]> asedeno.scripts.mit.edu Git - linux.git/blob - drivers/infiniband/hw/mlx5/devx.c
9d0a18cf9e5ebd28806dcd250d7b1c3903b574aa
[linux.git] / drivers / infiniband / hw / mlx5 / devx.c
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2018, Mellanox Technologies inc.  All rights reserved.
4  */
5
6 #include <rdma/ib_user_verbs.h>
7 #include <rdma/ib_verbs.h>
8 #include <rdma/uverbs_types.h>
9 #include <rdma/uverbs_ioctl.h>
10 #include <rdma/mlx5_user_ioctl_cmds.h>
11 #include <rdma/mlx5_user_ioctl_verbs.h>
12 #include <rdma/ib_umem.h>
13 #include <rdma/uverbs_std_types.h>
14 #include <linux/mlx5/driver.h>
15 #include <linux/mlx5/fs.h>
16 #include "mlx5_ib.h"
17 #include <linux/xarray.h>
18
19 #define UVERBS_MODULE_NAME mlx5_ib
20 #include <rdma/uverbs_named_ioctl.h>
21
22 static void dispatch_event_fd(struct list_head *fd_list, const void *data);
23
24 enum devx_obj_flags {
25         DEVX_OBJ_FLAGS_INDIRECT_MKEY = 1 << 0,
26         DEVX_OBJ_FLAGS_DCT = 1 << 1,
27         DEVX_OBJ_FLAGS_CQ = 1 << 2,
28 };
29
30 struct devx_async_data {
31         struct mlx5_ib_dev *mdev;
32         struct list_head list;
33         struct ib_uobject *fd_uobj;
34         struct mlx5_async_work cb_work;
35         u16 cmd_out_len;
36         /* must be last field in this structure */
37         struct mlx5_ib_uapi_devx_async_cmd_hdr hdr;
38 };
39
40 struct devx_async_event_data {
41         struct list_head list; /* headed in ev_file->event_list */
42         struct mlx5_ib_uapi_devx_async_event_hdr hdr;
43 };
44
45 /* first level XA value data structure */
46 struct devx_event {
47         struct xarray object_ids; /* second XA level, Key = object id */
48         struct list_head unaffiliated_list;
49 };
50
51 /* second level XA value data structure */
52 struct devx_obj_event {
53         struct rcu_head rcu;
54         struct list_head obj_sub_list;
55 };
56
57 struct devx_event_subscription {
58         struct list_head file_list; /* headed in ev_file->
59                                      * subscribed_events_list
60                                      */
61         struct list_head xa_list; /* headed in devx_event->unaffiliated_list or
62                                    * devx_obj_event->obj_sub_list
63                                    */
64         struct list_head obj_list; /* headed in devx_object */
65         struct list_head event_list; /* headed in ev_file->event_list or in
66                                       * temp list via subscription
67                                       */
68
69         u8 is_cleaned:1;
70         u32 xa_key_level1;
71         u32 xa_key_level2;
72         struct rcu_head rcu;
73         u64 cookie;
74         struct devx_async_event_file *ev_file;
75         struct file *filp; /* Upon hot unplug we need a direct access to */
76         struct eventfd_ctx *eventfd;
77 };
78
79 struct devx_async_event_file {
80         struct ib_uobject uobj;
81         /* Head of events that are subscribed to this FD */
82         struct list_head subscribed_events_list;
83         spinlock_t lock;
84         wait_queue_head_t poll_wait;
85         struct list_head event_list;
86         struct mlx5_ib_dev *dev;
87         u8 omit_data:1;
88         u8 is_overflow_err:1;
89         u8 is_destroyed:1;
90 };
91
92 #define MLX5_MAX_DESTROY_INBOX_SIZE_DW MLX5_ST_SZ_DW(delete_fte_in)
93 struct devx_obj {
94         struct mlx5_ib_dev      *ib_dev;
95         u64                     obj_id;
96         u32                     dinlen; /* destroy inbox length */
97         u32                     dinbox[MLX5_MAX_DESTROY_INBOX_SIZE_DW];
98         u32                     flags;
99         union {
100                 struct mlx5_ib_devx_mr  devx_mr;
101                 struct mlx5_core_dct    core_dct;
102                 struct mlx5_core_cq     core_cq;
103                 u32                     flow_counter_bulk_size;
104         };
105         struct list_head event_sub; /* holds devx_event_subscription entries */
106 };
107
108 struct devx_umem {
109         struct mlx5_core_dev            *mdev;
110         struct ib_umem                  *umem;
111         u32                             page_offset;
112         int                             page_shift;
113         int                             ncont;
114         u32                             dinlen;
115         u32                             dinbox[MLX5_ST_SZ_DW(general_obj_in_cmd_hdr)];
116 };
117
118 struct devx_umem_reg_cmd {
119         void                            *in;
120         u32                             inlen;
121         u32                             out[MLX5_ST_SZ_DW(general_obj_out_cmd_hdr)];
122 };
123
124 static struct mlx5_ib_ucontext *
125 devx_ufile2uctx(const struct uverbs_attr_bundle *attrs)
126 {
127         return to_mucontext(ib_uverbs_get_ucontext(attrs));
128 }
129
130 int mlx5_ib_devx_create(struct mlx5_ib_dev *dev, bool is_user)
131 {
132         u32 in[MLX5_ST_SZ_DW(create_uctx_in)] = {0};
133         u32 out[MLX5_ST_SZ_DW(general_obj_out_cmd_hdr)] = {0};
134         void *uctx;
135         int err;
136         u16 uid;
137         u32 cap = 0;
138
139         /* 0 means not supported */
140         if (!MLX5_CAP_GEN(dev->mdev, log_max_uctx))
141                 return -EINVAL;
142
143         uctx = MLX5_ADDR_OF(create_uctx_in, in, uctx);
144         if (is_user && capable(CAP_NET_RAW) &&
145             (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RAW_TX))
146                 cap |= MLX5_UCTX_CAP_RAW_TX;
147         if (is_user && capable(CAP_SYS_RAWIO) &&
148             (MLX5_CAP_GEN(dev->mdev, uctx_cap) &
149              MLX5_UCTX_CAP_INTERNAL_DEV_RES))
150                 cap |= MLX5_UCTX_CAP_INTERNAL_DEV_RES;
151
152         MLX5_SET(create_uctx_in, in, opcode, MLX5_CMD_OP_CREATE_UCTX);
153         MLX5_SET(uctx, uctx, cap, cap);
154
155         err = mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
156         if (err)
157                 return err;
158
159         uid = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
160         return uid;
161 }
162
163 void mlx5_ib_devx_destroy(struct mlx5_ib_dev *dev, u16 uid)
164 {
165         u32 in[MLX5_ST_SZ_DW(destroy_uctx_in)] = {0};
166         u32 out[MLX5_ST_SZ_DW(general_obj_out_cmd_hdr)] = {0};
167
168         MLX5_SET(destroy_uctx_in, in, opcode, MLX5_CMD_OP_DESTROY_UCTX);
169         MLX5_SET(destroy_uctx_in, in, uid, uid);
170
171         mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
172 }
173
174 bool mlx5_ib_devx_is_flow_dest(void *obj, int *dest_id, int *dest_type)
175 {
176         struct devx_obj *devx_obj = obj;
177         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, devx_obj->dinbox, opcode);
178
179         switch (opcode) {
180         case MLX5_CMD_OP_DESTROY_TIR:
181                 *dest_type = MLX5_FLOW_DESTINATION_TYPE_TIR;
182                 *dest_id = MLX5_GET(general_obj_in_cmd_hdr, devx_obj->dinbox,
183                                     obj_id);
184                 return true;
185
186         case MLX5_CMD_OP_DESTROY_FLOW_TABLE:
187                 *dest_type = MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE;
188                 *dest_id = MLX5_GET(destroy_flow_table_in, devx_obj->dinbox,
189                                     table_id);
190                 return true;
191         default:
192                 return false;
193         }
194 }
195
196 bool mlx5_ib_devx_is_flow_counter(void *obj, u32 offset, u32 *counter_id)
197 {
198         struct devx_obj *devx_obj = obj;
199         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, devx_obj->dinbox, opcode);
200
201         if (opcode == MLX5_CMD_OP_DEALLOC_FLOW_COUNTER) {
202
203                 if (offset && offset >= devx_obj->flow_counter_bulk_size)
204                         return false;
205
206                 *counter_id = MLX5_GET(dealloc_flow_counter_in,
207                                        devx_obj->dinbox,
208                                        flow_counter_id);
209                 *counter_id += offset;
210                 return true;
211         }
212
213         return false;
214 }
215
216 static bool is_legacy_unaffiliated_event_num(u16 event_num)
217 {
218         switch (event_num) {
219         case MLX5_EVENT_TYPE_PORT_CHANGE:
220                 return true;
221         default:
222                 return false;
223         }
224 }
225
226 static bool is_legacy_obj_event_num(u16 event_num)
227 {
228         switch (event_num) {
229         case MLX5_EVENT_TYPE_PATH_MIG:
230         case MLX5_EVENT_TYPE_COMM_EST:
231         case MLX5_EVENT_TYPE_SQ_DRAINED:
232         case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
233         case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
234         case MLX5_EVENT_TYPE_CQ_ERROR:
235         case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
236         case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
237         case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
238         case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
239         case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
240         case MLX5_EVENT_TYPE_DCT_DRAINED:
241         case MLX5_EVENT_TYPE_COMP:
242         case MLX5_EVENT_TYPE_DCT_KEY_VIOLATION:
243         case MLX5_EVENT_TYPE_XRQ_ERROR:
244                 return true;
245         default:
246                 return false;
247         }
248 }
249
250 static u16 get_legacy_obj_type(u16 opcode)
251 {
252         switch (opcode) {
253         case MLX5_CMD_OP_CREATE_RQ:
254                 return MLX5_EVENT_QUEUE_TYPE_RQ;
255         case MLX5_CMD_OP_CREATE_QP:
256                 return MLX5_EVENT_QUEUE_TYPE_QP;
257         case MLX5_CMD_OP_CREATE_SQ:
258                 return MLX5_EVENT_QUEUE_TYPE_SQ;
259         case MLX5_CMD_OP_CREATE_DCT:
260                 return MLX5_EVENT_QUEUE_TYPE_DCT;
261         default:
262                 return 0;
263         }
264 }
265
266 static u16 get_dec_obj_type(struct devx_obj *obj, u16 event_num)
267 {
268         u16 opcode;
269
270         opcode = (obj->obj_id >> 32) & 0xffff;
271
272         if (is_legacy_obj_event_num(event_num))
273                 return get_legacy_obj_type(opcode);
274
275         switch (opcode) {
276         case MLX5_CMD_OP_CREATE_GENERAL_OBJECT:
277                 return (obj->obj_id >> 48);
278         case MLX5_CMD_OP_CREATE_RQ:
279                 return MLX5_OBJ_TYPE_RQ;
280         case MLX5_CMD_OP_CREATE_QP:
281                 return MLX5_OBJ_TYPE_QP;
282         case MLX5_CMD_OP_CREATE_SQ:
283                 return MLX5_OBJ_TYPE_SQ;
284         case MLX5_CMD_OP_CREATE_DCT:
285                 return MLX5_OBJ_TYPE_DCT;
286         case MLX5_CMD_OP_CREATE_TIR:
287                 return MLX5_OBJ_TYPE_TIR;
288         case MLX5_CMD_OP_CREATE_TIS:
289                 return MLX5_OBJ_TYPE_TIS;
290         case MLX5_CMD_OP_CREATE_PSV:
291                 return MLX5_OBJ_TYPE_PSV;
292         case MLX5_OBJ_TYPE_MKEY:
293                 return MLX5_OBJ_TYPE_MKEY;
294         case MLX5_CMD_OP_CREATE_RMP:
295                 return MLX5_OBJ_TYPE_RMP;
296         case MLX5_CMD_OP_CREATE_XRC_SRQ:
297                 return MLX5_OBJ_TYPE_XRC_SRQ;
298         case MLX5_CMD_OP_CREATE_XRQ:
299                 return MLX5_OBJ_TYPE_XRQ;
300         case MLX5_CMD_OP_CREATE_RQT:
301                 return MLX5_OBJ_TYPE_RQT;
302         case MLX5_CMD_OP_ALLOC_FLOW_COUNTER:
303                 return MLX5_OBJ_TYPE_FLOW_COUNTER;
304         case MLX5_CMD_OP_CREATE_CQ:
305                 return MLX5_OBJ_TYPE_CQ;
306         default:
307                 return 0;
308         }
309 }
310
311 static u16 get_event_obj_type(unsigned long event_type, struct mlx5_eqe *eqe)
312 {
313         switch (event_type) {
314         case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
315         case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
316         case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
317         case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
318         case MLX5_EVENT_TYPE_PATH_MIG:
319         case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
320         case MLX5_EVENT_TYPE_COMM_EST:
321         case MLX5_EVENT_TYPE_SQ_DRAINED:
322         case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
323         case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
324                 return eqe->data.qp_srq.type;
325         case MLX5_EVENT_TYPE_CQ_ERROR:
326         case MLX5_EVENT_TYPE_XRQ_ERROR:
327                 return 0;
328         case MLX5_EVENT_TYPE_DCT_DRAINED:
329         case MLX5_EVENT_TYPE_DCT_KEY_VIOLATION:
330                 return MLX5_EVENT_QUEUE_TYPE_DCT;
331         default:
332                 return MLX5_GET(affiliated_event_header, &eqe->data, obj_type);
333         }
334 }
335
336 static u32 get_dec_obj_id(u64 obj_id)
337 {
338         return (obj_id & 0xffffffff);
339 }
340
341 /*
342  * As the obj_id in the firmware is not globally unique the object type
343  * must be considered upon checking for a valid object id.
344  * For that the opcode of the creator command is encoded as part of the obj_id.
345  */
346 static u64 get_enc_obj_id(u32 opcode, u32 obj_id)
347 {
348         return ((u64)opcode << 32) | obj_id;
349 }
350
351 static u64 devx_get_obj_id(const void *in)
352 {
353         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
354         u64 obj_id;
355
356         switch (opcode) {
357         case MLX5_CMD_OP_MODIFY_GENERAL_OBJECT:
358         case MLX5_CMD_OP_QUERY_GENERAL_OBJECT:
359                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_GENERAL_OBJECT |
360                                         MLX5_GET(general_obj_in_cmd_hdr, in,
361                                                  obj_type) << 16,
362                                         MLX5_GET(general_obj_in_cmd_hdr, in,
363                                                  obj_id));
364                 break;
365         case MLX5_CMD_OP_QUERY_MKEY:
366                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_MKEY,
367                                         MLX5_GET(query_mkey_in, in,
368                                                  mkey_index));
369                 break;
370         case MLX5_CMD_OP_QUERY_CQ:
371                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_CQ,
372                                         MLX5_GET(query_cq_in, in, cqn));
373                 break;
374         case MLX5_CMD_OP_MODIFY_CQ:
375                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_CQ,
376                                         MLX5_GET(modify_cq_in, in, cqn));
377                 break;
378         case MLX5_CMD_OP_QUERY_SQ:
379                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_SQ,
380                                         MLX5_GET(query_sq_in, in, sqn));
381                 break;
382         case MLX5_CMD_OP_MODIFY_SQ:
383                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_SQ,
384                                         MLX5_GET(modify_sq_in, in, sqn));
385                 break;
386         case MLX5_CMD_OP_QUERY_RQ:
387                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RQ,
388                                         MLX5_GET(query_rq_in, in, rqn));
389                 break;
390         case MLX5_CMD_OP_MODIFY_RQ:
391                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RQ,
392                                         MLX5_GET(modify_rq_in, in, rqn));
393                 break;
394         case MLX5_CMD_OP_QUERY_RMP:
395                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RMP,
396                                         MLX5_GET(query_rmp_in, in, rmpn));
397                 break;
398         case MLX5_CMD_OP_MODIFY_RMP:
399                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RMP,
400                                         MLX5_GET(modify_rmp_in, in, rmpn));
401                 break;
402         case MLX5_CMD_OP_QUERY_RQT:
403                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RQT,
404                                         MLX5_GET(query_rqt_in, in, rqtn));
405                 break;
406         case MLX5_CMD_OP_MODIFY_RQT:
407                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RQT,
408                                         MLX5_GET(modify_rqt_in, in, rqtn));
409                 break;
410         case MLX5_CMD_OP_QUERY_TIR:
411                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_TIR,
412                                         MLX5_GET(query_tir_in, in, tirn));
413                 break;
414         case MLX5_CMD_OP_MODIFY_TIR:
415                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_TIR,
416                                         MLX5_GET(modify_tir_in, in, tirn));
417                 break;
418         case MLX5_CMD_OP_QUERY_TIS:
419                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_TIS,
420                                         MLX5_GET(query_tis_in, in, tisn));
421                 break;
422         case MLX5_CMD_OP_MODIFY_TIS:
423                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_TIS,
424                                         MLX5_GET(modify_tis_in, in, tisn));
425                 break;
426         case MLX5_CMD_OP_QUERY_FLOW_TABLE:
427                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_FLOW_TABLE,
428                                         MLX5_GET(query_flow_table_in, in,
429                                                  table_id));
430                 break;
431         case MLX5_CMD_OP_MODIFY_FLOW_TABLE:
432                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_FLOW_TABLE,
433                                         MLX5_GET(modify_flow_table_in, in,
434                                                  table_id));
435                 break;
436         case MLX5_CMD_OP_QUERY_FLOW_GROUP:
437                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_FLOW_GROUP,
438                                         MLX5_GET(query_flow_group_in, in,
439                                                  group_id));
440                 break;
441         case MLX5_CMD_OP_QUERY_FLOW_TABLE_ENTRY:
442                 obj_id = get_enc_obj_id(MLX5_CMD_OP_SET_FLOW_TABLE_ENTRY,
443                                         MLX5_GET(query_fte_in, in,
444                                                  flow_index));
445                 break;
446         case MLX5_CMD_OP_SET_FLOW_TABLE_ENTRY:
447                 obj_id = get_enc_obj_id(MLX5_CMD_OP_SET_FLOW_TABLE_ENTRY,
448                                         MLX5_GET(set_fte_in, in, flow_index));
449                 break;
450         case MLX5_CMD_OP_QUERY_Q_COUNTER:
451                 obj_id = get_enc_obj_id(MLX5_CMD_OP_ALLOC_Q_COUNTER,
452                                         MLX5_GET(query_q_counter_in, in,
453                                                  counter_set_id));
454                 break;
455         case MLX5_CMD_OP_QUERY_FLOW_COUNTER:
456                 obj_id = get_enc_obj_id(MLX5_CMD_OP_ALLOC_FLOW_COUNTER,
457                                         MLX5_GET(query_flow_counter_in, in,
458                                                  flow_counter_id));
459                 break;
460         case MLX5_CMD_OP_QUERY_MODIFY_HEADER_CONTEXT:
461                 obj_id = get_enc_obj_id(MLX5_CMD_OP_ALLOC_MODIFY_HEADER_CONTEXT,
462                                         MLX5_GET(general_obj_in_cmd_hdr, in,
463                                                  obj_id));
464                 break;
465         case MLX5_CMD_OP_QUERY_SCHEDULING_ELEMENT:
466                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_SCHEDULING_ELEMENT,
467                                         MLX5_GET(query_scheduling_element_in,
468                                                  in, scheduling_element_id));
469                 break;
470         case MLX5_CMD_OP_MODIFY_SCHEDULING_ELEMENT:
471                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_SCHEDULING_ELEMENT,
472                                         MLX5_GET(modify_scheduling_element_in,
473                                                  in, scheduling_element_id));
474                 break;
475         case MLX5_CMD_OP_ADD_VXLAN_UDP_DPORT:
476                 obj_id = get_enc_obj_id(MLX5_CMD_OP_ADD_VXLAN_UDP_DPORT,
477                                         MLX5_GET(add_vxlan_udp_dport_in, in,
478                                                  vxlan_udp_port));
479                 break;
480         case MLX5_CMD_OP_QUERY_L2_TABLE_ENTRY:
481                 obj_id = get_enc_obj_id(MLX5_CMD_OP_SET_L2_TABLE_ENTRY,
482                                         MLX5_GET(query_l2_table_entry_in, in,
483                                                  table_index));
484                 break;
485         case MLX5_CMD_OP_SET_L2_TABLE_ENTRY:
486                 obj_id = get_enc_obj_id(MLX5_CMD_OP_SET_L2_TABLE_ENTRY,
487                                         MLX5_GET(set_l2_table_entry_in, in,
488                                                  table_index));
489                 break;
490         case MLX5_CMD_OP_QUERY_QP:
491                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
492                                         MLX5_GET(query_qp_in, in, qpn));
493                 break;
494         case MLX5_CMD_OP_RST2INIT_QP:
495                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
496                                         MLX5_GET(rst2init_qp_in, in, qpn));
497                 break;
498         case MLX5_CMD_OP_INIT2RTR_QP:
499                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
500                                         MLX5_GET(init2rtr_qp_in, in, qpn));
501                 break;
502         case MLX5_CMD_OP_RTR2RTS_QP:
503                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
504                                         MLX5_GET(rtr2rts_qp_in, in, qpn));
505                 break;
506         case MLX5_CMD_OP_RTS2RTS_QP:
507                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
508                                         MLX5_GET(rts2rts_qp_in, in, qpn));
509                 break;
510         case MLX5_CMD_OP_SQERR2RTS_QP:
511                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
512                                         MLX5_GET(sqerr2rts_qp_in, in, qpn));
513                 break;
514         case MLX5_CMD_OP_2ERR_QP:
515                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
516                                         MLX5_GET(qp_2err_in, in, qpn));
517                 break;
518         case MLX5_CMD_OP_2RST_QP:
519                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
520                                         MLX5_GET(qp_2rst_in, in, qpn));
521                 break;
522         case MLX5_CMD_OP_QUERY_DCT:
523                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_DCT,
524                                         MLX5_GET(query_dct_in, in, dctn));
525                 break;
526         case MLX5_CMD_OP_QUERY_XRQ:
527         case MLX5_CMD_OP_QUERY_XRQ_DC_PARAMS_ENTRY:
528         case MLX5_CMD_OP_QUERY_XRQ_ERROR_PARAMS:
529                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_XRQ,
530                                         MLX5_GET(query_xrq_in, in, xrqn));
531                 break;
532         case MLX5_CMD_OP_QUERY_XRC_SRQ:
533                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_XRC_SRQ,
534                                         MLX5_GET(query_xrc_srq_in, in,
535                                                  xrc_srqn));
536                 break;
537         case MLX5_CMD_OP_ARM_XRC_SRQ:
538                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_XRC_SRQ,
539                                         MLX5_GET(arm_xrc_srq_in, in, xrc_srqn));
540                 break;
541         case MLX5_CMD_OP_QUERY_SRQ:
542                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_SRQ,
543                                         MLX5_GET(query_srq_in, in, srqn));
544                 break;
545         case MLX5_CMD_OP_ARM_RQ:
546                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_RQ,
547                                         MLX5_GET(arm_rq_in, in, srq_number));
548                 break;
549         case MLX5_CMD_OP_ARM_DCT_FOR_KEY_VIOLATION:
550                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_DCT,
551                                         MLX5_GET(drain_dct_in, in, dctn));
552                 break;
553         case MLX5_CMD_OP_ARM_XRQ:
554         case MLX5_CMD_OP_SET_XRQ_DC_PARAMS_ENTRY:
555         case MLX5_CMD_OP_RELEASE_XRQ_ERROR:
556         case MLX5_CMD_OP_MODIFY_XRQ:
557                 obj_id = get_enc_obj_id(MLX5_CMD_OP_CREATE_XRQ,
558                                         MLX5_GET(arm_xrq_in, in, xrqn));
559                 break;
560         case MLX5_CMD_OP_QUERY_PACKET_REFORMAT_CONTEXT:
561                 obj_id = get_enc_obj_id
562                                 (MLX5_CMD_OP_ALLOC_PACKET_REFORMAT_CONTEXT,
563                                  MLX5_GET(query_packet_reformat_context_in,
564                                           in, packet_reformat_id));
565                 break;
566         default:
567                 obj_id = 0;
568         }
569
570         return obj_id;
571 }
572
573 static bool devx_is_valid_obj_id(struct uverbs_attr_bundle *attrs,
574                                  struct ib_uobject *uobj, const void *in)
575 {
576         struct mlx5_ib_dev *dev = mlx5_udata_to_mdev(&attrs->driver_udata);
577         u64 obj_id = devx_get_obj_id(in);
578
579         if (!obj_id)
580                 return false;
581
582         switch (uobj_get_object_id(uobj)) {
583         case UVERBS_OBJECT_CQ:
584                 return get_enc_obj_id(MLX5_CMD_OP_CREATE_CQ,
585                                       to_mcq(uobj->object)->mcq.cqn) ==
586                                       obj_id;
587
588         case UVERBS_OBJECT_SRQ:
589         {
590                 struct mlx5_core_srq *srq = &(to_msrq(uobj->object)->msrq);
591                 u16 opcode;
592
593                 switch (srq->common.res) {
594                 case MLX5_RES_XSRQ:
595                         opcode = MLX5_CMD_OP_CREATE_XRC_SRQ;
596                         break;
597                 case MLX5_RES_XRQ:
598                         opcode = MLX5_CMD_OP_CREATE_XRQ;
599                         break;
600                 default:
601                         if (!dev->mdev->issi)
602                                 opcode = MLX5_CMD_OP_CREATE_SRQ;
603                         else
604                                 opcode = MLX5_CMD_OP_CREATE_RMP;
605                 }
606
607                 return get_enc_obj_id(opcode,
608                                       to_msrq(uobj->object)->msrq.srqn) ==
609                                       obj_id;
610         }
611
612         case UVERBS_OBJECT_QP:
613         {
614                 struct mlx5_ib_qp *qp = to_mqp(uobj->object);
615                 enum ib_qp_type qp_type = qp->ibqp.qp_type;
616
617                 if (qp_type == IB_QPT_RAW_PACKET ||
618                     (qp->flags & MLX5_IB_QP_UNDERLAY)) {
619                         struct mlx5_ib_raw_packet_qp *raw_packet_qp =
620                                                          &qp->raw_packet_qp;
621                         struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
622                         struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
623
624                         return (get_enc_obj_id(MLX5_CMD_OP_CREATE_RQ,
625                                                rq->base.mqp.qpn) == obj_id ||
626                                 get_enc_obj_id(MLX5_CMD_OP_CREATE_SQ,
627                                                sq->base.mqp.qpn) == obj_id ||
628                                 get_enc_obj_id(MLX5_CMD_OP_CREATE_TIR,
629                                                rq->tirn) == obj_id ||
630                                 get_enc_obj_id(MLX5_CMD_OP_CREATE_TIS,
631                                                sq->tisn) == obj_id);
632                 }
633
634                 if (qp_type == MLX5_IB_QPT_DCT)
635                         return get_enc_obj_id(MLX5_CMD_OP_CREATE_DCT,
636                                               qp->dct.mdct.mqp.qpn) == obj_id;
637
638                 return get_enc_obj_id(MLX5_CMD_OP_CREATE_QP,
639                                       qp->ibqp.qp_num) == obj_id;
640         }
641
642         case UVERBS_OBJECT_WQ:
643                 return get_enc_obj_id(MLX5_CMD_OP_CREATE_RQ,
644                                       to_mrwq(uobj->object)->core_qp.qpn) ==
645                                       obj_id;
646
647         case UVERBS_OBJECT_RWQ_IND_TBL:
648                 return get_enc_obj_id(MLX5_CMD_OP_CREATE_RQT,
649                                       to_mrwq_ind_table(uobj->object)->rqtn) ==
650                                       obj_id;
651
652         case MLX5_IB_OBJECT_DEVX_OBJ:
653                 return ((struct devx_obj *)uobj->object)->obj_id == obj_id;
654
655         default:
656                 return false;
657         }
658 }
659
660 static void devx_set_umem_valid(const void *in)
661 {
662         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
663
664         switch (opcode) {
665         case MLX5_CMD_OP_CREATE_MKEY:
666                 MLX5_SET(create_mkey_in, in, mkey_umem_valid, 1);
667                 break;
668         case MLX5_CMD_OP_CREATE_CQ:
669         {
670                 void *cqc;
671
672                 MLX5_SET(create_cq_in, in, cq_umem_valid, 1);
673                 cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
674                 MLX5_SET(cqc, cqc, dbr_umem_valid, 1);
675                 break;
676         }
677         case MLX5_CMD_OP_CREATE_QP:
678         {
679                 void *qpc;
680
681                 qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
682                 MLX5_SET(qpc, qpc, dbr_umem_valid, 1);
683                 MLX5_SET(create_qp_in, in, wq_umem_valid, 1);
684                 break;
685         }
686
687         case MLX5_CMD_OP_CREATE_RQ:
688         {
689                 void *rqc, *wq;
690
691                 rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
692                 wq  = MLX5_ADDR_OF(rqc, rqc, wq);
693                 MLX5_SET(wq, wq, dbr_umem_valid, 1);
694                 MLX5_SET(wq, wq, wq_umem_valid, 1);
695                 break;
696         }
697
698         case MLX5_CMD_OP_CREATE_SQ:
699         {
700                 void *sqc, *wq;
701
702                 sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
703                 wq = MLX5_ADDR_OF(sqc, sqc, wq);
704                 MLX5_SET(wq, wq, dbr_umem_valid, 1);
705                 MLX5_SET(wq, wq, wq_umem_valid, 1);
706                 break;
707         }
708
709         case MLX5_CMD_OP_MODIFY_CQ:
710                 MLX5_SET(modify_cq_in, in, cq_umem_valid, 1);
711                 break;
712
713         case MLX5_CMD_OP_CREATE_RMP:
714         {
715                 void *rmpc, *wq;
716
717                 rmpc = MLX5_ADDR_OF(create_rmp_in, in, ctx);
718                 wq = MLX5_ADDR_OF(rmpc, rmpc, wq);
719                 MLX5_SET(wq, wq, dbr_umem_valid, 1);
720                 MLX5_SET(wq, wq, wq_umem_valid, 1);
721                 break;
722         }
723
724         case MLX5_CMD_OP_CREATE_XRQ:
725         {
726                 void *xrqc, *wq;
727
728                 xrqc = MLX5_ADDR_OF(create_xrq_in, in, xrq_context);
729                 wq = MLX5_ADDR_OF(xrqc, xrqc, wq);
730                 MLX5_SET(wq, wq, dbr_umem_valid, 1);
731                 MLX5_SET(wq, wq, wq_umem_valid, 1);
732                 break;
733         }
734
735         case MLX5_CMD_OP_CREATE_XRC_SRQ:
736         {
737                 void *xrc_srqc;
738
739                 MLX5_SET(create_xrc_srq_in, in, xrc_srq_umem_valid, 1);
740                 xrc_srqc = MLX5_ADDR_OF(create_xrc_srq_in, in,
741                                         xrc_srq_context_entry);
742                 MLX5_SET(xrc_srqc, xrc_srqc, dbr_umem_valid, 1);
743                 break;
744         }
745
746         default:
747                 return;
748         }
749 }
750
751 static bool devx_is_obj_create_cmd(const void *in, u16 *opcode)
752 {
753         *opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
754
755         switch (*opcode) {
756         case MLX5_CMD_OP_CREATE_GENERAL_OBJECT:
757         case MLX5_CMD_OP_CREATE_MKEY:
758         case MLX5_CMD_OP_CREATE_CQ:
759         case MLX5_CMD_OP_ALLOC_PD:
760         case MLX5_CMD_OP_ALLOC_TRANSPORT_DOMAIN:
761         case MLX5_CMD_OP_CREATE_RMP:
762         case MLX5_CMD_OP_CREATE_SQ:
763         case MLX5_CMD_OP_CREATE_RQ:
764         case MLX5_CMD_OP_CREATE_RQT:
765         case MLX5_CMD_OP_CREATE_TIR:
766         case MLX5_CMD_OP_CREATE_TIS:
767         case MLX5_CMD_OP_ALLOC_Q_COUNTER:
768         case MLX5_CMD_OP_CREATE_FLOW_TABLE:
769         case MLX5_CMD_OP_CREATE_FLOW_GROUP:
770         case MLX5_CMD_OP_ALLOC_FLOW_COUNTER:
771         case MLX5_CMD_OP_ALLOC_PACKET_REFORMAT_CONTEXT:
772         case MLX5_CMD_OP_ALLOC_MODIFY_HEADER_CONTEXT:
773         case MLX5_CMD_OP_CREATE_SCHEDULING_ELEMENT:
774         case MLX5_CMD_OP_ADD_VXLAN_UDP_DPORT:
775         case MLX5_CMD_OP_SET_L2_TABLE_ENTRY:
776         case MLX5_CMD_OP_CREATE_QP:
777         case MLX5_CMD_OP_CREATE_SRQ:
778         case MLX5_CMD_OP_CREATE_XRC_SRQ:
779         case MLX5_CMD_OP_CREATE_DCT:
780         case MLX5_CMD_OP_CREATE_XRQ:
781         case MLX5_CMD_OP_ATTACH_TO_MCG:
782         case MLX5_CMD_OP_ALLOC_XRCD:
783                 return true;
784         case MLX5_CMD_OP_SET_FLOW_TABLE_ENTRY:
785         {
786                 u16 op_mod = MLX5_GET(set_fte_in, in, op_mod);
787                 if (op_mod == 0)
788                         return true;
789                 return false;
790         }
791         case MLX5_CMD_OP_CREATE_PSV:
792         {
793                 u8 num_psv = MLX5_GET(create_psv_in, in, num_psv);
794
795                 if (num_psv == 1)
796                         return true;
797                 return false;
798         }
799         default:
800                 return false;
801         }
802 }
803
804 static bool devx_is_obj_modify_cmd(const void *in)
805 {
806         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
807
808         switch (opcode) {
809         case MLX5_CMD_OP_MODIFY_GENERAL_OBJECT:
810         case MLX5_CMD_OP_MODIFY_CQ:
811         case MLX5_CMD_OP_MODIFY_RMP:
812         case MLX5_CMD_OP_MODIFY_SQ:
813         case MLX5_CMD_OP_MODIFY_RQ:
814         case MLX5_CMD_OP_MODIFY_RQT:
815         case MLX5_CMD_OP_MODIFY_TIR:
816         case MLX5_CMD_OP_MODIFY_TIS:
817         case MLX5_CMD_OP_MODIFY_FLOW_TABLE:
818         case MLX5_CMD_OP_MODIFY_SCHEDULING_ELEMENT:
819         case MLX5_CMD_OP_ADD_VXLAN_UDP_DPORT:
820         case MLX5_CMD_OP_SET_L2_TABLE_ENTRY:
821         case MLX5_CMD_OP_RST2INIT_QP:
822         case MLX5_CMD_OP_INIT2RTR_QP:
823         case MLX5_CMD_OP_RTR2RTS_QP:
824         case MLX5_CMD_OP_RTS2RTS_QP:
825         case MLX5_CMD_OP_SQERR2RTS_QP:
826         case MLX5_CMD_OP_2ERR_QP:
827         case MLX5_CMD_OP_2RST_QP:
828         case MLX5_CMD_OP_ARM_XRC_SRQ:
829         case MLX5_CMD_OP_ARM_RQ:
830         case MLX5_CMD_OP_ARM_DCT_FOR_KEY_VIOLATION:
831         case MLX5_CMD_OP_ARM_XRQ:
832         case MLX5_CMD_OP_SET_XRQ_DC_PARAMS_ENTRY:
833         case MLX5_CMD_OP_RELEASE_XRQ_ERROR:
834         case MLX5_CMD_OP_MODIFY_XRQ:
835                 return true;
836         case MLX5_CMD_OP_SET_FLOW_TABLE_ENTRY:
837         {
838                 u16 op_mod = MLX5_GET(set_fte_in, in, op_mod);
839
840                 if (op_mod == 1)
841                         return true;
842                 return false;
843         }
844         default:
845                 return false;
846         }
847 }
848
849 static bool devx_is_obj_query_cmd(const void *in)
850 {
851         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
852
853         switch (opcode) {
854         case MLX5_CMD_OP_QUERY_GENERAL_OBJECT:
855         case MLX5_CMD_OP_QUERY_MKEY:
856         case MLX5_CMD_OP_QUERY_CQ:
857         case MLX5_CMD_OP_QUERY_RMP:
858         case MLX5_CMD_OP_QUERY_SQ:
859         case MLX5_CMD_OP_QUERY_RQ:
860         case MLX5_CMD_OP_QUERY_RQT:
861         case MLX5_CMD_OP_QUERY_TIR:
862         case MLX5_CMD_OP_QUERY_TIS:
863         case MLX5_CMD_OP_QUERY_Q_COUNTER:
864         case MLX5_CMD_OP_QUERY_FLOW_TABLE:
865         case MLX5_CMD_OP_QUERY_FLOW_GROUP:
866         case MLX5_CMD_OP_QUERY_FLOW_TABLE_ENTRY:
867         case MLX5_CMD_OP_QUERY_FLOW_COUNTER:
868         case MLX5_CMD_OP_QUERY_MODIFY_HEADER_CONTEXT:
869         case MLX5_CMD_OP_QUERY_SCHEDULING_ELEMENT:
870         case MLX5_CMD_OP_QUERY_L2_TABLE_ENTRY:
871         case MLX5_CMD_OP_QUERY_QP:
872         case MLX5_CMD_OP_QUERY_SRQ:
873         case MLX5_CMD_OP_QUERY_XRC_SRQ:
874         case MLX5_CMD_OP_QUERY_DCT:
875         case MLX5_CMD_OP_QUERY_XRQ:
876         case MLX5_CMD_OP_QUERY_XRQ_DC_PARAMS_ENTRY:
877         case MLX5_CMD_OP_QUERY_XRQ_ERROR_PARAMS:
878         case MLX5_CMD_OP_QUERY_PACKET_REFORMAT_CONTEXT:
879                 return true;
880         default:
881                 return false;
882         }
883 }
884
885 static bool devx_is_whitelist_cmd(void *in)
886 {
887         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
888
889         switch (opcode) {
890         case MLX5_CMD_OP_QUERY_HCA_CAP:
891         case MLX5_CMD_OP_QUERY_HCA_VPORT_CONTEXT:
892         case MLX5_CMD_OP_QUERY_ESW_VPORT_CONTEXT:
893                 return true;
894         default:
895                 return false;
896         }
897 }
898
899 static int devx_get_uid(struct mlx5_ib_ucontext *c, void *cmd_in)
900 {
901         if (devx_is_whitelist_cmd(cmd_in)) {
902                 struct mlx5_ib_dev *dev;
903
904                 if (c->devx_uid)
905                         return c->devx_uid;
906
907                 dev = to_mdev(c->ibucontext.device);
908                 if (dev->devx_whitelist_uid)
909                         return dev->devx_whitelist_uid;
910
911                 return -EOPNOTSUPP;
912         }
913
914         if (!c->devx_uid)
915                 return -EINVAL;
916
917         return c->devx_uid;
918 }
919
920 static bool devx_is_general_cmd(void *in, struct mlx5_ib_dev *dev)
921 {
922         u16 opcode = MLX5_GET(general_obj_in_cmd_hdr, in, opcode);
923
924         /* Pass all cmds for vhca_tunnel as general, tracking is done in FW */
925         if ((MLX5_CAP_GEN_64(dev->mdev, vhca_tunnel_commands) &&
926              MLX5_GET(general_obj_in_cmd_hdr, in, vhca_tunnel_id)) ||
927             (opcode >= MLX5_CMD_OP_GENERAL_START &&
928              opcode < MLX5_CMD_OP_GENERAL_END))
929                 return true;
930
931         switch (opcode) {
932         case MLX5_CMD_OP_QUERY_HCA_CAP:
933         case MLX5_CMD_OP_QUERY_HCA_VPORT_CONTEXT:
934         case MLX5_CMD_OP_QUERY_ESW_VPORT_CONTEXT:
935         case MLX5_CMD_OP_QUERY_VPORT_STATE:
936         case MLX5_CMD_OP_QUERY_ADAPTER:
937         case MLX5_CMD_OP_QUERY_ISSI:
938         case MLX5_CMD_OP_QUERY_NIC_VPORT_CONTEXT:
939         case MLX5_CMD_OP_QUERY_ROCE_ADDRESS:
940         case MLX5_CMD_OP_QUERY_VNIC_ENV:
941         case MLX5_CMD_OP_QUERY_VPORT_COUNTER:
942         case MLX5_CMD_OP_GET_DROPPED_PACKET_LOG:
943         case MLX5_CMD_OP_NOP:
944         case MLX5_CMD_OP_QUERY_CONG_STATUS:
945         case MLX5_CMD_OP_QUERY_CONG_PARAMS:
946         case MLX5_CMD_OP_QUERY_CONG_STATISTICS:
947         case MLX5_CMD_OP_QUERY_LAG:
948                 return true;
949         default:
950                 return false;
951         }
952 }
953
954 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_QUERY_EQN)(
955         struct uverbs_attr_bundle *attrs)
956 {
957         struct mlx5_ib_ucontext *c;
958         struct mlx5_ib_dev *dev;
959         int user_vector;
960         int dev_eqn;
961         unsigned int irqn;
962         int err;
963
964         if (uverbs_copy_from(&user_vector, attrs,
965                              MLX5_IB_ATTR_DEVX_QUERY_EQN_USER_VEC))
966                 return -EFAULT;
967
968         c = devx_ufile2uctx(attrs);
969         if (IS_ERR(c))
970                 return PTR_ERR(c);
971         dev = to_mdev(c->ibucontext.device);
972
973         err = mlx5_vector2eqn(dev->mdev, user_vector, &dev_eqn, &irqn);
974         if (err < 0)
975                 return err;
976
977         if (uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_QUERY_EQN_DEV_EQN,
978                            &dev_eqn, sizeof(dev_eqn)))
979                 return -EFAULT;
980
981         return 0;
982 }
983
984 /*
985  *Security note:
986  * The hardware protection mechanism works like this: Each device object that
987  * is subject to UAR doorbells (QP/SQ/CQ) gets a UAR ID (called uar_page in
988  * the device specification manual) upon its creation. Then upon doorbell,
989  * hardware fetches the object context for which the doorbell was rang, and
990  * validates that the UAR through which the DB was rang matches the UAR ID
991  * of the object.
992  * If no match the doorbell is silently ignored by the hardware. Of course,
993  * the user cannot ring a doorbell on a UAR that was not mapped to it.
994  * Now in devx, as the devx kernel does not manipulate the QP/SQ/CQ command
995  * mailboxes (except tagging them with UID), we expose to the user its UAR
996  * ID, so it can embed it in these objects in the expected specification
997  * format. So the only thing the user can do is hurt itself by creating a
998  * QP/SQ/CQ with a UAR ID other than his, and then in this case other users
999  * may ring a doorbell on its objects.
1000  * The consequence of that will be that another user can schedule a QP/SQ
1001  * of the buggy user for execution (just insert it to the hardware schedule
1002  * queue or arm its CQ for event generation), no further harm is expected.
1003  */
1004 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_QUERY_UAR)(
1005         struct uverbs_attr_bundle *attrs)
1006 {
1007         struct mlx5_ib_ucontext *c;
1008         struct mlx5_ib_dev *dev;
1009         u32 user_idx;
1010         s32 dev_idx;
1011
1012         c = devx_ufile2uctx(attrs);
1013         if (IS_ERR(c))
1014                 return PTR_ERR(c);
1015         dev = to_mdev(c->ibucontext.device);
1016
1017         if (uverbs_copy_from(&user_idx, attrs,
1018                              MLX5_IB_ATTR_DEVX_QUERY_UAR_USER_IDX))
1019                 return -EFAULT;
1020
1021         dev_idx = bfregn_to_uar_index(dev, &c->bfregi, user_idx, true);
1022         if (dev_idx < 0)
1023                 return dev_idx;
1024
1025         if (uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_QUERY_UAR_DEV_IDX,
1026                            &dev_idx, sizeof(dev_idx)))
1027                 return -EFAULT;
1028
1029         return 0;
1030 }
1031
1032 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_OTHER)(
1033         struct uverbs_attr_bundle *attrs)
1034 {
1035         struct mlx5_ib_ucontext *c;
1036         struct mlx5_ib_dev *dev;
1037         void *cmd_in = uverbs_attr_get_alloced_ptr(
1038                 attrs, MLX5_IB_ATTR_DEVX_OTHER_CMD_IN);
1039         int cmd_out_len = uverbs_attr_get_len(attrs,
1040                                         MLX5_IB_ATTR_DEVX_OTHER_CMD_OUT);
1041         void *cmd_out;
1042         int err;
1043         int uid;
1044
1045         c = devx_ufile2uctx(attrs);
1046         if (IS_ERR(c))
1047                 return PTR_ERR(c);
1048         dev = to_mdev(c->ibucontext.device);
1049
1050         uid = devx_get_uid(c, cmd_in);
1051         if (uid < 0)
1052                 return uid;
1053
1054         /* Only white list of some general HCA commands are allowed for this method. */
1055         if (!devx_is_general_cmd(cmd_in, dev))
1056                 return -EINVAL;
1057
1058         cmd_out = uverbs_zalloc(attrs, cmd_out_len);
1059         if (IS_ERR(cmd_out))
1060                 return PTR_ERR(cmd_out);
1061
1062         MLX5_SET(general_obj_in_cmd_hdr, cmd_in, uid, uid);
1063         err = mlx5_cmd_exec(dev->mdev, cmd_in,
1064                             uverbs_attr_get_len(attrs, MLX5_IB_ATTR_DEVX_OTHER_CMD_IN),
1065                             cmd_out, cmd_out_len);
1066         if (err)
1067                 return err;
1068
1069         return uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_OTHER_CMD_OUT, cmd_out,
1070                               cmd_out_len);
1071 }
1072
1073 static void devx_obj_build_destroy_cmd(void *in, void *out, void *din,
1074                                        u32 *dinlen,
1075                                        u32 *obj_id)
1076 {
1077         u16 obj_type = MLX5_GET(general_obj_in_cmd_hdr, in, obj_type);
1078         u16 uid = MLX5_GET(general_obj_in_cmd_hdr, in, uid);
1079
1080         *obj_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
1081         *dinlen = MLX5_ST_SZ_BYTES(general_obj_in_cmd_hdr);
1082
1083         MLX5_SET(general_obj_in_cmd_hdr, din, obj_id, *obj_id);
1084         MLX5_SET(general_obj_in_cmd_hdr, din, uid, uid);
1085
1086         switch (MLX5_GET(general_obj_in_cmd_hdr, in, opcode)) {
1087         case MLX5_CMD_OP_CREATE_GENERAL_OBJECT:
1088                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_GENERAL_OBJECT);
1089                 MLX5_SET(general_obj_in_cmd_hdr, din, obj_type, obj_type);
1090                 break;
1091
1092         case MLX5_CMD_OP_CREATE_UMEM:
1093                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1094                          MLX5_CMD_OP_DESTROY_UMEM);
1095                 break;
1096         case MLX5_CMD_OP_CREATE_MKEY:
1097                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_MKEY);
1098                 break;
1099         case MLX5_CMD_OP_CREATE_CQ:
1100                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_CQ);
1101                 break;
1102         case MLX5_CMD_OP_ALLOC_PD:
1103                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DEALLOC_PD);
1104                 break;
1105         case MLX5_CMD_OP_ALLOC_TRANSPORT_DOMAIN:
1106                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1107                          MLX5_CMD_OP_DEALLOC_TRANSPORT_DOMAIN);
1108                 break;
1109         case MLX5_CMD_OP_CREATE_RMP:
1110                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_RMP);
1111                 break;
1112         case MLX5_CMD_OP_CREATE_SQ:
1113                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_SQ);
1114                 break;
1115         case MLX5_CMD_OP_CREATE_RQ:
1116                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_RQ);
1117                 break;
1118         case MLX5_CMD_OP_CREATE_RQT:
1119                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_RQT);
1120                 break;
1121         case MLX5_CMD_OP_CREATE_TIR:
1122                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_TIR);
1123                 break;
1124         case MLX5_CMD_OP_CREATE_TIS:
1125                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_TIS);
1126                 break;
1127         case MLX5_CMD_OP_ALLOC_Q_COUNTER:
1128                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1129                          MLX5_CMD_OP_DEALLOC_Q_COUNTER);
1130                 break;
1131         case MLX5_CMD_OP_CREATE_FLOW_TABLE:
1132                 *dinlen = MLX5_ST_SZ_BYTES(destroy_flow_table_in);
1133                 *obj_id = MLX5_GET(create_flow_table_out, out, table_id);
1134                 MLX5_SET(destroy_flow_table_in, din, other_vport,
1135                          MLX5_GET(create_flow_table_in,  in, other_vport));
1136                 MLX5_SET(destroy_flow_table_in, din, vport_number,
1137                          MLX5_GET(create_flow_table_in,  in, vport_number));
1138                 MLX5_SET(destroy_flow_table_in, din, table_type,
1139                          MLX5_GET(create_flow_table_in,  in, table_type));
1140                 MLX5_SET(destroy_flow_table_in, din, table_id, *obj_id);
1141                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1142                          MLX5_CMD_OP_DESTROY_FLOW_TABLE);
1143                 break;
1144         case MLX5_CMD_OP_CREATE_FLOW_GROUP:
1145                 *dinlen = MLX5_ST_SZ_BYTES(destroy_flow_group_in);
1146                 *obj_id = MLX5_GET(create_flow_group_out, out, group_id);
1147                 MLX5_SET(destroy_flow_group_in, din, other_vport,
1148                          MLX5_GET(create_flow_group_in, in, other_vport));
1149                 MLX5_SET(destroy_flow_group_in, din, vport_number,
1150                          MLX5_GET(create_flow_group_in, in, vport_number));
1151                 MLX5_SET(destroy_flow_group_in, din, table_type,
1152                          MLX5_GET(create_flow_group_in, in, table_type));
1153                 MLX5_SET(destroy_flow_group_in, din, table_id,
1154                          MLX5_GET(create_flow_group_in, in, table_id));
1155                 MLX5_SET(destroy_flow_group_in, din, group_id, *obj_id);
1156                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1157                          MLX5_CMD_OP_DESTROY_FLOW_GROUP);
1158                 break;
1159         case MLX5_CMD_OP_SET_FLOW_TABLE_ENTRY:
1160                 *dinlen = MLX5_ST_SZ_BYTES(delete_fte_in);
1161                 *obj_id = MLX5_GET(set_fte_in, in, flow_index);
1162                 MLX5_SET(delete_fte_in, din, other_vport,
1163                          MLX5_GET(set_fte_in,  in, other_vport));
1164                 MLX5_SET(delete_fte_in, din, vport_number,
1165                          MLX5_GET(set_fte_in, in, vport_number));
1166                 MLX5_SET(delete_fte_in, din, table_type,
1167                          MLX5_GET(set_fte_in, in, table_type));
1168                 MLX5_SET(delete_fte_in, din, table_id,
1169                          MLX5_GET(set_fte_in, in, table_id));
1170                 MLX5_SET(delete_fte_in, din, flow_index, *obj_id);
1171                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1172                          MLX5_CMD_OP_DELETE_FLOW_TABLE_ENTRY);
1173                 break;
1174         case MLX5_CMD_OP_ALLOC_FLOW_COUNTER:
1175                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1176                          MLX5_CMD_OP_DEALLOC_FLOW_COUNTER);
1177                 break;
1178         case MLX5_CMD_OP_ALLOC_PACKET_REFORMAT_CONTEXT:
1179                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1180                          MLX5_CMD_OP_DEALLOC_PACKET_REFORMAT_CONTEXT);
1181                 break;
1182         case MLX5_CMD_OP_ALLOC_MODIFY_HEADER_CONTEXT:
1183                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1184                          MLX5_CMD_OP_DEALLOC_MODIFY_HEADER_CONTEXT);
1185                 break;
1186         case MLX5_CMD_OP_CREATE_SCHEDULING_ELEMENT:
1187                 *dinlen = MLX5_ST_SZ_BYTES(destroy_scheduling_element_in);
1188                 *obj_id = MLX5_GET(create_scheduling_element_out, out,
1189                                    scheduling_element_id);
1190                 MLX5_SET(destroy_scheduling_element_in, din,
1191                          scheduling_hierarchy,
1192                          MLX5_GET(create_scheduling_element_in, in,
1193                                   scheduling_hierarchy));
1194                 MLX5_SET(destroy_scheduling_element_in, din,
1195                          scheduling_element_id, *obj_id);
1196                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1197                          MLX5_CMD_OP_DESTROY_SCHEDULING_ELEMENT);
1198                 break;
1199         case MLX5_CMD_OP_ADD_VXLAN_UDP_DPORT:
1200                 *dinlen = MLX5_ST_SZ_BYTES(delete_vxlan_udp_dport_in);
1201                 *obj_id = MLX5_GET(add_vxlan_udp_dport_in, in, vxlan_udp_port);
1202                 MLX5_SET(delete_vxlan_udp_dport_in, din, vxlan_udp_port, *obj_id);
1203                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1204                          MLX5_CMD_OP_DELETE_VXLAN_UDP_DPORT);
1205                 break;
1206         case MLX5_CMD_OP_SET_L2_TABLE_ENTRY:
1207                 *dinlen = MLX5_ST_SZ_BYTES(delete_l2_table_entry_in);
1208                 *obj_id = MLX5_GET(set_l2_table_entry_in, in, table_index);
1209                 MLX5_SET(delete_l2_table_entry_in, din, table_index, *obj_id);
1210                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1211                          MLX5_CMD_OP_DELETE_L2_TABLE_ENTRY);
1212                 break;
1213         case MLX5_CMD_OP_CREATE_QP:
1214                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_QP);
1215                 break;
1216         case MLX5_CMD_OP_CREATE_SRQ:
1217                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_SRQ);
1218                 break;
1219         case MLX5_CMD_OP_CREATE_XRC_SRQ:
1220                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1221                          MLX5_CMD_OP_DESTROY_XRC_SRQ);
1222                 break;
1223         case MLX5_CMD_OP_CREATE_DCT:
1224                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_DCT);
1225                 break;
1226         case MLX5_CMD_OP_CREATE_XRQ:
1227                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DESTROY_XRQ);
1228                 break;
1229         case MLX5_CMD_OP_ATTACH_TO_MCG:
1230                 *dinlen = MLX5_ST_SZ_BYTES(detach_from_mcg_in);
1231                 MLX5_SET(detach_from_mcg_in, din, qpn,
1232                          MLX5_GET(attach_to_mcg_in, in, qpn));
1233                 memcpy(MLX5_ADDR_OF(detach_from_mcg_in, din, multicast_gid),
1234                        MLX5_ADDR_OF(attach_to_mcg_in, in, multicast_gid),
1235                        MLX5_FLD_SZ_BYTES(attach_to_mcg_in, multicast_gid));
1236                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DETACH_FROM_MCG);
1237                 break;
1238         case MLX5_CMD_OP_ALLOC_XRCD:
1239                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode, MLX5_CMD_OP_DEALLOC_XRCD);
1240                 break;
1241         case MLX5_CMD_OP_CREATE_PSV:
1242                 MLX5_SET(general_obj_in_cmd_hdr, din, opcode,
1243                          MLX5_CMD_OP_DESTROY_PSV);
1244                 MLX5_SET(destroy_psv_in, din, psvn,
1245                          MLX5_GET(create_psv_out, out, psv0_index));
1246                 break;
1247         default:
1248                 /* The entry must match to one of the devx_is_obj_create_cmd */
1249                 WARN_ON(true);
1250                 break;
1251         }
1252 }
1253
1254 static int devx_handle_mkey_indirect(struct devx_obj *obj,
1255                                      struct mlx5_ib_dev *dev,
1256                                      void *in, void *out)
1257 {
1258         struct mlx5_ib_devx_mr *devx_mr = &obj->devx_mr;
1259         struct mlx5_core_mkey *mkey;
1260         void *mkc;
1261         u8 key;
1262
1263         mkey = &devx_mr->mmkey;
1264         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1265         key = MLX5_GET(mkc, mkc, mkey_7_0);
1266         mkey->key = mlx5_idx_to_mkey(
1267                         MLX5_GET(create_mkey_out, out, mkey_index)) | key;
1268         mkey->type = MLX5_MKEY_INDIRECT_DEVX;
1269         mkey->iova = MLX5_GET64(mkc, mkc, start_addr);
1270         mkey->size = MLX5_GET64(mkc, mkc, len);
1271         mkey->pd = MLX5_GET(mkc, mkc, pd);
1272         devx_mr->ndescs = MLX5_GET(mkc, mkc, translations_octword_size);
1273
1274         return xa_err(xa_store(&dev->odp_mkeys, mlx5_base_mkey(mkey->key), mkey,
1275                                GFP_KERNEL));
1276 }
1277
1278 static int devx_handle_mkey_create(struct mlx5_ib_dev *dev,
1279                                    struct devx_obj *obj,
1280                                    void *in, int in_len)
1281 {
1282         int min_len = MLX5_BYTE_OFF(create_mkey_in, memory_key_mkey_entry) +
1283                         MLX5_FLD_SZ_BYTES(create_mkey_in,
1284                         memory_key_mkey_entry);
1285         void *mkc;
1286         u8 access_mode;
1287
1288         if (in_len < min_len)
1289                 return -EINVAL;
1290
1291         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1292
1293         access_mode = MLX5_GET(mkc, mkc, access_mode_1_0);
1294         access_mode |= MLX5_GET(mkc, mkc, access_mode_4_2) << 2;
1295
1296         if (access_mode == MLX5_MKC_ACCESS_MODE_KLMS ||
1297                 access_mode == MLX5_MKC_ACCESS_MODE_KSM) {
1298                 if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1299                         obj->flags |= DEVX_OBJ_FLAGS_INDIRECT_MKEY;
1300                 return 0;
1301         }
1302
1303         MLX5_SET(create_mkey_in, in, mkey_umem_valid, 1);
1304         return 0;
1305 }
1306
1307 static void devx_cleanup_subscription(struct mlx5_ib_dev *dev,
1308                                       struct devx_event_subscription *sub)
1309 {
1310         struct devx_event *event;
1311         struct devx_obj_event *xa_val_level2;
1312
1313         if (sub->is_cleaned)
1314                 return;
1315
1316         sub->is_cleaned = 1;
1317         list_del_rcu(&sub->xa_list);
1318
1319         if (list_empty(&sub->obj_list))
1320                 return;
1321
1322         list_del_rcu(&sub->obj_list);
1323         /* check whether key level 1 for this obj_sub_list is empty */
1324         event = xa_load(&dev->devx_event_table.event_xa,
1325                         sub->xa_key_level1);
1326         WARN_ON(!event);
1327
1328         xa_val_level2 = xa_load(&event->object_ids, sub->xa_key_level2);
1329         if (list_empty(&xa_val_level2->obj_sub_list)) {
1330                 xa_erase(&event->object_ids,
1331                          sub->xa_key_level2);
1332                 kfree_rcu(xa_val_level2, rcu);
1333         }
1334 }
1335
1336 static int devx_obj_cleanup(struct ib_uobject *uobject,
1337                             enum rdma_remove_reason why,
1338                             struct uverbs_attr_bundle *attrs)
1339 {
1340         u32 out[MLX5_ST_SZ_DW(general_obj_out_cmd_hdr)];
1341         struct mlx5_devx_event_table *devx_event_table;
1342         struct devx_obj *obj = uobject->object;
1343         struct devx_event_subscription *sub_entry, *tmp;
1344         struct mlx5_ib_dev *dev;
1345         int ret;
1346
1347         dev = mlx5_udata_to_mdev(&attrs->driver_udata);
1348         if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY) {
1349                 /*
1350                  * The pagefault_single_data_segment() does commands against
1351                  * the mmkey, we must wait for that to stop before freeing the
1352                  * mkey, as another allocation could get the same mkey #.
1353                  */
1354                 xa_erase(&obj->ib_dev->odp_mkeys,
1355                          mlx5_base_mkey(obj->devx_mr.mmkey.key));
1356                 synchronize_srcu(&dev->odp_srcu);
1357         }
1358
1359         if (obj->flags & DEVX_OBJ_FLAGS_DCT)
1360                 ret = mlx5_core_destroy_dct(obj->ib_dev->mdev, &obj->core_dct);
1361         else if (obj->flags & DEVX_OBJ_FLAGS_CQ)
1362                 ret = mlx5_core_destroy_cq(obj->ib_dev->mdev, &obj->core_cq);
1363         else
1364                 ret = mlx5_cmd_exec(obj->ib_dev->mdev, obj->dinbox,
1365                                     obj->dinlen, out, sizeof(out));
1366         if (ib_is_destroy_retryable(ret, why, uobject))
1367                 return ret;
1368
1369         devx_event_table = &dev->devx_event_table;
1370
1371         mutex_lock(&devx_event_table->event_xa_lock);
1372         list_for_each_entry_safe(sub_entry, tmp, &obj->event_sub, obj_list)
1373                 devx_cleanup_subscription(dev, sub_entry);
1374         mutex_unlock(&devx_event_table->event_xa_lock);
1375
1376         kfree(obj);
1377         return ret;
1378 }
1379
1380 static void devx_cq_comp(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe)
1381 {
1382         struct devx_obj *obj = container_of(mcq, struct devx_obj, core_cq);
1383         struct mlx5_devx_event_table *table;
1384         struct devx_event *event;
1385         struct devx_obj_event *obj_event;
1386         u32 obj_id = mcq->cqn;
1387
1388         table = &obj->ib_dev->devx_event_table;
1389         rcu_read_lock();
1390         event = xa_load(&table->event_xa, MLX5_EVENT_TYPE_COMP);
1391         if (!event)
1392                 goto out;
1393
1394         obj_event = xa_load(&event->object_ids, obj_id);
1395         if (!obj_event)
1396                 goto out;
1397
1398         dispatch_event_fd(&obj_event->obj_sub_list, eqe);
1399 out:
1400         rcu_read_unlock();
1401 }
1402
1403 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_OBJ_CREATE)(
1404         struct uverbs_attr_bundle *attrs)
1405 {
1406         void *cmd_in = uverbs_attr_get_alloced_ptr(attrs, MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_IN);
1407         int cmd_out_len =  uverbs_attr_get_len(attrs,
1408                                         MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_OUT);
1409         int cmd_in_len = uverbs_attr_get_len(attrs,
1410                                         MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_IN);
1411         void *cmd_out;
1412         struct ib_uobject *uobj = uverbs_attr_get_uobject(
1413                 attrs, MLX5_IB_ATTR_DEVX_OBJ_CREATE_HANDLE);
1414         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
1415                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
1416         struct mlx5_ib_dev *dev = to_mdev(c->ibucontext.device);
1417         u32 out[MLX5_ST_SZ_DW(general_obj_out_cmd_hdr)];
1418         struct devx_obj *obj;
1419         u16 obj_type = 0;
1420         int err;
1421         int uid;
1422         u32 obj_id;
1423         u16 opcode;
1424
1425         if (MLX5_GET(general_obj_in_cmd_hdr, cmd_in, vhca_tunnel_id))
1426                 return -EINVAL;
1427
1428         uid = devx_get_uid(c, cmd_in);
1429         if (uid < 0)
1430                 return uid;
1431
1432         if (!devx_is_obj_create_cmd(cmd_in, &opcode))
1433                 return -EINVAL;
1434
1435         cmd_out = uverbs_zalloc(attrs, cmd_out_len);
1436         if (IS_ERR(cmd_out))
1437                 return PTR_ERR(cmd_out);
1438
1439         obj = kzalloc(sizeof(struct devx_obj), GFP_KERNEL);
1440         if (!obj)
1441                 return -ENOMEM;
1442
1443         MLX5_SET(general_obj_in_cmd_hdr, cmd_in, uid, uid);
1444         if (opcode == MLX5_CMD_OP_CREATE_MKEY) {
1445                 err = devx_handle_mkey_create(dev, obj, cmd_in, cmd_in_len);
1446                 if (err)
1447                         goto obj_free;
1448         } else {
1449                 devx_set_umem_valid(cmd_in);
1450         }
1451
1452         if (opcode == MLX5_CMD_OP_CREATE_DCT) {
1453                 obj->flags |= DEVX_OBJ_FLAGS_DCT;
1454                 err = mlx5_core_create_dct(dev->mdev, &obj->core_dct,
1455                                            cmd_in, cmd_in_len,
1456                                            cmd_out, cmd_out_len);
1457         } else if (opcode == MLX5_CMD_OP_CREATE_CQ) {
1458                 obj->flags |= DEVX_OBJ_FLAGS_CQ;
1459                 obj->core_cq.comp = devx_cq_comp;
1460                 err = mlx5_core_create_cq(dev->mdev, &obj->core_cq,
1461                                           cmd_in, cmd_in_len, cmd_out,
1462                                           cmd_out_len);
1463         } else {
1464                 err = mlx5_cmd_exec(dev->mdev, cmd_in,
1465                                     cmd_in_len,
1466                                     cmd_out, cmd_out_len);
1467         }
1468
1469         if (err)
1470                 goto obj_free;
1471
1472         if (opcode == MLX5_CMD_OP_ALLOC_FLOW_COUNTER) {
1473                 u8 bulk = MLX5_GET(alloc_flow_counter_in,
1474                                    cmd_in,
1475                                    flow_counter_bulk);
1476                 obj->flow_counter_bulk_size = 128UL * bulk;
1477         }
1478
1479         uobj->object = obj;
1480         INIT_LIST_HEAD(&obj->event_sub);
1481         obj->ib_dev = dev;
1482         devx_obj_build_destroy_cmd(cmd_in, cmd_out, obj->dinbox, &obj->dinlen,
1483                                    &obj_id);
1484         WARN_ON(obj->dinlen > MLX5_MAX_DESTROY_INBOX_SIZE_DW * sizeof(u32));
1485
1486         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_OUT, cmd_out, cmd_out_len);
1487         if (err)
1488                 goto obj_destroy;
1489
1490         if (opcode == MLX5_CMD_OP_CREATE_GENERAL_OBJECT)
1491                 obj_type = MLX5_GET(general_obj_in_cmd_hdr, cmd_in, obj_type);
1492         obj->obj_id = get_enc_obj_id(opcode | obj_type << 16, obj_id);
1493
1494         if (obj->flags & DEVX_OBJ_FLAGS_INDIRECT_MKEY) {
1495                 err = devx_handle_mkey_indirect(obj, dev, cmd_in, cmd_out);
1496                 if (err)
1497                         goto obj_destroy;
1498         }
1499         return 0;
1500
1501 obj_destroy:
1502         if (obj->flags & DEVX_OBJ_FLAGS_DCT)
1503                 mlx5_core_destroy_dct(obj->ib_dev->mdev, &obj->core_dct);
1504         else if (obj->flags & DEVX_OBJ_FLAGS_CQ)
1505                 mlx5_core_destroy_cq(obj->ib_dev->mdev, &obj->core_cq);
1506         else
1507                 mlx5_cmd_exec(obj->ib_dev->mdev, obj->dinbox, obj->dinlen, out,
1508                               sizeof(out));
1509 obj_free:
1510         kfree(obj);
1511         return err;
1512 }
1513
1514 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_OBJ_MODIFY)(
1515         struct uverbs_attr_bundle *attrs)
1516 {
1517         void *cmd_in = uverbs_attr_get_alloced_ptr(attrs, MLX5_IB_ATTR_DEVX_OBJ_MODIFY_CMD_IN);
1518         int cmd_out_len = uverbs_attr_get_len(attrs,
1519                                         MLX5_IB_ATTR_DEVX_OBJ_MODIFY_CMD_OUT);
1520         struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs,
1521                                                           MLX5_IB_ATTR_DEVX_OBJ_MODIFY_HANDLE);
1522         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
1523                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
1524         struct mlx5_ib_dev *mdev = to_mdev(c->ibucontext.device);
1525         void *cmd_out;
1526         int err;
1527         int uid;
1528
1529         if (MLX5_GET(general_obj_in_cmd_hdr, cmd_in, vhca_tunnel_id))
1530                 return -EINVAL;
1531
1532         uid = devx_get_uid(c, cmd_in);
1533         if (uid < 0)
1534                 return uid;
1535
1536         if (!devx_is_obj_modify_cmd(cmd_in))
1537                 return -EINVAL;
1538
1539         if (!devx_is_valid_obj_id(attrs, uobj, cmd_in))
1540                 return -EINVAL;
1541
1542         cmd_out = uverbs_zalloc(attrs, cmd_out_len);
1543         if (IS_ERR(cmd_out))
1544                 return PTR_ERR(cmd_out);
1545
1546         MLX5_SET(general_obj_in_cmd_hdr, cmd_in, uid, uid);
1547         devx_set_umem_valid(cmd_in);
1548
1549         err = mlx5_cmd_exec(mdev->mdev, cmd_in,
1550                             uverbs_attr_get_len(attrs, MLX5_IB_ATTR_DEVX_OBJ_MODIFY_CMD_IN),
1551                             cmd_out, cmd_out_len);
1552         if (err)
1553                 return err;
1554
1555         return uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_OBJ_MODIFY_CMD_OUT,
1556                               cmd_out, cmd_out_len);
1557 }
1558
1559 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_OBJ_QUERY)(
1560         struct uverbs_attr_bundle *attrs)
1561 {
1562         void *cmd_in = uverbs_attr_get_alloced_ptr(attrs, MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_IN);
1563         int cmd_out_len = uverbs_attr_get_len(attrs,
1564                                               MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_OUT);
1565         struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs,
1566                                                           MLX5_IB_ATTR_DEVX_OBJ_QUERY_HANDLE);
1567         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
1568                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
1569         void *cmd_out;
1570         int err;
1571         int uid;
1572         struct mlx5_ib_dev *mdev = to_mdev(c->ibucontext.device);
1573
1574         if (MLX5_GET(general_obj_in_cmd_hdr, cmd_in, vhca_tunnel_id))
1575                 return -EINVAL;
1576
1577         uid = devx_get_uid(c, cmd_in);
1578         if (uid < 0)
1579                 return uid;
1580
1581         if (!devx_is_obj_query_cmd(cmd_in))
1582                 return -EINVAL;
1583
1584         if (!devx_is_valid_obj_id(attrs, uobj, cmd_in))
1585                 return -EINVAL;
1586
1587         cmd_out = uverbs_zalloc(attrs, cmd_out_len);
1588         if (IS_ERR(cmd_out))
1589                 return PTR_ERR(cmd_out);
1590
1591         MLX5_SET(general_obj_in_cmd_hdr, cmd_in, uid, uid);
1592         err = mlx5_cmd_exec(mdev->mdev, cmd_in,
1593                             uverbs_attr_get_len(attrs, MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_IN),
1594                             cmd_out, cmd_out_len);
1595         if (err)
1596                 return err;
1597
1598         return uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_OUT,
1599                               cmd_out, cmd_out_len);
1600 }
1601
1602 struct devx_async_event_queue {
1603         spinlock_t              lock;
1604         wait_queue_head_t       poll_wait;
1605         struct list_head        event_list;
1606         atomic_t                bytes_in_use;
1607         u8                      is_destroyed:1;
1608 };
1609
1610 struct devx_async_cmd_event_file {
1611         struct ib_uobject               uobj;
1612         struct devx_async_event_queue   ev_queue;
1613         struct mlx5_async_ctx           async_ctx;
1614 };
1615
1616 static void devx_init_event_queue(struct devx_async_event_queue *ev_queue)
1617 {
1618         spin_lock_init(&ev_queue->lock);
1619         INIT_LIST_HEAD(&ev_queue->event_list);
1620         init_waitqueue_head(&ev_queue->poll_wait);
1621         atomic_set(&ev_queue->bytes_in_use, 0);
1622         ev_queue->is_destroyed = 0;
1623 }
1624
1625 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_ASYNC_CMD_FD_ALLOC)(
1626         struct uverbs_attr_bundle *attrs)
1627 {
1628         struct devx_async_cmd_event_file *ev_file;
1629
1630         struct ib_uobject *uobj = uverbs_attr_get_uobject(
1631                 attrs, MLX5_IB_ATTR_DEVX_ASYNC_CMD_FD_ALLOC_HANDLE);
1632         struct mlx5_ib_dev *mdev = mlx5_udata_to_mdev(&attrs->driver_udata);
1633
1634         ev_file = container_of(uobj, struct devx_async_cmd_event_file,
1635                                uobj);
1636         devx_init_event_queue(&ev_file->ev_queue);
1637         mlx5_cmd_init_async_ctx(mdev->mdev, &ev_file->async_ctx);
1638         return 0;
1639 }
1640
1641 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_ASYNC_EVENT_FD_ALLOC)(
1642         struct uverbs_attr_bundle *attrs)
1643 {
1644         struct ib_uobject *uobj = uverbs_attr_get_uobject(
1645                 attrs, MLX5_IB_ATTR_DEVX_ASYNC_EVENT_FD_ALLOC_HANDLE);
1646         struct devx_async_event_file *ev_file;
1647         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
1648                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
1649         struct mlx5_ib_dev *dev = to_mdev(c->ibucontext.device);
1650         u32 flags;
1651         int err;
1652
1653         err = uverbs_get_flags32(&flags, attrs,
1654                 MLX5_IB_ATTR_DEVX_ASYNC_EVENT_FD_ALLOC_FLAGS,
1655                 MLX5_IB_UAPI_DEVX_CR_EV_CH_FLAGS_OMIT_DATA);
1656
1657         if (err)
1658                 return err;
1659
1660         ev_file = container_of(uobj, struct devx_async_event_file,
1661                                uobj);
1662         spin_lock_init(&ev_file->lock);
1663         INIT_LIST_HEAD(&ev_file->event_list);
1664         init_waitqueue_head(&ev_file->poll_wait);
1665         if (flags & MLX5_IB_UAPI_DEVX_CR_EV_CH_FLAGS_OMIT_DATA)
1666                 ev_file->omit_data = 1;
1667         INIT_LIST_HEAD(&ev_file->subscribed_events_list);
1668         ev_file->dev = dev;
1669         get_device(&dev->ib_dev.dev);
1670         return 0;
1671 }
1672
1673 static void devx_query_callback(int status, struct mlx5_async_work *context)
1674 {
1675         struct devx_async_data *async_data =
1676                 container_of(context, struct devx_async_data, cb_work);
1677         struct ib_uobject *fd_uobj = async_data->fd_uobj;
1678         struct devx_async_cmd_event_file *ev_file;
1679         struct devx_async_event_queue *ev_queue;
1680         unsigned long flags;
1681
1682         ev_file = container_of(fd_uobj, struct devx_async_cmd_event_file,
1683                                uobj);
1684         ev_queue = &ev_file->ev_queue;
1685
1686         spin_lock_irqsave(&ev_queue->lock, flags);
1687         list_add_tail(&async_data->list, &ev_queue->event_list);
1688         spin_unlock_irqrestore(&ev_queue->lock, flags);
1689
1690         wake_up_interruptible(&ev_queue->poll_wait);
1691         fput(fd_uobj->object);
1692 }
1693
1694 #define MAX_ASYNC_BYTES_IN_USE (1024 * 1024) /* 1MB */
1695
1696 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_OBJ_ASYNC_QUERY)(
1697         struct uverbs_attr_bundle *attrs)
1698 {
1699         void *cmd_in = uverbs_attr_get_alloced_ptr(attrs,
1700                                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_CMD_IN);
1701         struct ib_uobject *uobj = uverbs_attr_get_uobject(
1702                                 attrs,
1703                                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_HANDLE);
1704         u16 cmd_out_len;
1705         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
1706                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
1707         struct ib_uobject *fd_uobj;
1708         int err;
1709         int uid;
1710         struct mlx5_ib_dev *mdev = to_mdev(c->ibucontext.device);
1711         struct devx_async_cmd_event_file *ev_file;
1712         struct devx_async_data *async_data;
1713
1714         if (MLX5_GET(general_obj_in_cmd_hdr, cmd_in, vhca_tunnel_id))
1715                 return -EINVAL;
1716
1717         uid = devx_get_uid(c, cmd_in);
1718         if (uid < 0)
1719                 return uid;
1720
1721         if (!devx_is_obj_query_cmd(cmd_in))
1722                 return -EINVAL;
1723
1724         err = uverbs_get_const(&cmd_out_len, attrs,
1725                                MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_OUT_LEN);
1726         if (err)
1727                 return err;
1728
1729         if (!devx_is_valid_obj_id(attrs, uobj, cmd_in))
1730                 return -EINVAL;
1731
1732         fd_uobj = uverbs_attr_get_uobject(attrs,
1733                                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_FD);
1734         if (IS_ERR(fd_uobj))
1735                 return PTR_ERR(fd_uobj);
1736
1737         ev_file = container_of(fd_uobj, struct devx_async_cmd_event_file,
1738                                uobj);
1739
1740         if (atomic_add_return(cmd_out_len, &ev_file->ev_queue.bytes_in_use) >
1741                         MAX_ASYNC_BYTES_IN_USE) {
1742                 atomic_sub(cmd_out_len, &ev_file->ev_queue.bytes_in_use);
1743                 return -EAGAIN;
1744         }
1745
1746         async_data = kvzalloc(struct_size(async_data, hdr.out_data,
1747                                           cmd_out_len), GFP_KERNEL);
1748         if (!async_data) {
1749                 err = -ENOMEM;
1750                 goto sub_bytes;
1751         }
1752
1753         err = uverbs_copy_from(&async_data->hdr.wr_id, attrs,
1754                                MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_WR_ID);
1755         if (err)
1756                 goto free_async;
1757
1758         async_data->cmd_out_len = cmd_out_len;
1759         async_data->mdev = mdev;
1760         async_data->fd_uobj = fd_uobj;
1761
1762         get_file(fd_uobj->object);
1763         MLX5_SET(general_obj_in_cmd_hdr, cmd_in, uid, uid);
1764         err = mlx5_cmd_exec_cb(&ev_file->async_ctx, cmd_in,
1765                     uverbs_attr_get_len(attrs,
1766                                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_CMD_IN),
1767                     async_data->hdr.out_data,
1768                     async_data->cmd_out_len,
1769                     devx_query_callback, &async_data->cb_work);
1770
1771         if (err)
1772                 goto cb_err;
1773
1774         return 0;
1775
1776 cb_err:
1777         fput(fd_uobj->object);
1778 free_async:
1779         kvfree(async_data);
1780 sub_bytes:
1781         atomic_sub(cmd_out_len, &ev_file->ev_queue.bytes_in_use);
1782         return err;
1783 }
1784
1785 static void
1786 subscribe_event_xa_dealloc(struct mlx5_devx_event_table *devx_event_table,
1787                            u32 key_level1,
1788                            bool is_level2,
1789                            u32 key_level2)
1790 {
1791         struct devx_event *event;
1792         struct devx_obj_event *xa_val_level2;
1793
1794         /* Level 1 is valid for future use, no need to free */
1795         if (!is_level2)
1796                 return;
1797
1798         event = xa_load(&devx_event_table->event_xa, key_level1);
1799         WARN_ON(!event);
1800
1801         xa_val_level2 = xa_load(&event->object_ids,
1802                                 key_level2);
1803         if (list_empty(&xa_val_level2->obj_sub_list)) {
1804                 xa_erase(&event->object_ids,
1805                          key_level2);
1806                 kfree_rcu(xa_val_level2, rcu);
1807         }
1808 }
1809
1810 static int
1811 subscribe_event_xa_alloc(struct mlx5_devx_event_table *devx_event_table,
1812                          u32 key_level1,
1813                          bool is_level2,
1814                          u32 key_level2)
1815 {
1816         struct devx_obj_event *obj_event;
1817         struct devx_event *event;
1818         int err;
1819
1820         event = xa_load(&devx_event_table->event_xa, key_level1);
1821         if (!event) {
1822                 event = kzalloc(sizeof(*event), GFP_KERNEL);
1823                 if (!event)
1824                         return -ENOMEM;
1825
1826                 INIT_LIST_HEAD(&event->unaffiliated_list);
1827                 xa_init(&event->object_ids);
1828
1829                 err = xa_insert(&devx_event_table->event_xa,
1830                                 key_level1,
1831                                 event,
1832                                 GFP_KERNEL);
1833                 if (err) {
1834                         kfree(event);
1835                         return err;
1836                 }
1837         }
1838
1839         if (!is_level2)
1840                 return 0;
1841
1842         obj_event = xa_load(&event->object_ids, key_level2);
1843         if (!obj_event) {
1844                 obj_event = kzalloc(sizeof(*obj_event), GFP_KERNEL);
1845                 if (!obj_event)
1846                         /* Level1 is valid for future use, no need to free */
1847                         return -ENOMEM;
1848
1849                 err = xa_insert(&event->object_ids,
1850                                 key_level2,
1851                                 obj_event,
1852                                 GFP_KERNEL);
1853                 if (err)
1854                         return err;
1855                 INIT_LIST_HEAD(&obj_event->obj_sub_list);
1856         }
1857
1858         return 0;
1859 }
1860
1861 static bool is_valid_events_legacy(int num_events, u16 *event_type_num_list,
1862                                    struct devx_obj *obj)
1863 {
1864         int i;
1865
1866         for (i = 0; i < num_events; i++) {
1867                 if (obj) {
1868                         if (!is_legacy_obj_event_num(event_type_num_list[i]))
1869                                 return false;
1870                 } else if (!is_legacy_unaffiliated_event_num(
1871                                 event_type_num_list[i])) {
1872                         return false;
1873                 }
1874         }
1875
1876         return true;
1877 }
1878
1879 #define MAX_SUPP_EVENT_NUM 255
1880 static bool is_valid_events(struct mlx5_core_dev *dev,
1881                             int num_events, u16 *event_type_num_list,
1882                             struct devx_obj *obj)
1883 {
1884         __be64 *aff_events;
1885         __be64 *unaff_events;
1886         int mask_entry;
1887         int mask_bit;
1888         int i;
1889
1890         if (MLX5_CAP_GEN(dev, event_cap)) {
1891                 aff_events = MLX5_CAP_DEV_EVENT(dev,
1892                                                 user_affiliated_events);
1893                 unaff_events = MLX5_CAP_DEV_EVENT(dev,
1894                                                   user_unaffiliated_events);
1895         } else {
1896                 return is_valid_events_legacy(num_events, event_type_num_list,
1897                                               obj);
1898         }
1899
1900         for (i = 0; i < num_events; i++) {
1901                 if (event_type_num_list[i] > MAX_SUPP_EVENT_NUM)
1902                         return false;
1903
1904                 mask_entry = event_type_num_list[i] / 64;
1905                 mask_bit = event_type_num_list[i] % 64;
1906
1907                 if (obj) {
1908                         /* CQ completion */
1909                         if (event_type_num_list[i] == 0)
1910                                 continue;
1911
1912                         if (!(be64_to_cpu(aff_events[mask_entry]) &
1913                                         (1ull << mask_bit)))
1914                                 return false;
1915
1916                         continue;
1917                 }
1918
1919                 if (!(be64_to_cpu(unaff_events[mask_entry]) &
1920                                 (1ull << mask_bit)))
1921                         return false;
1922         }
1923
1924         return true;
1925 }
1926
1927 #define MAX_NUM_EVENTS 16
1928 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_SUBSCRIBE_EVENT)(
1929         struct uverbs_attr_bundle *attrs)
1930 {
1931         struct ib_uobject *devx_uobj = uverbs_attr_get_uobject(
1932                                 attrs,
1933                                 MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_OBJ_HANDLE);
1934         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
1935                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
1936         struct mlx5_ib_dev *dev = to_mdev(c->ibucontext.device);
1937         struct ib_uobject *fd_uobj;
1938         struct devx_obj *obj = NULL;
1939         struct devx_async_event_file *ev_file;
1940         struct mlx5_devx_event_table *devx_event_table = &dev->devx_event_table;
1941         u16 *event_type_num_list;
1942         struct devx_event_subscription *event_sub, *tmp_sub;
1943         struct list_head sub_list;
1944         int redirect_fd;
1945         bool use_eventfd = false;
1946         int num_events;
1947         int num_alloc_xa_entries = 0;
1948         u16 obj_type = 0;
1949         u64 cookie = 0;
1950         u32 obj_id = 0;
1951         int err;
1952         int i;
1953
1954         if (!c->devx_uid)
1955                 return -EINVAL;
1956
1957         if (!IS_ERR(devx_uobj)) {
1958                 obj = (struct devx_obj *)devx_uobj->object;
1959                 if (obj)
1960                         obj_id = get_dec_obj_id(obj->obj_id);
1961         }
1962
1963         fd_uobj = uverbs_attr_get_uobject(attrs,
1964                                 MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_FD_HANDLE);
1965         if (IS_ERR(fd_uobj))
1966                 return PTR_ERR(fd_uobj);
1967
1968         ev_file = container_of(fd_uobj, struct devx_async_event_file,
1969                                uobj);
1970
1971         if (uverbs_attr_is_valid(attrs,
1972                                  MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_FD_NUM)) {
1973                 err = uverbs_copy_from(&redirect_fd, attrs,
1974                                MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_FD_NUM);
1975                 if (err)
1976                         return err;
1977
1978                 use_eventfd = true;
1979         }
1980
1981         if (uverbs_attr_is_valid(attrs,
1982                                  MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_COOKIE)) {
1983                 if (use_eventfd)
1984                         return -EINVAL;
1985
1986                 err = uverbs_copy_from(&cookie, attrs,
1987                                 MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_COOKIE);
1988                 if (err)
1989                         return err;
1990         }
1991
1992         num_events = uverbs_attr_ptr_get_array_size(
1993                 attrs, MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_TYPE_NUM_LIST,
1994                 sizeof(u16));
1995
1996         if (num_events < 0)
1997                 return num_events;
1998
1999         if (num_events > MAX_NUM_EVENTS)
2000                 return -EINVAL;
2001
2002         event_type_num_list = uverbs_attr_get_alloced_ptr(attrs,
2003                         MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_TYPE_NUM_LIST);
2004
2005         if (!is_valid_events(dev->mdev, num_events, event_type_num_list, obj))
2006                 return -EINVAL;
2007
2008         INIT_LIST_HEAD(&sub_list);
2009
2010         /* Protect from concurrent subscriptions to same XA entries to allow
2011          * both to succeed
2012          */
2013         mutex_lock(&devx_event_table->event_xa_lock);
2014         for (i = 0; i < num_events; i++) {
2015                 u32 key_level1;
2016
2017                 if (obj)
2018                         obj_type = get_dec_obj_type(obj,
2019                                                     event_type_num_list[i]);
2020                 key_level1 = event_type_num_list[i] | obj_type << 16;
2021
2022                 err = subscribe_event_xa_alloc(devx_event_table,
2023                                                key_level1,
2024                                                obj,
2025                                                obj_id);
2026                 if (err)
2027                         goto err;
2028
2029                 num_alloc_xa_entries++;
2030                 event_sub = kzalloc(sizeof(*event_sub), GFP_KERNEL);
2031                 if (!event_sub)
2032                         goto err;
2033
2034                 list_add_tail(&event_sub->event_list, &sub_list);
2035                 if (use_eventfd) {
2036                         event_sub->eventfd =
2037                                 eventfd_ctx_fdget(redirect_fd);
2038
2039                         if (IS_ERR(event_sub->eventfd)) {
2040                                 err = PTR_ERR(event_sub->eventfd);
2041                                 event_sub->eventfd = NULL;
2042                                 goto err;
2043                         }
2044                 }
2045
2046                 event_sub->cookie = cookie;
2047                 event_sub->ev_file = ev_file;
2048                 event_sub->filp = fd_uobj->object;
2049                 /* May be needed upon cleanup the devx object/subscription */
2050                 event_sub->xa_key_level1 = key_level1;
2051                 event_sub->xa_key_level2 = obj_id;
2052                 INIT_LIST_HEAD(&event_sub->obj_list);
2053         }
2054
2055         /* Once all the allocations and the XA data insertions were done we
2056          * can go ahead and add all the subscriptions to the relevant lists
2057          * without concern of a failure.
2058          */
2059         list_for_each_entry_safe(event_sub, tmp_sub, &sub_list, event_list) {
2060                 struct devx_event *event;
2061                 struct devx_obj_event *obj_event;
2062
2063                 list_del_init(&event_sub->event_list);
2064
2065                 spin_lock_irq(&ev_file->lock);
2066                 list_add_tail_rcu(&event_sub->file_list,
2067                                   &ev_file->subscribed_events_list);
2068                 spin_unlock_irq(&ev_file->lock);
2069
2070                 event = xa_load(&devx_event_table->event_xa,
2071                                 event_sub->xa_key_level1);
2072                 WARN_ON(!event);
2073
2074                 if (!obj) {
2075                         list_add_tail_rcu(&event_sub->xa_list,
2076                                           &event->unaffiliated_list);
2077                         continue;
2078                 }
2079
2080                 obj_event = xa_load(&event->object_ids, obj_id);
2081                 WARN_ON(!obj_event);
2082                 list_add_tail_rcu(&event_sub->xa_list,
2083                                   &obj_event->obj_sub_list);
2084                 list_add_tail_rcu(&event_sub->obj_list,
2085                                   &obj->event_sub);
2086         }
2087
2088         mutex_unlock(&devx_event_table->event_xa_lock);
2089         return 0;
2090
2091 err:
2092         list_for_each_entry_safe(event_sub, tmp_sub, &sub_list, event_list) {
2093                 list_del(&event_sub->event_list);
2094
2095                 subscribe_event_xa_dealloc(devx_event_table,
2096                                            event_sub->xa_key_level1,
2097                                            obj,
2098                                            obj_id);
2099
2100                 if (event_sub->eventfd)
2101                         eventfd_ctx_put(event_sub->eventfd);
2102
2103                 kfree(event_sub);
2104         }
2105
2106         mutex_unlock(&devx_event_table->event_xa_lock);
2107         return err;
2108 }
2109
2110 static int devx_umem_get(struct mlx5_ib_dev *dev, struct ib_ucontext *ucontext,
2111                          struct uverbs_attr_bundle *attrs,
2112                          struct devx_umem *obj)
2113 {
2114         u64 addr;
2115         size_t size;
2116         u32 access;
2117         int npages;
2118         int err;
2119         u32 page_mask;
2120
2121         if (uverbs_copy_from(&addr, attrs, MLX5_IB_ATTR_DEVX_UMEM_REG_ADDR) ||
2122             uverbs_copy_from(&size, attrs, MLX5_IB_ATTR_DEVX_UMEM_REG_LEN))
2123                 return -EFAULT;
2124
2125         err = uverbs_get_flags32(&access, attrs,
2126                                  MLX5_IB_ATTR_DEVX_UMEM_REG_ACCESS,
2127                                  IB_ACCESS_LOCAL_WRITE |
2128                                  IB_ACCESS_REMOTE_WRITE |
2129                                  IB_ACCESS_REMOTE_READ);
2130         if (err)
2131                 return err;
2132
2133         err = ib_check_mr_access(access);
2134         if (err)
2135                 return err;
2136
2137         obj->umem = ib_umem_get(&attrs->driver_udata, addr, size, access);
2138         if (IS_ERR(obj->umem))
2139                 return PTR_ERR(obj->umem);
2140
2141         mlx5_ib_cont_pages(obj->umem, obj->umem->address,
2142                            MLX5_MKEY_PAGE_SHIFT_MASK, &npages,
2143                            &obj->page_shift, &obj->ncont, NULL);
2144
2145         if (!npages) {
2146                 ib_umem_release(obj->umem);
2147                 return -EINVAL;
2148         }
2149
2150         page_mask = (1 << obj->page_shift) - 1;
2151         obj->page_offset = obj->umem->address & page_mask;
2152
2153         return 0;
2154 }
2155
2156 static int devx_umem_reg_cmd_alloc(struct uverbs_attr_bundle *attrs,
2157                                    struct devx_umem *obj,
2158                                    struct devx_umem_reg_cmd *cmd)
2159 {
2160         cmd->inlen = MLX5_ST_SZ_BYTES(create_umem_in) +
2161                     (MLX5_ST_SZ_BYTES(mtt) * obj->ncont);
2162         cmd->in = uverbs_zalloc(attrs, cmd->inlen);
2163         return PTR_ERR_OR_ZERO(cmd->in);
2164 }
2165
2166 static void devx_umem_reg_cmd_build(struct mlx5_ib_dev *dev,
2167                                     struct devx_umem *obj,
2168                                     struct devx_umem_reg_cmd *cmd)
2169 {
2170         void *umem;
2171         __be64 *mtt;
2172
2173         umem = MLX5_ADDR_OF(create_umem_in, cmd->in, umem);
2174         mtt = (__be64 *)MLX5_ADDR_OF(umem, umem, mtt);
2175
2176         MLX5_SET(create_umem_in, cmd->in, opcode, MLX5_CMD_OP_CREATE_UMEM);
2177         MLX5_SET64(umem, umem, num_of_mtt, obj->ncont);
2178         MLX5_SET(umem, umem, log_page_size, obj->page_shift -
2179                                             MLX5_ADAPTER_PAGE_SHIFT);
2180         MLX5_SET(umem, umem, page_offset, obj->page_offset);
2181         mlx5_ib_populate_pas(dev, obj->umem, obj->page_shift, mtt,
2182                              (obj->umem->writable ? MLX5_IB_MTT_WRITE : 0) |
2183                              MLX5_IB_MTT_READ);
2184 }
2185
2186 static int UVERBS_HANDLER(MLX5_IB_METHOD_DEVX_UMEM_REG)(
2187         struct uverbs_attr_bundle *attrs)
2188 {
2189         struct devx_umem_reg_cmd cmd;
2190         struct devx_umem *obj;
2191         struct ib_uobject *uobj = uverbs_attr_get_uobject(
2192                 attrs, MLX5_IB_ATTR_DEVX_UMEM_REG_HANDLE);
2193         u32 obj_id;
2194         struct mlx5_ib_ucontext *c = rdma_udata_to_drv_context(
2195                 &attrs->driver_udata, struct mlx5_ib_ucontext, ibucontext);
2196         struct mlx5_ib_dev *dev = to_mdev(c->ibucontext.device);
2197         int err;
2198
2199         if (!c->devx_uid)
2200                 return -EINVAL;
2201
2202         obj = kzalloc(sizeof(struct devx_umem), GFP_KERNEL);
2203         if (!obj)
2204                 return -ENOMEM;
2205
2206         err = devx_umem_get(dev, &c->ibucontext, attrs, obj);
2207         if (err)
2208                 goto err_obj_free;
2209
2210         err = devx_umem_reg_cmd_alloc(attrs, obj, &cmd);
2211         if (err)
2212                 goto err_umem_release;
2213
2214         devx_umem_reg_cmd_build(dev, obj, &cmd);
2215
2216         MLX5_SET(create_umem_in, cmd.in, uid, c->devx_uid);
2217         err = mlx5_cmd_exec(dev->mdev, cmd.in, cmd.inlen, cmd.out,
2218                             sizeof(cmd.out));
2219         if (err)
2220                 goto err_umem_release;
2221
2222         obj->mdev = dev->mdev;
2223         uobj->object = obj;
2224         devx_obj_build_destroy_cmd(cmd.in, cmd.out, obj->dinbox, &obj->dinlen, &obj_id);
2225         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_DEVX_UMEM_REG_OUT_ID, &obj_id, sizeof(obj_id));
2226         if (err)
2227                 goto err_umem_destroy;
2228
2229         return 0;
2230
2231 err_umem_destroy:
2232         mlx5_cmd_exec(obj->mdev, obj->dinbox, obj->dinlen, cmd.out, sizeof(cmd.out));
2233 err_umem_release:
2234         ib_umem_release(obj->umem);
2235 err_obj_free:
2236         kfree(obj);
2237         return err;
2238 }
2239
2240 static int devx_umem_cleanup(struct ib_uobject *uobject,
2241                              enum rdma_remove_reason why,
2242                              struct uverbs_attr_bundle *attrs)
2243 {
2244         struct devx_umem *obj = uobject->object;
2245         u32 out[MLX5_ST_SZ_DW(general_obj_out_cmd_hdr)];
2246         int err;
2247
2248         err = mlx5_cmd_exec(obj->mdev, obj->dinbox, obj->dinlen, out, sizeof(out));
2249         if (ib_is_destroy_retryable(err, why, uobject))
2250                 return err;
2251
2252         ib_umem_release(obj->umem);
2253         kfree(obj);
2254         return 0;
2255 }
2256
2257 static bool is_unaffiliated_event(struct mlx5_core_dev *dev,
2258                                   unsigned long event_type)
2259 {
2260         __be64 *unaff_events;
2261         int mask_entry;
2262         int mask_bit;
2263
2264         if (!MLX5_CAP_GEN(dev, event_cap))
2265                 return is_legacy_unaffiliated_event_num(event_type);
2266
2267         unaff_events = MLX5_CAP_DEV_EVENT(dev,
2268                                           user_unaffiliated_events);
2269         WARN_ON(event_type > MAX_SUPP_EVENT_NUM);
2270
2271         mask_entry = event_type / 64;
2272         mask_bit = event_type % 64;
2273
2274         if (!(be64_to_cpu(unaff_events[mask_entry]) & (1ull << mask_bit)))
2275                 return false;
2276
2277         return true;
2278 }
2279
2280 static u32 devx_get_obj_id_from_event(unsigned long event_type, void *data)
2281 {
2282         struct mlx5_eqe *eqe = data;
2283         u32 obj_id = 0;
2284
2285         switch (event_type) {
2286         case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
2287         case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
2288         case MLX5_EVENT_TYPE_PATH_MIG:
2289         case MLX5_EVENT_TYPE_COMM_EST:
2290         case MLX5_EVENT_TYPE_SQ_DRAINED:
2291         case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
2292         case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
2293         case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
2294         case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
2295         case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
2296                 obj_id = be32_to_cpu(eqe->data.qp_srq.qp_srq_n) & 0xffffff;
2297                 break;
2298         case MLX5_EVENT_TYPE_XRQ_ERROR:
2299                 obj_id = be32_to_cpu(eqe->data.xrq_err.type_xrqn) & 0xffffff;
2300                 break;
2301         case MLX5_EVENT_TYPE_DCT_DRAINED:
2302         case MLX5_EVENT_TYPE_DCT_KEY_VIOLATION:
2303                 obj_id = be32_to_cpu(eqe->data.dct.dctn) & 0xffffff;
2304                 break;
2305         case MLX5_EVENT_TYPE_CQ_ERROR:
2306                 obj_id = be32_to_cpu(eqe->data.cq_err.cqn) & 0xffffff;
2307                 break;
2308         default:
2309                 obj_id = MLX5_GET(affiliated_event_header, &eqe->data, obj_id);
2310                 break;
2311         }
2312
2313         return obj_id;
2314 }
2315
2316 static int deliver_event(struct devx_event_subscription *event_sub,
2317                          const void *data)
2318 {
2319         struct devx_async_event_file *ev_file;
2320         struct devx_async_event_data *event_data;
2321         unsigned long flags;
2322
2323         ev_file = event_sub->ev_file;
2324
2325         if (ev_file->omit_data) {
2326                 spin_lock_irqsave(&ev_file->lock, flags);
2327                 if (!list_empty(&event_sub->event_list)) {
2328                         spin_unlock_irqrestore(&ev_file->lock, flags);
2329                         return 0;
2330                 }
2331
2332                 list_add_tail(&event_sub->event_list, &ev_file->event_list);
2333                 spin_unlock_irqrestore(&ev_file->lock, flags);
2334                 wake_up_interruptible(&ev_file->poll_wait);
2335                 return 0;
2336         }
2337
2338         event_data = kzalloc(sizeof(*event_data) + sizeof(struct mlx5_eqe),
2339                              GFP_ATOMIC);
2340         if (!event_data) {
2341                 spin_lock_irqsave(&ev_file->lock, flags);
2342                 ev_file->is_overflow_err = 1;
2343                 spin_unlock_irqrestore(&ev_file->lock, flags);
2344                 return -ENOMEM;
2345         }
2346
2347         event_data->hdr.cookie = event_sub->cookie;
2348         memcpy(event_data->hdr.out_data, data, sizeof(struct mlx5_eqe));
2349
2350         spin_lock_irqsave(&ev_file->lock, flags);
2351         list_add_tail(&event_data->list, &ev_file->event_list);
2352         spin_unlock_irqrestore(&ev_file->lock, flags);
2353         wake_up_interruptible(&ev_file->poll_wait);
2354
2355         return 0;
2356 }
2357
2358 static void dispatch_event_fd(struct list_head *fd_list,
2359                               const void *data)
2360 {
2361         struct devx_event_subscription *item;
2362
2363         list_for_each_entry_rcu(item, fd_list, xa_list) {
2364                 if (!get_file_rcu(item->filp))
2365                         continue;
2366
2367                 if (item->eventfd) {
2368                         eventfd_signal(item->eventfd, 1);
2369                         fput(item->filp);
2370                         continue;
2371                 }
2372
2373                 deliver_event(item, data);
2374                 fput(item->filp);
2375         }
2376 }
2377
2378 static int devx_event_notifier(struct notifier_block *nb,
2379                                unsigned long event_type, void *data)
2380 {
2381         struct mlx5_devx_event_table *table;
2382         struct mlx5_ib_dev *dev;
2383         struct devx_event *event;
2384         struct devx_obj_event *obj_event;
2385         u16 obj_type = 0;
2386         bool is_unaffiliated;
2387         u32 obj_id;
2388
2389         /* Explicit filtering to kernel events which may occur frequently */
2390         if (event_type == MLX5_EVENT_TYPE_CMD ||
2391             event_type == MLX5_EVENT_TYPE_PAGE_REQUEST)
2392                 return NOTIFY_OK;
2393
2394         table = container_of(nb, struct mlx5_devx_event_table, devx_nb.nb);
2395         dev = container_of(table, struct mlx5_ib_dev, devx_event_table);
2396         is_unaffiliated = is_unaffiliated_event(dev->mdev, event_type);
2397
2398         if (!is_unaffiliated)
2399                 obj_type = get_event_obj_type(event_type, data);
2400
2401         rcu_read_lock();
2402         event = xa_load(&table->event_xa, event_type | (obj_type << 16));
2403         if (!event) {
2404                 rcu_read_unlock();
2405                 return NOTIFY_DONE;
2406         }
2407
2408         if (is_unaffiliated) {
2409                 dispatch_event_fd(&event->unaffiliated_list, data);
2410                 rcu_read_unlock();
2411                 return NOTIFY_OK;
2412         }
2413
2414         obj_id = devx_get_obj_id_from_event(event_type, data);
2415         obj_event = xa_load(&event->object_ids, obj_id);
2416         if (!obj_event) {
2417                 rcu_read_unlock();
2418                 return NOTIFY_DONE;
2419         }
2420
2421         dispatch_event_fd(&obj_event->obj_sub_list, data);
2422
2423         rcu_read_unlock();
2424         return NOTIFY_OK;
2425 }
2426
2427 void mlx5_ib_devx_init_event_table(struct mlx5_ib_dev *dev)
2428 {
2429         struct mlx5_devx_event_table *table = &dev->devx_event_table;
2430
2431         xa_init(&table->event_xa);
2432         mutex_init(&table->event_xa_lock);
2433         MLX5_NB_INIT(&table->devx_nb, devx_event_notifier, NOTIFY_ANY);
2434         mlx5_eq_notifier_register(dev->mdev, &table->devx_nb);
2435 }
2436
2437 void mlx5_ib_devx_cleanup_event_table(struct mlx5_ib_dev *dev)
2438 {
2439         struct mlx5_devx_event_table *table = &dev->devx_event_table;
2440         struct devx_event_subscription *sub, *tmp;
2441         struct devx_event *event;
2442         void *entry;
2443         unsigned long id;
2444
2445         mlx5_eq_notifier_unregister(dev->mdev, &table->devx_nb);
2446         mutex_lock(&dev->devx_event_table.event_xa_lock);
2447         xa_for_each(&table->event_xa, id, entry) {
2448                 event = entry;
2449                 list_for_each_entry_safe(sub, tmp, &event->unaffiliated_list,
2450                                          xa_list)
2451                         devx_cleanup_subscription(dev, sub);
2452                 kfree(entry);
2453         }
2454         mutex_unlock(&dev->devx_event_table.event_xa_lock);
2455         xa_destroy(&table->event_xa);
2456 }
2457
2458 static ssize_t devx_async_cmd_event_read(struct file *filp, char __user *buf,
2459                                          size_t count, loff_t *pos)
2460 {
2461         struct devx_async_cmd_event_file *comp_ev_file = filp->private_data;
2462         struct devx_async_event_queue *ev_queue = &comp_ev_file->ev_queue;
2463         struct devx_async_data *event;
2464         int ret = 0;
2465         size_t eventsz;
2466
2467         spin_lock_irq(&ev_queue->lock);
2468
2469         while (list_empty(&ev_queue->event_list)) {
2470                 spin_unlock_irq(&ev_queue->lock);
2471
2472                 if (filp->f_flags & O_NONBLOCK)
2473                         return -EAGAIN;
2474
2475                 if (wait_event_interruptible(
2476                             ev_queue->poll_wait,
2477                             (!list_empty(&ev_queue->event_list) ||
2478                              ev_queue->is_destroyed))) {
2479                         return -ERESTARTSYS;
2480                 }
2481
2482                 if (list_empty(&ev_queue->event_list) &&
2483                     ev_queue->is_destroyed)
2484                         return -EIO;
2485
2486                 spin_lock_irq(&ev_queue->lock);
2487         }
2488
2489         event = list_entry(ev_queue->event_list.next,
2490                            struct devx_async_data, list);
2491         eventsz = event->cmd_out_len +
2492                         sizeof(struct mlx5_ib_uapi_devx_async_cmd_hdr);
2493
2494         if (eventsz > count) {
2495                 spin_unlock_irq(&ev_queue->lock);
2496                 return -ENOSPC;
2497         }
2498
2499         list_del(ev_queue->event_list.next);
2500         spin_unlock_irq(&ev_queue->lock);
2501
2502         if (copy_to_user(buf, &event->hdr, eventsz))
2503                 ret = -EFAULT;
2504         else
2505                 ret = eventsz;
2506
2507         atomic_sub(event->cmd_out_len, &ev_queue->bytes_in_use);
2508         kvfree(event);
2509         return ret;
2510 }
2511
2512 static int devx_async_cmd_event_close(struct inode *inode, struct file *filp)
2513 {
2514         struct ib_uobject *uobj = filp->private_data;
2515         struct devx_async_cmd_event_file *comp_ev_file = container_of(
2516                 uobj, struct devx_async_cmd_event_file, uobj);
2517         struct devx_async_data *entry, *tmp;
2518
2519         spin_lock_irq(&comp_ev_file->ev_queue.lock);
2520         list_for_each_entry_safe(entry, tmp,
2521                                  &comp_ev_file->ev_queue.event_list, list)
2522                 kvfree(entry);
2523         spin_unlock_irq(&comp_ev_file->ev_queue.lock);
2524
2525         uverbs_close_fd(filp);
2526         return 0;
2527 }
2528
2529 static __poll_t devx_async_cmd_event_poll(struct file *filp,
2530                                               struct poll_table_struct *wait)
2531 {
2532         struct devx_async_cmd_event_file *comp_ev_file = filp->private_data;
2533         struct devx_async_event_queue *ev_queue = &comp_ev_file->ev_queue;
2534         __poll_t pollflags = 0;
2535
2536         poll_wait(filp, &ev_queue->poll_wait, wait);
2537
2538         spin_lock_irq(&ev_queue->lock);
2539         if (ev_queue->is_destroyed)
2540                 pollflags = EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2541         else if (!list_empty(&ev_queue->event_list))
2542                 pollflags = EPOLLIN | EPOLLRDNORM;
2543         spin_unlock_irq(&ev_queue->lock);
2544
2545         return pollflags;
2546 }
2547
2548 static const struct file_operations devx_async_cmd_event_fops = {
2549         .owner   = THIS_MODULE,
2550         .read    = devx_async_cmd_event_read,
2551         .poll    = devx_async_cmd_event_poll,
2552         .release = devx_async_cmd_event_close,
2553         .llseek  = no_llseek,
2554 };
2555
2556 static ssize_t devx_async_event_read(struct file *filp, char __user *buf,
2557                                      size_t count, loff_t *pos)
2558 {
2559         struct devx_async_event_file *ev_file = filp->private_data;
2560         struct devx_event_subscription *event_sub;
2561         struct devx_async_event_data *uninitialized_var(event);
2562         int ret = 0;
2563         size_t eventsz;
2564         bool omit_data;
2565         void *event_data;
2566
2567         omit_data = ev_file->omit_data;
2568
2569         spin_lock_irq(&ev_file->lock);
2570
2571         if (ev_file->is_overflow_err) {
2572                 ev_file->is_overflow_err = 0;
2573                 spin_unlock_irq(&ev_file->lock);
2574                 return -EOVERFLOW;
2575         }
2576
2577         if (ev_file->is_destroyed) {
2578                 spin_unlock_irq(&ev_file->lock);
2579                 return -EIO;
2580         }
2581
2582         while (list_empty(&ev_file->event_list)) {
2583                 spin_unlock_irq(&ev_file->lock);
2584
2585                 if (filp->f_flags & O_NONBLOCK)
2586                         return -EAGAIN;
2587
2588                 if (wait_event_interruptible(ev_file->poll_wait,
2589                             (!list_empty(&ev_file->event_list) ||
2590                              ev_file->is_destroyed))) {
2591                         return -ERESTARTSYS;
2592                 }
2593
2594                 spin_lock_irq(&ev_file->lock);
2595                 if (ev_file->is_destroyed) {
2596                         spin_unlock_irq(&ev_file->lock);
2597                         return -EIO;
2598                 }
2599         }
2600
2601         if (omit_data) {
2602                 event_sub = list_first_entry(&ev_file->event_list,
2603                                         struct devx_event_subscription,
2604                                         event_list);
2605                 eventsz = sizeof(event_sub->cookie);
2606                 event_data = &event_sub->cookie;
2607         } else {
2608                 event = list_first_entry(&ev_file->event_list,
2609                                       struct devx_async_event_data, list);
2610                 eventsz = sizeof(struct mlx5_eqe) +
2611                         sizeof(struct mlx5_ib_uapi_devx_async_event_hdr);
2612                 event_data = &event->hdr;
2613         }
2614
2615         if (eventsz > count) {
2616                 spin_unlock_irq(&ev_file->lock);
2617                 return -EINVAL;
2618         }
2619
2620         if (omit_data)
2621                 list_del_init(&event_sub->event_list);
2622         else
2623                 list_del(&event->list);
2624
2625         spin_unlock_irq(&ev_file->lock);
2626
2627         if (copy_to_user(buf, event_data, eventsz))
2628                 /* This points to an application issue, not a kernel concern */
2629                 ret = -EFAULT;
2630         else
2631                 ret = eventsz;
2632
2633         if (!omit_data)
2634                 kfree(event);
2635         return ret;
2636 }
2637
2638 static __poll_t devx_async_event_poll(struct file *filp,
2639                                       struct poll_table_struct *wait)
2640 {
2641         struct devx_async_event_file *ev_file = filp->private_data;
2642         __poll_t pollflags = 0;
2643
2644         poll_wait(filp, &ev_file->poll_wait, wait);
2645
2646         spin_lock_irq(&ev_file->lock);
2647         if (ev_file->is_destroyed)
2648                 pollflags = EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2649         else if (!list_empty(&ev_file->event_list))
2650                 pollflags = EPOLLIN | EPOLLRDNORM;
2651         spin_unlock_irq(&ev_file->lock);
2652
2653         return pollflags;
2654 }
2655
2656 static int devx_async_event_close(struct inode *inode, struct file *filp)
2657 {
2658         struct devx_async_event_file *ev_file = filp->private_data;
2659         struct devx_event_subscription *event_sub, *event_sub_tmp;
2660         struct devx_async_event_data *entry, *tmp;
2661         struct mlx5_ib_dev *dev = ev_file->dev;
2662
2663         mutex_lock(&dev->devx_event_table.event_xa_lock);
2664         /* delete the subscriptions which are related to this FD */
2665         list_for_each_entry_safe(event_sub, event_sub_tmp,
2666                                  &ev_file->subscribed_events_list, file_list) {
2667                 devx_cleanup_subscription(dev, event_sub);
2668                 if (event_sub->eventfd)
2669                         eventfd_ctx_put(event_sub->eventfd);
2670
2671                 list_del_rcu(&event_sub->file_list);
2672                 /* subscription may not be used by the read API any more */
2673                 kfree_rcu(event_sub, rcu);
2674         }
2675
2676         mutex_unlock(&dev->devx_event_table.event_xa_lock);
2677
2678         /* free the pending events allocation */
2679         if (!ev_file->omit_data) {
2680                 spin_lock_irq(&ev_file->lock);
2681                 list_for_each_entry_safe(entry, tmp,
2682                                          &ev_file->event_list, list)
2683                         kfree(entry); /* read can't come any more */
2684                 spin_unlock_irq(&ev_file->lock);
2685         }
2686
2687         uverbs_close_fd(filp);
2688         put_device(&dev->ib_dev.dev);
2689         return 0;
2690 }
2691
2692 static const struct file_operations devx_async_event_fops = {
2693         .owner   = THIS_MODULE,
2694         .read    = devx_async_event_read,
2695         .poll    = devx_async_event_poll,
2696         .release = devx_async_event_close,
2697         .llseek  = no_llseek,
2698 };
2699
2700 static int devx_hot_unplug_async_cmd_event_file(struct ib_uobject *uobj,
2701                                                    enum rdma_remove_reason why)
2702 {
2703         struct devx_async_cmd_event_file *comp_ev_file =
2704                 container_of(uobj, struct devx_async_cmd_event_file,
2705                              uobj);
2706         struct devx_async_event_queue *ev_queue = &comp_ev_file->ev_queue;
2707
2708         spin_lock_irq(&ev_queue->lock);
2709         ev_queue->is_destroyed = 1;
2710         spin_unlock_irq(&ev_queue->lock);
2711
2712         if (why == RDMA_REMOVE_DRIVER_REMOVE)
2713                 wake_up_interruptible(&ev_queue->poll_wait);
2714
2715         mlx5_cmd_cleanup_async_ctx(&comp_ev_file->async_ctx);
2716         return 0;
2717 };
2718
2719 static int devx_hot_unplug_async_event_file(struct ib_uobject *uobj,
2720                                             enum rdma_remove_reason why)
2721 {
2722         struct devx_async_event_file *ev_file =
2723                 container_of(uobj, struct devx_async_event_file,
2724                              uobj);
2725
2726         spin_lock_irq(&ev_file->lock);
2727         ev_file->is_destroyed = 1;
2728         spin_unlock_irq(&ev_file->lock);
2729
2730         wake_up_interruptible(&ev_file->poll_wait);
2731         return 0;
2732 };
2733
2734 DECLARE_UVERBS_NAMED_METHOD(
2735         MLX5_IB_METHOD_DEVX_UMEM_REG,
2736         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_UMEM_REG_HANDLE,
2737                         MLX5_IB_OBJECT_DEVX_UMEM,
2738                         UVERBS_ACCESS_NEW,
2739                         UA_MANDATORY),
2740         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_UMEM_REG_ADDR,
2741                            UVERBS_ATTR_TYPE(u64),
2742                            UA_MANDATORY),
2743         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_UMEM_REG_LEN,
2744                            UVERBS_ATTR_TYPE(u64),
2745                            UA_MANDATORY),
2746         UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_DEVX_UMEM_REG_ACCESS,
2747                              enum ib_access_flags),
2748         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_DEVX_UMEM_REG_OUT_ID,
2749                             UVERBS_ATTR_TYPE(u32),
2750                             UA_MANDATORY));
2751
2752 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
2753         MLX5_IB_METHOD_DEVX_UMEM_DEREG,
2754         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_UMEM_DEREG_HANDLE,
2755                         MLX5_IB_OBJECT_DEVX_UMEM,
2756                         UVERBS_ACCESS_DESTROY,
2757                         UA_MANDATORY));
2758
2759 DECLARE_UVERBS_NAMED_METHOD(
2760         MLX5_IB_METHOD_DEVX_QUERY_EQN,
2761         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_QUERY_EQN_USER_VEC,
2762                            UVERBS_ATTR_TYPE(u32),
2763                            UA_MANDATORY),
2764         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_DEVX_QUERY_EQN_DEV_EQN,
2765                             UVERBS_ATTR_TYPE(u32),
2766                             UA_MANDATORY));
2767
2768 DECLARE_UVERBS_NAMED_METHOD(
2769         MLX5_IB_METHOD_DEVX_QUERY_UAR,
2770         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_QUERY_UAR_USER_IDX,
2771                            UVERBS_ATTR_TYPE(u32),
2772                            UA_MANDATORY),
2773         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_DEVX_QUERY_UAR_DEV_IDX,
2774                             UVERBS_ATTR_TYPE(u32),
2775                             UA_MANDATORY));
2776
2777 DECLARE_UVERBS_NAMED_METHOD(
2778         MLX5_IB_METHOD_DEVX_OTHER,
2779         UVERBS_ATTR_PTR_IN(
2780                 MLX5_IB_ATTR_DEVX_OTHER_CMD_IN,
2781                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_in_cmd_hdr)),
2782                 UA_MANDATORY,
2783                 UA_ALLOC_AND_COPY),
2784         UVERBS_ATTR_PTR_OUT(
2785                 MLX5_IB_ATTR_DEVX_OTHER_CMD_OUT,
2786                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_out_cmd_hdr)),
2787                 UA_MANDATORY));
2788
2789 DECLARE_UVERBS_NAMED_METHOD(
2790         MLX5_IB_METHOD_DEVX_OBJ_CREATE,
2791         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_OBJ_CREATE_HANDLE,
2792                         MLX5_IB_OBJECT_DEVX_OBJ,
2793                         UVERBS_ACCESS_NEW,
2794                         UA_MANDATORY),
2795         UVERBS_ATTR_PTR_IN(
2796                 MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_IN,
2797                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_in_cmd_hdr)),
2798                 UA_MANDATORY,
2799                 UA_ALLOC_AND_COPY),
2800         UVERBS_ATTR_PTR_OUT(
2801                 MLX5_IB_ATTR_DEVX_OBJ_CREATE_CMD_OUT,
2802                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_out_cmd_hdr)),
2803                 UA_MANDATORY));
2804
2805 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
2806         MLX5_IB_METHOD_DEVX_OBJ_DESTROY,
2807         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_OBJ_DESTROY_HANDLE,
2808                         MLX5_IB_OBJECT_DEVX_OBJ,
2809                         UVERBS_ACCESS_DESTROY,
2810                         UA_MANDATORY));
2811
2812 DECLARE_UVERBS_NAMED_METHOD(
2813         MLX5_IB_METHOD_DEVX_OBJ_MODIFY,
2814         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_OBJ_MODIFY_HANDLE,
2815                         UVERBS_IDR_ANY_OBJECT,
2816                         UVERBS_ACCESS_WRITE,
2817                         UA_MANDATORY),
2818         UVERBS_ATTR_PTR_IN(
2819                 MLX5_IB_ATTR_DEVX_OBJ_MODIFY_CMD_IN,
2820                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_in_cmd_hdr)),
2821                 UA_MANDATORY,
2822                 UA_ALLOC_AND_COPY),
2823         UVERBS_ATTR_PTR_OUT(
2824                 MLX5_IB_ATTR_DEVX_OBJ_MODIFY_CMD_OUT,
2825                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_out_cmd_hdr)),
2826                 UA_MANDATORY));
2827
2828 DECLARE_UVERBS_NAMED_METHOD(
2829         MLX5_IB_METHOD_DEVX_OBJ_QUERY,
2830         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_OBJ_QUERY_HANDLE,
2831                         UVERBS_IDR_ANY_OBJECT,
2832                         UVERBS_ACCESS_READ,
2833                         UA_MANDATORY),
2834         UVERBS_ATTR_PTR_IN(
2835                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_IN,
2836                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_in_cmd_hdr)),
2837                 UA_MANDATORY,
2838                 UA_ALLOC_AND_COPY),
2839         UVERBS_ATTR_PTR_OUT(
2840                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_OUT,
2841                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_out_cmd_hdr)),
2842                 UA_MANDATORY));
2843
2844 DECLARE_UVERBS_NAMED_METHOD(
2845         MLX5_IB_METHOD_DEVX_OBJ_ASYNC_QUERY,
2846         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_OBJ_QUERY_HANDLE,
2847                         UVERBS_IDR_ANY_OBJECT,
2848                         UVERBS_ACCESS_READ,
2849                         UA_MANDATORY),
2850         UVERBS_ATTR_PTR_IN(
2851                 MLX5_IB_ATTR_DEVX_OBJ_QUERY_CMD_IN,
2852                 UVERBS_ATTR_MIN_SIZE(MLX5_ST_SZ_BYTES(general_obj_in_cmd_hdr)),
2853                 UA_MANDATORY,
2854                 UA_ALLOC_AND_COPY),
2855         UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_OUT_LEN,
2856                 u16, UA_MANDATORY),
2857         UVERBS_ATTR_FD(MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_FD,
2858                 MLX5_IB_OBJECT_DEVX_ASYNC_CMD_FD,
2859                 UVERBS_ACCESS_READ,
2860                 UA_MANDATORY),
2861         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_OBJ_QUERY_ASYNC_WR_ID,
2862                 UVERBS_ATTR_TYPE(u64),
2863                 UA_MANDATORY));
2864
2865 DECLARE_UVERBS_NAMED_METHOD(
2866         MLX5_IB_METHOD_DEVX_SUBSCRIBE_EVENT,
2867         UVERBS_ATTR_FD(MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_FD_HANDLE,
2868                 MLX5_IB_OBJECT_DEVX_ASYNC_EVENT_FD,
2869                 UVERBS_ACCESS_READ,
2870                 UA_MANDATORY),
2871         UVERBS_ATTR_IDR(MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_OBJ_HANDLE,
2872                 MLX5_IB_OBJECT_DEVX_OBJ,
2873                 UVERBS_ACCESS_READ,
2874                 UA_OPTIONAL),
2875         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_TYPE_NUM_LIST,
2876                 UVERBS_ATTR_MIN_SIZE(sizeof(u16)),
2877                 UA_MANDATORY,
2878                 UA_ALLOC_AND_COPY),
2879         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_COOKIE,
2880                 UVERBS_ATTR_TYPE(u64),
2881                 UA_OPTIONAL),
2882         UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_DEVX_SUBSCRIBE_EVENT_FD_NUM,
2883                 UVERBS_ATTR_TYPE(u32),
2884                 UA_OPTIONAL));
2885
2886 DECLARE_UVERBS_GLOBAL_METHODS(MLX5_IB_OBJECT_DEVX,
2887                               &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_OTHER),
2888                               &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_QUERY_UAR),
2889                               &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_QUERY_EQN),
2890                               &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_SUBSCRIBE_EVENT));
2891
2892 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_DEVX_OBJ,
2893                             UVERBS_TYPE_ALLOC_IDR(devx_obj_cleanup),
2894                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_OBJ_CREATE),
2895                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_OBJ_DESTROY),
2896                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_OBJ_MODIFY),
2897                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_OBJ_QUERY),
2898                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_OBJ_ASYNC_QUERY));
2899
2900 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_DEVX_UMEM,
2901                             UVERBS_TYPE_ALLOC_IDR(devx_umem_cleanup),
2902                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_UMEM_REG),
2903                             &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_UMEM_DEREG));
2904
2905
2906 DECLARE_UVERBS_NAMED_METHOD(
2907         MLX5_IB_METHOD_DEVX_ASYNC_CMD_FD_ALLOC,
2908         UVERBS_ATTR_FD(MLX5_IB_ATTR_DEVX_ASYNC_CMD_FD_ALLOC_HANDLE,
2909                         MLX5_IB_OBJECT_DEVX_ASYNC_CMD_FD,
2910                         UVERBS_ACCESS_NEW,
2911                         UA_MANDATORY));
2912
2913 DECLARE_UVERBS_NAMED_OBJECT(
2914         MLX5_IB_OBJECT_DEVX_ASYNC_CMD_FD,
2915         UVERBS_TYPE_ALLOC_FD(sizeof(struct devx_async_cmd_event_file),
2916                              devx_hot_unplug_async_cmd_event_file,
2917                              &devx_async_cmd_event_fops, "[devx_async_cmd]",
2918                              O_RDONLY),
2919         &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_ASYNC_CMD_FD_ALLOC));
2920
2921 DECLARE_UVERBS_NAMED_METHOD(
2922         MLX5_IB_METHOD_DEVX_ASYNC_EVENT_FD_ALLOC,
2923         UVERBS_ATTR_FD(MLX5_IB_ATTR_DEVX_ASYNC_EVENT_FD_ALLOC_HANDLE,
2924                         MLX5_IB_OBJECT_DEVX_ASYNC_EVENT_FD,
2925                         UVERBS_ACCESS_NEW,
2926                         UA_MANDATORY),
2927         UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_DEVX_ASYNC_EVENT_FD_ALLOC_FLAGS,
2928                         enum mlx5_ib_uapi_devx_create_event_channel_flags,
2929                         UA_MANDATORY));
2930
2931 DECLARE_UVERBS_NAMED_OBJECT(
2932         MLX5_IB_OBJECT_DEVX_ASYNC_EVENT_FD,
2933         UVERBS_TYPE_ALLOC_FD(sizeof(struct devx_async_event_file),
2934                              devx_hot_unplug_async_event_file,
2935                              &devx_async_event_fops, "[devx_async_event]",
2936                              O_RDONLY),
2937         &UVERBS_METHOD(MLX5_IB_METHOD_DEVX_ASYNC_EVENT_FD_ALLOC));
2938
2939 static bool devx_is_supported(struct ib_device *device)
2940 {
2941         struct mlx5_ib_dev *dev = to_mdev(device);
2942
2943         return MLX5_CAP_GEN(dev->mdev, log_max_uctx);
2944 }
2945
2946 const struct uapi_definition mlx5_ib_devx_defs[] = {
2947         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
2948                 MLX5_IB_OBJECT_DEVX,
2949                 UAPI_DEF_IS_OBJ_SUPPORTED(devx_is_supported)),
2950         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
2951                 MLX5_IB_OBJECT_DEVX_OBJ,
2952                 UAPI_DEF_IS_OBJ_SUPPORTED(devx_is_supported)),
2953         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
2954                 MLX5_IB_OBJECT_DEVX_UMEM,
2955                 UAPI_DEF_IS_OBJ_SUPPORTED(devx_is_supported)),
2956         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
2957                 MLX5_IB_OBJECT_DEVX_ASYNC_CMD_FD,
2958                 UAPI_DEF_IS_OBJ_SUPPORTED(devx_is_supported)),
2959         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
2960                 MLX5_IB_OBJECT_DEVX_ASYNC_EVENT_FD,
2961                 UAPI_DEF_IS_OBJ_SUPPORTED(devx_is_supported)),
2962         {},
2963 };