mlx4: Use port management change event instead of smp_snoop
[cascardo/linux.git] / drivers / infiniband / hw / mlx4 / mad.c
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
36 #include <linux/mlx4/cmd.h>
37 #include <linux/gfp.h>
38 #include <rdma/ib_pma.h>
39
40 #include "mlx4_ib.h"
41
42 enum {
43         MLX4_IB_VENDOR_CLASS1 = 0x9,
44         MLX4_IB_VENDOR_CLASS2 = 0xa
45 };
46
47 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int ignore_mkey, int ignore_bkey,
48                  int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
49                  void *in_mad, void *response_mad)
50 {
51         struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
52         void *inbox;
53         int err;
54         u32 in_modifier = port;
55         u8 op_modifier = 0;
56
57         inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
58         if (IS_ERR(inmailbox))
59                 return PTR_ERR(inmailbox);
60         inbox = inmailbox->buf;
61
62         outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
63         if (IS_ERR(outmailbox)) {
64                 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
65                 return PTR_ERR(outmailbox);
66         }
67
68         memcpy(inbox, in_mad, 256);
69
70         /*
71          * Key check traps can't be generated unless we have in_wc to
72          * tell us where to send the trap.
73          */
74         if (ignore_mkey || !in_wc)
75                 op_modifier |= 0x1;
76         if (ignore_bkey || !in_wc)
77                 op_modifier |= 0x2;
78
79         if (in_wc) {
80                 struct {
81                         __be32          my_qpn;
82                         u32             reserved1;
83                         __be32          rqpn;
84                         u8              sl;
85                         u8              g_path;
86                         u16             reserved2[2];
87                         __be16          pkey;
88                         u32             reserved3[11];
89                         u8              grh[40];
90                 } *ext_info;
91
92                 memset(inbox + 256, 0, 256);
93                 ext_info = inbox + 256;
94
95                 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
96                 ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
97                 ext_info->sl     = in_wc->sl << 4;
98                 ext_info->g_path = in_wc->dlid_path_bits |
99                         (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
100                 ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
101
102                 if (in_grh)
103                         memcpy(ext_info->grh, in_grh, 40);
104
105                 op_modifier |= 0x4;
106
107                 in_modifier |= in_wc->slid << 16;
108         }
109
110         err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma,
111                            in_modifier, op_modifier,
112                            MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
113                            MLX4_CMD_NATIVE);
114
115         if (!err)
116                 memcpy(response_mad, outmailbox->buf, 256);
117
118         mlx4_free_cmd_mailbox(dev->dev, inmailbox);
119         mlx4_free_cmd_mailbox(dev->dev, outmailbox);
120
121         return err;
122 }
123
124 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
125 {
126         struct ib_ah *new_ah;
127         struct ib_ah_attr ah_attr;
128
129         if (!dev->send_agent[port_num - 1][0])
130                 return;
131
132         memset(&ah_attr, 0, sizeof ah_attr);
133         ah_attr.dlid     = lid;
134         ah_attr.sl       = sl;
135         ah_attr.port_num = port_num;
136
137         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
138                               &ah_attr);
139         if (IS_ERR(new_ah))
140                 return;
141
142         spin_lock(&dev->sm_lock);
143         if (dev->sm_ah[port_num - 1])
144                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
145         dev->sm_ah[port_num - 1] = new_ah;
146         spin_unlock(&dev->sm_lock);
147 }
148
149 /*
150  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
151  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
152  */
153 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
154                       u16 prev_lid)
155 {
156         struct ib_port_info *pinfo;
157         u16 lid;
158
159         struct mlx4_ib_dev *dev = to_mdev(ibdev);
160         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
161              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
162             mad->mad_hdr.method == IB_MGMT_METHOD_SET)
163                 switch (mad->mad_hdr.attr_id) {
164                 case IB_SMP_ATTR_PORT_INFO:
165                         pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
166                         lid = be16_to_cpu(pinfo->lid);
167
168                         update_sm_ah(dev, port_num,
169                                      be16_to_cpu(pinfo->sm_lid),
170                                      pinfo->neighbormtu_mastersmsl & 0xf);
171
172                         if (pinfo->clientrereg_resv_subnetto & 0x80)
173                                 mlx4_ib_dispatch_event(dev, port_num,
174                                                        IB_EVENT_CLIENT_REREGISTER);
175
176                         if (prev_lid != lid)
177                                 mlx4_ib_dispatch_event(dev, port_num,
178                                                        IB_EVENT_LID_CHANGE);
179                         break;
180
181                 case IB_SMP_ATTR_PKEY_TABLE:
182                         mlx4_ib_dispatch_event(dev, port_num,
183                                                IB_EVENT_PKEY_CHANGE);
184                         break;
185
186                 case IB_SMP_ATTR_GUID_INFO:
187                         mlx4_ib_dispatch_event(dev, port_num,
188                                                IB_EVENT_GID_CHANGE);
189                         break;
190                 default:
191                         break;
192                 }
193 }
194
195 static void node_desc_override(struct ib_device *dev,
196                                struct ib_mad *mad)
197 {
198         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
199              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
200             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
201             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
202                 spin_lock(&to_mdev(dev)->sm_lock);
203                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
204                 spin_unlock(&to_mdev(dev)->sm_lock);
205         }
206 }
207
208 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
209 {
210         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
211         struct ib_mad_send_buf *send_buf;
212         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
213         int ret;
214
215         if (agent) {
216                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
217                                               IB_MGMT_MAD_DATA, GFP_ATOMIC);
218                 if (IS_ERR(send_buf))
219                         return;
220                 /*
221                  * We rely here on the fact that MLX QPs don't use the
222                  * address handle after the send is posted (this is
223                  * wrong following the IB spec strictly, but we know
224                  * it's OK for our devices).
225                  */
226                 spin_lock(&dev->sm_lock);
227                 memcpy(send_buf->mad, mad, sizeof *mad);
228                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
229                         ret = ib_post_send_mad(send_buf, NULL);
230                 else
231                         ret = -EINVAL;
232                 spin_unlock(&dev->sm_lock);
233
234                 if (ret)
235                         ib_free_send_mad(send_buf);
236         }
237 }
238
239 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
240                         struct ib_wc *in_wc, struct ib_grh *in_grh,
241                         struct ib_mad *in_mad, struct ib_mad *out_mad)
242 {
243         u16 slid, prev_lid = 0;
244         int err;
245         struct ib_port_attr pattr;
246
247         if (in_wc && in_wc->qp->qp_num) {
248                 pr_debug("received MAD: slid:%d sqpn:%d "
249                         "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
250                         in_wc->slid, in_wc->src_qp,
251                         in_wc->dlid_path_bits,
252                         in_wc->qp->qp_num,
253                         in_wc->wc_flags,
254                         in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
255                         be16_to_cpu(in_mad->mad_hdr.attr_id));
256                 if (in_wc->wc_flags & IB_WC_GRH) {
257                         pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
258                                  be64_to_cpu(in_grh->sgid.global.subnet_prefix),
259                                  be64_to_cpu(in_grh->sgid.global.interface_id));
260                         pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
261                                  be64_to_cpu(in_grh->dgid.global.subnet_prefix),
262                                  be64_to_cpu(in_grh->dgid.global.interface_id));
263                 }
264         }
265
266         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
267
268         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
269                 forward_trap(to_mdev(ibdev), port_num, in_mad);
270                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
271         }
272
273         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
274             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
275                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
276                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
277                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
278                         return IB_MAD_RESULT_SUCCESS;
279
280                 /*
281                  * Don't process SMInfo queries -- the SMA can't handle them.
282                  */
283                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
284                         return IB_MAD_RESULT_SUCCESS;
285         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
286                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
287                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
288                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
289                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
290                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
291                         return IB_MAD_RESULT_SUCCESS;
292         } else
293                 return IB_MAD_RESULT_SUCCESS;
294
295         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
296              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
297             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
298             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
299             !ib_query_port(ibdev, port_num, &pattr))
300                 prev_lid = pattr.lid;
301
302         err = mlx4_MAD_IFC(to_mdev(ibdev),
303                            mad_flags & IB_MAD_IGNORE_MKEY,
304                            mad_flags & IB_MAD_IGNORE_BKEY,
305                            port_num, in_wc, in_grh, in_mad, out_mad);
306         if (err)
307                 return IB_MAD_RESULT_FAILURE;
308
309         if (!out_mad->mad_hdr.status) {
310                 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
311                         smp_snoop(ibdev, port_num, in_mad, prev_lid);
312                 node_desc_override(ibdev, out_mad);
313         }
314
315         /* set return bit in status of directed route responses */
316         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
317                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
318
319         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
320                 /* no response for trap repress */
321                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
322
323         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
324 }
325
326 static void edit_counter(struct mlx4_counter *cnt,
327                                         struct ib_pma_portcounters *pma_cnt)
328 {
329         pma_cnt->port_xmit_data = cpu_to_be32((be64_to_cpu(cnt->tx_bytes)>>2));
330         pma_cnt->port_rcv_data  = cpu_to_be32((be64_to_cpu(cnt->rx_bytes)>>2));
331         pma_cnt->port_xmit_packets = cpu_to_be32(be64_to_cpu(cnt->tx_frames));
332         pma_cnt->port_rcv_packets  = cpu_to_be32(be64_to_cpu(cnt->rx_frames));
333 }
334
335 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
336                         struct ib_wc *in_wc, struct ib_grh *in_grh,
337                         struct ib_mad *in_mad, struct ib_mad *out_mad)
338 {
339         struct mlx4_cmd_mailbox *mailbox;
340         struct mlx4_ib_dev *dev = to_mdev(ibdev);
341         int err;
342         u32 inmod = dev->counters[port_num - 1] & 0xffff;
343         u8 mode;
344
345         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
346                 return -EINVAL;
347
348         mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
349         if (IS_ERR(mailbox))
350                 return IB_MAD_RESULT_FAILURE;
351
352         err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
353                            MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
354                            MLX4_CMD_WRAPPED);
355         if (err)
356                 err = IB_MAD_RESULT_FAILURE;
357         else {
358                 memset(out_mad->data, 0, sizeof out_mad->data);
359                 mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
360                 switch (mode & 0xf) {
361                 case 0:
362                         edit_counter(mailbox->buf,
363                                                 (void *)(out_mad->data + 40));
364                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
365                         break;
366                 default:
367                         err = IB_MAD_RESULT_FAILURE;
368                 }
369         }
370
371         mlx4_free_cmd_mailbox(dev->dev, mailbox);
372
373         return err;
374 }
375
376 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
377                         struct ib_wc *in_wc, struct ib_grh *in_grh,
378                         struct ib_mad *in_mad, struct ib_mad *out_mad)
379 {
380         switch (rdma_port_get_link_layer(ibdev, port_num)) {
381         case IB_LINK_LAYER_INFINIBAND:
382                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
383                                       in_grh, in_mad, out_mad);
384         case IB_LINK_LAYER_ETHERNET:
385                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
386                                           in_grh, in_mad, out_mad);
387         default:
388                 return -EINVAL;
389         }
390 }
391
392 static void send_handler(struct ib_mad_agent *agent,
393                          struct ib_mad_send_wc *mad_send_wc)
394 {
395         ib_free_send_mad(mad_send_wc->send_buf);
396 }
397
398 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
399 {
400         struct ib_mad_agent *agent;
401         int p, q;
402         int ret;
403         enum rdma_link_layer ll;
404
405         for (p = 0; p < dev->num_ports; ++p) {
406                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
407                 for (q = 0; q <= 1; ++q) {
408                         if (ll == IB_LINK_LAYER_INFINIBAND) {
409                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
410                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
411                                                               NULL, 0, send_handler,
412                                                               NULL, NULL);
413                                 if (IS_ERR(agent)) {
414                                         ret = PTR_ERR(agent);
415                                         goto err;
416                                 }
417                                 dev->send_agent[p][q] = agent;
418                         } else
419                                 dev->send_agent[p][q] = NULL;
420                 }
421         }
422
423         return 0;
424
425 err:
426         for (p = 0; p < dev->num_ports; ++p)
427                 for (q = 0; q <= 1; ++q)
428                         if (dev->send_agent[p][q])
429                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
430
431         return ret;
432 }
433
434 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
435 {
436         struct ib_mad_agent *agent;
437         int p, q;
438
439         for (p = 0; p < dev->num_ports; ++p) {
440                 for (q = 0; q <= 1; ++q) {
441                         agent = dev->send_agent[p][q];
442                         if (agent) {
443                                 dev->send_agent[p][q] = NULL;
444                                 ib_unregister_mad_agent(agent);
445                         }
446                 }
447
448                 if (dev->sm_ah[p])
449                         ib_destroy_ah(dev->sm_ah[p]);
450         }
451 }
452
453 void handle_port_mgmt_change_event(struct work_struct *work)
454 {
455         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
456         struct mlx4_ib_dev *dev = ew->ib_dev;
457         struct mlx4_eqe *eqe = &(ew->ib_eqe);
458         u8 port = eqe->event.port_mgmt_change.port;
459         u32 changed_attr;
460
461         switch (eqe->subtype) {
462         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
463                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
464
465                 /* Update the SM ah - This should be done before handling
466                    the other changed attributes so that MADs can be sent to the SM */
467                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
468                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
469                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
470                         update_sm_ah(dev, port, lid, sl);
471                 }
472
473                 /* Check if it is a lid change event */
474                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
475                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_LID_CHANGE);
476
477                 /* Generate GUID changed event */
478                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK)
479                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
480
481                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
482                         mlx4_ib_dispatch_event(dev, port,
483                                                IB_EVENT_CLIENT_REREGISTER);
484                 break;
485
486         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
487                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
488                 break;
489         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
490                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
491                 break;
492         default:
493                 pr_warn("Unsupported subtype 0x%x for "
494                         "Port Management Change event\n", eqe->subtype);
495         }
496
497         kfree(ew);
498 }
499
500 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
501                             enum ib_event_type type)
502 {
503         struct ib_event event;
504
505         event.device            = &dev->ib_dev;
506         event.element.port_num  = port_num;
507         event.event             = type;
508
509         ib_dispatch_event(&event);
510 }