1 /*- 2 * Copyright (c) 2015, Mellanox Technologies, Inc. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 */ 25 26 #ifndef ICL_ISER_H 27 #define ICL_ISER_H 28 29 /* 30 * iSCSI Common Layer for RDMA. 31 */ 32 33 #include <sys/cdefs.h> 34 #include <sys/param.h> 35 #include <sys/capsicum.h> 36 #include <sys/condvar.h> 37 #include <sys/conf.h> 38 #include <sys/file.h> 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/lock.h> 42 #include <sys/mbuf.h> 43 #include <sys/mutex.h> 44 #include <sys/module.h> 45 #include <sys/protosw.h> 46 #include <sys/socket.h> 47 #include <sys/socketvar.h> 48 #include <sys/sysctl.h> 49 #include <sys/systm.h> 50 #include <sys/sx.h> 51 #include <sys/uio.h> 52 #include <sys/taskqueue.h> 53 #include <sys/bio.h> 54 #include <vm/uma.h> 55 #include <netinet/in.h> 56 #include <netinet/tcp.h> 57 #include <dev/iscsi/icl.h> 58 #include <dev/iscsi/iscsi_proto.h> 59 #include <icl_conn_if.h> 60 #include <cam/cam.h> 61 #include <cam/cam_ccb.h> 62 #include <rdma/ib_verbs.h> 63 #include <rdma/ib_fmr_pool.h> 64 #include <rdma/rdma_cm.h> 65 66 67 #define ISER_DBG(X, ...) \ 68 do { \ 69 if (unlikely(iser_debug > 2)) \ 70 printf("DEBUG: %s: " X "\n", \ 71 __func__, ## __VA_ARGS__); \ 72 } while (0) 73 74 #define ISER_INFO(X, ...) \ 75 do { \ 76 if (unlikely(iser_debug > 1)) \ 77 printf("INFO: %s: " X "\n", \ 78 __func__, ## __VA_ARGS__); \ 79 } while (0) 80 81 #define ISER_WARN(X, ...) \ 82 do { \ 83 if (unlikely(iser_debug > 0)) { \ 84 printf("WARNING: %s: " X "\n", \ 85 __func__, ## __VA_ARGS__); \ 86 } \ 87 } while (0) 88 89 #define ISER_ERR(X, ...) \ 90 printf("ERROR: %s: " X "\n", __func__, ## __VA_ARGS__) 91 92 #define ISER_VER 0x10 93 #define ISER_WSV 0x08 94 #define ISER_RSV 0x04 95 96 #define ISER_FASTREG_LI_WRID 0xffffffffffffffffULL 97 #define ISER_BEACON_WRID 0xfffffffffffffffeULL 98 99 #define SHIFT_4K 12 100 #define SIZE_4K (1ULL << SHIFT_4K) 101 #define MASK_4K (~(SIZE_4K-1)) 102 103 /* support up to 512KB in one RDMA */ 104 #define ISCSI_ISER_SG_TABLESIZE (0x80000 >> SHIFT_4K) 105 #define ISER_DEF_XMIT_CMDS_MAX 256 106 107 /* the max RX (recv) WR supported by the iSER QP is defined by * 108 * max_recv_wr = commands_max + recv_beacon */ 109 #define ISER_QP_MAX_RECV_DTOS (ISER_DEF_XMIT_CMDS_MAX + 1) 110 #define ISER_MIN_POSTED_RX (ISER_DEF_XMIT_CMDS_MAX >> 2) 111 112 /* QP settings */ 113 /* Maximal bounds on received asynchronous PDUs */ 114 #define ISER_MAX_RX_MISC_PDUS 4 /* NOOP_IN(2) , ASYNC_EVENT(2) */ 115 #define ISER_MAX_TX_MISC_PDUS 6 /* NOOP_OUT(2), TEXT(1), SCSI_TMFUNC(2), LOGOUT(1) */ 116 117 /* the max TX (send) WR supported by the iSER QP is defined by * 118 * max_send_wr = T * (1 + D) + C ; D is how many inflight dataouts we expect * 119 * to have at max for SCSI command. The tx posting & completion handling code * 120 * supports -EAGAIN scheme where tx is suspended till the QP has room for more * 121 * send WR. D=8 comes from 64K/8K */ 122 123 #define ISER_INFLIGHT_DATAOUTS 8 124 125 /* the send_beacon increase the max_send_wr by 1 */ 126 #define ISER_QP_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \ 127 (1 + ISER_INFLIGHT_DATAOUTS) + \ 128 ISER_MAX_TX_MISC_PDUS + \ 129 ISER_MAX_RX_MISC_PDUS + 1) 130 131 #define ISER_GET_MAX_XMIT_CMDS(send_wr) ((send_wr \ 132 - ISER_MAX_TX_MISC_PDUS \ 133 - ISER_MAX_RX_MISC_PDUS - 1) / \ 134 (1 + ISER_INFLIGHT_DATAOUTS)) 135 136 #define ISER_WC_BATCH_COUNT 16 137 #define ISER_SIGNAL_CMD_COUNT 32 138 139 /* Maximal QP's recommended per CQ. In case we use more QP's per CQ we might * 140 * encounter a CQ overrun state. */ 141 #define ISCSI_ISER_MAX_CONN 8 142 #define ISER_MAX_RX_LEN (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN) 143 #define ISER_MAX_TX_LEN (ISER_QP_MAX_REQ_DTOS * ISCSI_ISER_MAX_CONN) 144 #define ISER_MAX_CQ_LEN (ISER_MAX_RX_LEN + ISER_MAX_TX_LEN + \ 145 ISCSI_ISER_MAX_CONN) 146 147 #define ISER_ZBVA_NOT_SUPPORTED 0x80 148 #define ISER_SEND_W_INV_NOT_SUPPORTED 0x40 149 150 #define ISCSI_DEF_MAX_RECV_SEG_LEN 8192 151 #define ISCSI_OPCODE_MASK 0x3f 152 153 #define icl_to_iser_conn(ic) \ 154 container_of(ic, struct iser_conn, icl_conn) 155 #define icl_to_iser_pdu(ip) \ 156 container_of(ip, struct icl_iser_pdu, icl_pdu) 157 158 /** 159 * struct iser_hdr - iSER header 160 * 161 * @flags: flags support (zbva, remote_inv) 162 * @rsvd: reserved 163 * @write_stag: write rkey 164 * @write_va: write virtual address 165 * @reaf_stag: read rkey 166 * @read_va: read virtual address 167 */ 168 struct iser_hdr { 169 u8 flags; 170 u8 rsvd[3]; 171 __be32 write_stag; 172 __be64 write_va; 173 __be32 read_stag; 174 __be64 read_va; 175 } __attribute__((packed)); 176 177 struct iser_cm_hdr { 178 u8 flags; 179 u8 rsvd[3]; 180 } __packed; 181 182 /* Constant PDU lengths calculations */ 183 #define ISER_HEADERS_LEN (sizeof(struct iser_hdr) + ISCSI_BHS_SIZE) 184 185 #define ISER_RECV_DATA_SEG_LEN 128 186 #define ISER_RX_PAYLOAD_SIZE (ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN) 187 188 #define ISER_RX_LOGIN_SIZE (ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN) 189 190 enum iser_conn_state { 191 ISER_CONN_INIT, /* descriptor allocd, no conn */ 192 ISER_CONN_PENDING, /* in the process of being established */ 193 ISER_CONN_UP, /* up and running */ 194 ISER_CONN_TERMINATING, /* in the process of being terminated */ 195 ISER_CONN_DOWN, /* shut down */ 196 ISER_CONN_STATES_NUM 197 }; 198 199 enum iser_task_status { 200 ISER_TASK_STATUS_INIT = 0, 201 ISER_TASK_STATUS_STARTED, 202 ISER_TASK_STATUS_COMPLETED 203 }; 204 205 enum iser_data_dir { 206 ISER_DIR_IN = 0, /* to initiator */ 207 ISER_DIR_OUT, /* from initiator */ 208 ISER_DIRS_NUM 209 }; 210 211 /** 212 * struct iser_mem_reg - iSER memory registration info 213 * 214 * @sge: memory region sg element 215 * @rkey: memory region remote key 216 * @mem_h: pointer to registration context (FMR/Fastreg) 217 */ 218 struct iser_mem_reg { 219 struct ib_sge sge; 220 u32 rkey; 221 void *mem_h; 222 }; 223 224 enum iser_desc_type { 225 ISCSI_TX_CONTROL , 226 ISCSI_TX_SCSI_COMMAND, 227 ISCSI_TX_DATAOUT 228 }; 229 230 /** 231 * struct iser_data_buf - iSER data buffer 232 * 233 * @sg: pointer to the sg list 234 * @size: num entries of this sg 235 * @data_len: total beffer byte len 236 * @dma_nents: returned by dma_map_sg 237 * @copy_buf: allocated copy buf for SGs unaligned 238 * for rdma which are copied 239 * @orig_sg: pointer to the original sg list (in case 240 * we used a copy) 241 * @sg_single: SG-ified clone of a non SG SC or 242 * unaligned SG 243 */ 244 struct iser_data_buf { 245 struct scatterlist sgl[ISCSI_ISER_SG_TABLESIZE]; 246 void *sg; 247 int size; 248 unsigned long data_len; 249 unsigned int dma_nents; 250 char *copy_buf; 251 struct scatterlist *orig_sg; 252 struct scatterlist sg_single; 253 }; 254 255 /* fwd declarations */ 256 struct iser_conn; 257 struct ib_conn; 258 struct iser_device; 259 260 /** 261 * struct iser_tx_desc - iSER TX descriptor (for send wr_id) 262 * 263 * @iser_header: iser header 264 * @iscsi_header: iscsi header (bhs) 265 * @type: command/control/dataout 266 * @dma_addr: header buffer dma_address 267 * @tx_sg: sg[0] points to iser/iscsi headers 268 * sg[1] optionally points to either of immediate data 269 * unsolicited data-out or control 270 * @num_sge: number sges used on this TX task 271 * @mapped: indicates if the descriptor is dma mapped 272 */ 273 struct iser_tx_desc { 274 struct iser_hdr iser_header; 275 struct iscsi_bhs iscsi_header __attribute__((packed)); 276 enum iser_desc_type type; 277 u64 dma_addr; 278 struct ib_sge tx_sg[2]; 279 int num_sge; 280 bool mapped; 281 }; 282 283 #define ISER_RX_PAD_SIZE (256 - (ISER_RX_PAYLOAD_SIZE + \ 284 sizeof(u64) + sizeof(struct ib_sge))) 285 /** 286 * struct iser_rx_desc - iSER RX descriptor (for recv wr_id) 287 * 288 * @iser_header: iser header 289 * @iscsi_header: iscsi header 290 * @data: received data segment 291 * @dma_addr: receive buffer dma address 292 * @rx_sg: ib_sge of receive buffer 293 * @pad: for sense data TODO: Modify to maximum sense length supported 294 */ 295 struct iser_rx_desc { 296 struct iser_hdr iser_header; 297 struct iscsi_bhs iscsi_header; 298 char data[ISER_RECV_DATA_SEG_LEN]; 299 u64 dma_addr; 300 struct ib_sge rx_sg; 301 char pad[ISER_RX_PAD_SIZE]; 302 } __attribute__((packed)); 303 304 struct icl_iser_pdu { 305 struct icl_pdu icl_pdu; 306 struct iser_tx_desc desc; 307 struct iser_conn *iser_conn; 308 enum iser_task_status status; 309 struct ccb_scsiio *csio; 310 int command_sent; 311 int dir[ISER_DIRS_NUM]; 312 struct iser_mem_reg rdma_reg[ISER_DIRS_NUM]; 313 struct iser_data_buf data[ISER_DIRS_NUM]; 314 }; 315 316 /** 317 * struct iser_comp - iSER completion context 318 * 319 * @device: pointer to device handle 320 * @cq: completion queue 321 * @wcs: work completion array 322 * @tq: taskqueue handle 323 * @task: task to run task_fn 324 * @active_qps: Number of active QPs attached 325 * to completion context 326 */ 327 struct iser_comp { 328 struct iser_device *device; 329 struct ib_cq *cq; 330 struct ib_wc wcs[ISER_WC_BATCH_COUNT]; 331 struct taskqueue *tq; 332 struct task task; 333 int active_qps; 334 }; 335 336 /** 337 * struct iser_device - iSER device handle 338 * 339 * @ib_device: RDMA device 340 * @pd: Protection Domain for this device 341 * @dev_attr: Device attributes container 342 * @mr: Global DMA memory region 343 * @event_handler: IB events handle routine 344 * @ig_list: entry in devices list 345 * @refcount: Reference counter, dominated by open iser connections 346 * @comps_used: Number of completion contexts used, Min between online 347 * cpus and device max completion vectors 348 * @comps: Dinamically allocated array of completion handlers 349 */ 350 struct iser_device { 351 struct ib_device *ib_device; 352 struct ib_pd *pd; 353 struct ib_device_attr dev_attr; 354 struct ib_mr *mr; 355 struct ib_event_handler event_handler; 356 struct list_head ig_list; 357 int refcount; 358 int comps_used; 359 struct iser_comp *comps; 360 }; 361 362 /** 363 * struct iser_reg_resources - Fast registration recources 364 * 365 * @mr: memory region 366 * @mr_valid: is mr valid indicator 367 */ 368 struct iser_reg_resources { 369 struct ib_mr *mr; 370 u8 mr_valid:1; 371 }; 372 373 /** 374 * struct fast_reg_descriptor - Fast registration descriptor 375 * 376 * @list: entry in connection fastreg pool 377 * @rsc: data buffer registration resources 378 */ 379 struct fast_reg_descriptor { 380 struct list_head list; 381 struct iser_reg_resources rsc; 382 }; 383 384 385 /** 386 * struct iser_beacon - beacon to signal all flush errors were drained 387 * 388 * @send: send wr 389 * @recv: recv wr 390 * @flush_lock: protects flush_cv 391 * @flush_cv: condition variable for beacon flush 392 */ 393 struct iser_beacon { 394 union { 395 struct ib_send_wr send; 396 struct ib_recv_wr recv; 397 }; 398 struct mtx flush_lock; 399 struct cv flush_cv; 400 }; 401 402 /** 403 * struct ib_conn - Infiniband related objects 404 * 405 * @cma_id: rdma_cm connection maneger handle 406 * @qp: Connection Queue-pair 407 * @device: reference to iser device 408 * @comp: iser completion context 409 */ 410 struct ib_conn { 411 struct rdma_cm_id *cma_id; 412 struct ib_qp *qp; 413 int post_recv_buf_count; 414 u8 sig_count; 415 struct ib_recv_wr rx_wr[ISER_MIN_POSTED_RX]; 416 struct iser_device *device; 417 struct iser_comp *comp; 418 struct iser_beacon beacon; 419 struct mtx lock; 420 union { 421 struct { 422 struct ib_fmr_pool *pool; 423 struct iser_page_vec *page_vec; 424 } fmr; 425 struct { 426 struct list_head pool; 427 int pool_size; 428 } fastreg; 429 }; 430 }; 431 432 struct iser_conn { 433 struct icl_conn icl_conn; 434 struct ib_conn ib_conn; 435 struct cv up_cv; 436 struct list_head conn_list; 437 struct sx state_mutex; 438 enum iser_conn_state state; 439 int qp_max_recv_dtos; 440 int min_posted_rx; 441 u16 max_cmds; 442 char *login_buf; 443 char *login_req_buf, *login_resp_buf; 444 u64 login_req_dma, login_resp_dma; 445 unsigned int rx_desc_head; 446 struct iser_rx_desc *rx_descs; 447 u32 num_rx_descs; 448 bool handoff_done; 449 }; 450 451 /** 452 * struct iser_global: iSER global context 453 * 454 * @device_list_mutex: protects device_list 455 * @device_list: iser devices global list 456 * @connlist_mutex: protects connlist 457 * @connlist: iser connections global list 458 * @desc_cache: kmem cache for tx dataout 459 * @close_conns_mutex: serializes conns closure 460 */ 461 struct iser_global { 462 struct sx device_list_mutex; 463 struct list_head device_list; 464 struct mtx connlist_mutex; 465 struct list_head connlist; 466 struct sx close_conns_mutex; 467 }; 468 469 extern struct iser_global ig; 470 extern int iser_debug; 471 472 void 473 iser_create_send_desc(struct iser_conn *, struct iser_tx_desc *); 474 475 int 476 iser_post_recvl(struct iser_conn *); 477 478 int 479 iser_post_recvm(struct iser_conn *, int); 480 481 int 482 iser_alloc_login_buf(struct iser_conn *iser_conn); 483 484 void 485 iser_free_login_buf(struct iser_conn *iser_conn); 486 487 int 488 iser_post_send(struct ib_conn *, struct iser_tx_desc *, bool); 489 490 void 491 iser_snd_completion(struct iser_tx_desc *, struct ib_conn *); 492 493 void 494 iser_rcv_completion(struct iser_rx_desc *, unsigned long, 495 struct ib_conn *); 496 497 void 498 iser_pdu_free(struct icl_conn *, struct icl_pdu *); 499 500 struct icl_pdu * 501 iser_new_pdu(struct icl_conn *ic, int flags); 502 503 int 504 iser_alloc_rx_descriptors(struct iser_conn *, int); 505 506 void 507 iser_free_rx_descriptors(struct iser_conn *); 508 509 int 510 iser_initialize_headers(struct icl_iser_pdu *, struct iser_conn *); 511 512 int 513 iser_send_control(struct iser_conn *, struct icl_iser_pdu *); 514 515 int 516 iser_send_command(struct iser_conn *, struct icl_iser_pdu *); 517 518 int 519 iser_reg_rdma_mem(struct icl_iser_pdu *, enum iser_data_dir); 520 521 void 522 iser_unreg_rdma_mem(struct icl_iser_pdu *, enum iser_data_dir); 523 524 int 525 iser_create_fastreg_pool(struct ib_conn *, unsigned); 526 527 void 528 iser_free_fastreg_pool(struct ib_conn *); 529 530 int 531 iser_dma_map_task_data(struct icl_iser_pdu *, 532 struct iser_data_buf *, enum iser_data_dir, 533 enum dma_data_direction); 534 535 int 536 iser_conn_terminate(struct iser_conn *); 537 538 void 539 iser_free_ib_conn_res(struct iser_conn *, bool); 540 541 void 542 iser_dma_unmap_task_data(struct icl_iser_pdu *, struct iser_data_buf *, 543 enum dma_data_direction); 544 545 int 546 iser_cma_handler(struct rdma_cm_id *, struct rdma_cm_event *); 547 548 #endif /* !ICL_ISER_H */ 549