16(pattr.lid);
864
865 err = mlx4_MAD_IFC(to_mdev(ibdev),
866 (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
867 (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
868 MLX4_MAD_IFC_NET_VIEW,
869 port_num, in_wc, in_grh, in_mad, out_mad);
870 if (err)
871 return IB_MAD_RESULT_FAILURE;
872
873 if (!out_mad->mad_hdr.status) {
874 smp_snoop(ibdev, port_num, in_mad, prev_lid);
875 /* slaves get node desc from FW */
876 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
877 node_desc_override(ibdev, out_mad);
878 }
879
880 /* set return bit in status of directed route responses */
881 if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
882 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
883
884 if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
885 /* no response for trap repress */
886 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
887
888 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
889 }
890
edit_counter(struct mlx4_counter * cnt,void * counters,__be16 attr_id)891 static void edit_counter(struct mlx4_counter *cnt, void *counters,
892 __be16 attr_id)
893 {
894 switch (attr_id) {
895 case IB_PMA_PORT_COUNTERS:
896 {
897 struct ib_pma_portcounters *pma_cnt =
898 (struct ib_pma_portcounters *)counters;
899
900 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
901 (be64_to_cpu(cnt->tx_bytes) >> 2));
902 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
903 (be64_to_cpu(cnt->rx_bytes) >> 2));
904 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
905 be64_to_cpu(cnt->tx_frames));
906 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
907 be64_to_cpu(cnt->rx_frames));
908 break;
909 }
910 case IB_PMA_PORT_COUNTERS_EXT:
911 {
912 struct ib_pma_portcounters_ext *pma_cnt_ext =
913 (struct ib_pma_portcounters_ext *)counters;
914
915 pma_cnt_ext->port_xmit_data =
916 cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
917 pma_cnt_ext->port_rcv_data =
918 cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
919 pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
920 pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
921 break;
922 }
923 }
924 }
925
iboe_process_mad_port_info(void * out_mad)926 static int iboe_process_mad_port_info(void *out_mad)
927 {
928 struct ib_class_port_info cpi = {};
929
930 cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
931 memcpy(out_mad, &cpi, sizeof(cpi));
932 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
933 }
934
iboe_process_mad(struct ib_device * ibdev,int mad_flags,u8 port_num,const struct ib_wc * in_wc,const struct ib_grh * in_grh,const struct ib_mad * in_mad,struct ib_mad * out_mad)935 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
936 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
937 const struct ib_mad *in_mad, struct ib_mad *out_mad)
938 {
939 struct mlx4_counter counter_stats;
940 struct mlx4_ib_dev *dev = to_mdev(ibdev);
941 struct counter_index *tmp_counter;
942 int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
943
944 if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
945 return -EINVAL;
946
947 if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
948 return iboe_process_mad_port_info((void *)(out_mad->data + 40));
949
950 memset(&counter_stats, 0, sizeof(counter_stats));
951 mutex_lock(&dev->counters_table[port_num - 1].mutex);
952 list_for_each_entry(tmp_counter,
953 &dev->counters_table[port_num - 1].counters_list,
954 list) {
955 err = mlx4_get_counter_stats(dev->dev,
956 tmp_counter->index,
957 &counter_stats, 0);
958 if (err) {
959 err = IB_MAD_RESULT_FAILURE;
960 stats_avail = 0;
961 break;
962 }
963 stats_avail = 1;
964 }
965 mutex_unlock(&dev->counters_table[port_num - 1].mutex);
966 if (stats_avail) {
967 memset(out_mad->data, 0, sizeof out_mad->data);
968 switch (counter_stats.counter_mode & 0xf) {
969 case 0:
970 edit_counter(&counter_stats,
971 (void *)(out_mad->data + 40),
972 in_mad->mad_hdr.attr_id);
973 err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
974 break;
975 default:
976 err = IB_MAD_RESULT_FAILURE;
977 }
978 }
979
980 return err;
981 }
982
mlx4_ib_process_mad(struct ib_device * ibdev,int mad_flags,u8 port_num,const struct ib_wc * in_wc,const struct ib_grh * in_grh,const struct ib_mad_hdr * in,size_t in_mad_size,struct ib_mad_hdr * out,size_t * out_mad_size,u16 * out_mad_pkey_index)983 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
984 const struct ib_wc *in_wc, const struct ib_grh *in_grh,
985 const struct ib_mad_hdr *in, size_t in_mad_size,
986 struct ib_mad_hdr *out, size_t *out_mad_size,
987 u16 *out_mad_pkey_index)
988 {
989 struct mlx4_ib_dev *dev = to_mdev(ibdev);
990 const struct ib_mad *in_mad = (const struct ib_mad *)in;
991 struct ib_mad *out_mad = (struct ib_mad *)out;
992 enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
993
994 if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
995 *out_mad_size != sizeof(*out_mad)))
996 return IB_MAD_RESULT_FAILURE;
997
998 /* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
999 * queries, should be called only by VFs and for that specific purpose
1000 */
1001 if (link == IB_LINK_LAYER_INFINIBAND) {
1002 if (mlx4_is_slave(dev->dev) &&
1003 (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
1004 (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
1005 in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
1006 in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
1007 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1008 in_grh, in_mad, out_mad);
1009
1010 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
1011 in_grh, in_mad, out_mad);
1012 }
1013
1014 if (link == IB_LINK_LAYER_ETHERNET)
1015 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1016 in_grh, in_mad, out_mad);
1017
1018 return -EINVAL;
1019 }
1020
send_handler(struct ib_mad_agent * agent,struct ib_mad_send_wc * mad_send_wc)1021 static void send_handler(struct ib_mad_agent *agent,
1022 struct ib_mad_send_wc *mad_send_wc)
1023 {
1024 if (mad_send_wc->send_buf->context[0])
1025 rdma_destroy_ah(mad_send_wc->send_buf->context[0]);
1026 ib_free_send_mad(mad_send_wc->send_buf);
1027 }
1028
mlx4_ib_mad_init(struct mlx4_ib_dev * dev)1029 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
1030 {
1031 struct ib_mad_agent *agent;
1032 int p, q;
1033 int ret;
1034 enum rdma_link_layer ll;
1035
1036 for (p = 0; p < dev->num_ports; ++p) {
1037 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1038 for (q = 0; q <= 1; ++q) {
1039 if (ll == IB_LINK_LAYER_INFINIBAND) {
1040 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1041 q ? IB_QPT_GSI : IB_QPT_SMI,
1042 NULL, 0, send_handler,
1043 NULL, NULL, 0);
1044 if (IS_ERR(agent)) {
1045 ret = PTR_ERR(agent);
1046 goto err;
1047 }
1048 dev->send_agent[p][q] = agent;
1049 } else
1050 dev->send_agent[p][q] = NULL;
1051 }
1052 }
1053
1054 return 0;
1055
1056 err:
1057 for (p = 0; p < dev->num_ports; ++p)
1058 for (q = 0; q <= 1; ++q)
1059 if (dev->send_agent[p][q])
1060 ib_unregister_mad_agent(dev->send_agent[p][q]);
1061
1062 return ret;
1063 }
1064
mlx4_ib_mad_cleanup(struct mlx4_ib_dev * dev)1065 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1066 {
1067 struct ib_mad_agent *agent;
1068 int p, q;
1069
1070 for (p = 0; p < dev->num_ports; ++p) {
1071 for (q = 0; q <= 1; ++q) {
1072 agent = dev->send_agent[p][q];
1073 if (agent) {
1074 dev->send_agent[p][q] = NULL;
1075 ib_unregister_mad_agent(agent);
1076 }
1077 }
1078
1079 if (dev->sm_ah[p])
1080 rdma_destroy_ah(dev->sm_ah[p]);
1081 }
1082 }
1083
handle_lid_change_event(struct mlx4_ib_dev * dev,u8 port_num)1084 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1085 {
1086 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1087
1088 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1089 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1090 MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1091 }
1092
handle_client_rereg_event(struct mlx4_ib_dev * dev,u8 port_num)1093 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1094 {
1095 /* re-configure the alias-guid and mcg's */
1096 if (mlx4_is_master(dev->dev)) {
1097 mlx4_ib_invalidate_all_guid_record(dev, port_num);
1098
1099 if (!dev->sriov.is_going_down) {
1100 mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1101 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1102 MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1103 }
1104 }
1105
1106 /* Update the sl to vl table from inside client rereg
1107 * only if in secure-host mode (snooping is not possible)
1108 * and the sl-to-vl change event is not generated by FW.
1109 */
1110 if (!mlx4_is_slave(dev->dev) &&
1111 dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1112 !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1113 if (mlx4_is_master(dev->dev))
1114 /* already in work queue from mlx4_ib_event queueing
1115 * mlx4_handle_port_mgmt_change_event, which calls
1116 * this procedure. Therefore, call sl2vl_update directly.
1117 */
1118 mlx4_ib_sl2vl_update(dev, port_num);
1119 else
1120 mlx4_sched_ib_sl2vl_update_work(dev, port_num);
1121 }
1122 mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1123 }
1124
propagate_pkey_ev(struct mlx4_ib_dev * dev,int port_num,struct mlx4_eqe * eqe)1125 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1126 struct mlx4_eqe *eqe)
1127 {
1128 __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1129 GET_MASK_FROM_EQE(eqe));
1130 }
1131
handle_slaves_guid_change(struct mlx4_ib_dev * dev,u8 port_num,u32 guid_tbl_blk_num,u32 change_bitmap)1132 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1133 u32 guid_tbl_blk_num, u32 change_bitmap)
1134 {
1135 struct ib_smp *in_mad = NULL;
1136 struct ib_smp *out_mad = NULL;
1137 u16 i;
1138
1139 if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1140 return;
1141
1142 in_mad = kmalloc(sizeof *in_mad, GFP_KERNEL);
1143 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1144 if (!in_mad || !out_mad)
1145 goto out;
1146
1147 guid_tbl_blk_num *= 4;
1148
1149 for (i = 0; i < 4; i++) {
1150 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1151 continue;
1152 memset(in_mad, 0, sizeof *in_mad);
1153 memset(out_mad, 0, sizeof *out_mad);
1154
1155 in_mad->base_version = 1;
1156 in_mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1157 in_mad->class_version = 1;
1158 in_mad->method = IB_MGMT_METHOD_GET;
1159 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
1160 in_mad->attr_mod = cpu_to_be32(guid_tbl_blk_num + i);
1161
1162 if (mlx4_MAD_IFC(dev,
1163 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1164 port_num, NULL, NULL, in_mad, out_mad)) {
1165 mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1166 goto out;
1167 }
1168
1169 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1170 port_num,
1171 (u8 *)(&((struct ib_smp *)out_mad)->data));
1172 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1173 port_num,
1174 (u8 *)(&((struct ib_smp *)out_mad)->data));
1175 }
1176
1177 out:
1178 kfree(in_mad);
1179 kfree(out_mad);
1180 return;
1181 }
1182
handle_port_mgmt_change_event(struct work_struct * work)1183 void handle_port_mgmt_change_event(struct work_struct *work)
1184 {
1185 struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1186 struct mlx4_ib_dev *dev = ew->ib_dev;
1187 struct mlx4_eqe *eqe = &(ew->ib_eqe);
1188 u8 port = eqe->event.port_mgmt_change.port;
1189 u32 changed_attr;
1190 u32 tbl_block;
1191 u32 change_bitmap;
1192
1193 switch (eqe->subtype) {
1194 case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1195 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1196
1197 /* Update the SM ah - This should be done before handling
1198 the other changed attributes so that MADs can be sent to the SM */
1199 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1200 u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1201 u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1202 update_sm_ah(dev, port, lid, sl);
1203 }
1204
1205 /* Check if it is a lid change event */
1206 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1207 handle_lid_change_event(dev, port);
1208
1209 /* Generate GUID changed event */
1210 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1211 if (mlx4_is_master(dev->dev)) {
1212 union ib_gid gid;
1213 int err = 0;
1214
1215 if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1216 err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1217 else
1218 gid.global.subnet_prefix =
1219 eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1220 if (err) {
1221 pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1222 port, err);
1223 } else {
1224 pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1225 port,
1226 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1227 be64_to_cpu(gid.global.subnet_prefix));
1228 atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1229 be64_to_cpu(gid.global.subnet_prefix));
1230 }
1231 }
1232 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1233 /*if master, notify all slaves*/
1234 if (mlx4_is_master(dev->dev))
1235 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1236 MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1237 }
1238
1239 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1240 handle_client_rereg_event(dev, port);
1241 break;
1242
1243 case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1244 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1245 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1246 propagate_pkey_ev(dev, port, eqe);
1247 break;
1248 case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1249 /* paravirtualized master's guid is guid 0 -- does not change */
1250 if (!mlx4_is_master(dev->dev))
1251 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1252 /*if master, notify relevant slaves*/
1253 else if (!dev->sriov.is_going_down) {
1254 tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1255 change_bitmap = GET_MASK_FROM_EQE(eqe);
1256 handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1257 }
1258 break;
1259
1260 case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1261 /* cache sl to vl mapping changes for use in
1262 * filling QP1 LRH VL field when sending packets
1263 */
1264 if (!mlx4_is_slave(dev->dev)) {
1265 union sl2vl_tbl_to_u64 sl2vl64;
1266 int jj;
1267
1268 for (jj = 0; jj < 8; jj++) {
1269 sl2vl64.sl8[jj] =
1270 eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1271 pr_debug("port %u, sl2vl[%d] = %02x\n",
1272 port, jj, sl2vl64.sl8[jj]);
1273 }
1274 atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1275 }
1276 break;
1277 default:
1278 pr_warn("Unsupported subtype 0x%x for "
1279 "Port Management Change event\n", eqe->subtype);
1280 }
1281
1282 kfree(ew);
1283 }
1284
mlx4_ib_dispatch_event(struct mlx4_ib_dev * dev,u8 port_num,enum ib_event_type type)1285 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1286 enum ib_event_type type)
1287 {
1288 struct ib_event event;
1289
1290 event.device = &dev->ib_dev;
1291 event.element.port_num = port_num;
1292 event.event = type;
1293
1294 ib_dispatch_event(&event);
1295 }
1296
mlx4_ib_tunnel_comp_handler(struct ib_cq * cq,void * arg)1297 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1298 {
1299 unsigned long flags;
1300 struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1301 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1302 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1303 if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1304 queue_work(ctx->wq, &ctx->work);
1305 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1306 }
1307
mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx * ctx,struct mlx4_ib_demux_pv_qp * tun_qp,int index)1308 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1309 struct mlx4_ib_demux_pv_qp *tun_qp,
1310 int index)
1311 {
1312 struct ib_sge sg_list;
1313 struct ib_recv_wr recv_wr;
1314 const struct ib_recv_wr *bad_recv_wr;
1315 int size;
1316
1317 size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1318 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1319
1320 sg_list.addr = tun_qp->ring[index].map;
1321 sg_list.length = size;
1322 sg_list.lkey = ctx->pd->local_dma_lkey;
1323
1324 recv_wr.next = NULL;
1325 recv_wr.sg_list = &sg_list;
1326 recv_wr.num_sge = 1;
1327 recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1328 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1329 ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1330 size, DMA_FROM_DEVICE);
1331 return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1332 }
1333
mlx4_ib_multiplex_sa_handler(struct ib_device * ibdev,int port,int slave,struct ib_sa_mad * sa_mad)1334 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1335 int slave, struct ib_sa_mad *sa_mad)
1336 {
1337 int ret = 0;
1338
1339 /* dispatch to different sa handlers */
1340 switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1341 case IB_SA_ATTR_MC_MEMBER_REC:
1342 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1343 break;
1344 default:
1345 break;
1346 }
1347 return ret;
1348 }
1349
is_proxy_qp0(struct mlx4_ib_dev * dev,int qpn,int slave)1350 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1351 {
1352 int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1353
1354 return (qpn >= proxy_start && qpn <= proxy_start + 1);
1355 }
1356
1357
mlx4_ib_send_to_wire(struct mlx4_ib_dev * dev,int slave,u8 port,enum ib_qp_type dest_qpt,u16 pkey_index,u32 remote_qpn,u32 qkey,struct rdma_ah_attr * attr,u8 * s_mac,u16 vlan_id,struct ib_mad * mad)1358 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1359 enum ib_qp_type dest_qpt, u16 pkey_index,
1360 u32 remote_qpn, u32 qkey, struct rdma_ah_attr *attr,
1361 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1362 {
1363 struct ib_sge list;
1364 struct ib_ud_wr wr;
1365 const struct ib_send_wr *bad_wr;
1366 struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1367 struct mlx4_ib_demux_pv_qp *sqp;
1368 struct mlx4_mad_snd_buf *sqp_mad;
1369 struct ib_ah *ah;
1370 struct ib_qp *send_qp = NULL;
1371 unsigned wire_tx_ix = 0;
1372 int ret = 0;
1373 u16 wire_pkey_ix;
1374 int src_qpnum;
1375
1376 sqp_ctx = dev->sriov.sqps[port-1];
1377
1378 /* check if proxy qp created */
1379 if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1380 return -EAGAIN;
1381
1382 if (dest_qpt == IB_QPT_SMI) {
1383 src_qpnum = 0;
1384 sqp = &sqp_ctx->qp[0];
1385 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1386 } else {
1387 src_qpnum = 1;
1388 sqp = &sqp_ctx->qp[1];
1389 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1390 }
1391
1392 send_qp = sqp->qp;
1393
1394 /* create ah */
1395 ah = mlx4_ib_create_ah_slave(sqp_ctx->pd, attr,
1396 rdma_ah_retrieve_grh(attr)->sgid_index,
1397 s_mac, vlan_id);
1398 if (IS_ERR(ah))
1399 return -ENOMEM;
1400 spin_lock(&sqp->tx_lock);
1401 if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1402 (MLX4_NUM_TUNNEL_BUFS - 1))
1403 ret = -EAGAIN;
1404 else
1405 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1406 spin_unlock(&sqp->tx_lock);
1407 if (ret)
1408 goto out;
1409
1410 sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1411 if (sqp->tx_ring[wire_tx_ix].ah)
1412 rdma_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1413 sqp->tx_ring[wire_tx_ix].ah = ah;
1414 ib_dma_sync_single_for_cpu(&dev->ib_dev,
1415 sqp->tx_ring[wire_tx_ix].buf.map,
1416 sizeof (struct mlx4_mad_snd_buf),
1417 DMA_TO_DEVICE);
1418
1419 memcpy(&sqp_mad->payload, mad, sizeof *mad);
1420
1421 ib_dma_sync_single_for_device(&dev->ib_dev,
1422 sqp->tx_ring[wire_tx_ix].buf.map,
1423 sizeof (struct mlx4_mad_snd_buf),
1424 DMA_TO_DEVICE);
1425
1426 list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1427 list.length = sizeof (struct mlx4_mad_snd_buf);
1428 list.lkey = sqp_ctx->pd->local_dma_lkey;
1429
1430 wr.ah = ah;
1431 wr.port_num = port;
1432 wr.pkey_index = wire_pkey_ix;
1433 wr.remote_qkey = qkey;
1434 wr.remote_qpn = remote_qpn;
1435 wr.wr.next = NULL;
1436 wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1437 wr.wr.sg_list = &list;
1438 wr.wr.num_sge = 1;
1439 wr.wr.opcode = IB_WR_SEND;
1440 wr.wr.send_flags = IB_SEND_SIGNALED;
1441
1442 ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1443 if (!ret)
1444 return 0;
1445
1446 spin_lock(&sqp->tx_lock);
1447 sqp->tx_ix_tail++;
1448 spin_unlock(&sqp->tx_lock);
1449 sqp->tx_ring[wire_tx_ix].ah = NULL;
1450 out:
1451 mlx4_ib_destroy_ah(ah);
1452 return ret;
1453 }
1454
get_slave_base_gid_ix(struct mlx4_ib_dev * dev,int slave,int port)1455 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1456 {
1457 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1458 return slave;
1459 return mlx4_get_base_gid_ix(dev->dev, slave, port);
1460 }
1461
fill_in_real_sgid_index(struct mlx4_ib_dev * dev,int slave,int port,struct rdma_ah_attr * ah_attr)1462 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1463 struct rdma_ah_attr *ah_attr)
1464 {
1465 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
1466 if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1467 grh->sgid_index = slave;
1468 else
1469 grh->sgid_index += get_slave_base_gid_ix(dev, slave, port);
1470 }
1471
mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx * ctx,struct ib_wc * wc)1472 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1473 {
1474 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1475 struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1476 int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1477 struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1478 struct mlx4_ib_ah ah;
1479 struct rdma_ah_attr ah_attr;
1480 u8 *slave_id;
1481 int slave;
1482 int port;
1483 u16 vlan_id;
1484 u8 qos;
1485 u8 *dmac;
1486
1487 /* Get slave that sent this packet */
1488 if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1489 wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1490 (wc->src_qp & 0x1) != ctx->port - 1 ||
1491 wc->src_qp & 0x4) {
1492 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1493 return;
1494 }
1495 slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1496 if (slave != ctx->slave) {
1497 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1498 "belongs to another slave\n", wc->src_qp);
1499 return;
1500 }
1501
1502 /* Map transaction ID */
1503 ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1504 sizeof (struct mlx4_tunnel_mad),
1505 DMA_FROM_DEVICE);
1506 switch (tunnel->mad.mad_hdr.method) {
1507 case IB_MGMT_METHOD_SET:
1508 case IB_MGMT_METHOD_GET:
1509 case IB_MGMT_METHOD_REPORT:
1510 case IB_SA_METHOD_GET_TABLE:
1511 case IB_SA_METHOD_DELETE:
1512 case IB_SA_METHOD_GET_MULTI:
1513 case IB_SA_METHOD_GET_TRACE_TBL:
1514 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1515 if (*slave_id) {
1516 mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1517 "class:%d slave:%d\n", *slave_id,
1518 tunnel->mad.mad_hdr.mgmt_class, slave);
1519 return;
1520 } else
1521 *slave_id = slave;
1522 default:
1523 /* nothing */;
1524 }
1525
1526 /* Class-specific handling */
1527 switch (tunnel->mad.mad_hdr.mgmt_class) {
1528 case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1529 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1530 if (slave != mlx4_master_func_num(dev->dev) &&
1531 !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1532 return;
1533 break;
1534 case IB_MGMT_CLASS_SUBN_ADM:
1535 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1536 (struct ib_sa_mad *) &tunnel->mad))
1537 return;
1538 break;
1539 case IB_MGMT_CLASS_CM:
1540 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1541 (struct ib_mad *) &tunnel->mad))
1542 return;
1543 break;
1544 case IB_MGMT_CLASS_DEVICE_MGMT:
1545 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1546 tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1547 return;
1548 break;
1549 default:
1550 /* Drop unsupported classes for slaves in tunnel mode */
1551 if (slave != mlx4_master_func_num(dev->dev)) {
1552 mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1553 "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1554 return;
1555 }
1556 }
1557
1558 /* We are using standard ib_core services to send the mad, so generate a
1559 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1560 memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1561 ah.ibah.device = ctx->ib_dev;
1562
1563 port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1564 port = mlx4_slave_convert_port(dev->dev, slave, port);
1565 if (port < 0)
1566 return;
1567 ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1568 ah.ibah.type = rdma_ah_find_type(&dev->ib_dev, port);
1569
1570 mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1571 if (rdma_ah_get_ah_flags(&ah_attr) & IB_AH_GRH)
1572 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1573 dmac = rdma_ah_retrieve_dmac(&ah_attr);
1574 if (dmac)
1575 memcpy(dmac, tunnel->hdr.mac, ETH_ALEN);
1576 vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1577 /* if slave have default vlan use it */
1578 if (mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1579 &vlan_id, &qos))
1580 rdma_ah_set_sl(&ah_attr, qos);
1581
1582 mlx4_ib_send_to_wire(dev, slave, ctx->port,
1583 is_proxy_qp0(dev, wc->src_qp, slave) ?
1584 IB_QPT_SMI : IB_QPT_GSI,
1585 be16_to_cpu(tunnel->hdr.pkey_index),
1586 be32_to_cpu(tunnel->hdr.remote_qpn),
1587 be32_to_cpu(tunnel->hdr.qkey),
1588 &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1589 }
1590
mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx * ctx,enum ib_qp_type qp_type,int is_tun)1591 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1592 enum ib_qp_type qp_type, int is_tun)
1593 {
1594 int i;
1595 struct mlx4_ib_demux_pv_qp *tun_qp;
1596 int rx_buf_size, tx_buf_size;
1597
1598 if (qp_type > IB_QPT_GSI)
1599 return -EINVAL;
1600
1601 tun_qp = &ctx->qp[qp_type];
1602
1603 tun_qp->ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1604 sizeof(struct mlx4_ib_buf),
1605 GFP_KERNEL);
1606 if (!tun_qp->ring)
1607 return -ENOMEM;
1608
1609 tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1610 sizeof (struct mlx4_ib_tun_tx_buf),
1611 GFP_KERNEL);
1612 if (!tun_qp->tx_ring) {
1613 kfree(tun_qp->ring);
1614 tun_qp->ring = NULL;
1615 return -ENOMEM;
1616 }
1617
1618 if (is_tun) {
1619 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1620 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1621 } else {
1622 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1623 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1624 }
1625
1626 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1627 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1628 if (!tun_qp->ring[i].addr)
1629 goto err;
1630 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1631 tun_qp->ring[i].addr,
1632 rx_buf_size,
1633 DMA_FROM_DEVICE);
1634 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1635 kfree(tun_qp->ring[i].addr);
1636 goto err;
1637 }
1638 }
1639
1640 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1641 tun_qp->tx_ring[i].buf.addr =
1642 kmalloc(tx_buf_size, GFP_KERNEL);
1643 if (!tun_qp->tx_ring[i].buf.addr)
1644 goto tx_err;
1645 tun_qp->tx_ring[i].buf.map =
1646 ib_dma_map_single(ctx->ib_dev,
1647 tun_qp->tx_ring[i].buf.addr,
1648 tx_buf_size,
1649 DMA_TO_DEVICE);
1650 if (ib_dma_mapping_error(ctx->ib_dev,
1651 tun_qp->tx_ring[i].buf.map)) {
1652 kfree(tun_qp->tx_ring[i].buf.addr);
1653 goto tx_err;
1654 }
1655 tun_qp->tx_ring[i].ah = NULL;
1656 }
1657 spin_lock_init(&tun_qp->tx_lock);
1658 tun_qp->tx_ix_head = 0;
1659 tun_qp->tx_ix_tail = 0;
1660 tun_qp->proxy_qpt = qp_type;
1661
1662 return 0;
1663
1664 tx_err:
1665 while (i > 0) {
1666 --i;
1667 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1668 tx_buf_size, DMA_TO_DEVICE);
1669 kfree(tun_qp->tx_ring[i].buf.addr);
1670 }
1671 kfree(tun_qp->tx_ring);
1672 tun_qp->tx_ring = NULL;
1673 i = MLX4_NUM_TUNNEL_BUFS;
1674 err:
1675 while (i > 0) {
1676 --i;
1677 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1678 rx_buf_size, DMA_FROM_DEVICE);
1679 kfree(tun_qp->ring[i].addr);
1680 }
1681 kfree(tun_qp->ring);
1682 tun_qp->ring = NULL;
1683 return -ENOMEM;
1684 }
1685
mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx * ctx,enum ib_qp_type qp_type,int is_tun)1686 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1687 enum ib_qp_type qp_type, int is_tun)
1688 {
1689 int i;
1690 struct mlx4_ib_demux_pv_qp *tun_qp;
1691 int rx_buf_size, tx_buf_size;
1692
1693 if (qp_type > IB_QPT_GSI)
1694 return;
1695
1696 tun_qp = &ctx->qp[qp_type];
1697 if (is_tun) {
1698 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1699 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1700 } else {
1701 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1702 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1703 }
1704
1705
1706 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1707 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1708 rx_buf_size, DMA_FROM_DEVICE);
1709 kfree(tun_qp->ring[i].addr);
1710 }
1711
1712 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1713 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1714 tx_buf_size, DMA_TO_DEVICE);
1715 kfree(tun_qp->tx_ring[i].buf.addr);
1716 if (tun_qp->tx_ring[i].ah)
1717 rdma_destroy_ah(tun_qp->tx_ring[i].ah);
1718 }
1719 kfree(tun_qp->tx_ring);
1720 kfree(tun_qp->ring);
1721 }
1722
mlx4_ib_tunnel_comp_worker(struct work_struct * work)1723 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1724 {
1725 struct mlx4_ib_demux_pv_ctx *ctx;
1726 struct mlx4_ib_demux_pv_qp *tun_qp;
1727 struct ib_wc wc;
1728 int ret;
1729 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1730 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1731
1732 while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1733 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1734 if (wc.status == IB_WC_SUCCESS) {
1735 switch (wc.opcode) {
1736 case IB_WC_RECV:
1737 mlx4_ib_multiplex_mad(ctx, &wc);
1738 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1739 wc.wr_id &
1740 (MLX4_NUM_TUNNEL_BUFS - 1));
1741 if (ret)
1742 pr_err("Failed reposting tunnel "
1743 "buf:%lld\n", wc.wr_id);
1744 break;
1745 case IB_WC_SEND:
1746 pr_debug("received tunnel send completion:"
1747 "wrid=0x%llx, status=0x%x\n",
1748 wc.wr_id, wc.status);
1749 rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1750 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1751 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1752 = NULL;
1753 spin_lock(&tun_qp->tx_lock);
1754 tun_qp->tx_ix_tail++;
1755 spin_unlock(&tun_qp->tx_lock);
1756
1757 break;
1758 default:
1759 break;
1760 }
1761 } else {
1762 pr_debug("mlx4_ib: completion error in tunnel: %d."
1763 " status = %d, wrid = 0x%llx\n",
1764 ctx->slave, wc.status, wc.wr_id);
1765 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1766 rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1767 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1768 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1769 = NULL;
1770 spin_lock(&tun_qp->tx_lock);
1771 tun_qp->tx_ix_tail++;
1772 spin_unlock(&tun_qp->tx_lock);
1773 }
1774 }
1775 }
1776 }
1777
pv_qp_event_handler(struct ib_event * event,void * qp_context)1778 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1779 {
1780 struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1781
1782 /* It's worse than that! He's dead, Jim! */
1783 pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1784 event->event, sqp->port);
1785 }
1786
create_pv_sqp(struct mlx4_ib_demux_pv_ctx * ctx,enum ib_qp_type qp_type,int create_tun)1787 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1788 enum ib_qp_type qp_type, int create_tun)
1789 {
1790 int i, ret;
1791 struct mlx4_ib_demux_pv_qp *tun_qp;
1792 struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1793 struct ib_qp_attr attr;
1794 int qp_attr_mask_INIT;
1795
1796 if (qp_type > IB_QPT_GSI)
1797 return -EINVAL;
1798
1799 tun_qp = &ctx->qp[qp_type];
1800
1801 memset(&qp_init_attr, 0, sizeof qp_init_attr);
1802 qp_init_attr.init_attr.send_cq = ctx->cq;
1803 qp_init_attr.init_attr.recv_cq = ctx->cq;
1804 qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1805 qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1806 qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1807 qp_init_attr.init_attr.cap.max_send_sge = 1;
1808 qp_init_attr.init_attr.cap.max_recv_sge = 1;
1809 if (create_tun) {
1810 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1811 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1812 qp_init_attr.port = ctx->port;
1813 qp_init_attr.slave = ctx->slave;
1814 qp_init_attr.proxy_qp_type = qp_type;
1815 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1816 IB_QP_QKEY | IB_QP_PORT;
1817 } else {
1818 qp_init_attr.init_attr.qp_type = qp_type;
1819 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1820 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1821 }
1822 qp_init_attr.init_attr.port_num = ctx->port;
1823 qp_init_attr.init_attr.qp_context = ctx;
1824 qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1825 tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1826 if (IS_ERR(tun_qp->qp)) {
1827 ret = PTR_ERR(tun_qp->qp);
1828 tun_qp->qp = NULL;
1829 pr_err("Couldn't create %s QP (%d)\n",
1830 create_tun ? "tunnel" : "special", ret);
1831 return ret;
1832 }
1833
1834 memset(&attr, 0, sizeof attr);
1835 attr.qp_state = IB_QPS_INIT;
1836 ret = 0;
1837 if (create_tun)
1838 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1839 ctx->port, IB_DEFAULT_PKEY_FULL,
1840 &attr.pkey_index);
1841 if (ret || !create_tun)
1842 attr.pkey_index =
1843 to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1844 attr.qkey = IB_QP1_QKEY;
1845 attr.port_num = ctx->port;
1846 ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1847 if (ret) {
1848 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1849 create_tun ? "tunnel" : "special", ret);
1850 goto err_qp;
1851 }
1852 attr.qp_state = IB_QPS_RTR;
1853 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1854 if (ret) {
1855 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1856 create_tun ? "tunnel" : "special", ret);
1857 goto err_qp;
1858 }
1859 attr.qp_state = IB_QPS_RTS;
1860 attr.sq_psn = 0;
1861 ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1862 if (ret) {
1863 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1864 create_tun ? "tunnel" : "special", ret);
1865 goto err_qp;
1866 }
1867
1868 for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1869 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1870 if (ret) {
1871 pr_err(" mlx4_ib_post_pv_buf error"
1872 " (err = %d, i = %d)\n", ret, i);
1873 goto err_qp;
1874 }
1875 }
1876 return 0;
1877
1878 err_qp:
1879 ib_destroy_qp(tun_qp->qp);
1880 tun_qp->qp = NULL;
1881 return ret;
1882 }
1883
1884 /*
1885 * IB MAD completion callback for real SQPs
1886 */
mlx4_ib_sqp_comp_worker(struct work_struct * work)1887 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1888 {
1889 struct mlx4_ib_demux_pv_ctx *ctx;
1890 struct mlx4_ib_demux_pv_qp *sqp;
1891 struct ib_wc wc;
1892 struct ib_grh *grh;
1893 struct ib_mad *mad;
1894
1895 ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1896 ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1897
1898 while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1899 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1900 if (wc.status == IB_WC_SUCCESS) {
1901 switch (wc.opcode) {
1902 case IB_WC_SEND:
1903 rdma_destroy_ah(sqp->tx_ring[wc.wr_id &
1904 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1905 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1906 = NULL;
1907 spin_lock(&sqp->tx_lock);
1908 sqp->tx_ix_tail++;
1909 spin_unlock(&sqp->tx_lock);
1910 break;
1911 case IB_WC_RECV:
1912 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1913 (sqp->ring[wc.wr_id &
1914 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1915 grh = &(((struct mlx4_mad_rcv_buf *)
1916 (sqp->ring[wc.wr_id &
1917 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1918 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1919 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1920 (MLX4_NUM_TUNNEL_BUFS - 1)))
1921 pr_err("Failed reposting SQP "
1922 "buf:%lld\n", wc.wr_id);
1923 break;
1924 default:
1925 break;
1926 }
1927 } else {
1928 pr_debug("mlx4_ib: completion error in tunnel: %d."
1929 " status = %d, wrid = 0x%llx\n",
1930 ctx->slave, wc.status, wc.wr_id);
1931 if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1932 rdma_destroy_ah(sqp->tx_ring[wc.wr_id &
1933 (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1934 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1935 = NULL;
1936 spin_lock(&sqp->tx_lock);
1937 sqp->tx_ix_tail++;
1938 spin_unlock(&sqp->tx_lock);
1939 }
1940 }
1941 }
1942 }
1943
alloc_pv_object(struct mlx4_ib_dev * dev,int slave,int port,struct mlx4_ib_demux_pv_ctx ** ret_ctx)1944 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1945 struct mlx4_ib_demux_pv_ctx **ret_ctx)
1946 {
1947 struct mlx4_ib_demux_pv_ctx *ctx;
1948
1949 *ret_ctx = NULL;
1950 ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1951 if (!ctx)
1952 return -ENOMEM;
1953
1954 ctx->ib_dev = &dev->ib_dev;
1955 ctx->port = port;
1956 ctx->slave = slave;
1957 *ret_ctx = ctx;
1958 return 0;
1959 }
1960
free_pv_object(struct mlx4_ib_dev * dev,int slave,int port)1961 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1962 {
1963 if (dev->sriov.demux[port - 1].tun[slave]) {
1964 kfree(dev->sriov.demux[port - 1].tun[slave]);
1965 dev->sriov.demux[port - 1].tun[slave] = NULL;
1966 }
1967 }
1968
create_pv_resources(struct ib_device * ibdev,int slave,int port,int create_tun,struct mlx4_ib_demux_pv_ctx * ctx)1969 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1970 int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1971 {
1972 int ret, cq_size;
1973 struct ib_cq_init_attr cq_attr = {};
1974
1975 if (ctx->state != DEMUX_PV_STATE_DOWN)
1976 return -EEXIST;
1977
1978 ctx->state = DEMUX_PV_STATE_STARTING;
1979 /* have QP0 only if link layer is IB */
1980 if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1981 IB_LINK_LAYER_INFINIBAND)
1982 ctx->has_smi = 1;
1983
1984 if (ctx->has_smi) {
1985 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1986 if (ret) {
1987 pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1988 goto err_out;
1989 }
1990 }
1991
1992 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1993 if (ret) {
1994 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1995 goto err_out_qp0;
1996 }
1997
1998 cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1999 if (ctx->has_smi)
2000 cq_size *= 2;
2001
2002 cq_attr.cqe = cq_size;
2003 ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
2004 NULL, ctx, &cq_attr);
2005 if (IS_ERR(ctx->cq)) {
2006 ret = PTR_ERR(ctx->cq);
2007 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
2008 goto err_buf;
2009 }
2010
2011 ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
2012 if (IS_ERR(ctx->pd)) {
2013 ret = PTR_ERR(ctx->pd);
2014 pr_err("Couldn't create tunnel PD (%d)\n", ret);
2015 goto err_cq;
2016 }
2017
2018 if (ctx->has_smi) {
2019 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
2020 if (ret) {
2021 pr_err("Couldn't create %s QP0 (%d)\n",
2022 create_tun ? "tunnel for" : "", ret);
2023 goto err_pd;
2024 }
2025 }
2026
2027 ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
2028 if (ret) {
2029 pr_err("Couldn't create %s QP1 (%d)\n",
2030 create_tun ? "tunnel for" : "", ret);
2031 goto err_qp0;
2032 }
2033
2034 if (create_tun)
2035 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2036 else
2037 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2038
2039 ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2040
2041 ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2042 if (ret) {
2043 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2044 goto err_wq;
2045 }
2046 ctx->state = DEMUX_PV_STATE_ACTIVE;
2047 return 0;
2048
2049 err_wq:
2050 ctx->wq = NULL;
2051 ib_destroy_qp(ctx->qp[1].qp);
2052 ctx->qp[1].qp = NULL;
2053
2054
2055 err_qp0:
2056 if (ctx->has_smi)
2057 ib_destroy_qp(ctx->qp[0].qp);
2058 ctx->qp[0].qp = NULL;
2059
2060 err_pd:
2061 ib_dealloc_pd(ctx->pd);
2062 ctx->pd = NULL;
2063
2064 err_cq:
2065 ib_destroy_cq(ctx->cq);
2066 ctx->cq = NULL;
2067
2068 err_buf:
2069 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2070
2071 err_out_qp0:
2072 if (ctx->has_smi)
2073 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2074 err_out:
2075 ctx->state = DEMUX_PV_STATE_DOWN;
2076 return ret;
2077 }
2078
destroy_pv_resources(struct mlx4_ib_dev * dev,int slave,int port,struct mlx4_ib_demux_pv_ctx * ctx,int flush)2079 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2080 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2081 {
2082 if (!ctx)
2083 return;
2084 if (ctx->state > DEMUX_PV_STATE_DOWN) {
2085 ctx->state = DEMUX_PV_STATE_DOWNING;
2086 if (flush)
2087 flush_workqueue(ctx->wq);
2088 if (ctx->has_smi) {
2089 ib_destroy_qp(ctx->qp[0].qp);
2090 ctx->qp[0].qp = NULL;
2091 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2092 }
2093 ib_destroy_qp(ctx->qp[1].qp);
2094 ctx->qp[1].qp = NULL;
2095 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2096 ib_dealloc_pd(ctx->pd);
2097 ctx->pd = NULL;
2098 ib_destroy_cq(ctx->cq);
2099 ctx->cq = NULL;
2100 ctx->state = DEMUX_PV_STATE_DOWN;
2101 }
2102 }
2103
mlx4_ib_tunnels_update(struct mlx4_ib_dev * dev,int slave,int port,int do_init)2104 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2105 int port, int do_init)
2106 {
2107 int ret = 0;
2108
2109 if (!do_init) {
2110 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2111 /* for master, destroy real sqp resources */
2112 if (slave == mlx4_master_func_num(dev->dev))
2113 destroy_pv_resources(dev, slave, port,
2114 dev->sriov.sqps[port - 1], 1);
2115 /* destroy the tunnel qp resources */
2116 destroy_pv_resources(dev, slave, port,
2117 dev->sriov.demux[port - 1].tun[slave], 1);
2118 return 0;
2119 }
2120
2121 /* create the tunnel qp resources */
2122 ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2123 dev->sriov.demux[port - 1].tun[slave]);
2124
2125 /* for master, create the real sqp resources */
2126 if (!ret && slave == mlx4_master_func_num(dev->dev))
2127 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2128 dev->sriov.sqps[port - 1]);
2129 return ret;
2130 }
2131
mlx4_ib_tunnels_update_work(struct work_struct * work)2132 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2133 {
2134 struct mlx4_ib_demux_work *dmxw;
2135
2136 dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2137 mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2138 dmxw->do_init);
2139 kfree(dmxw);
2140 return;
2141 }
2142
mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev * dev,struct mlx4_ib_demux_ctx * ctx,int port)2143 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2144 struct mlx4_ib_demux_ctx *ctx,
2145 int port)
2146 {
2147 char name[12];
2148 int ret = 0;
2149 int i;
2150
2151 ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2152 sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2153 if (!ctx->tun)
2154 return -ENOMEM;
2155
2156 ctx->dev = dev;
2157 ctx->port = port;
2158 ctx->ib_dev = &dev->ib_dev;
2159
2160 for (i = 0;
2161 i < min(dev->dev->caps.sqp_demux,
2162 (u16)(dev->dev->persist->num_vfs + 1));
2163 i++) {
2164 struct mlx4_active_ports actv_ports =
2165 mlx4_get_active_ports(dev->dev, i);
2166
2167 if (!test_bit(port - 1, actv_ports.ports))
2168 continue;
2169
2170 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2171 if (ret) {
2172 ret = -ENOMEM;
2173 goto err_mcg;
2174 }
2175 }
2176
2177 ret = mlx4_ib_mcg_port_init(ctx);
2178 if (ret) {
2179 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2180 goto err_mcg;
2181 }
2182
2183 snprintf(name, sizeof name, "mlx4_ibt%d", port);
2184 ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2185 if (!ctx->wq) {
2186 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2187 ret = -ENOMEM;
2188 goto err_wq;
2189 }
2190
2191 snprintf(name, sizeof name, "mlx4_ibud%d", port);
2192 ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2193 if (!ctx->ud_wq) {
2194 pr_err("Failed to create up/down WQ for port %d\n", port);
2195 ret = -ENOMEM;
2196 goto err_udwq;
2197 }
2198
2199 return 0;
2200
2201 err_udwq:
2202 destroy_workqueue(ctx->wq);
2203 ctx->wq = NULL;
2204
2205 err_wq:
2206 mlx4_ib_mcg_port_cleanup(ctx, 1);
2207 err_mcg:
2208 for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2209 free_pv_object(dev, i, port);
2210 kfree(ctx->tun);
2211 ctx->tun = NULL;
2212 return ret;
2213 }
2214
mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx * sqp_ctx)2215 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2216 {
2217 if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2218 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2219 flush_workqueue(sqp_ctx->wq);
2220 if (sqp_ctx->has_smi) {
2221 ib_destroy_qp(sqp_ctx->qp[0].qp);
2222 sqp_ctx->qp[0].qp = NULL;
2223 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2224 }
2225 ib_destroy_qp(sqp_ctx->qp[1].qp);
2226 sqp_ctx->qp[1].qp = NULL;
2227 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2228 ib_dealloc_pd(sqp_ctx->pd);
2229 sqp_ctx->pd = NULL;
2230 ib_destroy_cq(sqp_ctx->cq);
2231 sqp_ctx->cq = NULL;
2232 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2233 }
2234 }
2235
mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx * ctx)2236 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2237 {
2238 int i;
2239 if (ctx) {
2240 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2241 mlx4_ib_mcg_port_cleanup(ctx, 1);
2242 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2243 if (!ctx->tun[i])
2244 continue;
2245 if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2246 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2247 }
2248 flush_workqueue(ctx->wq);
2249 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2250 destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2251 free_pv_object(dev, i, ctx->port);
2252 }
2253 kfree(ctx->tun);
2254 destroy_workqueue(ctx->ud_wq);
2255 destroy_workqueue(ctx->wq);
2256 }
2257 }
2258
mlx4_ib_master_tunnels(struct mlx4_ib_dev * dev,int do_init)2259 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2260 {
2261 int i;
2262
2263 if (!mlx4_is_master(dev->dev))
2264 return;
2265 /* initialize or tear down tunnel QPs for the master */
2266 for (i = 0; i < dev->dev->caps.num_ports; i++)
2267 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2268 return;
2269 }
2270
mlx4_ib_init_sriov(struct mlx4_ib_dev * dev)2271 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2272 {
2273 int i = 0;
2274 int err;
2275
2276 if (!mlx4_is_mfunc(dev->dev))
2277 return 0;
2278
2279 dev->sriov.is_going_down = 0;
2280 spin_lock_init(&dev->sriov.going_down_lock);
2281 mlx4_ib_cm_paravirt_init(dev);
2282
2283 mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2284
2285 if (mlx4_is_slave(dev->dev)) {
2286 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2287 return 0;
2288 }
2289
2290 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2291 if (i == mlx4_master_func_num(dev->dev))
2292 mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2293 else
2294 mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2295 }
2296
2297 err = mlx4_ib_init_alias_guid_service(dev);
2298 if (err) {
2299 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2300 goto paravirt_err;
2301 }
2302 err = mlx4_ib_device_register_sysfs(dev);
2303 if (err) {
2304 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2305 goto sysfs_err;
2306 }
2307
2308 mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2309 dev->dev->caps.sqp_demux);
2310 for (i = 0; i < dev->num_ports; i++) {
2311 union ib_gid gid;
2312 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2313 if (err)
2314 goto demux_err;
2315 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2316 atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2317 be64_to_cpu(gid.global.subnet_prefix));
2318 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2319 &dev->sriov.sqps[i]);
2320 if (err)
2321 goto demux_err;
2322 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2323 if (err)
2324 goto free_pv;
2325 }
2326 mlx4_ib_master_tunnels(dev, 1);
2327 return 0;
2328
2329 free_pv:
2330 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2331 demux_err:
2332 while (--i >= 0) {
2333 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2334 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2335 }
2336 mlx4_ib_device_unregister_sysfs(dev);
2337
2338 sysfs_err:
2339 mlx4_ib_destroy_alias_guid_service(dev);
2340
2341 paravirt_err:
2342 mlx4_ib_cm_paravirt_clean(dev, -1);
2343
2344 return err;
2345 }
2346
mlx4_ib_close_sriov(struct mlx4_ib_dev * dev)2347 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2348 {
2349 int i;
2350 unsigned long flags;
2351
2352 if (!mlx4_is_mfunc(dev->dev))
2353 return;
2354
2355 spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2356 dev->sriov.is_going_down = 1;
2357 spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2358 if (mlx4_is_master(dev->dev)) {
2359 for (i = 0; i < dev->num_ports; i++) {
2360 flush_workqueue(dev->sriov.demux[i].ud_wq);
2361 mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2362 kfree(dev->sriov.sqps[i]);
2363 dev->sriov.sqps[i] = NULL;
2364 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2365 }
2366
2367 mlx4_ib_cm_paravirt_clean(dev, -1);
2368 mlx4_ib_destroy_alias_guid_service(dev);
2369 mlx4_ib_device_unregister_sysfs(dev);
2370 }
2371 }
2372
