2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <rdma/ib_mad.h>
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_sa.h>
36 #include <rdma/ib_cache.h>
38 #include <linux/random.h>
39 #include <linux/mlx4/cmd.h>
40 #include <linux/gfp.h>
41 #include <rdma/ib_pma.h>
46 MLX4_IB_VENDOR_CLASS1
= 0x9,
47 MLX4_IB_VENDOR_CLASS2
= 0xa
50 #define MLX4_TUN_SEND_WRID_SHIFT 34
51 #define MLX4_TUN_QPN_SHIFT 32
52 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
53 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
55 #define MLX4_TUN_IS_RECV(a) (((a) >> MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
56 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
58 /* Port mgmt change event handling */
60 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
61 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
62 #define NUM_IDX_IN_PKEY_TBL_BLK 32
63 #define GUID_TBL_ENTRY_SIZE 8 /* size in bytes */
64 #define GUID_TBL_BLK_NUM_ENTRIES 8
65 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
67 /* Counters should be saturate once they reach their maximum value */
68 #define ASSIGN_32BIT_COUNTER(counter, value) do {\
69 if ((value) > U32_MAX) \
70 counter = cpu_to_be32(U32_MAX); \
72 counter = cpu_to_be32(value); \
75 struct mlx4_mad_rcv_buf
{
80 struct mlx4_mad_snd_buf
{
84 struct mlx4_tunnel_mad
{
86 struct mlx4_ib_tunnel_header hdr
;
90 struct mlx4_rcv_tunnel_mad
{
91 struct mlx4_rcv_tunnel_hdr hdr
;
96 static void handle_client_rereg_event(struct mlx4_ib_dev
*dev
, u8 port_num
);
97 static void handle_lid_change_event(struct mlx4_ib_dev
*dev
, u8 port_num
);
98 static void __propagate_pkey_ev(struct mlx4_ib_dev
*dev
, int port_num
,
99 int block
, u32 change_bitmap
);
101 __be64
mlx4_ib_gen_node_guid(void)
103 #define NODE_GUID_HI ((u64) (((u64)IB_OPENIB_OUI) << 40))
104 return cpu_to_be64(NODE_GUID_HI
| prandom_u32());
107 __be64
mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx
*ctx
)
109 return cpu_to_be64(atomic_inc_return(&ctx
->tid
)) |
110 cpu_to_be64(0xff00000000000000LL
);
113 int mlx4_MAD_IFC(struct mlx4_ib_dev
*dev
, int mad_ifc_flags
,
114 int port
, const struct ib_wc
*in_wc
,
115 const struct ib_grh
*in_grh
,
116 const void *in_mad
, void *response_mad
)
118 struct mlx4_cmd_mailbox
*inmailbox
, *outmailbox
;
121 u32 in_modifier
= port
;
124 inmailbox
= mlx4_alloc_cmd_mailbox(dev
->dev
);
125 if (IS_ERR(inmailbox
))
126 return PTR_ERR(inmailbox
);
127 inbox
= inmailbox
->buf
;
129 outmailbox
= mlx4_alloc_cmd_mailbox(dev
->dev
);
130 if (IS_ERR(outmailbox
)) {
131 mlx4_free_cmd_mailbox(dev
->dev
, inmailbox
);
132 return PTR_ERR(outmailbox
);
135 memcpy(inbox
, in_mad
, 256);
138 * Key check traps can't be generated unless we have in_wc to
139 * tell us where to send the trap.
141 if ((mad_ifc_flags
& MLX4_MAD_IFC_IGNORE_MKEY
) || !in_wc
)
143 if ((mad_ifc_flags
& MLX4_MAD_IFC_IGNORE_BKEY
) || !in_wc
)
145 if (mlx4_is_mfunc(dev
->dev
) &&
146 (mad_ifc_flags
& MLX4_MAD_IFC_NET_VIEW
|| in_wc
))
162 memset(inbox
+ 256, 0, 256);
163 ext_info
= inbox
+ 256;
165 ext_info
->my_qpn
= cpu_to_be32(in_wc
->qp
->qp_num
);
166 ext_info
->rqpn
= cpu_to_be32(in_wc
->src_qp
);
167 ext_info
->sl
= in_wc
->sl
<< 4;
168 ext_info
->g_path
= in_wc
->dlid_path_bits
|
169 (in_wc
->wc_flags
& IB_WC_GRH
? 0x80 : 0);
170 ext_info
->pkey
= cpu_to_be16(in_wc
->pkey_index
);
173 memcpy(ext_info
->grh
, in_grh
, 40);
177 in_modifier
|= in_wc
->slid
<< 16;
180 err
= mlx4_cmd_box(dev
->dev
, inmailbox
->dma
, outmailbox
->dma
, in_modifier
,
181 mlx4_is_master(dev
->dev
) ? (op_modifier
& ~0x8) : op_modifier
,
182 MLX4_CMD_MAD_IFC
, MLX4_CMD_TIME_CLASS_C
,
183 (op_modifier
& 0x8) ? MLX4_CMD_NATIVE
: MLX4_CMD_WRAPPED
);
186 memcpy(response_mad
, outmailbox
->buf
, 256);
188 mlx4_free_cmd_mailbox(dev
->dev
, inmailbox
);
189 mlx4_free_cmd_mailbox(dev
->dev
, outmailbox
);
194 static void update_sm_ah(struct mlx4_ib_dev
*dev
, u8 port_num
, u16 lid
, u8 sl
)
196 struct ib_ah
*new_ah
;
197 struct ib_ah_attr ah_attr
;
200 if (!dev
->send_agent
[port_num
- 1][0])
203 memset(&ah_attr
, 0, sizeof ah_attr
);
206 ah_attr
.port_num
= port_num
;
208 new_ah
= ib_create_ah(dev
->send_agent
[port_num
- 1][0]->qp
->pd
,
213 spin_lock_irqsave(&dev
->sm_lock
, flags
);
214 if (dev
->sm_ah
[port_num
- 1])
215 ib_destroy_ah(dev
->sm_ah
[port_num
- 1]);
216 dev
->sm_ah
[port_num
- 1] = new_ah
;
217 spin_unlock_irqrestore(&dev
->sm_lock
, flags
);
221 * Snoop SM MADs for port info, GUID info, and P_Key table sets, so we can
222 * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
224 static void smp_snoop(struct ib_device
*ibdev
, u8 port_num
, const struct ib_mad
*mad
,
227 struct ib_port_info
*pinfo
;
230 u32 bn
, pkey_change_bitmap
;
234 struct mlx4_ib_dev
*dev
= to_mdev(ibdev
);
235 if ((mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_LID_ROUTED
||
236 mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
) &&
237 mad
->mad_hdr
.method
== IB_MGMT_METHOD_SET
)
238 switch (mad
->mad_hdr
.attr_id
) {
239 case IB_SMP_ATTR_PORT_INFO
:
240 pinfo
= (struct ib_port_info
*) ((struct ib_smp
*) mad
)->data
;
241 lid
= be16_to_cpu(pinfo
->lid
);
243 update_sm_ah(dev
, port_num
,
244 be16_to_cpu(pinfo
->sm_lid
),
245 pinfo
->neighbormtu_mastersmsl
& 0xf);
247 if (pinfo
->clientrereg_resv_subnetto
& 0x80)
248 handle_client_rereg_event(dev
, port_num
);
251 handle_lid_change_event(dev
, port_num
);
254 case IB_SMP_ATTR_PKEY_TABLE
:
255 if (!mlx4_is_mfunc(dev
->dev
)) {
256 mlx4_ib_dispatch_event(dev
, port_num
,
257 IB_EVENT_PKEY_CHANGE
);
261 /* at this point, we are running in the master.
262 * Slaves do not receive SMPs.
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
]);
277 pr_debug("PKEY Change event: port=%d, "
278 "block=0x%x, change_bitmap=0x%x\n",
279 port_num
, bn
, pkey_change_bitmap
);
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
,
290 case IB_SMP_ATTR_GUID_INFO
:
291 /* paravirtualized master's guid is guid 0 -- does not change */
292 if (!mlx4_is_master(dev
->dev
))
293 mlx4_ib_dispatch_event(dev
, port_num
,
294 IB_EVENT_GID_CHANGE
);
295 /*if master, notify relevant slaves*/
296 if (mlx4_is_master(dev
->dev
) &&
297 !dev
->sriov
.is_going_down
) {
298 bn
= be32_to_cpu(((struct ib_smp
*)mad
)->attr_mod
);
299 mlx4_ib_update_cache_on_guid_change(dev
, bn
, port_num
,
300 (u8
*)(&((struct ib_smp
*)mad
)->data
));
301 mlx4_ib_notify_slaves_on_guid_change(dev
, bn
, port_num
,
302 (u8
*)(&((struct ib_smp
*)mad
)->data
));
311 static void __propagate_pkey_ev(struct mlx4_ib_dev
*dev
, int port_num
,
312 int block
, u32 change_bitmap
)
314 int i
, ix
, slave
, err
;
317 for (slave
= 0; slave
< dev
->dev
->caps
.sqp_demux
; slave
++) {
318 if (slave
== mlx4_master_func_num(dev
->dev
))
320 if (!mlx4_is_slave_active(dev
->dev
, slave
))
324 for (i
= 0; i
< 32; i
++) {
325 if (!(change_bitmap
& (1 << i
)))
328 ix
< dev
->dev
->caps
.pkey_table_len
[port_num
]; ix
++) {
329 if (dev
->pkeys
.virt2phys_pkey
[slave
][port_num
- 1]
330 [ix
] == i
+ 32 * block
) {
331 err
= mlx4_gen_pkey_eqe(dev
->dev
, slave
, port_num
);
332 pr_debug("propagate_pkey_ev: slave %d,"
333 " port %d, ix %d (%d)\n",
334 slave
, port_num
, ix
, err
);
345 static void node_desc_override(struct ib_device
*dev
,
350 if ((mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_LID_ROUTED
||
351 mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
) &&
352 mad
->mad_hdr
.method
== IB_MGMT_METHOD_GET_RESP
&&
353 mad
->mad_hdr
.attr_id
== IB_SMP_ATTR_NODE_DESC
) {
354 spin_lock_irqsave(&to_mdev(dev
)->sm_lock
, flags
);
355 memcpy(((struct ib_smp
*) mad
)->data
, dev
->node_desc
, 64);
356 spin_unlock_irqrestore(&to_mdev(dev
)->sm_lock
, flags
);
360 static void forward_trap(struct mlx4_ib_dev
*dev
, u8 port_num
, const struct ib_mad
*mad
)
362 int qpn
= mad
->mad_hdr
.mgmt_class
!= IB_MGMT_CLASS_SUBN_LID_ROUTED
;
363 struct ib_mad_send_buf
*send_buf
;
364 struct ib_mad_agent
*agent
= dev
->send_agent
[port_num
- 1][qpn
];
369 send_buf
= ib_create_send_mad(agent
, qpn
, 0, 0, IB_MGMT_MAD_HDR
,
370 IB_MGMT_MAD_DATA
, GFP_ATOMIC
,
371 IB_MGMT_BASE_VERSION
);
372 if (IS_ERR(send_buf
))
375 * We rely here on the fact that MLX QPs don't use the
376 * address handle after the send is posted (this is
377 * wrong following the IB spec strictly, but we know
378 * it's OK for our devices).
380 spin_lock_irqsave(&dev
->sm_lock
, flags
);
381 memcpy(send_buf
->mad
, mad
, sizeof *mad
);
382 if ((send_buf
->ah
= dev
->sm_ah
[port_num
- 1]))
383 ret
= ib_post_send_mad(send_buf
, NULL
);
386 spin_unlock_irqrestore(&dev
->sm_lock
, flags
);
389 ib_free_send_mad(send_buf
);
393 static int mlx4_ib_demux_sa_handler(struct ib_device
*ibdev
, int port
, int slave
,
394 struct ib_sa_mad
*sa_mad
)
398 /* dispatch to different sa handlers */
399 switch (be16_to_cpu(sa_mad
->mad_hdr
.attr_id
)) {
400 case IB_SA_ATTR_MC_MEMBER_REC
:
401 ret
= mlx4_ib_mcg_demux_handler(ibdev
, port
, slave
, sa_mad
);
409 int mlx4_ib_find_real_gid(struct ib_device
*ibdev
, u8 port
, __be64 guid
)
411 struct mlx4_ib_dev
*dev
= to_mdev(ibdev
);
414 for (i
= 0; i
< dev
->dev
->caps
.sqp_demux
; i
++) {
415 if (dev
->sriov
.demux
[port
- 1].guid_cache
[i
] == guid
)
422 static int find_slave_port_pkey_ix(struct mlx4_ib_dev
*dev
, int slave
,
423 u8 port
, u16 pkey
, u16
*ix
)
426 u8 unassigned_pkey_ix
, pkey_ix
, partial_ix
= 0xFF;
429 if (slave
== mlx4_master_func_num(dev
->dev
))
430 return ib_find_cached_pkey(&dev
->ib_dev
, port
, pkey
, ix
);
432 unassigned_pkey_ix
= dev
->dev
->phys_caps
.pkey_phys_table_len
[port
] - 1;
434 for (i
= 0; i
< dev
->dev
->caps
.pkey_table_len
[port
]; i
++) {
435 if (dev
->pkeys
.virt2phys_pkey
[slave
][port
- 1][i
] == unassigned_pkey_ix
)
438 pkey_ix
= dev
->pkeys
.virt2phys_pkey
[slave
][port
- 1][i
];
440 ret
= ib_get_cached_pkey(&dev
->ib_dev
, port
, pkey_ix
, &slot_pkey
);
443 if ((slot_pkey
& 0x7FFF) == (pkey
& 0x7FFF)) {
444 if (slot_pkey
& 0x8000) {
448 /* take first partial pkey index found */
449 if (partial_ix
== 0xFF)
450 partial_ix
= pkey_ix
;
455 if (partial_ix
< 0xFF) {
456 *ix
= (u16
) partial_ix
;
463 int mlx4_ib_send_to_slave(struct mlx4_ib_dev
*dev
, int slave
, u8 port
,
464 enum ib_qp_type dest_qpt
, struct ib_wc
*wc
,
465 struct ib_grh
*grh
, struct ib_mad
*mad
)
468 struct ib_send_wr wr
, *bad_wr
;
469 struct mlx4_ib_demux_pv_ctx
*tun_ctx
;
470 struct mlx4_ib_demux_pv_qp
*tun_qp
;
471 struct mlx4_rcv_tunnel_mad
*tun_mad
;
472 struct ib_ah_attr attr
;
474 struct ib_qp
*src_qp
= NULL
;
475 unsigned tun_tx_ix
= 0;
480 u8 is_eth
= dev
->dev
->caps
.port_type
[port
] == MLX4_PORT_TYPE_ETH
;
482 if (dest_qpt
> IB_QPT_GSI
)
485 tun_ctx
= dev
->sriov
.demux
[port
-1].tun
[slave
];
487 /* check if proxy qp created */
488 if (!tun_ctx
|| tun_ctx
->state
!= DEMUX_PV_STATE_ACTIVE
)
492 tun_qp
= &tun_ctx
->qp
[0];
494 tun_qp
= &tun_ctx
->qp
[1];
496 /* compute P_Key index to put in tunnel header for slave */
499 ret
= ib_get_cached_pkey(&dev
->ib_dev
, port
, wc
->pkey_index
, &cached_pkey
);
503 ret
= find_slave_port_pkey_ix(dev
, slave
, port
, cached_pkey
, &pkey_ix
);
506 tun_pkey_ix
= pkey_ix
;
508 tun_pkey_ix
= dev
->pkeys
.virt2phys_pkey
[slave
][port
- 1][0];
510 dqpn
= dev
->dev
->phys_caps
.base_proxy_sqpn
+ 8 * slave
+ port
+ (dest_qpt
* 2) - 1;
512 /* get tunnel tx data buf for slave */
515 /* create ah. Just need an empty one with the port num for the post send.
516 * The driver will set the force loopback bit in post_send */
517 memset(&attr
, 0, sizeof attr
);
518 attr
.port_num
= port
;
520 memcpy(&attr
.grh
.dgid
.raw
[0], &grh
->dgid
.raw
[0], 16);
521 attr
.ah_flags
= IB_AH_GRH
;
523 ah
= ib_create_ah(tun_ctx
->pd
, &attr
);
527 /* allocate tunnel tx buf after pass failure returns */
528 spin_lock(&tun_qp
->tx_lock
);
529 if (tun_qp
->tx_ix_head
- tun_qp
->tx_ix_tail
>=
530 (MLX4_NUM_TUNNEL_BUFS
- 1))
533 tun_tx_ix
= (++tun_qp
->tx_ix_head
) & (MLX4_NUM_TUNNEL_BUFS
- 1);
534 spin_unlock(&tun_qp
->tx_lock
);
538 tun_mad
= (struct mlx4_rcv_tunnel_mad
*) (tun_qp
->tx_ring
[tun_tx_ix
].buf
.addr
);
539 if (tun_qp
->tx_ring
[tun_tx_ix
].ah
)
540 ib_destroy_ah(tun_qp
->tx_ring
[tun_tx_ix
].ah
);
541 tun_qp
->tx_ring
[tun_tx_ix
].ah
= ah
;
542 ib_dma_sync_single_for_cpu(&dev
->ib_dev
,
543 tun_qp
->tx_ring
[tun_tx_ix
].buf
.map
,
544 sizeof (struct mlx4_rcv_tunnel_mad
),
547 /* copy over to tunnel buffer */
549 memcpy(&tun_mad
->grh
, grh
, sizeof *grh
);
550 memcpy(&tun_mad
->mad
, mad
, sizeof *mad
);
552 /* adjust tunnel data */
553 tun_mad
->hdr
.pkey_index
= cpu_to_be16(tun_pkey_ix
);
554 tun_mad
->hdr
.flags_src_qp
= cpu_to_be32(wc
->src_qp
& 0xFFFFFF);
555 tun_mad
->hdr
.g_ml_path
= (grh
&& (wc
->wc_flags
& IB_WC_GRH
)) ? 0x80 : 0;
559 if (mlx4_get_slave_default_vlan(dev
->dev
, port
, slave
, &vlan
,
562 if (vlan
!= wc
->vlan_id
)
563 /* Packet vlan is not the VST-assigned vlan.
568 /* Remove the vlan tag before forwarding
569 * the packet to the VF.
576 tun_mad
->hdr
.sl_vid
= cpu_to_be16(vlan
);
577 memcpy((char *)&tun_mad
->hdr
.mac_31_0
, &(wc
->smac
[0]), 4);
578 memcpy((char *)&tun_mad
->hdr
.slid_mac_47_32
, &(wc
->smac
[4]), 2);
580 tun_mad
->hdr
.sl_vid
= cpu_to_be16(((u16
)(wc
->sl
)) << 12);
581 tun_mad
->hdr
.slid_mac_47_32
= cpu_to_be16(wc
->slid
);
584 ib_dma_sync_single_for_device(&dev
->ib_dev
,
585 tun_qp
->tx_ring
[tun_tx_ix
].buf
.map
,
586 sizeof (struct mlx4_rcv_tunnel_mad
),
589 list
.addr
= tun_qp
->tx_ring
[tun_tx_ix
].buf
.map
;
590 list
.length
= sizeof (struct mlx4_rcv_tunnel_mad
);
591 list
.lkey
= tun_ctx
->mr
->lkey
;
594 wr
.wr
.ud
.port_num
= port
;
595 wr
.wr
.ud
.remote_qkey
= IB_QP_SET_QKEY
;
596 wr
.wr
.ud
.remote_qpn
= dqpn
;
598 wr
.wr_id
= ((u64
) tun_tx_ix
) | MLX4_TUN_SET_WRID_QPN(dest_qpt
);
601 wr
.opcode
= IB_WR_SEND
;
602 wr
.send_flags
= IB_SEND_SIGNALED
;
604 ret
= ib_post_send(src_qp
, &wr
, &bad_wr
);
611 static int mlx4_ib_demux_mad(struct ib_device
*ibdev
, u8 port
,
612 struct ib_wc
*wc
, struct ib_grh
*grh
,
615 struct mlx4_ib_dev
*dev
= to_mdev(ibdev
);
621 if (rdma_port_get_link_layer(ibdev
, port
) == IB_LINK_LAYER_INFINIBAND
)
627 if (!(wc
->wc_flags
& IB_WC_GRH
)) {
628 mlx4_ib_warn(ibdev
, "RoCE grh not present.\n");
631 if (mad
->mad_hdr
.mgmt_class
!= IB_MGMT_CLASS_CM
) {
632 mlx4_ib_warn(ibdev
, "RoCE mgmt class is not CM\n");
635 if (mlx4_get_slave_from_roce_gid(dev
->dev
, port
, grh
->dgid
.raw
, &slave
)) {
636 mlx4_ib_warn(ibdev
, "failed matching grh\n");
639 if (slave
>= dev
->dev
->caps
.sqp_demux
) {
640 mlx4_ib_warn(ibdev
, "slave id: %d is bigger than allowed:%d\n",
641 slave
, dev
->dev
->caps
.sqp_demux
);
645 if (mlx4_ib_demux_cm_handler(ibdev
, port
, NULL
, mad
))
648 err
= mlx4_ib_send_to_slave(dev
, slave
, port
, wc
->qp
->qp_type
, wc
, grh
, mad
);
650 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
655 /* Initially assume that this mad is for us */
656 slave
= mlx4_master_func_num(dev
->dev
);
658 /* See if the slave id is encoded in a response mad */
659 if (mad
->mad_hdr
.method
& 0x80) {
660 slave_id
= (u8
*) &mad
->mad_hdr
.tid
;
662 if (slave
!= 255) /*255 indicates the dom0*/
663 *slave_id
= 0; /* remap tid */
666 /* If a grh is present, we demux according to it */
667 if (wc
->wc_flags
& IB_WC_GRH
) {
668 slave
= mlx4_ib_find_real_gid(ibdev
, port
, grh
->dgid
.global
.interface_id
);
670 mlx4_ib_warn(ibdev
, "failed matching grh\n");
674 /* Class-specific handling */
675 switch (mad
->mad_hdr
.mgmt_class
) {
676 case IB_MGMT_CLASS_SUBN_LID_ROUTED
:
677 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
:
678 /* 255 indicates the dom0 */
679 if (slave
!= 255 && slave
!= mlx4_master_func_num(dev
->dev
)) {
680 if (!mlx4_vf_smi_enabled(dev
->dev
, slave
, port
))
682 /* for a VF. drop unsolicited MADs */
683 if (!(mad
->mad_hdr
.method
& IB_MGMT_METHOD_RESP
)) {
684 mlx4_ib_warn(ibdev
, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
685 slave
, mad
->mad_hdr
.mgmt_class
,
686 mad
->mad_hdr
.method
);
691 case IB_MGMT_CLASS_SUBN_ADM
:
692 if (mlx4_ib_demux_sa_handler(ibdev
, port
, slave
,
693 (struct ib_sa_mad
*) mad
))
696 case IB_MGMT_CLASS_CM
:
697 if (mlx4_ib_demux_cm_handler(ibdev
, port
, &slave
, mad
))
700 case IB_MGMT_CLASS_DEVICE_MGMT
:
701 if (mad
->mad_hdr
.method
!= IB_MGMT_METHOD_GET_RESP
)
705 /* Drop unsupported classes for slaves in tunnel mode */
706 if (slave
!= mlx4_master_func_num(dev
->dev
)) {
707 pr_debug("dropping unsupported ingress mad from class:%d "
708 "for slave:%d\n", mad
->mad_hdr
.mgmt_class
, slave
);
712 /*make sure that no slave==255 was not handled yet.*/
713 if (slave
>= dev
->dev
->caps
.sqp_demux
) {
714 mlx4_ib_warn(ibdev
, "slave id: %d is bigger than allowed:%d\n",
715 slave
, dev
->dev
->caps
.sqp_demux
);
719 err
= mlx4_ib_send_to_slave(dev
, slave
, port
, wc
->qp
->qp_type
, wc
, grh
, mad
);
721 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
726 static int ib_process_mad(struct ib_device
*ibdev
, int mad_flags
, u8 port_num
,
727 const struct ib_wc
*in_wc
, const struct ib_grh
*in_grh
,
728 const struct ib_mad
*in_mad
, struct ib_mad
*out_mad
)
730 u16 slid
, prev_lid
= 0;
732 struct ib_port_attr pattr
;
734 if (in_wc
&& in_wc
->qp
->qp_num
) {
735 pr_debug("received MAD: slid:%d sqpn:%d "
736 "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
737 in_wc
->slid
, in_wc
->src_qp
,
738 in_wc
->dlid_path_bits
,
741 in_mad
->mad_hdr
.mgmt_class
, in_mad
->mad_hdr
.method
,
742 be16_to_cpu(in_mad
->mad_hdr
.attr_id
));
743 if (in_wc
->wc_flags
& IB_WC_GRH
) {
744 pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
745 be64_to_cpu(in_grh
->sgid
.global
.subnet_prefix
),
746 be64_to_cpu(in_grh
->sgid
.global
.interface_id
));
747 pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
748 be64_to_cpu(in_grh
->dgid
.global
.subnet_prefix
),
749 be64_to_cpu(in_grh
->dgid
.global
.interface_id
));
753 slid
= in_wc
? in_wc
->slid
: be16_to_cpu(IB_LID_PERMISSIVE
);
755 if (in_mad
->mad_hdr
.method
== IB_MGMT_METHOD_TRAP
&& slid
== 0) {
756 forward_trap(to_mdev(ibdev
), port_num
, in_mad
);
757 return IB_MAD_RESULT_SUCCESS
| IB_MAD_RESULT_CONSUMED
;
760 if (in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_LID_ROUTED
||
761 in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
) {
762 if (in_mad
->mad_hdr
.method
!= IB_MGMT_METHOD_GET
&&
763 in_mad
->mad_hdr
.method
!= IB_MGMT_METHOD_SET
&&
764 in_mad
->mad_hdr
.method
!= IB_MGMT_METHOD_TRAP_REPRESS
)
765 return IB_MAD_RESULT_SUCCESS
;
768 * Don't process SMInfo queries -- the SMA can't handle them.
770 if (in_mad
->mad_hdr
.attr_id
== IB_SMP_ATTR_SM_INFO
)
771 return IB_MAD_RESULT_SUCCESS
;
772 } else if (in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_PERF_MGMT
||
773 in_mad
->mad_hdr
.mgmt_class
== MLX4_IB_VENDOR_CLASS1
||
774 in_mad
->mad_hdr
.mgmt_class
== MLX4_IB_VENDOR_CLASS2
||
775 in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_CONG_MGMT
) {
776 if (in_mad
->mad_hdr
.method
!= IB_MGMT_METHOD_GET
&&
777 in_mad
->mad_hdr
.method
!= IB_MGMT_METHOD_SET
)
778 return IB_MAD_RESULT_SUCCESS
;
780 return IB_MAD_RESULT_SUCCESS
;
782 if ((in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_LID_ROUTED
||
783 in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
) &&
784 in_mad
->mad_hdr
.method
== IB_MGMT_METHOD_SET
&&
785 in_mad
->mad_hdr
.attr_id
== IB_SMP_ATTR_PORT_INFO
&&
786 !ib_query_port(ibdev
, port_num
, &pattr
))
787 prev_lid
= pattr
.lid
;
789 err
= mlx4_MAD_IFC(to_mdev(ibdev
),
790 (mad_flags
& IB_MAD_IGNORE_MKEY
? MLX4_MAD_IFC_IGNORE_MKEY
: 0) |
791 (mad_flags
& IB_MAD_IGNORE_BKEY
? MLX4_MAD_IFC_IGNORE_BKEY
: 0) |
792 MLX4_MAD_IFC_NET_VIEW
,
793 port_num
, in_wc
, in_grh
, in_mad
, out_mad
);
795 return IB_MAD_RESULT_FAILURE
;
797 if (!out_mad
->mad_hdr
.status
) {
798 if (!(to_mdev(ibdev
)->dev
->caps
.flags
& MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV
))
799 smp_snoop(ibdev
, port_num
, in_mad
, prev_lid
);
800 /* slaves get node desc from FW */
801 if (!mlx4_is_slave(to_mdev(ibdev
)->dev
))
802 node_desc_override(ibdev
, out_mad
);
805 /* set return bit in status of directed route responses */
806 if (in_mad
->mad_hdr
.mgmt_class
== IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
)
807 out_mad
->mad_hdr
.status
|= cpu_to_be16(1 << 15);
809 if (in_mad
->mad_hdr
.method
== IB_MGMT_METHOD_TRAP_REPRESS
)
810 /* no response for trap repress */
811 return IB_MAD_RESULT_SUCCESS
| IB_MAD_RESULT_CONSUMED
;
813 return IB_MAD_RESULT_SUCCESS
| IB_MAD_RESULT_REPLY
;
816 static void edit_counter(struct mlx4_counter
*cnt
,
817 struct ib_pma_portcounters
*pma_cnt
)
819 ASSIGN_32BIT_COUNTER(pma_cnt
->port_xmit_data
,
820 (be64_to_cpu(cnt
->tx_bytes
) >> 2));
821 ASSIGN_32BIT_COUNTER(pma_cnt
->port_rcv_data
,
822 (be64_to_cpu(cnt
->rx_bytes
) >> 2));
823 ASSIGN_32BIT_COUNTER(pma_cnt
->port_xmit_packets
,
824 be64_to_cpu(cnt
->tx_frames
));
825 ASSIGN_32BIT_COUNTER(pma_cnt
->port_rcv_packets
,
826 be64_to_cpu(cnt
->rx_frames
));
829 static int iboe_process_mad(struct ib_device
*ibdev
, int mad_flags
, u8 port_num
,
830 const struct ib_wc
*in_wc
, const struct ib_grh
*in_grh
,
831 const struct ib_mad
*in_mad
, struct ib_mad
*out_mad
)
833 struct mlx4_cmd_mailbox
*mailbox
;
834 struct mlx4_ib_dev
*dev
= to_mdev(ibdev
);
836 u32 inmod
= dev
->counters
[port_num
- 1] & 0xffff;
839 if (in_mad
->mad_hdr
.mgmt_class
!= IB_MGMT_CLASS_PERF_MGMT
)
842 mailbox
= mlx4_alloc_cmd_mailbox(dev
->dev
);
844 return IB_MAD_RESULT_FAILURE
;
846 err
= mlx4_cmd_box(dev
->dev
, 0, mailbox
->dma
, inmod
, 0,
847 MLX4_CMD_QUERY_IF_STAT
, MLX4_CMD_TIME_CLASS_C
,
850 err
= IB_MAD_RESULT_FAILURE
;
852 memset(out_mad
->data
, 0, sizeof out_mad
->data
);
853 mode
= ((struct mlx4_counter
*)mailbox
->buf
)->counter_mode
;
854 switch (mode
& 0xf) {
856 edit_counter(mailbox
->buf
,
857 (void *)(out_mad
->data
+ 40));
858 err
= IB_MAD_RESULT_SUCCESS
| IB_MAD_RESULT_REPLY
;
861 err
= IB_MAD_RESULT_FAILURE
;
865 mlx4_free_cmd_mailbox(dev
->dev
, mailbox
);
870 int mlx4_ib_process_mad(struct ib_device
*ibdev
, int mad_flags
, u8 port_num
,
871 const struct ib_wc
*in_wc
, const struct ib_grh
*in_grh
,
872 const struct ib_mad_hdr
*in
, size_t in_mad_size
,
873 struct ib_mad_hdr
*out
, size_t *out_mad_size
,
874 u16
*out_mad_pkey_index
)
876 const struct ib_mad
*in_mad
= (const struct ib_mad
*)in
;
877 struct ib_mad
*out_mad
= (struct ib_mad
*)out
;
879 BUG_ON(in_mad_size
!= sizeof(*in_mad
) ||
880 *out_mad_size
!= sizeof(*out_mad
));
882 switch (rdma_port_get_link_layer(ibdev
, port_num
)) {
883 case IB_LINK_LAYER_INFINIBAND
:
884 return ib_process_mad(ibdev
, mad_flags
, port_num
, in_wc
,
885 in_grh
, in_mad
, out_mad
);
886 case IB_LINK_LAYER_ETHERNET
:
887 return iboe_process_mad(ibdev
, mad_flags
, port_num
, in_wc
,
888 in_grh
, in_mad
, out_mad
);
894 static void send_handler(struct ib_mad_agent
*agent
,
895 struct ib_mad_send_wc
*mad_send_wc
)
897 if (mad_send_wc
->send_buf
->context
[0])
898 ib_destroy_ah(mad_send_wc
->send_buf
->context
[0]);
899 ib_free_send_mad(mad_send_wc
->send_buf
);
902 int mlx4_ib_mad_init(struct mlx4_ib_dev
*dev
)
904 struct ib_mad_agent
*agent
;
907 enum rdma_link_layer ll
;
909 for (p
= 0; p
< dev
->num_ports
; ++p
) {
910 ll
= rdma_port_get_link_layer(&dev
->ib_dev
, p
+ 1);
911 for (q
= 0; q
<= 1; ++q
) {
912 if (ll
== IB_LINK_LAYER_INFINIBAND
) {
913 agent
= ib_register_mad_agent(&dev
->ib_dev
, p
+ 1,
914 q
? IB_QPT_GSI
: IB_QPT_SMI
,
915 NULL
, 0, send_handler
,
918 ret
= PTR_ERR(agent
);
921 dev
->send_agent
[p
][q
] = agent
;
923 dev
->send_agent
[p
][q
] = NULL
;
930 for (p
= 0; p
< dev
->num_ports
; ++p
)
931 for (q
= 0; q
<= 1; ++q
)
932 if (dev
->send_agent
[p
][q
])
933 ib_unregister_mad_agent(dev
->send_agent
[p
][q
]);
938 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev
*dev
)
940 struct ib_mad_agent
*agent
;
943 for (p
= 0; p
< dev
->num_ports
; ++p
) {
944 for (q
= 0; q
<= 1; ++q
) {
945 agent
= dev
->send_agent
[p
][q
];
947 dev
->send_agent
[p
][q
] = NULL
;
948 ib_unregister_mad_agent(agent
);
953 ib_destroy_ah(dev
->sm_ah
[p
]);
957 static void handle_lid_change_event(struct mlx4_ib_dev
*dev
, u8 port_num
)
959 mlx4_ib_dispatch_event(dev
, port_num
, IB_EVENT_LID_CHANGE
);
961 if (mlx4_is_master(dev
->dev
) && !dev
->sriov
.is_going_down
)
962 mlx4_gen_slaves_port_mgt_ev(dev
->dev
, port_num
,
963 MLX4_EQ_PORT_INFO_LID_CHANGE_MASK
);
966 static void handle_client_rereg_event(struct mlx4_ib_dev
*dev
, u8 port_num
)
968 /* re-configure the alias-guid and mcg's */
969 if (mlx4_is_master(dev
->dev
)) {
970 mlx4_ib_invalidate_all_guid_record(dev
, port_num
);
972 if (!dev
->sriov
.is_going_down
) {
973 mlx4_ib_mcg_port_cleanup(&dev
->sriov
.demux
[port_num
- 1], 0);
974 mlx4_gen_slaves_port_mgt_ev(dev
->dev
, port_num
,
975 MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK
);
978 mlx4_ib_dispatch_event(dev
, port_num
, IB_EVENT_CLIENT_REREGISTER
);
981 static void propagate_pkey_ev(struct mlx4_ib_dev
*dev
, int port_num
,
982 struct mlx4_eqe
*eqe
)
984 __propagate_pkey_ev(dev
, port_num
, GET_BLK_PTR_FROM_EQE(eqe
),
985 GET_MASK_FROM_EQE(eqe
));
988 static void handle_slaves_guid_change(struct mlx4_ib_dev
*dev
, u8 port_num
,
989 u32 guid_tbl_blk_num
, u32 change_bitmap
)
991 struct ib_smp
*in_mad
= NULL
;
992 struct ib_smp
*out_mad
= NULL
;
995 if (!mlx4_is_mfunc(dev
->dev
) || !mlx4_is_master(dev
->dev
))
998 in_mad
= kmalloc(sizeof *in_mad
, GFP_KERNEL
);
999 out_mad
= kmalloc(sizeof *out_mad
, GFP_KERNEL
);
1000 if (!in_mad
|| !out_mad
) {
1001 mlx4_ib_warn(&dev
->ib_dev
, "failed to allocate memory for guid info mads\n");
1005 guid_tbl_blk_num
*= 4;
1007 for (i
= 0; i
< 4; i
++) {
1008 if (change_bitmap
&& (!((change_bitmap
>> (8 * i
)) & 0xff)))
1010 memset(in_mad
, 0, sizeof *in_mad
);
1011 memset(out_mad
, 0, sizeof *out_mad
);
1013 in_mad
->base_version
= 1;
1014 in_mad
->mgmt_class
= IB_MGMT_CLASS_SUBN_LID_ROUTED
;
1015 in_mad
->class_version
= 1;
1016 in_mad
->method
= IB_MGMT_METHOD_GET
;
1017 in_mad
->attr_id
= IB_SMP_ATTR_GUID_INFO
;
1018 in_mad
->attr_mod
= cpu_to_be32(guid_tbl_blk_num
+ i
);
1020 if (mlx4_MAD_IFC(dev
,
1021 MLX4_MAD_IFC_IGNORE_KEYS
| MLX4_MAD_IFC_NET_VIEW
,
1022 port_num
, NULL
, NULL
, in_mad
, out_mad
)) {
1023 mlx4_ib_warn(&dev
->ib_dev
, "Failed in get GUID INFO MAD_IFC\n");
1027 mlx4_ib_update_cache_on_guid_change(dev
, guid_tbl_blk_num
+ i
,
1029 (u8
*)(&((struct ib_smp
*)out_mad
)->data
));
1030 mlx4_ib_notify_slaves_on_guid_change(dev
, guid_tbl_blk_num
+ i
,
1032 (u8
*)(&((struct ib_smp
*)out_mad
)->data
));
1041 void handle_port_mgmt_change_event(struct work_struct
*work
)
1043 struct ib_event_work
*ew
= container_of(work
, struct ib_event_work
, work
);
1044 struct mlx4_ib_dev
*dev
= ew
->ib_dev
;
1045 struct mlx4_eqe
*eqe
= &(ew
->ib_eqe
);
1046 u8 port
= eqe
->event
.port_mgmt_change
.port
;
1051 switch (eqe
->subtype
) {
1052 case MLX4_DEV_PMC_SUBTYPE_PORT_INFO
:
1053 changed_attr
= be32_to_cpu(eqe
->event
.port_mgmt_change
.params
.port_info
.changed_attr
);
1055 /* Update the SM ah - This should be done before handling
1056 the other changed attributes so that MADs can be sent to the SM */
1057 if (changed_attr
& MSTR_SM_CHANGE_MASK
) {
1058 u16 lid
= be16_to_cpu(eqe
->event
.port_mgmt_change
.params
.port_info
.mstr_sm_lid
);
1059 u8 sl
= eqe
->event
.port_mgmt_change
.params
.port_info
.mstr_sm_sl
& 0xf;
1060 update_sm_ah(dev
, port
, lid
, sl
);
1063 /* Check if it is a lid change event */
1064 if (changed_attr
& MLX4_EQ_PORT_INFO_LID_CHANGE_MASK
)
1065 handle_lid_change_event(dev
, port
);
1067 /* Generate GUID changed event */
1068 if (changed_attr
& MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK
) {
1069 mlx4_ib_dispatch_event(dev
, port
, IB_EVENT_GID_CHANGE
);
1070 /*if master, notify all slaves*/
1071 if (mlx4_is_master(dev
->dev
))
1072 mlx4_gen_slaves_port_mgt_ev(dev
->dev
, port
,
1073 MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK
);
1076 if (changed_attr
& MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK
)
1077 handle_client_rereg_event(dev
, port
);
1080 case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE
:
1081 mlx4_ib_dispatch_event(dev
, port
, IB_EVENT_PKEY_CHANGE
);
1082 if (mlx4_is_master(dev
->dev
) && !dev
->sriov
.is_going_down
)
1083 propagate_pkey_ev(dev
, port
, eqe
);
1085 case MLX4_DEV_PMC_SUBTYPE_GUID_INFO
:
1086 /* paravirtualized master's guid is guid 0 -- does not change */
1087 if (!mlx4_is_master(dev
->dev
))
1088 mlx4_ib_dispatch_event(dev
, port
, IB_EVENT_GID_CHANGE
);
1089 /*if master, notify relevant slaves*/
1090 else if (!dev
->sriov
.is_going_down
) {
1091 tbl_block
= GET_BLK_PTR_FROM_EQE(eqe
);
1092 change_bitmap
= GET_MASK_FROM_EQE(eqe
);
1093 handle_slaves_guid_change(dev
, port
, tbl_block
, change_bitmap
);
1097 pr_warn("Unsupported subtype 0x%x for "
1098 "Port Management Change event\n", eqe
->subtype
);
1104 void mlx4_ib_dispatch_event(struct mlx4_ib_dev
*dev
, u8 port_num
,
1105 enum ib_event_type type
)
1107 struct ib_event event
;
1109 event
.device
= &dev
->ib_dev
;
1110 event
.element
.port_num
= port_num
;
1113 ib_dispatch_event(&event
);
1116 static void mlx4_ib_tunnel_comp_handler(struct ib_cq
*cq
, void *arg
)
1118 unsigned long flags
;
1119 struct mlx4_ib_demux_pv_ctx
*ctx
= cq
->cq_context
;
1120 struct mlx4_ib_dev
*dev
= to_mdev(ctx
->ib_dev
);
1121 spin_lock_irqsave(&dev
->sriov
.going_down_lock
, flags
);
1122 if (!dev
->sriov
.is_going_down
&& ctx
->state
== DEMUX_PV_STATE_ACTIVE
)
1123 queue_work(ctx
->wq
, &ctx
->work
);
1124 spin_unlock_irqrestore(&dev
->sriov
.going_down_lock
, flags
);
1127 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx
*ctx
,
1128 struct mlx4_ib_demux_pv_qp
*tun_qp
,
1131 struct ib_sge sg_list
;
1132 struct ib_recv_wr recv_wr
, *bad_recv_wr
;
1135 size
= (tun_qp
->qp
->qp_type
== IB_QPT_UD
) ?
1136 sizeof (struct mlx4_tunnel_mad
) : sizeof (struct mlx4_mad_rcv_buf
);
1138 sg_list
.addr
= tun_qp
->ring
[index
].map
;
1139 sg_list
.length
= size
;
1140 sg_list
.lkey
= ctx
->mr
->lkey
;
1142 recv_wr
.next
= NULL
;
1143 recv_wr
.sg_list
= &sg_list
;
1144 recv_wr
.num_sge
= 1;
1145 recv_wr
.wr_id
= (u64
) index
| MLX4_TUN_WRID_RECV
|
1146 MLX4_TUN_SET_WRID_QPN(tun_qp
->proxy_qpt
);
1147 ib_dma_sync_single_for_device(ctx
->ib_dev
, tun_qp
->ring
[index
].map
,
1148 size
, DMA_FROM_DEVICE
);
1149 return ib_post_recv(tun_qp
->qp
, &recv_wr
, &bad_recv_wr
);
1152 static int mlx4_ib_multiplex_sa_handler(struct ib_device
*ibdev
, int port
,
1153 int slave
, struct ib_sa_mad
*sa_mad
)
1157 /* dispatch to different sa handlers */
1158 switch (be16_to_cpu(sa_mad
->mad_hdr
.attr_id
)) {
1159 case IB_SA_ATTR_MC_MEMBER_REC
:
1160 ret
= mlx4_ib_mcg_multiplex_handler(ibdev
, port
, slave
, sa_mad
);
1168 static int is_proxy_qp0(struct mlx4_ib_dev
*dev
, int qpn
, int slave
)
1170 int proxy_start
= dev
->dev
->phys_caps
.base_proxy_sqpn
+ 8 * slave
;
1172 return (qpn
>= proxy_start
&& qpn
<= proxy_start
+ 1);
1176 int mlx4_ib_send_to_wire(struct mlx4_ib_dev
*dev
, int slave
, u8 port
,
1177 enum ib_qp_type dest_qpt
, u16 pkey_index
,
1178 u32 remote_qpn
, u32 qkey
, struct ib_ah_attr
*attr
,
1179 u8
*s_mac
, struct ib_mad
*mad
)
1182 struct ib_send_wr wr
, *bad_wr
;
1183 struct mlx4_ib_demux_pv_ctx
*sqp_ctx
;
1184 struct mlx4_ib_demux_pv_qp
*sqp
;
1185 struct mlx4_mad_snd_buf
*sqp_mad
;
1187 struct ib_qp
*send_qp
= NULL
;
1188 unsigned wire_tx_ix
= 0;
1195 sqp_ctx
= dev
->sriov
.sqps
[port
-1];
1197 /* check if proxy qp created */
1198 if (!sqp_ctx
|| sqp_ctx
->state
!= DEMUX_PV_STATE_ACTIVE
)
1201 if (dest_qpt
== IB_QPT_SMI
) {
1203 sqp
= &sqp_ctx
->qp
[0];
1204 wire_pkey_ix
= dev
->pkeys
.virt2phys_pkey
[slave
][port
- 1][0];
1207 sqp
= &sqp_ctx
->qp
[1];
1208 wire_pkey_ix
= dev
->pkeys
.virt2phys_pkey
[slave
][port
- 1][pkey_index
];
1214 sgid_index
= attr
->grh
.sgid_index
;
1215 attr
->grh
.sgid_index
= 0;
1216 ah
= ib_create_ah(sqp_ctx
->pd
, attr
);
1219 attr
->grh
.sgid_index
= sgid_index
;
1220 to_mah(ah
)->av
.ib
.gid_index
= sgid_index
;
1221 /* get rid of force-loopback bit */
1222 to_mah(ah
)->av
.ib
.port_pd
&= cpu_to_be32(0x7FFFFFFF);
1223 spin_lock(&sqp
->tx_lock
);
1224 if (sqp
->tx_ix_head
- sqp
->tx_ix_tail
>=
1225 (MLX4_NUM_TUNNEL_BUFS
- 1))
1228 wire_tx_ix
= (++sqp
->tx_ix_head
) & (MLX4_NUM_TUNNEL_BUFS
- 1);
1229 spin_unlock(&sqp
->tx_lock
);
1233 sqp_mad
= (struct mlx4_mad_snd_buf
*) (sqp
->tx_ring
[wire_tx_ix
].buf
.addr
);
1234 if (sqp
->tx_ring
[wire_tx_ix
].ah
)
1235 ib_destroy_ah(sqp
->tx_ring
[wire_tx_ix
].ah
);
1236 sqp
->tx_ring
[wire_tx_ix
].ah
= ah
;
1237 ib_dma_sync_single_for_cpu(&dev
->ib_dev
,
1238 sqp
->tx_ring
[wire_tx_ix
].buf
.map
,
1239 sizeof (struct mlx4_mad_snd_buf
),
1242 memcpy(&sqp_mad
->payload
, mad
, sizeof *mad
);
1244 ib_dma_sync_single_for_device(&dev
->ib_dev
,
1245 sqp
->tx_ring
[wire_tx_ix
].buf
.map
,
1246 sizeof (struct mlx4_mad_snd_buf
),
1249 list
.addr
= sqp
->tx_ring
[wire_tx_ix
].buf
.map
;
1250 list
.length
= sizeof (struct mlx4_mad_snd_buf
);
1251 list
.lkey
= sqp_ctx
->mr
->lkey
;
1254 wr
.wr
.ud
.port_num
= port
;
1255 wr
.wr
.ud
.pkey_index
= wire_pkey_ix
;
1256 wr
.wr
.ud
.remote_qkey
= qkey
;
1257 wr
.wr
.ud
.remote_qpn
= remote_qpn
;
1259 wr
.wr_id
= ((u64
) wire_tx_ix
) | MLX4_TUN_SET_WRID_QPN(src_qpnum
);
1262 wr
.opcode
= IB_WR_SEND
;
1263 wr
.send_flags
= IB_SEND_SIGNALED
;
1265 memcpy(to_mah(ah
)->av
.eth
.s_mac
, s_mac
, 6);
1268 ret
= ib_post_send(send_qp
, &wr
, &bad_wr
);
1275 static int get_slave_base_gid_ix(struct mlx4_ib_dev
*dev
, int slave
, int port
)
1277 if (rdma_port_get_link_layer(&dev
->ib_dev
, port
) == IB_LINK_LAYER_INFINIBAND
)
1279 return mlx4_get_base_gid_ix(dev
->dev
, slave
, port
);
1282 static void fill_in_real_sgid_index(struct mlx4_ib_dev
*dev
, int slave
, int port
,
1283 struct ib_ah_attr
*ah_attr
)
1285 if (rdma_port_get_link_layer(&dev
->ib_dev
, port
) == IB_LINK_LAYER_INFINIBAND
)
1286 ah_attr
->grh
.sgid_index
= slave
;
1288 ah_attr
->grh
.sgid_index
+= get_slave_base_gid_ix(dev
, slave
, port
);
1291 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx
*ctx
, struct ib_wc
*wc
)
1293 struct mlx4_ib_dev
*dev
= to_mdev(ctx
->ib_dev
);
1294 struct mlx4_ib_demux_pv_qp
*tun_qp
= &ctx
->qp
[MLX4_TUN_WRID_QPN(wc
->wr_id
)];
1295 int wr_ix
= wc
->wr_id
& (MLX4_NUM_TUNNEL_BUFS
- 1);
1296 struct mlx4_tunnel_mad
*tunnel
= tun_qp
->ring
[wr_ix
].addr
;
1297 struct mlx4_ib_ah ah
;
1298 struct ib_ah_attr ah_attr
;
1303 /* Get slave that sent this packet */
1304 if (wc
->src_qp
< dev
->dev
->phys_caps
.base_proxy_sqpn
||
1305 wc
->src_qp
>= dev
->dev
->phys_caps
.base_proxy_sqpn
+ 8 * MLX4_MFUNC_MAX
||
1306 (wc
->src_qp
& 0x1) != ctx
->port
- 1 ||
1308 mlx4_ib_warn(ctx
->ib_dev
, "can't multiplex bad sqp:%d\n", wc
->src_qp
);
1311 slave
= ((wc
->src_qp
& ~0x7) - dev
->dev
->phys_caps
.base_proxy_sqpn
) / 8;
1312 if (slave
!= ctx
->slave
) {
1313 mlx4_ib_warn(ctx
->ib_dev
, "can't multiplex bad sqp:%d: "
1314 "belongs to another slave\n", wc
->src_qp
);
1318 /* Map transaction ID */
1319 ib_dma_sync_single_for_cpu(ctx
->ib_dev
, tun_qp
->ring
[wr_ix
].map
,
1320 sizeof (struct mlx4_tunnel_mad
),
1322 switch (tunnel
->mad
.mad_hdr
.method
) {
1323 case IB_MGMT_METHOD_SET
:
1324 case IB_MGMT_METHOD_GET
:
1325 case IB_MGMT_METHOD_REPORT
:
1326 case IB_SA_METHOD_GET_TABLE
:
1327 case IB_SA_METHOD_DELETE
:
1328 case IB_SA_METHOD_GET_MULTI
:
1329 case IB_SA_METHOD_GET_TRACE_TBL
:
1330 slave_id
= (u8
*) &tunnel
->mad
.mad_hdr
.tid
;
1332 mlx4_ib_warn(ctx
->ib_dev
, "egress mad has non-null tid msb:%d "
1333 "class:%d slave:%d\n", *slave_id
,
1334 tunnel
->mad
.mad_hdr
.mgmt_class
, slave
);
1342 /* Class-specific handling */
1343 switch (tunnel
->mad
.mad_hdr
.mgmt_class
) {
1344 case IB_MGMT_CLASS_SUBN_LID_ROUTED
:
1345 case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
:
1346 if (slave
!= mlx4_master_func_num(dev
->dev
) &&
1347 !mlx4_vf_smi_enabled(dev
->dev
, slave
, ctx
->port
))
1350 case IB_MGMT_CLASS_SUBN_ADM
:
1351 if (mlx4_ib_multiplex_sa_handler(ctx
->ib_dev
, ctx
->port
, slave
,
1352 (struct ib_sa_mad
*) &tunnel
->mad
))
1355 case IB_MGMT_CLASS_CM
:
1356 if (mlx4_ib_multiplex_cm_handler(ctx
->ib_dev
, ctx
->port
, slave
,
1357 (struct ib_mad
*) &tunnel
->mad
))
1360 case IB_MGMT_CLASS_DEVICE_MGMT
:
1361 if (tunnel
->mad
.mad_hdr
.method
!= IB_MGMT_METHOD_GET
&&
1362 tunnel
->mad
.mad_hdr
.method
!= IB_MGMT_METHOD_SET
)
1366 /* Drop unsupported classes for slaves in tunnel mode */
1367 if (slave
!= mlx4_master_func_num(dev
->dev
)) {
1368 mlx4_ib_warn(ctx
->ib_dev
, "dropping unsupported egress mad from class:%d "
1369 "for slave:%d\n", tunnel
->mad
.mad_hdr
.mgmt_class
, slave
);
1374 /* We are using standard ib_core services to send the mad, so generate a
1375 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1376 memcpy(&ah
.av
, &tunnel
->hdr
.av
, sizeof (struct mlx4_av
));
1377 ah
.ibah
.device
= ctx
->ib_dev
;
1378 mlx4_ib_query_ah(&ah
.ibah
, &ah_attr
);
1379 if (ah_attr
.ah_flags
& IB_AH_GRH
)
1380 fill_in_real_sgid_index(dev
, slave
, ctx
->port
, &ah_attr
);
1382 port
= mlx4_slave_convert_port(dev
->dev
, slave
, ah_attr
.port_num
);
1385 ah_attr
.port_num
= port
;
1386 memcpy(ah_attr
.dmac
, tunnel
->hdr
.mac
, 6);
1387 ah_attr
.vlan_id
= be16_to_cpu(tunnel
->hdr
.vlan
);
1388 /* if slave have default vlan use it */
1389 mlx4_get_slave_default_vlan(dev
->dev
, ctx
->port
, slave
,
1390 &ah_attr
.vlan_id
, &ah_attr
.sl
);
1392 mlx4_ib_send_to_wire(dev
, slave
, ctx
->port
,
1393 is_proxy_qp0(dev
, wc
->src_qp
, slave
) ?
1394 IB_QPT_SMI
: IB_QPT_GSI
,
1395 be16_to_cpu(tunnel
->hdr
.pkey_index
),
1396 be32_to_cpu(tunnel
->hdr
.remote_qpn
),
1397 be32_to_cpu(tunnel
->hdr
.qkey
),
1398 &ah_attr
, wc
->smac
, &tunnel
->mad
);
1401 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx
*ctx
,
1402 enum ib_qp_type qp_type
, int is_tun
)
1405 struct mlx4_ib_demux_pv_qp
*tun_qp
;
1406 int rx_buf_size
, tx_buf_size
;
1408 if (qp_type
> IB_QPT_GSI
)
1411 tun_qp
= &ctx
->qp
[qp_type
];
1413 tun_qp
->ring
= kzalloc(sizeof (struct mlx4_ib_buf
) * MLX4_NUM_TUNNEL_BUFS
,
1418 tun_qp
->tx_ring
= kcalloc(MLX4_NUM_TUNNEL_BUFS
,
1419 sizeof (struct mlx4_ib_tun_tx_buf
),
1421 if (!tun_qp
->tx_ring
) {
1422 kfree(tun_qp
->ring
);
1423 tun_qp
->ring
= NULL
;
1428 rx_buf_size
= sizeof (struct mlx4_tunnel_mad
);
1429 tx_buf_size
= sizeof (struct mlx4_rcv_tunnel_mad
);
1431 rx_buf_size
= sizeof (struct mlx4_mad_rcv_buf
);
1432 tx_buf_size
= sizeof (struct mlx4_mad_snd_buf
);
1435 for (i
= 0; i
< MLX4_NUM_TUNNEL_BUFS
; i
++) {
1436 tun_qp
->ring
[i
].addr
= kmalloc(rx_buf_size
, GFP_KERNEL
);
1437 if (!tun_qp
->ring
[i
].addr
)
1439 tun_qp
->ring
[i
].map
= ib_dma_map_single(ctx
->ib_dev
,
1440 tun_qp
->ring
[i
].addr
,
1443 if (ib_dma_mapping_error(ctx
->ib_dev
, tun_qp
->ring
[i
].map
)) {
1444 kfree(tun_qp
->ring
[i
].addr
);
1449 for (i
= 0; i
< MLX4_NUM_TUNNEL_BUFS
; i
++) {
1450 tun_qp
->tx_ring
[i
].buf
.addr
=
1451 kmalloc(tx_buf_size
, GFP_KERNEL
);
1452 if (!tun_qp
->tx_ring
[i
].buf
.addr
)
1454 tun_qp
->tx_ring
[i
].buf
.map
=
1455 ib_dma_map_single(ctx
->ib_dev
,
1456 tun_qp
->tx_ring
[i
].buf
.addr
,
1459 if (ib_dma_mapping_error(ctx
->ib_dev
,
1460 tun_qp
->tx_ring
[i
].buf
.map
)) {
1461 kfree(tun_qp
->tx_ring
[i
].buf
.addr
);
1464 tun_qp
->tx_ring
[i
].ah
= NULL
;
1466 spin_lock_init(&tun_qp
->tx_lock
);
1467 tun_qp
->tx_ix_head
= 0;
1468 tun_qp
->tx_ix_tail
= 0;
1469 tun_qp
->proxy_qpt
= qp_type
;
1476 ib_dma_unmap_single(ctx
->ib_dev
, tun_qp
->tx_ring
[i
].buf
.map
,
1477 tx_buf_size
, DMA_TO_DEVICE
);
1478 kfree(tun_qp
->tx_ring
[i
].buf
.addr
);
1480 kfree(tun_qp
->tx_ring
);
1481 tun_qp
->tx_ring
= NULL
;
1482 i
= MLX4_NUM_TUNNEL_BUFS
;
1486 ib_dma_unmap_single(ctx
->ib_dev
, tun_qp
->ring
[i
].map
,
1487 rx_buf_size
, DMA_FROM_DEVICE
);
1488 kfree(tun_qp
->ring
[i
].addr
);
1490 kfree(tun_qp
->ring
);
1491 tun_qp
->ring
= NULL
;
1495 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx
*ctx
,
1496 enum ib_qp_type qp_type
, int is_tun
)
1499 struct mlx4_ib_demux_pv_qp
*tun_qp
;
1500 int rx_buf_size
, tx_buf_size
;
1502 if (qp_type
> IB_QPT_GSI
)
1505 tun_qp
= &ctx
->qp
[qp_type
];
1507 rx_buf_size
= sizeof (struct mlx4_tunnel_mad
);
1508 tx_buf_size
= sizeof (struct mlx4_rcv_tunnel_mad
);
1510 rx_buf_size
= sizeof (struct mlx4_mad_rcv_buf
);
1511 tx_buf_size
= sizeof (struct mlx4_mad_snd_buf
);
1515 for (i
= 0; i
< MLX4_NUM_TUNNEL_BUFS
; i
++) {
1516 ib_dma_unmap_single(ctx
->ib_dev
, tun_qp
->ring
[i
].map
,
1517 rx_buf_size
, DMA_FROM_DEVICE
);
1518 kfree(tun_qp
->ring
[i
].addr
);
1521 for (i
= 0; i
< MLX4_NUM_TUNNEL_BUFS
; i
++) {
1522 ib_dma_unmap_single(ctx
->ib_dev
, tun_qp
->tx_ring
[i
].buf
.map
,
1523 tx_buf_size
, DMA_TO_DEVICE
);
1524 kfree(tun_qp
->tx_ring
[i
].buf
.addr
);
1525 if (tun_qp
->tx_ring
[i
].ah
)
1526 ib_destroy_ah(tun_qp
->tx_ring
[i
].ah
);
1528 kfree(tun_qp
->tx_ring
);
1529 kfree(tun_qp
->ring
);
1532 static void mlx4_ib_tunnel_comp_worker(struct work_struct
*work
)
1534 struct mlx4_ib_demux_pv_ctx
*ctx
;
1535 struct mlx4_ib_demux_pv_qp
*tun_qp
;
1538 ctx
= container_of(work
, struct mlx4_ib_demux_pv_ctx
, work
);
1539 ib_req_notify_cq(ctx
->cq
, IB_CQ_NEXT_COMP
);
1541 while (ib_poll_cq(ctx
->cq
, 1, &wc
) == 1) {
1542 tun_qp
= &ctx
->qp
[MLX4_TUN_WRID_QPN(wc
.wr_id
)];
1543 if (wc
.status
== IB_WC_SUCCESS
) {
1544 switch (wc
.opcode
) {
1546 mlx4_ib_multiplex_mad(ctx
, &wc
);
1547 ret
= mlx4_ib_post_pv_qp_buf(ctx
, tun_qp
,
1549 (MLX4_NUM_TUNNEL_BUFS
- 1));
1551 pr_err("Failed reposting tunnel "
1552 "buf:%lld\n", wc
.wr_id
);
1555 pr_debug("received tunnel send completion:"
1556 "wrid=0x%llx, status=0x%x\n",
1557 wc
.wr_id
, wc
.status
);
1558 ib_destroy_ah(tun_qp
->tx_ring
[wc
.wr_id
&
1559 (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
);
1560 tun_qp
->tx_ring
[wc
.wr_id
& (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
1562 spin_lock(&tun_qp
->tx_lock
);
1563 tun_qp
->tx_ix_tail
++;
1564 spin_unlock(&tun_qp
->tx_lock
);
1571 pr_debug("mlx4_ib: completion error in tunnel: %d."
1572 " status = %d, wrid = 0x%llx\n",
1573 ctx
->slave
, wc
.status
, wc
.wr_id
);
1574 if (!MLX4_TUN_IS_RECV(wc
.wr_id
)) {
1575 ib_destroy_ah(tun_qp
->tx_ring
[wc
.wr_id
&
1576 (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
);
1577 tun_qp
->tx_ring
[wc
.wr_id
& (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
1579 spin_lock(&tun_qp
->tx_lock
);
1580 tun_qp
->tx_ix_tail
++;
1581 spin_unlock(&tun_qp
->tx_lock
);
1587 static void pv_qp_event_handler(struct ib_event
*event
, void *qp_context
)
1589 struct mlx4_ib_demux_pv_ctx
*sqp
= qp_context
;
1591 /* It's worse than that! He's dead, Jim! */
1592 pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1593 event
->event
, sqp
->port
);
1596 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx
*ctx
,
1597 enum ib_qp_type qp_type
, int create_tun
)
1600 struct mlx4_ib_demux_pv_qp
*tun_qp
;
1601 struct mlx4_ib_qp_tunnel_init_attr qp_init_attr
;
1602 struct ib_qp_attr attr
;
1603 int qp_attr_mask_INIT
;
1605 if (qp_type
> IB_QPT_GSI
)
1608 tun_qp
= &ctx
->qp
[qp_type
];
1610 memset(&qp_init_attr
, 0, sizeof qp_init_attr
);
1611 qp_init_attr
.init_attr
.send_cq
= ctx
->cq
;
1612 qp_init_attr
.init_attr
.recv_cq
= ctx
->cq
;
1613 qp_init_attr
.init_attr
.sq_sig_type
= IB_SIGNAL_ALL_WR
;
1614 qp_init_attr
.init_attr
.cap
.max_send_wr
= MLX4_NUM_TUNNEL_BUFS
;
1615 qp_init_attr
.init_attr
.cap
.max_recv_wr
= MLX4_NUM_TUNNEL_BUFS
;
1616 qp_init_attr
.init_attr
.cap
.max_send_sge
= 1;
1617 qp_init_attr
.init_attr
.cap
.max_recv_sge
= 1;
1619 qp_init_attr
.init_attr
.qp_type
= IB_QPT_UD
;
1620 qp_init_attr
.init_attr
.create_flags
= MLX4_IB_SRIOV_TUNNEL_QP
;
1621 qp_init_attr
.port
= ctx
->port
;
1622 qp_init_attr
.slave
= ctx
->slave
;
1623 qp_init_attr
.proxy_qp_type
= qp_type
;
1624 qp_attr_mask_INIT
= IB_QP_STATE
| IB_QP_PKEY_INDEX
|
1625 IB_QP_QKEY
| IB_QP_PORT
;
1627 qp_init_attr
.init_attr
.qp_type
= qp_type
;
1628 qp_init_attr
.init_attr
.create_flags
= MLX4_IB_SRIOV_SQP
;
1629 qp_attr_mask_INIT
= IB_QP_STATE
| IB_QP_PKEY_INDEX
| IB_QP_QKEY
;
1631 qp_init_attr
.init_attr
.port_num
= ctx
->port
;
1632 qp_init_attr
.init_attr
.qp_context
= ctx
;
1633 qp_init_attr
.init_attr
.event_handler
= pv_qp_event_handler
;
1634 tun_qp
->qp
= ib_create_qp(ctx
->pd
, &qp_init_attr
.init_attr
);
1635 if (IS_ERR(tun_qp
->qp
)) {
1636 ret
= PTR_ERR(tun_qp
->qp
);
1638 pr_err("Couldn't create %s QP (%d)\n",
1639 create_tun
? "tunnel" : "special", ret
);
1643 memset(&attr
, 0, sizeof attr
);
1644 attr
.qp_state
= IB_QPS_INIT
;
1647 ret
= find_slave_port_pkey_ix(to_mdev(ctx
->ib_dev
), ctx
->slave
,
1648 ctx
->port
, IB_DEFAULT_PKEY_FULL
,
1650 if (ret
|| !create_tun
)
1652 to_mdev(ctx
->ib_dev
)->pkeys
.virt2phys_pkey
[ctx
->slave
][ctx
->port
- 1][0];
1653 attr
.qkey
= IB_QP1_QKEY
;
1654 attr
.port_num
= ctx
->port
;
1655 ret
= ib_modify_qp(tun_qp
->qp
, &attr
, qp_attr_mask_INIT
);
1657 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1658 create_tun
? "tunnel" : "special", ret
);
1661 attr
.qp_state
= IB_QPS_RTR
;
1662 ret
= ib_modify_qp(tun_qp
->qp
, &attr
, IB_QP_STATE
);
1664 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1665 create_tun
? "tunnel" : "special", ret
);
1668 attr
.qp_state
= IB_QPS_RTS
;
1670 ret
= ib_modify_qp(tun_qp
->qp
, &attr
, IB_QP_STATE
| IB_QP_SQ_PSN
);
1672 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1673 create_tun
? "tunnel" : "special", ret
);
1677 for (i
= 0; i
< MLX4_NUM_TUNNEL_BUFS
; i
++) {
1678 ret
= mlx4_ib_post_pv_qp_buf(ctx
, tun_qp
, i
);
1680 pr_err(" mlx4_ib_post_pv_buf error"
1681 " (err = %d, i = %d)\n", ret
, i
);
1688 ib_destroy_qp(tun_qp
->qp
);
1694 * IB MAD completion callback for real SQPs
1696 static void mlx4_ib_sqp_comp_worker(struct work_struct
*work
)
1698 struct mlx4_ib_demux_pv_ctx
*ctx
;
1699 struct mlx4_ib_demux_pv_qp
*sqp
;
1704 ctx
= container_of(work
, struct mlx4_ib_demux_pv_ctx
, work
);
1705 ib_req_notify_cq(ctx
->cq
, IB_CQ_NEXT_COMP
);
1707 while (mlx4_ib_poll_cq(ctx
->cq
, 1, &wc
) == 1) {
1708 sqp
= &ctx
->qp
[MLX4_TUN_WRID_QPN(wc
.wr_id
)];
1709 if (wc
.status
== IB_WC_SUCCESS
) {
1710 switch (wc
.opcode
) {
1712 ib_destroy_ah(sqp
->tx_ring
[wc
.wr_id
&
1713 (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
);
1714 sqp
->tx_ring
[wc
.wr_id
& (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
1716 spin_lock(&sqp
->tx_lock
);
1718 spin_unlock(&sqp
->tx_lock
);
1721 mad
= (struct ib_mad
*) &(((struct mlx4_mad_rcv_buf
*)
1722 (sqp
->ring
[wc
.wr_id
&
1723 (MLX4_NUM_TUNNEL_BUFS
- 1)].addr
))->payload
);
1724 grh
= &(((struct mlx4_mad_rcv_buf
*)
1725 (sqp
->ring
[wc
.wr_id
&
1726 (MLX4_NUM_TUNNEL_BUFS
- 1)].addr
))->grh
);
1727 mlx4_ib_demux_mad(ctx
->ib_dev
, ctx
->port
, &wc
, grh
, mad
);
1728 if (mlx4_ib_post_pv_qp_buf(ctx
, sqp
, wc
.wr_id
&
1729 (MLX4_NUM_TUNNEL_BUFS
- 1)))
1730 pr_err("Failed reposting SQP "
1731 "buf:%lld\n", wc
.wr_id
);
1738 pr_debug("mlx4_ib: completion error in tunnel: %d."
1739 " status = %d, wrid = 0x%llx\n",
1740 ctx
->slave
, wc
.status
, wc
.wr_id
);
1741 if (!MLX4_TUN_IS_RECV(wc
.wr_id
)) {
1742 ib_destroy_ah(sqp
->tx_ring
[wc
.wr_id
&
1743 (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
);
1744 sqp
->tx_ring
[wc
.wr_id
& (MLX4_NUM_TUNNEL_BUFS
- 1)].ah
1746 spin_lock(&sqp
->tx_lock
);
1748 spin_unlock(&sqp
->tx_lock
);
1754 static int alloc_pv_object(struct mlx4_ib_dev
*dev
, int slave
, int port
,
1755 struct mlx4_ib_demux_pv_ctx
**ret_ctx
)
1757 struct mlx4_ib_demux_pv_ctx
*ctx
;
1760 ctx
= kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx
), GFP_KERNEL
);
1762 pr_err("failed allocating pv resource context "
1763 "for port %d, slave %d\n", port
, slave
);
1767 ctx
->ib_dev
= &dev
->ib_dev
;
1774 static void free_pv_object(struct mlx4_ib_dev
*dev
, int slave
, int port
)
1776 if (dev
->sriov
.demux
[port
- 1].tun
[slave
]) {
1777 kfree(dev
->sriov
.demux
[port
- 1].tun
[slave
]);
1778 dev
->sriov
.demux
[port
- 1].tun
[slave
] = NULL
;
1782 static int create_pv_resources(struct ib_device
*ibdev
, int slave
, int port
,
1783 int create_tun
, struct mlx4_ib_demux_pv_ctx
*ctx
)
1786 struct ib_cq_init_attr cq_attr
= {};
1788 if (ctx
->state
!= DEMUX_PV_STATE_DOWN
)
1791 ctx
->state
= DEMUX_PV_STATE_STARTING
;
1792 /* have QP0 only if link layer is IB */
1793 if (rdma_port_get_link_layer(ibdev
, ctx
->port
) ==
1794 IB_LINK_LAYER_INFINIBAND
)
1798 ret
= mlx4_ib_alloc_pv_bufs(ctx
, IB_QPT_SMI
, create_tun
);
1800 pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret
);
1805 ret
= mlx4_ib_alloc_pv_bufs(ctx
, IB_QPT_GSI
, create_tun
);
1807 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret
);
1811 cq_size
= 2 * MLX4_NUM_TUNNEL_BUFS
;
1815 cq_attr
.cqe
= cq_size
;
1816 ctx
->cq
= ib_create_cq(ctx
->ib_dev
, mlx4_ib_tunnel_comp_handler
,
1817 NULL
, ctx
, &cq_attr
);
1818 if (IS_ERR(ctx
->cq
)) {
1819 ret
= PTR_ERR(ctx
->cq
);
1820 pr_err("Couldn't create tunnel CQ (%d)\n", ret
);
1824 ctx
->pd
= ib_alloc_pd(ctx
->ib_dev
);
1825 if (IS_ERR(ctx
->pd
)) {
1826 ret
= PTR_ERR(ctx
->pd
);
1827 pr_err("Couldn't create tunnel PD (%d)\n", ret
);
1831 ctx
->mr
= ib_get_dma_mr(ctx
->pd
, IB_ACCESS_LOCAL_WRITE
);
1832 if (IS_ERR(ctx
->mr
)) {
1833 ret
= PTR_ERR(ctx
->mr
);
1834 pr_err("Couldn't get tunnel DMA MR (%d)\n", ret
);
1839 ret
= create_pv_sqp(ctx
, IB_QPT_SMI
, create_tun
);
1841 pr_err("Couldn't create %s QP0 (%d)\n",
1842 create_tun
? "tunnel for" : "", ret
);
1847 ret
= create_pv_sqp(ctx
, IB_QPT_GSI
, create_tun
);
1849 pr_err("Couldn't create %s QP1 (%d)\n",
1850 create_tun
? "tunnel for" : "", ret
);
1855 INIT_WORK(&ctx
->work
, mlx4_ib_tunnel_comp_worker
);
1857 INIT_WORK(&ctx
->work
, mlx4_ib_sqp_comp_worker
);
1859 ctx
->wq
= to_mdev(ibdev
)->sriov
.demux
[port
- 1].wq
;
1861 ret
= ib_req_notify_cq(ctx
->cq
, IB_CQ_NEXT_COMP
);
1863 pr_err("Couldn't arm tunnel cq (%d)\n", ret
);
1866 ctx
->state
= DEMUX_PV_STATE_ACTIVE
;
1871 ib_destroy_qp(ctx
->qp
[1].qp
);
1872 ctx
->qp
[1].qp
= NULL
;
1877 ib_destroy_qp(ctx
->qp
[0].qp
);
1878 ctx
->qp
[0].qp
= NULL
;
1881 ib_dereg_mr(ctx
->mr
);
1885 ib_dealloc_pd(ctx
->pd
);
1889 ib_destroy_cq(ctx
->cq
);
1893 mlx4_ib_free_pv_qp_bufs(ctx
, IB_QPT_GSI
, create_tun
);
1897 mlx4_ib_free_pv_qp_bufs(ctx
, IB_QPT_SMI
, create_tun
);
1899 ctx
->state
= DEMUX_PV_STATE_DOWN
;
1903 static void destroy_pv_resources(struct mlx4_ib_dev
*dev
, int slave
, int port
,
1904 struct mlx4_ib_demux_pv_ctx
*ctx
, int flush
)
1908 if (ctx
->state
> DEMUX_PV_STATE_DOWN
) {
1909 ctx
->state
= DEMUX_PV_STATE_DOWNING
;
1911 flush_workqueue(ctx
->wq
);
1913 ib_destroy_qp(ctx
->qp
[0].qp
);
1914 ctx
->qp
[0].qp
= NULL
;
1915 mlx4_ib_free_pv_qp_bufs(ctx
, IB_QPT_SMI
, 1);
1917 ib_destroy_qp(ctx
->qp
[1].qp
);
1918 ctx
->qp
[1].qp
= NULL
;
1919 mlx4_ib_free_pv_qp_bufs(ctx
, IB_QPT_GSI
, 1);
1920 ib_dereg_mr(ctx
->mr
);
1922 ib_dealloc_pd(ctx
->pd
);
1924 ib_destroy_cq(ctx
->cq
);
1926 ctx
->state
= DEMUX_PV_STATE_DOWN
;
1930 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev
*dev
, int slave
,
1931 int port
, int do_init
)
1936 clean_vf_mcast(&dev
->sriov
.demux
[port
- 1], slave
);
1937 /* for master, destroy real sqp resources */
1938 if (slave
== mlx4_master_func_num(dev
->dev
))
1939 destroy_pv_resources(dev
, slave
, port
,
1940 dev
->sriov
.sqps
[port
- 1], 1);
1941 /* destroy the tunnel qp resources */
1942 destroy_pv_resources(dev
, slave
, port
,
1943 dev
->sriov
.demux
[port
- 1].tun
[slave
], 1);
1947 /* create the tunnel qp resources */
1948 ret
= create_pv_resources(&dev
->ib_dev
, slave
, port
, 1,
1949 dev
->sriov
.demux
[port
- 1].tun
[slave
]);
1951 /* for master, create the real sqp resources */
1952 if (!ret
&& slave
== mlx4_master_func_num(dev
->dev
))
1953 ret
= create_pv_resources(&dev
->ib_dev
, slave
, port
, 0,
1954 dev
->sriov
.sqps
[port
- 1]);
1958 void mlx4_ib_tunnels_update_work(struct work_struct
*work
)
1960 struct mlx4_ib_demux_work
*dmxw
;
1962 dmxw
= container_of(work
, struct mlx4_ib_demux_work
, work
);
1963 mlx4_ib_tunnels_update(dmxw
->dev
, dmxw
->slave
, (int) dmxw
->port
,
1969 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev
*dev
,
1970 struct mlx4_ib_demux_ctx
*ctx
,
1977 ctx
->tun
= kcalloc(dev
->dev
->caps
.sqp_demux
,
1978 sizeof (struct mlx4_ib_demux_pv_ctx
*), GFP_KERNEL
);
1984 ctx
->ib_dev
= &dev
->ib_dev
;
1987 i
< min(dev
->dev
->caps
.sqp_demux
,
1988 (u16
)(dev
->dev
->persist
->num_vfs
+ 1));
1990 struct mlx4_active_ports actv_ports
=
1991 mlx4_get_active_ports(dev
->dev
, i
);
1993 if (!test_bit(port
- 1, actv_ports
.ports
))
1996 ret
= alloc_pv_object(dev
, i
, port
, &ctx
->tun
[i
]);
2003 ret
= mlx4_ib_mcg_port_init(ctx
);
2005 pr_err("Failed initializing mcg para-virt (%d)\n", ret
);
2009 snprintf(name
, sizeof name
, "mlx4_ibt%d", port
);
2010 ctx
->wq
= create_singlethread_workqueue(name
);
2012 pr_err("Failed to create tunnelling WQ for port %d\n", port
);
2017 snprintf(name
, sizeof name
, "mlx4_ibud%d", port
);
2018 ctx
->ud_wq
= create_singlethread_workqueue(name
);
2020 pr_err("Failed to create up/down WQ for port %d\n", port
);
2028 destroy_workqueue(ctx
->wq
);
2032 mlx4_ib_mcg_port_cleanup(ctx
, 1);
2034 for (i
= 0; i
< dev
->dev
->caps
.sqp_demux
; i
++)
2035 free_pv_object(dev
, i
, port
);
2041 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx
*sqp_ctx
)
2043 if (sqp_ctx
->state
> DEMUX_PV_STATE_DOWN
) {
2044 sqp_ctx
->state
= DEMUX_PV_STATE_DOWNING
;
2045 flush_workqueue(sqp_ctx
->wq
);
2046 if (sqp_ctx
->has_smi
) {
2047 ib_destroy_qp(sqp_ctx
->qp
[0].qp
);
2048 sqp_ctx
->qp
[0].qp
= NULL
;
2049 mlx4_ib_free_pv_qp_bufs(sqp_ctx
, IB_QPT_SMI
, 0);
2051 ib_destroy_qp(sqp_ctx
->qp
[1].qp
);
2052 sqp_ctx
->qp
[1].qp
= NULL
;
2053 mlx4_ib_free_pv_qp_bufs(sqp_ctx
, IB_QPT_GSI
, 0);
2054 ib_dereg_mr(sqp_ctx
->mr
);
2056 ib_dealloc_pd(sqp_ctx
->pd
);
2058 ib_destroy_cq(sqp_ctx
->cq
);
2060 sqp_ctx
->state
= DEMUX_PV_STATE_DOWN
;
2064 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx
*ctx
)
2068 struct mlx4_ib_dev
*dev
= to_mdev(ctx
->ib_dev
);
2069 mlx4_ib_mcg_port_cleanup(ctx
, 1);
2070 for (i
= 0; i
< dev
->dev
->caps
.sqp_demux
; i
++) {
2073 if (ctx
->tun
[i
]->state
> DEMUX_PV_STATE_DOWN
)
2074 ctx
->tun
[i
]->state
= DEMUX_PV_STATE_DOWNING
;
2076 flush_workqueue(ctx
->wq
);
2077 for (i
= 0; i
< dev
->dev
->caps
.sqp_demux
; i
++) {
2078 destroy_pv_resources(dev
, i
, ctx
->port
, ctx
->tun
[i
], 0);
2079 free_pv_object(dev
, i
, ctx
->port
);
2082 destroy_workqueue(ctx
->ud_wq
);
2083 destroy_workqueue(ctx
->wq
);
2087 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev
*dev
, int do_init
)
2091 if (!mlx4_is_master(dev
->dev
))
2093 /* initialize or tear down tunnel QPs for the master */
2094 for (i
= 0; i
< dev
->dev
->caps
.num_ports
; i
++)
2095 mlx4_ib_tunnels_update(dev
, mlx4_master_func_num(dev
->dev
), i
+ 1, do_init
);
2099 int mlx4_ib_init_sriov(struct mlx4_ib_dev
*dev
)
2104 if (!mlx4_is_mfunc(dev
->dev
))
2107 dev
->sriov
.is_going_down
= 0;
2108 spin_lock_init(&dev
->sriov
.going_down_lock
);
2109 mlx4_ib_cm_paravirt_init(dev
);
2111 mlx4_ib_warn(&dev
->ib_dev
, "multi-function enabled\n");
2113 if (mlx4_is_slave(dev
->dev
)) {
2114 mlx4_ib_warn(&dev
->ib_dev
, "operating in qp1 tunnel mode\n");
2118 for (i
= 0; i
< dev
->dev
->caps
.sqp_demux
; i
++) {
2119 if (i
== mlx4_master_func_num(dev
->dev
))
2120 mlx4_put_slave_node_guid(dev
->dev
, i
, dev
->ib_dev
.node_guid
);
2122 mlx4_put_slave_node_guid(dev
->dev
, i
, mlx4_ib_gen_node_guid());
2125 err
= mlx4_ib_init_alias_guid_service(dev
);
2127 mlx4_ib_warn(&dev
->ib_dev
, "Failed init alias guid process.\n");
2130 err
= mlx4_ib_device_register_sysfs(dev
);
2132 mlx4_ib_warn(&dev
->ib_dev
, "Failed to register sysfs\n");
2136 mlx4_ib_warn(&dev
->ib_dev
, "initializing demux service for %d qp1 clients\n",
2137 dev
->dev
->caps
.sqp_demux
);
2138 for (i
= 0; i
< dev
->num_ports
; i
++) {
2140 err
= __mlx4_ib_query_gid(&dev
->ib_dev
, i
+ 1, 0, &gid
, 1);
2143 dev
->sriov
.demux
[i
].guid_cache
[0] = gid
.global
.interface_id
;
2144 err
= alloc_pv_object(dev
, mlx4_master_func_num(dev
->dev
), i
+ 1,
2145 &dev
->sriov
.sqps
[i
]);
2148 err
= mlx4_ib_alloc_demux_ctx(dev
, &dev
->sriov
.demux
[i
], i
+ 1);
2152 mlx4_ib_master_tunnels(dev
, 1);
2156 free_pv_object(dev
, mlx4_master_func_num(dev
->dev
), i
+ 1);
2159 free_pv_object(dev
, mlx4_master_func_num(dev
->dev
), i
+ 1);
2160 mlx4_ib_free_demux_ctx(&dev
->sriov
.demux
[i
]);
2162 mlx4_ib_device_unregister_sysfs(dev
);
2165 mlx4_ib_destroy_alias_guid_service(dev
);
2168 mlx4_ib_cm_paravirt_clean(dev
, -1);
2173 void mlx4_ib_close_sriov(struct mlx4_ib_dev
*dev
)
2176 unsigned long flags
;
2178 if (!mlx4_is_mfunc(dev
->dev
))
2181 spin_lock_irqsave(&dev
->sriov
.going_down_lock
, flags
);
2182 dev
->sriov
.is_going_down
= 1;
2183 spin_unlock_irqrestore(&dev
->sriov
.going_down_lock
, flags
);
2184 if (mlx4_is_master(dev
->dev
)) {
2185 for (i
= 0; i
< dev
->num_ports
; i
++) {
2186 flush_workqueue(dev
->sriov
.demux
[i
].ud_wq
);
2187 mlx4_ib_free_sqp_ctx(dev
->sriov
.sqps
[i
]);
2188 kfree(dev
->sriov
.sqps
[i
]);
2189 dev
->sriov
.sqps
[i
] = NULL
;
2190 mlx4_ib_free_demux_ctx(&dev
->sriov
.demux
[i
]);
2193 mlx4_ib_cm_paravirt_clean(dev
, -1);
2194 mlx4_ib_destroy_alias_guid_service(dev
);
2195 mlx4_ib_device_unregister_sysfs(dev
);