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