| 1 | /* |
| 2 | * librpma_gpspm: IO engine that uses PMDK librpma to write data, |
| 3 | * based on General Purpose Server Persistency Method |
| 4 | * |
| 5 | * Copyright 2020-2021, Intel Corporation |
| 6 | * |
| 7 | * This program is free software; you can redistribute it and/or |
| 8 | * modify it under the terms of the GNU General Public License, |
| 9 | * version 2 as published by the Free Software Foundation.. |
| 10 | * |
| 11 | * This program is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | * GNU General Public License for more details. |
| 15 | */ |
| 16 | |
| 17 | #include "librpma_fio.h" |
| 18 | |
| 19 | #include <libpmem.h> |
| 20 | |
| 21 | /* Generated by the protocol buffer compiler from: librpma_gpspm_flush.proto */ |
| 22 | #include "librpma_gpspm_flush.pb-c.h" |
| 23 | |
| 24 | #define MAX_MSG_SIZE (512) |
| 25 | #define IO_U_BUF_LEN (2 * MAX_MSG_SIZE) |
| 26 | #define SEND_OFFSET (0) |
| 27 | #define RECV_OFFSET (SEND_OFFSET + MAX_MSG_SIZE) |
| 28 | |
| 29 | #define GPSPM_FLUSH_REQUEST__LAST \ |
| 30 | { PROTOBUF_C_MESSAGE_INIT(&gpspm_flush_request__descriptor), 0, 0, 0 } |
| 31 | |
| 32 | /* |
| 33 | * 'Flush_req_last' is the last flush request |
| 34 | * the client has to send to server to indicate |
| 35 | * that the client is done. |
| 36 | */ |
| 37 | static const GPSPMFlushRequest Flush_req_last = GPSPM_FLUSH_REQUEST__LAST; |
| 38 | |
| 39 | #define IS_NOT_THE_LAST_MESSAGE(flush_req) \ |
| 40 | (flush_req->length != Flush_req_last.length || \ |
| 41 | flush_req->offset != Flush_req_last.offset) |
| 42 | |
| 43 | /* client side implementation */ |
| 44 | |
| 45 | /* get next io_u message buffer in the round-robin fashion */ |
| 46 | #define IO_U_NEXT_BUF_OFF_CLIENT(cd) \ |
| 47 | (IO_U_BUF_LEN * ((cd->msg_curr++) % cd->msg_num)) |
| 48 | |
| 49 | struct client_data { |
| 50 | /* memory for sending and receiving buffered */ |
| 51 | char *io_us_msgs; |
| 52 | |
| 53 | /* resources for messaging buffer */ |
| 54 | uint32_t msg_num; |
| 55 | uint32_t msg_curr; |
| 56 | struct rpma_mr_local *msg_mr; |
| 57 | }; |
| 58 | |
| 59 | static inline int client_io_flush(struct thread_data *td, |
| 60 | struct io_u *first_io_u, struct io_u *last_io_u, |
| 61 | unsigned long long int len); |
| 62 | |
| 63 | static int client_get_io_u_index(struct ibv_wc *wc, unsigned int *io_u_index); |
| 64 | |
| 65 | static int client_init(struct thread_data *td) |
| 66 | { |
| 67 | struct librpma_fio_client_data *ccd; |
| 68 | struct client_data *cd; |
| 69 | uint32_t write_num; |
| 70 | struct rpma_conn_cfg *cfg = NULL; |
| 71 | int ret; |
| 72 | |
| 73 | /* |
| 74 | * not supported: |
| 75 | * - readwrite = read / trim / randread / randtrim / |
| 76 | * / rw / randrw / trimwrite |
| 77 | */ |
| 78 | if (td_read(td) || td_trim(td)) { |
| 79 | td_verror(td, EINVAL, "Not supported mode."); |
| 80 | return -1; |
| 81 | } |
| 82 | |
| 83 | /* allocate client's data */ |
| 84 | cd = calloc(1, sizeof(*cd)); |
| 85 | if (cd == NULL) { |
| 86 | td_verror(td, errno, "calloc"); |
| 87 | return -1; |
| 88 | } |
| 89 | |
| 90 | /* |
| 91 | * Calculate the required number of WRITEs and FLUSHes. |
| 92 | * |
| 93 | * Note: Each flush is a request (SEND) and response (RECV) pair. |
| 94 | */ |
| 95 | if (td_random(td)) { |
| 96 | write_num = td->o.iodepth; /* WRITE * N */ |
| 97 | cd->msg_num = td->o.iodepth; /* FLUSH * N */ |
| 98 | } else { |
| 99 | if (td->o.sync_io) { |
| 100 | write_num = 1; /* WRITE */ |
| 101 | cd->msg_num = 1; /* FLUSH */ |
| 102 | } else { |
| 103 | write_num = td->o.iodepth; /* WRITE * N */ |
| 104 | /* |
| 105 | * FLUSH * B where: |
| 106 | * - B == ceil(iodepth / iodepth_batch) |
| 107 | * which is the number of batches for N writes |
| 108 | */ |
| 109 | cd->msg_num = LIBRPMA_FIO_CEIL(td->o.iodepth, |
| 110 | td->o.iodepth_batch); |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | /* create a connection configuration object */ |
| 115 | if ((ret = rpma_conn_cfg_new(&cfg))) { |
| 116 | librpma_td_verror(td, ret, "rpma_conn_cfg_new"); |
| 117 | goto err_free_cd; |
| 118 | } |
| 119 | |
| 120 | /* |
| 121 | * Calculate the required queue sizes where: |
| 122 | * - the send queue (SQ) has to be big enough to accommodate |
| 123 | * all io_us (WRITEs) and all flush requests (SENDs) |
| 124 | * - the receive queue (RQ) has to be big enough to accommodate |
| 125 | * all flush responses (RECVs) |
| 126 | * - the completion queue (CQ) has to be big enough to accommodate all |
| 127 | * success and error completions (sq_size + rq_size) |
| 128 | */ |
| 129 | if ((ret = rpma_conn_cfg_set_sq_size(cfg, write_num + cd->msg_num))) { |
| 130 | librpma_td_verror(td, ret, "rpma_conn_cfg_set_sq_size"); |
| 131 | goto err_cfg_delete; |
| 132 | } |
| 133 | if ((ret = rpma_conn_cfg_set_rq_size(cfg, cd->msg_num))) { |
| 134 | librpma_td_verror(td, ret, "rpma_conn_cfg_set_rq_size"); |
| 135 | goto err_cfg_delete; |
| 136 | } |
| 137 | if ((ret = rpma_conn_cfg_set_cq_size(cfg, write_num + cd->msg_num * 2))) { |
| 138 | librpma_td_verror(td, ret, "rpma_conn_cfg_set_cq_size"); |
| 139 | goto err_cfg_delete; |
| 140 | } |
| 141 | |
| 142 | if (librpma_fio_client_init(td, cfg)) |
| 143 | goto err_cfg_delete; |
| 144 | |
| 145 | ccd = td->io_ops_data; |
| 146 | |
| 147 | if (ccd->ws->direct_write_to_pmem && |
| 148 | ccd->server_mr_flush_type == RPMA_FLUSH_TYPE_PERSISTENT && |
| 149 | td->thread_number == 1) { |
| 150 | /* XXX log_info mixes with the JSON output */ |
| 151 | log_err( |
| 152 | "Note: The server side supports Direct Write to PMem and it is equipped with PMem (direct_write_to_pmem).\n" |
| 153 | "You can use librpma_client and librpma_server engines for better performance instead of GPSPM.\n"); |
| 154 | } |
| 155 | |
| 156 | /* validate the server's RQ capacity */ |
| 157 | if (cd->msg_num > ccd->ws->max_msg_num) { |
| 158 | log_err( |
| 159 | "server's RQ size (iodepth) too small to handle the client's workspace requirements (%u < %u)\n", |
| 160 | ccd->ws->max_msg_num, cd->msg_num); |
| 161 | goto err_cleanup_common; |
| 162 | } |
| 163 | |
| 164 | if ((ret = rpma_conn_cfg_delete(&cfg))) { |
| 165 | librpma_td_verror(td, ret, "rpma_conn_cfg_delete"); |
| 166 | /* non fatal error - continue */ |
| 167 | } |
| 168 | |
| 169 | ccd->flush = client_io_flush; |
| 170 | ccd->get_io_u_index = client_get_io_u_index; |
| 171 | ccd->client_data = cd; |
| 172 | |
| 173 | return 0; |
| 174 | |
| 175 | err_cleanup_common: |
| 176 | librpma_fio_client_cleanup(td); |
| 177 | |
| 178 | err_cfg_delete: |
| 179 | (void) rpma_conn_cfg_delete(&cfg); |
| 180 | |
| 181 | err_free_cd: |
| 182 | free(cd); |
| 183 | |
| 184 | return -1; |
| 185 | } |
| 186 | |
| 187 | static int client_post_init(struct thread_data *td) |
| 188 | { |
| 189 | struct librpma_fio_client_data *ccd = td->io_ops_data; |
| 190 | struct client_data *cd = ccd->client_data; |
| 191 | unsigned int io_us_msgs_size; |
| 192 | int ret; |
| 193 | |
| 194 | /* message buffers initialization and registration */ |
| 195 | io_us_msgs_size = cd->msg_num * IO_U_BUF_LEN; |
| 196 | if ((ret = posix_memalign((void **)&cd->io_us_msgs, page_size, |
| 197 | io_us_msgs_size))) { |
| 198 | td_verror(td, ret, "posix_memalign"); |
| 199 | return ret; |
| 200 | } |
| 201 | if ((ret = rpma_mr_reg(ccd->peer, cd->io_us_msgs, io_us_msgs_size, |
| 202 | RPMA_MR_USAGE_SEND | RPMA_MR_USAGE_RECV, |
| 203 | &cd->msg_mr))) { |
| 204 | librpma_td_verror(td, ret, "rpma_mr_reg"); |
| 205 | return ret; |
| 206 | } |
| 207 | |
| 208 | return librpma_fio_client_post_init(td); |
| 209 | } |
| 210 | |
| 211 | static void client_cleanup(struct thread_data *td) |
| 212 | { |
| 213 | struct librpma_fio_client_data *ccd = td->io_ops_data; |
| 214 | struct client_data *cd; |
| 215 | size_t flush_req_size; |
| 216 | size_t io_u_buf_off; |
| 217 | size_t send_offset; |
| 218 | void *send_ptr; |
| 219 | int ret; |
| 220 | |
| 221 | if (ccd == NULL) |
| 222 | return; |
| 223 | |
| 224 | cd = ccd->client_data; |
| 225 | if (cd == NULL) { |
| 226 | librpma_fio_client_cleanup(td); |
| 227 | return; |
| 228 | } |
| 229 | |
| 230 | /* |
| 231 | * Make sure all SEND completions are collected ergo there are free |
| 232 | * slots in the SQ for the last SEND message. |
| 233 | * |
| 234 | * Note: If any operation will fail we still can send the termination |
| 235 | * notice. |
| 236 | */ |
| 237 | (void) librpma_fio_client_io_complete_all_sends(td); |
| 238 | |
| 239 | /* prepare the last flush message and pack it to the send buffer */ |
| 240 | flush_req_size = gpspm_flush_request__get_packed_size(&Flush_req_last); |
| 241 | if (flush_req_size > MAX_MSG_SIZE) { |
| 242 | log_err( |
| 243 | "Packed flush request size is bigger than available send buffer space (%zu > %d\n", |
| 244 | flush_req_size, MAX_MSG_SIZE); |
| 245 | } else { |
| 246 | io_u_buf_off = IO_U_NEXT_BUF_OFF_CLIENT(cd); |
| 247 | send_offset = io_u_buf_off + SEND_OFFSET; |
| 248 | send_ptr = cd->io_us_msgs + send_offset; |
| 249 | (void) gpspm_flush_request__pack(&Flush_req_last, send_ptr); |
| 250 | |
| 251 | /* send the flush message */ |
| 252 | if ((ret = rpma_send(ccd->conn, cd->msg_mr, send_offset, |
| 253 | flush_req_size, RPMA_F_COMPLETION_ALWAYS, |
| 254 | NULL))) |
| 255 | librpma_td_verror(td, ret, "rpma_send"); |
| 256 | |
| 257 | ++ccd->op_send_posted; |
| 258 | |
| 259 | /* Wait for the SEND to complete */ |
| 260 | (void) librpma_fio_client_io_complete_all_sends(td); |
| 261 | } |
| 262 | |
| 263 | /* deregister the messaging buffer memory */ |
| 264 | if ((ret = rpma_mr_dereg(&cd->msg_mr))) |
| 265 | librpma_td_verror(td, ret, "rpma_mr_dereg"); |
| 266 | |
| 267 | free(ccd->client_data); |
| 268 | |
| 269 | librpma_fio_client_cleanup(td); |
| 270 | } |
| 271 | |
| 272 | static inline int client_io_flush(struct thread_data *td, |
| 273 | struct io_u *first_io_u, struct io_u *last_io_u, |
| 274 | unsigned long long int len) |
| 275 | { |
| 276 | struct librpma_fio_client_data *ccd = td->io_ops_data; |
| 277 | struct client_data *cd = ccd->client_data; |
| 278 | size_t io_u_buf_off = IO_U_NEXT_BUF_OFF_CLIENT(cd); |
| 279 | size_t send_offset = io_u_buf_off + SEND_OFFSET; |
| 280 | size_t recv_offset = io_u_buf_off + RECV_OFFSET; |
| 281 | void *send_ptr = cd->io_us_msgs + send_offset; |
| 282 | void *recv_ptr = cd->io_us_msgs + recv_offset; |
| 283 | GPSPMFlushRequest flush_req = GPSPM_FLUSH_REQUEST__INIT; |
| 284 | size_t flush_req_size = 0; |
| 285 | int ret; |
| 286 | |
| 287 | /* prepare a response buffer */ |
| 288 | if ((ret = rpma_recv(ccd->conn, cd->msg_mr, recv_offset, MAX_MSG_SIZE, |
| 289 | recv_ptr))) { |
| 290 | librpma_td_verror(td, ret, "rpma_recv"); |
| 291 | return -1; |
| 292 | } |
| 293 | |
| 294 | /* prepare a flush message and pack it to a send buffer */ |
| 295 | flush_req.offset = first_io_u->offset; |
| 296 | flush_req.length = len; |
| 297 | flush_req.op_context = last_io_u->index; |
| 298 | flush_req_size = gpspm_flush_request__get_packed_size(&flush_req); |
| 299 | if (flush_req_size > MAX_MSG_SIZE) { |
| 300 | log_err( |
| 301 | "Packed flush request size is bigger than available send buffer space (%" |
| 302 | PRIu64 " > %d\n", flush_req_size, MAX_MSG_SIZE); |
| 303 | return -1; |
| 304 | } |
| 305 | (void) gpspm_flush_request__pack(&flush_req, send_ptr); |
| 306 | |
| 307 | /* send the flush message */ |
| 308 | if ((ret = rpma_send(ccd->conn, cd->msg_mr, send_offset, flush_req_size, |
| 309 | RPMA_F_COMPLETION_ALWAYS, NULL))) { |
| 310 | librpma_td_verror(td, ret, "rpma_send"); |
| 311 | return -1; |
| 312 | } |
| 313 | |
| 314 | ++ccd->op_send_posted; |
| 315 | |
| 316 | return 0; |
| 317 | } |
| 318 | |
| 319 | static int client_get_io_u_index(struct ibv_wc *wc, unsigned int *io_u_index) |
| 320 | { |
| 321 | GPSPMFlushResponse *flush_resp; |
| 322 | |
| 323 | if (wc->opcode != IBV_WC_RECV) |
| 324 | return 0; |
| 325 | |
| 326 | /* unpack a response from the received buffer */ |
| 327 | flush_resp = gpspm_flush_response__unpack(NULL, |
| 328 | wc->byte_len, (void *)wc->wr_id); |
| 329 | if (flush_resp == NULL) { |
| 330 | log_err("Cannot unpack the flush response buffer\n"); |
| 331 | return -1; |
| 332 | } |
| 333 | |
| 334 | memcpy(io_u_index, &flush_resp->op_context, sizeof(*io_u_index)); |
| 335 | |
| 336 | gpspm_flush_response__free_unpacked(flush_resp, NULL); |
| 337 | |
| 338 | return 1; |
| 339 | } |
| 340 | |
| 341 | FIO_STATIC struct ioengine_ops ioengine_client = { |
| 342 | .name = "librpma_gpspm_client", |
| 343 | .version = FIO_IOOPS_VERSION, |
| 344 | .init = client_init, |
| 345 | .post_init = client_post_init, |
| 346 | .get_file_size = librpma_fio_client_get_file_size, |
| 347 | .open_file = librpma_fio_file_nop, |
| 348 | .queue = librpma_fio_client_queue, |
| 349 | .commit = librpma_fio_client_commit, |
| 350 | .getevents = librpma_fio_client_getevents, |
| 351 | .event = librpma_fio_client_event, |
| 352 | .errdetails = librpma_fio_client_errdetails, |
| 353 | .close_file = librpma_fio_file_nop, |
| 354 | .cleanup = client_cleanup, |
| 355 | .flags = FIO_DISKLESSIO, |
| 356 | .options = librpma_fio_options, |
| 357 | .option_struct_size = sizeof(struct librpma_fio_options_values), |
| 358 | }; |
| 359 | |
| 360 | /* server side implementation */ |
| 361 | |
| 362 | #define IO_U_BUFF_OFF_SERVER(i) (i * IO_U_BUF_LEN) |
| 363 | |
| 364 | struct server_data { |
| 365 | /* aligned td->orig_buffer */ |
| 366 | char *orig_buffer_aligned; |
| 367 | |
| 368 | /* resources for messaging buffer from DRAM allocated by fio */ |
| 369 | struct rpma_mr_local *msg_mr; |
| 370 | |
| 371 | uint32_t msg_sqe_available; /* # of free SQ slots */ |
| 372 | |
| 373 | /* in-memory queues */ |
| 374 | struct ibv_wc *msgs_queued; |
| 375 | uint32_t msg_queued_nr; |
| 376 | }; |
| 377 | |
| 378 | static int server_init(struct thread_data *td) |
| 379 | { |
| 380 | struct librpma_fio_server_data *csd; |
| 381 | struct server_data *sd; |
| 382 | int ret = -1; |
| 383 | |
| 384 | if ((ret = librpma_fio_server_init(td))) |
| 385 | return ret; |
| 386 | |
| 387 | csd = td->io_ops_data; |
| 388 | |
| 389 | /* allocate server's data */ |
| 390 | sd = calloc(1, sizeof(*sd)); |
| 391 | if (sd == NULL) { |
| 392 | td_verror(td, errno, "calloc"); |
| 393 | goto err_server_cleanup; |
| 394 | } |
| 395 | |
| 396 | /* allocate in-memory queue */ |
| 397 | sd->msgs_queued = calloc(td->o.iodepth, sizeof(*sd->msgs_queued)); |
| 398 | if (sd->msgs_queued == NULL) { |
| 399 | td_verror(td, errno, "calloc"); |
| 400 | goto err_free_sd; |
| 401 | } |
| 402 | |
| 403 | /* |
| 404 | * Assure a single io_u buffer can store both SEND and RECV messages and |
| 405 | * an io_us buffer allocation is page-size-aligned which is required |
| 406 | * to register for RDMA. User-provided values are intentionally ignored. |
| 407 | */ |
| 408 | td->o.max_bs[DDIR_READ] = IO_U_BUF_LEN; |
| 409 | td->o.mem_align = page_size; |
| 410 | |
| 411 | csd->server_data = sd; |
| 412 | |
| 413 | return 0; |
| 414 | |
| 415 | err_free_sd: |
| 416 | free(sd); |
| 417 | |
| 418 | err_server_cleanup: |
| 419 | librpma_fio_server_cleanup(td); |
| 420 | |
| 421 | return -1; |
| 422 | } |
| 423 | |
| 424 | static int server_post_init(struct thread_data *td) |
| 425 | { |
| 426 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 427 | struct server_data *sd = csd->server_data; |
| 428 | size_t io_us_size; |
| 429 | size_t io_u_buflen; |
| 430 | int ret; |
| 431 | |
| 432 | /* |
| 433 | * td->orig_buffer is not aligned. The engine requires aligned io_us |
| 434 | * so FIO aligns up the address using the formula below. |
| 435 | */ |
| 436 | sd->orig_buffer_aligned = PTR_ALIGN(td->orig_buffer, page_mask) + |
| 437 | td->o.mem_align; |
| 438 | |
| 439 | /* |
| 440 | * XXX |
| 441 | * Each io_u message buffer contains recv and send messages. |
| 442 | * Aligning each of those buffers may potentially give |
| 443 | * some performance benefits. |
| 444 | */ |
| 445 | io_u_buflen = td_max_bs(td); |
| 446 | |
| 447 | /* check whether io_u buffer is big enough */ |
| 448 | if (io_u_buflen < IO_U_BUF_LEN) { |
| 449 | log_err( |
| 450 | "blocksize too small to accommodate assumed maximal request/response pair size (%" PRIu64 " < %d)\n", |
| 451 | io_u_buflen, IO_U_BUF_LEN); |
| 452 | return -1; |
| 453 | } |
| 454 | |
| 455 | /* |
| 456 | * td->orig_buffer_size beside the space really consumed by io_us |
| 457 | * has paddings which can be omitted for the memory registration. |
| 458 | */ |
| 459 | io_us_size = (unsigned long long)io_u_buflen * |
| 460 | (unsigned long long)td->o.iodepth; |
| 461 | |
| 462 | if ((ret = rpma_mr_reg(csd->peer, sd->orig_buffer_aligned, io_us_size, |
| 463 | RPMA_MR_USAGE_SEND | RPMA_MR_USAGE_RECV, |
| 464 | &sd->msg_mr))) { |
| 465 | librpma_td_verror(td, ret, "rpma_mr_reg"); |
| 466 | return -1; |
| 467 | } |
| 468 | |
| 469 | return 0; |
| 470 | } |
| 471 | |
| 472 | static void server_cleanup(struct thread_data *td) |
| 473 | { |
| 474 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 475 | struct server_data *sd; |
| 476 | int ret; |
| 477 | |
| 478 | if (csd == NULL) |
| 479 | return; |
| 480 | |
| 481 | sd = csd->server_data; |
| 482 | |
| 483 | if (sd != NULL) { |
| 484 | /* rpma_mr_dereg(messaging buffer from DRAM) */ |
| 485 | if ((ret = rpma_mr_dereg(&sd->msg_mr))) |
| 486 | librpma_td_verror(td, ret, "rpma_mr_dereg"); |
| 487 | |
| 488 | free(sd->msgs_queued); |
| 489 | free(sd); |
| 490 | } |
| 491 | |
| 492 | librpma_fio_server_cleanup(td); |
| 493 | } |
| 494 | |
| 495 | static int prepare_connection(struct thread_data *td, |
| 496 | struct rpma_conn_req *conn_req) |
| 497 | { |
| 498 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 499 | struct server_data *sd = csd->server_data; |
| 500 | int ret; |
| 501 | int i; |
| 502 | |
| 503 | /* prepare buffers for a flush requests */ |
| 504 | sd->msg_sqe_available = td->o.iodepth; |
| 505 | for (i = 0; i < td->o.iodepth; i++) { |
| 506 | size_t offset_recv_msg = IO_U_BUFF_OFF_SERVER(i) + RECV_OFFSET; |
| 507 | if ((ret = rpma_conn_req_recv(conn_req, sd->msg_mr, |
| 508 | offset_recv_msg, MAX_MSG_SIZE, |
| 509 | (const void *)(uintptr_t)i))) { |
| 510 | librpma_td_verror(td, ret, "rpma_conn_req_recv"); |
| 511 | return ret; |
| 512 | } |
| 513 | } |
| 514 | |
| 515 | return 0; |
| 516 | } |
| 517 | |
| 518 | static int server_open_file(struct thread_data *td, struct fio_file *f) |
| 519 | { |
| 520 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 521 | struct rpma_conn_cfg *cfg = NULL; |
| 522 | uint16_t max_msg_num = td->o.iodepth; |
| 523 | int ret; |
| 524 | |
| 525 | csd->prepare_connection = prepare_connection; |
| 526 | |
| 527 | /* create a connection configuration object */ |
| 528 | if ((ret = rpma_conn_cfg_new(&cfg))) { |
| 529 | librpma_td_verror(td, ret, "rpma_conn_cfg_new"); |
| 530 | return -1; |
| 531 | } |
| 532 | |
| 533 | /* |
| 534 | * Calculate the required queue sizes where: |
| 535 | * - the send queue (SQ) has to be big enough to accommodate |
| 536 | * all possible flush requests (SENDs) |
| 537 | * - the receive queue (RQ) has to be big enough to accommodate |
| 538 | * all flush responses (RECVs) |
| 539 | * - the completion queue (CQ) has to be big enough to accommodate |
| 540 | * all success and error completions (sq_size + rq_size) |
| 541 | */ |
| 542 | if ((ret = rpma_conn_cfg_set_sq_size(cfg, max_msg_num))) { |
| 543 | librpma_td_verror(td, ret, "rpma_conn_cfg_set_sq_size"); |
| 544 | goto err_cfg_delete; |
| 545 | } |
| 546 | if ((ret = rpma_conn_cfg_set_rq_size(cfg, max_msg_num))) { |
| 547 | librpma_td_verror(td, ret, "rpma_conn_cfg_set_rq_size"); |
| 548 | goto err_cfg_delete; |
| 549 | } |
| 550 | if ((ret = rpma_conn_cfg_set_cq_size(cfg, max_msg_num * 2))) { |
| 551 | librpma_td_verror(td, ret, "rpma_conn_cfg_set_cq_size"); |
| 552 | goto err_cfg_delete; |
| 553 | } |
| 554 | |
| 555 | ret = librpma_fio_server_open_file(td, f, cfg); |
| 556 | |
| 557 | err_cfg_delete: |
| 558 | (void) rpma_conn_cfg_delete(&cfg); |
| 559 | |
| 560 | return ret; |
| 561 | } |
| 562 | |
| 563 | static int server_qe_process(struct thread_data *td, struct ibv_wc *wc) |
| 564 | { |
| 565 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 566 | struct server_data *sd = csd->server_data; |
| 567 | GPSPMFlushRequest *flush_req; |
| 568 | GPSPMFlushResponse flush_resp = GPSPM_FLUSH_RESPONSE__INIT; |
| 569 | size_t flush_resp_size = 0; |
| 570 | size_t send_buff_offset; |
| 571 | size_t recv_buff_offset; |
| 572 | size_t io_u_buff_offset; |
| 573 | void *send_buff_ptr; |
| 574 | void *recv_buff_ptr; |
| 575 | void *op_ptr; |
| 576 | int msg_index; |
| 577 | int ret; |
| 578 | |
| 579 | /* calculate SEND/RECV pair parameters */ |
| 580 | msg_index = (int)(uintptr_t)wc->wr_id; |
| 581 | io_u_buff_offset = IO_U_BUFF_OFF_SERVER(msg_index); |
| 582 | send_buff_offset = io_u_buff_offset + SEND_OFFSET; |
| 583 | recv_buff_offset = io_u_buff_offset + RECV_OFFSET; |
| 584 | send_buff_ptr = sd->orig_buffer_aligned + send_buff_offset; |
| 585 | recv_buff_ptr = sd->orig_buffer_aligned + recv_buff_offset; |
| 586 | |
| 587 | /* unpack a flush request from the received buffer */ |
| 588 | flush_req = gpspm_flush_request__unpack(NULL, wc->byte_len, |
| 589 | recv_buff_ptr); |
| 590 | if (flush_req == NULL) { |
| 591 | log_err("cannot unpack the flush request buffer\n"); |
| 592 | goto err_terminate; |
| 593 | } |
| 594 | |
| 595 | if (IS_NOT_THE_LAST_MESSAGE(flush_req)) { |
| 596 | op_ptr = csd->ws_ptr + flush_req->offset; |
| 597 | pmem_persist(op_ptr, flush_req->length); |
| 598 | } else { |
| 599 | /* |
| 600 | * This is the last message - the client is done. |
| 601 | */ |
| 602 | gpspm_flush_request__free_unpacked(flush_req, NULL); |
| 603 | td->done = true; |
| 604 | return 0; |
| 605 | } |
| 606 | |
| 607 | /* initiate the next receive operation */ |
| 608 | if ((ret = rpma_recv(csd->conn, sd->msg_mr, recv_buff_offset, |
| 609 | MAX_MSG_SIZE, |
| 610 | (const void *)(uintptr_t)msg_index))) { |
| 611 | librpma_td_verror(td, ret, "rpma_recv"); |
| 612 | goto err_free_unpacked; |
| 613 | } |
| 614 | |
| 615 | /* prepare a flush response and pack it to a send buffer */ |
| 616 | flush_resp.op_context = flush_req->op_context; |
| 617 | flush_resp_size = gpspm_flush_response__get_packed_size(&flush_resp); |
| 618 | if (flush_resp_size > MAX_MSG_SIZE) { |
| 619 | log_err( |
| 620 | "Size of the packed flush response is bigger than the available space of the send buffer (%" |
| 621 | PRIu64 " > %i\n", flush_resp_size, MAX_MSG_SIZE); |
| 622 | goto err_free_unpacked; |
| 623 | } |
| 624 | |
| 625 | (void) gpspm_flush_response__pack(&flush_resp, send_buff_ptr); |
| 626 | |
| 627 | /* send the flush response */ |
| 628 | if ((ret = rpma_send(csd->conn, sd->msg_mr, send_buff_offset, |
| 629 | flush_resp_size, RPMA_F_COMPLETION_ALWAYS, NULL))) { |
| 630 | librpma_td_verror(td, ret, "rpma_send"); |
| 631 | goto err_free_unpacked; |
| 632 | } |
| 633 | --sd->msg_sqe_available; |
| 634 | |
| 635 | gpspm_flush_request__free_unpacked(flush_req, NULL); |
| 636 | |
| 637 | return 0; |
| 638 | |
| 639 | err_free_unpacked: |
| 640 | gpspm_flush_request__free_unpacked(flush_req, NULL); |
| 641 | |
| 642 | err_terminate: |
| 643 | td->terminate = true; |
| 644 | |
| 645 | return -1; |
| 646 | } |
| 647 | |
| 648 | static inline int server_queue_process(struct thread_data *td) |
| 649 | { |
| 650 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 651 | struct server_data *sd = csd->server_data; |
| 652 | int ret; |
| 653 | int i; |
| 654 | |
| 655 | /* min(# of queue entries, # of SQ entries available) */ |
| 656 | uint32_t qes_to_process = min(sd->msg_queued_nr, sd->msg_sqe_available); |
| 657 | if (qes_to_process == 0) |
| 658 | return 0; |
| 659 | |
| 660 | /* process queued completions */ |
| 661 | for (i = 0; i < qes_to_process; ++i) { |
| 662 | if ((ret = server_qe_process(td, &sd->msgs_queued[i]))) |
| 663 | return ret; |
| 664 | } |
| 665 | |
| 666 | /* progress the queue */ |
| 667 | for (i = 0; i < sd->msg_queued_nr - qes_to_process; ++i) { |
| 668 | memcpy(&sd->msgs_queued[i], |
| 669 | &sd->msgs_queued[qes_to_process + i], |
| 670 | sizeof(sd->msgs_queued[i])); |
| 671 | } |
| 672 | |
| 673 | sd->msg_queued_nr -= qes_to_process; |
| 674 | |
| 675 | return 0; |
| 676 | } |
| 677 | |
| 678 | static int server_cmpl_process(struct thread_data *td) |
| 679 | { |
| 680 | struct librpma_fio_server_data *csd = td->io_ops_data; |
| 681 | struct server_data *sd = csd->server_data; |
| 682 | struct ibv_wc *wc = &sd->msgs_queued[sd->msg_queued_nr]; |
| 683 | struct librpma_fio_options_values *o = td->eo; |
| 684 | int ret; |
| 685 | |
| 686 | ret = rpma_cq_get_wc(csd->cq, 1, wc, NULL); |
| 687 | if (ret == RPMA_E_NO_COMPLETION) { |
| 688 | if (o->busy_wait_polling == 0) { |
| 689 | ret = rpma_cq_wait(csd->cq); |
| 690 | if (ret == RPMA_E_NO_COMPLETION) { |
| 691 | /* lack of completion is not an error */ |
| 692 | return 0; |
| 693 | } else if (ret != 0) { |
| 694 | librpma_td_verror(td, ret, "rpma_cq_wait"); |
| 695 | goto err_terminate; |
| 696 | } |
| 697 | |
| 698 | ret = rpma_cq_get_wc(csd->cq, 1, wc, NULL); |
| 699 | if (ret == RPMA_E_NO_COMPLETION) { |
| 700 | /* lack of completion is not an error */ |
| 701 | return 0; |
| 702 | } else if (ret != 0) { |
| 703 | librpma_td_verror(td, ret, "rpma_cq_get_wc"); |
| 704 | goto err_terminate; |
| 705 | } |
| 706 | } else { |
| 707 | /* lack of completion is not an error */ |
| 708 | return 0; |
| 709 | } |
| 710 | } else if (ret != 0) { |
| 711 | librpma_td_verror(td, ret, "rpma_cq_get_wc"); |
| 712 | goto err_terminate; |
| 713 | } |
| 714 | |
| 715 | /* validate the completion */ |
| 716 | if (wc->status != IBV_WC_SUCCESS) |
| 717 | goto err_terminate; |
| 718 | |
| 719 | if (wc->opcode == IBV_WC_RECV) |
| 720 | ++sd->msg_queued_nr; |
| 721 | else if (wc->opcode == IBV_WC_SEND) |
| 722 | ++sd->msg_sqe_available; |
| 723 | |
| 724 | return 0; |
| 725 | |
| 726 | err_terminate: |
| 727 | td->terminate = true; |
| 728 | |
| 729 | return -1; |
| 730 | } |
| 731 | |
| 732 | static enum fio_q_status server_queue(struct thread_data *td, struct io_u *io_u) |
| 733 | { |
| 734 | do { |
| 735 | if (server_cmpl_process(td)) |
| 736 | return FIO_Q_BUSY; |
| 737 | |
| 738 | if (server_queue_process(td)) |
| 739 | return FIO_Q_BUSY; |
| 740 | |
| 741 | } while (!td->done); |
| 742 | |
| 743 | return FIO_Q_COMPLETED; |
| 744 | } |
| 745 | |
| 746 | FIO_STATIC struct ioengine_ops ioengine_server = { |
| 747 | .name = "librpma_gpspm_server", |
| 748 | .version = FIO_IOOPS_VERSION, |
| 749 | .init = server_init, |
| 750 | .post_init = server_post_init, |
| 751 | .open_file = server_open_file, |
| 752 | .close_file = librpma_fio_server_close_file, |
| 753 | .queue = server_queue, |
| 754 | .invalidate = librpma_fio_file_nop, |
| 755 | .cleanup = server_cleanup, |
| 756 | .flags = FIO_SYNCIO, |
| 757 | .options = librpma_fio_options, |
| 758 | .option_struct_size = sizeof(struct librpma_fio_options_values), |
| 759 | }; |
| 760 | |
| 761 | /* register both engines */ |
| 762 | |
| 763 | static void fio_init fio_librpma_gpspm_register(void) |
| 764 | { |
| 765 | register_ioengine(&ioengine_client); |
| 766 | register_ioengine(&ioengine_server); |
| 767 | } |
| 768 | |
| 769 | static void fio_exit fio_librpma_gpspm_unregister(void) |
| 770 | { |
| 771 | unregister_ioengine(&ioengine_client); |
| 772 | unregister_ioengine(&ioengine_server); |
| 773 | } |