Merge branch 'master' of git://1984.lsi.us.es/nf
[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         unsigned long flags;
129
130         if (!dev->send_agent[port_num - 1][0])
131                 return;
132
133         memset(&ah_attr, 0, sizeof ah_attr);
134         ah_attr.dlid     = lid;
135         ah_attr.sl       = sl;
136         ah_attr.port_num = port_num;
137
138         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
139                               &ah_attr);
140         if (IS_ERR(new_ah))
141                 return;
142
143         spin_lock_irqsave(&dev->sm_lock, flags);
144         if (dev->sm_ah[port_num - 1])
145                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
146         dev->sm_ah[port_num - 1] = new_ah;
147         spin_unlock_irqrestore(&dev->sm_lock, flags);
148 }
149
150 /*
151  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
152  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
153  */
154 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
155                       u16 prev_lid)
156 {
157         struct ib_port_info *pinfo;
158         u16 lid;
159
160         struct mlx4_ib_dev *dev = to_mdev(ibdev);
161         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
162              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
163             mad->mad_hdr.method == IB_MGMT_METHOD_SET)
164                 switch (mad->mad_hdr.attr_id) {
165                 case IB_SMP_ATTR_PORT_INFO:
166                         pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
167                         lid = be16_to_cpu(pinfo->lid);
168
169                         update_sm_ah(dev, port_num,
170                                      be16_to_cpu(pinfo->sm_lid),
171                                      pinfo->neighbormtu_mastersmsl & 0xf);
172
173                         if (pinfo->clientrereg_resv_subnetto & 0x80)
174                                 mlx4_ib_dispatch_event(dev, port_num,
175                                                        IB_EVENT_CLIENT_REREGISTER);
176
177                         if (prev_lid != lid)
178                                 mlx4_ib_dispatch_event(dev, port_num,
179                                                        IB_EVENT_LID_CHANGE);
180                         break;
181
182                 case IB_SMP_ATTR_PKEY_TABLE:
183                         mlx4_ib_dispatch_event(dev, port_num,
184                                                IB_EVENT_PKEY_CHANGE);
185                         break;
186
187                 case IB_SMP_ATTR_GUID_INFO:
188                         /* paravirtualized master's guid is guid 0 -- does not change */
189                         if (!mlx4_is_master(dev->dev))
190                                 mlx4_ib_dispatch_event(dev, port_num,
191                                                        IB_EVENT_GID_CHANGE);
192                         break;
193                 default:
194                         break;
195                 }
196 }
197
198 static void node_desc_override(struct ib_device *dev,
199                                struct ib_mad *mad)
200 {
201         unsigned long flags;
202
203         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
204              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
205             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
206             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
207                 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
208                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
209                 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
210         }
211 }
212
213 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
214 {
215         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
216         struct ib_mad_send_buf *send_buf;
217         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
218         int ret;
219         unsigned long flags;
220
221         if (agent) {
222                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
223                                               IB_MGMT_MAD_DATA, GFP_ATOMIC);
224                 if (IS_ERR(send_buf))
225                         return;
226                 /*
227                  * We rely here on the fact that MLX QPs don't use the
228                  * address handle after the send is posted (this is
229                  * wrong following the IB spec strictly, but we know
230                  * it's OK for our devices).
231                  */
232                 spin_lock_irqsave(&dev->sm_lock, flags);
233                 memcpy(send_buf->mad, mad, sizeof *mad);
234                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
235                         ret = ib_post_send_mad(send_buf, NULL);
236                 else
237                         ret = -EINVAL;
238                 spin_unlock_irqrestore(&dev->sm_lock, flags);
239
240                 if (ret)
241                         ib_free_send_mad(send_buf);
242         }
243 }
244
245 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
246                         struct ib_wc *in_wc, struct ib_grh *in_grh,
247                         struct ib_mad *in_mad, struct ib_mad *out_mad)
248 {
249         u16 slid, prev_lid = 0;
250         int err;
251         struct ib_port_attr pattr;
252
253         if (in_wc && in_wc->qp->qp_num) {
254                 pr_debug("received MAD: slid:%d sqpn:%d "
255                         "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
256                         in_wc->slid, in_wc->src_qp,
257                         in_wc->dlid_path_bits,
258                         in_wc->qp->qp_num,
259                         in_wc->wc_flags,
260                         in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
261                         be16_to_cpu(in_mad->mad_hdr.attr_id));
262                 if (in_wc->wc_flags & IB_WC_GRH) {
263                         pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
264                                  be64_to_cpu(in_grh->sgid.global.subnet_prefix),
265                                  be64_to_cpu(in_grh->sgid.global.interface_id));
266                         pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
267                                  be64_to_cpu(in_grh->dgid.global.subnet_prefix),
268                                  be64_to_cpu(in_grh->dgid.global.interface_id));
269                 }
270         }
271
272         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
273
274         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
275                 forward_trap(to_mdev(ibdev), port_num, in_mad);
276                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
277         }
278
279         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
280             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
281                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
282                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
283                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
284                         return IB_MAD_RESULT_SUCCESS;
285
286                 /*
287                  * Don't process SMInfo queries -- the SMA can't handle them.
288                  */
289                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
290                         return IB_MAD_RESULT_SUCCESS;
291         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
292                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
293                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
294                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
295                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
296                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
297                         return IB_MAD_RESULT_SUCCESS;
298         } else
299                 return IB_MAD_RESULT_SUCCESS;
300
301         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
302              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
303             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
304             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
305             !ib_query_port(ibdev, port_num, &pattr))
306                 prev_lid = pattr.lid;
307
308         err = mlx4_MAD_IFC(to_mdev(ibdev),
309                            mad_flags & IB_MAD_IGNORE_MKEY,
310                            mad_flags & IB_MAD_IGNORE_BKEY,
311                            port_num, in_wc, in_grh, in_mad, out_mad);
312         if (err)
313                 return IB_MAD_RESULT_FAILURE;
314
315         if (!out_mad->mad_hdr.status) {
316                 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
317                         smp_snoop(ibdev, port_num, in_mad, prev_lid);
318                 node_desc_override(ibdev, out_mad);
319         }
320
321         /* set return bit in status of directed route responses */
322         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
323                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
324
325         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
326                 /* no response for trap repress */
327                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
328
329         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
330 }
331
332 static void edit_counter(struct mlx4_counter *cnt,
333                                         struct ib_pma_portcounters *pma_cnt)
334 {
335         pma_cnt->port_xmit_data = cpu_to_be32((be64_to_cpu(cnt->tx_bytes)>>2));
336         pma_cnt->port_rcv_data  = cpu_to_be32((be64_to_cpu(cnt->rx_bytes)>>2));
337         pma_cnt->port_xmit_packets = cpu_to_be32(be64_to_cpu(cnt->tx_frames));
338         pma_cnt->port_rcv_packets  = cpu_to_be32(be64_to_cpu(cnt->rx_frames));
339 }
340
341 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
342                         struct ib_wc *in_wc, struct ib_grh *in_grh,
343                         struct ib_mad *in_mad, struct ib_mad *out_mad)
344 {
345         struct mlx4_cmd_mailbox *mailbox;
346         struct mlx4_ib_dev *dev = to_mdev(ibdev);
347         int err;
348         u32 inmod = dev->counters[port_num - 1] & 0xffff;
349         u8 mode;
350
351         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
352                 return -EINVAL;
353
354         mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
355         if (IS_ERR(mailbox))
356                 return IB_MAD_RESULT_FAILURE;
357
358         err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
359                            MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
360                            MLX4_CMD_WRAPPED);
361         if (err)
362                 err = IB_MAD_RESULT_FAILURE;
363         else {
364                 memset(out_mad->data, 0, sizeof out_mad->data);
365                 mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
366                 switch (mode & 0xf) {
367                 case 0:
368                         edit_counter(mailbox->buf,
369                                                 (void *)(out_mad->data + 40));
370                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
371                         break;
372                 default:
373                         err = IB_MAD_RESULT_FAILURE;
374                 }
375         }
376
377         mlx4_free_cmd_mailbox(dev->dev, mailbox);
378
379         return err;
380 }
381
382 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
383                         struct ib_wc *in_wc, struct ib_grh *in_grh,
384                         struct ib_mad *in_mad, struct ib_mad *out_mad)
385 {
386         switch (rdma_port_get_link_layer(ibdev, port_num)) {
387         case IB_LINK_LAYER_INFINIBAND:
388                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
389                                       in_grh, in_mad, out_mad);
390         case IB_LINK_LAYER_ETHERNET:
391                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
392                                           in_grh, in_mad, out_mad);
393         default:
394                 return -EINVAL;
395         }
396 }
397
398 static void send_handler(struct ib_mad_agent *agent,
399                          struct ib_mad_send_wc *mad_send_wc)
400 {
401         ib_free_send_mad(mad_send_wc->send_buf);
402 }
403
404 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
405 {
406         struct ib_mad_agent *agent;
407         int p, q;
408         int ret;
409         enum rdma_link_layer ll;
410
411         for (p = 0; p < dev->num_ports; ++p) {
412                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
413                 for (q = 0; q <= 1; ++q) {
414                         if (ll == IB_LINK_LAYER_INFINIBAND) {
415                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
416                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
417                                                               NULL, 0, send_handler,
418                                                               NULL, NULL);
419                                 if (IS_ERR(agent)) {
420                                         ret = PTR_ERR(agent);
421                                         goto err;
422                                 }
423                                 dev->send_agent[p][q] = agent;
424                         } else
425                                 dev->send_agent[p][q] = NULL;
426                 }
427         }
428
429         return 0;
430
431 err:
432         for (p = 0; p < dev->num_ports; ++p)
433                 for (q = 0; q <= 1; ++q)
434                         if (dev->send_agent[p][q])
435                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
436
437         return ret;
438 }
439
440 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
441 {
442         struct ib_mad_agent *agent;
443         int p, q;
444
445         for (p = 0; p < dev->num_ports; ++p) {
446                 for (q = 0; q <= 1; ++q) {
447                         agent = dev->send_agent[p][q];
448                         if (agent) {
449                                 dev->send_agent[p][q] = NULL;
450                                 ib_unregister_mad_agent(agent);
451                         }
452                 }
453
454                 if (dev->sm_ah[p])
455                         ib_destroy_ah(dev->sm_ah[p]);
456         }
457 }
458
459 void handle_port_mgmt_change_event(struct work_struct *work)
460 {
461         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
462         struct mlx4_ib_dev *dev = ew->ib_dev;
463         struct mlx4_eqe *eqe = &(ew->ib_eqe);
464         u8 port = eqe->event.port_mgmt_change.port;
465         u32 changed_attr;
466
467         switch (eqe->subtype) {
468         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
469                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
470
471                 /* Update the SM ah - This should be done before handling
472                    the other changed attributes so that MADs can be sent to the SM */
473                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
474                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
475                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
476                         update_sm_ah(dev, port, lid, sl);
477                 }
478
479                 /* Check if it is a lid change event */
480                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
481                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_LID_CHANGE);
482
483                 /* Generate GUID changed event */
484                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK)
485                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
486
487                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
488                         mlx4_ib_dispatch_event(dev, port,
489                                                IB_EVENT_CLIENT_REREGISTER);
490                 break;
491
492         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
493                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
494                 break;
495         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
496                 /* paravirtualized master's guid is guid 0 -- does not change */
497                 if (!mlx4_is_master(dev->dev))
498                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
499                 break;
500         default:
501                 pr_warn("Unsupported subtype 0x%x for "
502                         "Port Management Change event\n", eqe->subtype);
503         }
504
505         kfree(ew);
506 }
507
508 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
509                             enum ib_event_type type)
510 {
511         struct ib_event event;
512
513         event.device            = &dev->ib_dev;
514         event.element.port_num  = port_num;
515         event.event             = type;
516
517         ib_dispatch_event(&event);
518 }