Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 1 | /* |
| 2 | * RDMA protocol and interfaces |
| 3 | * |
| 4 | * Copyright IBM, Corp. 2010-2013 |
| 5 | * |
| 6 | * Authors: |
| 7 | * Michael R. Hines <mrhines@us.ibm.com> |
| 8 | * Jiuxing Liu <jl@us.ibm.com> |
| 9 | * |
| 10 | * This work is licensed under the terms of the GNU GPL, version 2 or |
| 11 | * later. See the COPYING file in the top-level directory. |
| 12 | * |
| 13 | */ |
| 14 | #include "qemu-common.h" |
| 15 | #include "migration/migration.h" |
| 16 | #include "migration/qemu-file.h" |
| 17 | #include "exec/cpu-common.h" |
| 18 | #include "qemu/main-loop.h" |
| 19 | #include "qemu/sockets.h" |
| 20 | #include "qemu/bitmap.h" |
| 21 | #include "block/coroutine.h" |
| 22 | #include <stdio.h> |
| 23 | #include <sys/types.h> |
| 24 | #include <sys/socket.h> |
| 25 | #include <netdb.h> |
| 26 | #include <arpa/inet.h> |
| 27 | #include <string.h> |
| 28 | #include <rdma/rdma_cma.h> |
| 29 | |
Michael R. Hines | 8cd31ad | 2013-08-03 22:54:49 -0400 | [diff] [blame] | 30 | //#define DEBUG_RDMA |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 31 | //#define DEBUG_RDMA_VERBOSE |
| 32 | //#define DEBUG_RDMA_REALLY_VERBOSE |
| 33 | |
| 34 | #ifdef DEBUG_RDMA |
| 35 | #define DPRINTF(fmt, ...) \ |
| 36 | do { printf("rdma: " fmt, ## __VA_ARGS__); } while (0) |
| 37 | #else |
| 38 | #define DPRINTF(fmt, ...) \ |
| 39 | do { } while (0) |
| 40 | #endif |
| 41 | |
| 42 | #ifdef DEBUG_RDMA_VERBOSE |
| 43 | #define DDPRINTF(fmt, ...) \ |
| 44 | do { printf("rdma: " fmt, ## __VA_ARGS__); } while (0) |
| 45 | #else |
| 46 | #define DDPRINTF(fmt, ...) \ |
| 47 | do { } while (0) |
| 48 | #endif |
| 49 | |
| 50 | #ifdef DEBUG_RDMA_REALLY_VERBOSE |
| 51 | #define DDDPRINTF(fmt, ...) \ |
| 52 | do { printf("rdma: " fmt, ## __VA_ARGS__); } while (0) |
| 53 | #else |
| 54 | #define DDDPRINTF(fmt, ...) \ |
| 55 | do { } while (0) |
| 56 | #endif |
| 57 | |
| 58 | /* |
| 59 | * Print and error on both the Monitor and the Log file. |
| 60 | */ |
| 61 | #define ERROR(errp, fmt, ...) \ |
| 62 | do { \ |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 63 | fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \ |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 64 | if (errp && (*(errp) == NULL)) { \ |
| 65 | error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \ |
| 66 | } \ |
| 67 | } while (0) |
| 68 | |
| 69 | #define RDMA_RESOLVE_TIMEOUT_MS 10000 |
| 70 | |
| 71 | /* Do not merge data if larger than this. */ |
| 72 | #define RDMA_MERGE_MAX (2 * 1024 * 1024) |
| 73 | #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096) |
| 74 | |
| 75 | #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */ |
| 76 | |
| 77 | /* |
| 78 | * This is only for non-live state being migrated. |
| 79 | * Instead of RDMA_WRITE messages, we use RDMA_SEND |
| 80 | * messages for that state, which requires a different |
| 81 | * delivery design than main memory. |
| 82 | */ |
| 83 | #define RDMA_SEND_INCREMENT 32768 |
| 84 | |
| 85 | /* |
| 86 | * Maximum size infiniband SEND message |
| 87 | */ |
| 88 | #define RDMA_CONTROL_MAX_BUFFER (512 * 1024) |
| 89 | #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096 |
| 90 | |
| 91 | #define RDMA_CONTROL_VERSION_CURRENT 1 |
| 92 | /* |
| 93 | * Capabilities for negotiation. |
| 94 | */ |
| 95 | #define RDMA_CAPABILITY_PIN_ALL 0x01 |
| 96 | |
| 97 | /* |
| 98 | * Add the other flags above to this list of known capabilities |
| 99 | * as they are introduced. |
| 100 | */ |
| 101 | static uint32_t known_capabilities = RDMA_CAPABILITY_PIN_ALL; |
| 102 | |
| 103 | #define CHECK_ERROR_STATE() \ |
| 104 | do { \ |
| 105 | if (rdma->error_state) { \ |
| 106 | if (!rdma->error_reported) { \ |
| 107 | fprintf(stderr, "RDMA is in an error state waiting migration" \ |
| 108 | " to abort!\n"); \ |
| 109 | rdma->error_reported = 1; \ |
| 110 | } \ |
| 111 | return rdma->error_state; \ |
| 112 | } \ |
| 113 | } while (0); |
| 114 | |
| 115 | /* |
| 116 | * A work request ID is 64-bits and we split up these bits |
| 117 | * into 3 parts: |
| 118 | * |
| 119 | * bits 0-15 : type of control message, 2^16 |
| 120 | * bits 16-29: ram block index, 2^14 |
| 121 | * bits 30-63: ram block chunk number, 2^34 |
| 122 | * |
| 123 | * The last two bit ranges are only used for RDMA writes, |
| 124 | * in order to track their completion and potentially |
| 125 | * also track unregistration status of the message. |
| 126 | */ |
| 127 | #define RDMA_WRID_TYPE_SHIFT 0UL |
| 128 | #define RDMA_WRID_BLOCK_SHIFT 16UL |
| 129 | #define RDMA_WRID_CHUNK_SHIFT 30UL |
| 130 | |
| 131 | #define RDMA_WRID_TYPE_MASK \ |
| 132 | ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL) |
| 133 | |
| 134 | #define RDMA_WRID_BLOCK_MASK \ |
| 135 | (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL)) |
| 136 | |
| 137 | #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK) |
| 138 | |
| 139 | /* |
| 140 | * RDMA migration protocol: |
| 141 | * 1. RDMA Writes (data messages, i.e. RAM) |
| 142 | * 2. IB Send/Recv (control channel messages) |
| 143 | */ |
| 144 | enum { |
| 145 | RDMA_WRID_NONE = 0, |
| 146 | RDMA_WRID_RDMA_WRITE = 1, |
| 147 | RDMA_WRID_SEND_CONTROL = 2000, |
| 148 | RDMA_WRID_RECV_CONTROL = 4000, |
| 149 | }; |
| 150 | |
| 151 | const char *wrid_desc[] = { |
| 152 | [RDMA_WRID_NONE] = "NONE", |
| 153 | [RDMA_WRID_RDMA_WRITE] = "WRITE RDMA", |
| 154 | [RDMA_WRID_SEND_CONTROL] = "CONTROL SEND", |
| 155 | [RDMA_WRID_RECV_CONTROL] = "CONTROL RECV", |
| 156 | }; |
| 157 | |
| 158 | /* |
| 159 | * Work request IDs for IB SEND messages only (not RDMA writes). |
| 160 | * This is used by the migration protocol to transmit |
| 161 | * control messages (such as device state and registration commands) |
| 162 | * |
| 163 | * We could use more WRs, but we have enough for now. |
| 164 | */ |
| 165 | enum { |
| 166 | RDMA_WRID_READY = 0, |
| 167 | RDMA_WRID_DATA, |
| 168 | RDMA_WRID_CONTROL, |
| 169 | RDMA_WRID_MAX, |
| 170 | }; |
| 171 | |
| 172 | /* |
| 173 | * SEND/RECV IB Control Messages. |
| 174 | */ |
| 175 | enum { |
| 176 | RDMA_CONTROL_NONE = 0, |
| 177 | RDMA_CONTROL_ERROR, |
| 178 | RDMA_CONTROL_READY, /* ready to receive */ |
| 179 | RDMA_CONTROL_QEMU_FILE, /* QEMUFile-transmitted bytes */ |
| 180 | RDMA_CONTROL_RAM_BLOCKS_REQUEST, /* RAMBlock synchronization */ |
| 181 | RDMA_CONTROL_RAM_BLOCKS_RESULT, /* RAMBlock synchronization */ |
| 182 | RDMA_CONTROL_COMPRESS, /* page contains repeat values */ |
| 183 | RDMA_CONTROL_REGISTER_REQUEST, /* dynamic page registration */ |
| 184 | RDMA_CONTROL_REGISTER_RESULT, /* key to use after registration */ |
| 185 | RDMA_CONTROL_REGISTER_FINISHED, /* current iteration finished */ |
| 186 | RDMA_CONTROL_UNREGISTER_REQUEST, /* dynamic UN-registration */ |
| 187 | RDMA_CONTROL_UNREGISTER_FINISHED, /* unpinning finished */ |
| 188 | }; |
| 189 | |
| 190 | const char *control_desc[] = { |
| 191 | [RDMA_CONTROL_NONE] = "NONE", |
| 192 | [RDMA_CONTROL_ERROR] = "ERROR", |
| 193 | [RDMA_CONTROL_READY] = "READY", |
| 194 | [RDMA_CONTROL_QEMU_FILE] = "QEMU FILE", |
| 195 | [RDMA_CONTROL_RAM_BLOCKS_REQUEST] = "RAM BLOCKS REQUEST", |
| 196 | [RDMA_CONTROL_RAM_BLOCKS_RESULT] = "RAM BLOCKS RESULT", |
| 197 | [RDMA_CONTROL_COMPRESS] = "COMPRESS", |
| 198 | [RDMA_CONTROL_REGISTER_REQUEST] = "REGISTER REQUEST", |
| 199 | [RDMA_CONTROL_REGISTER_RESULT] = "REGISTER RESULT", |
| 200 | [RDMA_CONTROL_REGISTER_FINISHED] = "REGISTER FINISHED", |
| 201 | [RDMA_CONTROL_UNREGISTER_REQUEST] = "UNREGISTER REQUEST", |
| 202 | [RDMA_CONTROL_UNREGISTER_FINISHED] = "UNREGISTER FINISHED", |
| 203 | }; |
| 204 | |
| 205 | /* |
| 206 | * Memory and MR structures used to represent an IB Send/Recv work request. |
| 207 | * This is *not* used for RDMA writes, only IB Send/Recv. |
| 208 | */ |
| 209 | typedef struct { |
| 210 | uint8_t control[RDMA_CONTROL_MAX_BUFFER]; /* actual buffer to register */ |
| 211 | struct ibv_mr *control_mr; /* registration metadata */ |
| 212 | size_t control_len; /* length of the message */ |
| 213 | uint8_t *control_curr; /* start of unconsumed bytes */ |
| 214 | } RDMAWorkRequestData; |
| 215 | |
| 216 | /* |
| 217 | * Negotiate RDMA capabilities during connection-setup time. |
| 218 | */ |
| 219 | typedef struct { |
| 220 | uint32_t version; |
| 221 | uint32_t flags; |
| 222 | } RDMACapabilities; |
| 223 | |
| 224 | static void caps_to_network(RDMACapabilities *cap) |
| 225 | { |
| 226 | cap->version = htonl(cap->version); |
| 227 | cap->flags = htonl(cap->flags); |
| 228 | } |
| 229 | |
| 230 | static void network_to_caps(RDMACapabilities *cap) |
| 231 | { |
| 232 | cap->version = ntohl(cap->version); |
| 233 | cap->flags = ntohl(cap->flags); |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * Representation of a RAMBlock from an RDMA perspective. |
| 238 | * This is not transmitted, only local. |
| 239 | * This and subsequent structures cannot be linked lists |
| 240 | * because we're using a single IB message to transmit |
| 241 | * the information. It's small anyway, so a list is overkill. |
| 242 | */ |
| 243 | typedef struct RDMALocalBlock { |
| 244 | uint8_t *local_host_addr; /* local virtual address */ |
| 245 | uint64_t remote_host_addr; /* remote virtual address */ |
| 246 | uint64_t offset; |
| 247 | uint64_t length; |
| 248 | struct ibv_mr **pmr; /* MRs for chunk-level registration */ |
| 249 | struct ibv_mr *mr; /* MR for non-chunk-level registration */ |
| 250 | uint32_t *remote_keys; /* rkeys for chunk-level registration */ |
| 251 | uint32_t remote_rkey; /* rkeys for non-chunk-level registration */ |
| 252 | int index; /* which block are we */ |
| 253 | bool is_ram_block; |
| 254 | int nb_chunks; |
| 255 | unsigned long *transit_bitmap; |
| 256 | unsigned long *unregister_bitmap; |
| 257 | } RDMALocalBlock; |
| 258 | |
| 259 | /* |
| 260 | * Also represents a RAMblock, but only on the dest. |
| 261 | * This gets transmitted by the dest during connection-time |
| 262 | * to the source VM and then is used to populate the |
| 263 | * corresponding RDMALocalBlock with |
| 264 | * the information needed to perform the actual RDMA. |
| 265 | */ |
| 266 | typedef struct QEMU_PACKED RDMARemoteBlock { |
| 267 | uint64_t remote_host_addr; |
| 268 | uint64_t offset; |
| 269 | uint64_t length; |
| 270 | uint32_t remote_rkey; |
| 271 | uint32_t padding; |
| 272 | } RDMARemoteBlock; |
| 273 | |
| 274 | static uint64_t htonll(uint64_t v) |
| 275 | { |
| 276 | union { uint32_t lv[2]; uint64_t llv; } u; |
| 277 | u.lv[0] = htonl(v >> 32); |
| 278 | u.lv[1] = htonl(v & 0xFFFFFFFFULL); |
| 279 | return u.llv; |
| 280 | } |
| 281 | |
| 282 | static uint64_t ntohll(uint64_t v) { |
| 283 | union { uint32_t lv[2]; uint64_t llv; } u; |
| 284 | u.llv = v; |
| 285 | return ((uint64_t)ntohl(u.lv[0]) << 32) | (uint64_t) ntohl(u.lv[1]); |
| 286 | } |
| 287 | |
| 288 | static void remote_block_to_network(RDMARemoteBlock *rb) |
| 289 | { |
| 290 | rb->remote_host_addr = htonll(rb->remote_host_addr); |
| 291 | rb->offset = htonll(rb->offset); |
| 292 | rb->length = htonll(rb->length); |
| 293 | rb->remote_rkey = htonl(rb->remote_rkey); |
| 294 | } |
| 295 | |
| 296 | static void network_to_remote_block(RDMARemoteBlock *rb) |
| 297 | { |
| 298 | rb->remote_host_addr = ntohll(rb->remote_host_addr); |
| 299 | rb->offset = ntohll(rb->offset); |
| 300 | rb->length = ntohll(rb->length); |
| 301 | rb->remote_rkey = ntohl(rb->remote_rkey); |
| 302 | } |
| 303 | |
| 304 | /* |
| 305 | * Virtual address of the above structures used for transmitting |
| 306 | * the RAMBlock descriptions at connection-time. |
| 307 | * This structure is *not* transmitted. |
| 308 | */ |
| 309 | typedef struct RDMALocalBlocks { |
| 310 | int nb_blocks; |
| 311 | bool init; /* main memory init complete */ |
| 312 | RDMALocalBlock *block; |
| 313 | } RDMALocalBlocks; |
| 314 | |
| 315 | /* |
| 316 | * Main data structure for RDMA state. |
| 317 | * While there is only one copy of this structure being allocated right now, |
| 318 | * this is the place where one would start if you wanted to consider |
| 319 | * having more than one RDMA connection open at the same time. |
| 320 | */ |
| 321 | typedef struct RDMAContext { |
| 322 | char *host; |
| 323 | int port; |
| 324 | |
Isaku Yamahata | 1f22364 | 2013-08-03 22:54:52 -0400 | [diff] [blame] | 325 | RDMAWorkRequestData wr_data[RDMA_WRID_MAX]; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 326 | |
| 327 | /* |
| 328 | * This is used by *_exchange_send() to figure out whether or not |
| 329 | * the initial "READY" message has already been received or not. |
| 330 | * This is because other functions may potentially poll() and detect |
| 331 | * the READY message before send() does, in which case we need to |
| 332 | * know if it completed. |
| 333 | */ |
| 334 | int control_ready_expected; |
| 335 | |
| 336 | /* number of outstanding writes */ |
| 337 | int nb_sent; |
| 338 | |
| 339 | /* store info about current buffer so that we can |
| 340 | merge it with future sends */ |
| 341 | uint64_t current_addr; |
| 342 | uint64_t current_length; |
| 343 | /* index of ram block the current buffer belongs to */ |
| 344 | int current_index; |
| 345 | /* index of the chunk in the current ram block */ |
| 346 | int current_chunk; |
| 347 | |
| 348 | bool pin_all; |
| 349 | |
| 350 | /* |
| 351 | * infiniband-specific variables for opening the device |
| 352 | * and maintaining connection state and so forth. |
| 353 | * |
| 354 | * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in |
| 355 | * cm_id->verbs, cm_id->channel, and cm_id->qp. |
| 356 | */ |
| 357 | struct rdma_cm_id *cm_id; /* connection manager ID */ |
| 358 | struct rdma_cm_id *listen_id; |
| 359 | |
| 360 | struct ibv_context *verbs; |
| 361 | struct rdma_event_channel *channel; |
| 362 | struct ibv_qp *qp; /* queue pair */ |
| 363 | struct ibv_comp_channel *comp_channel; /* completion channel */ |
| 364 | struct ibv_pd *pd; /* protection domain */ |
| 365 | struct ibv_cq *cq; /* completion queue */ |
| 366 | |
| 367 | /* |
| 368 | * If a previous write failed (perhaps because of a failed |
| 369 | * memory registration, then do not attempt any future work |
| 370 | * and remember the error state. |
| 371 | */ |
| 372 | int error_state; |
| 373 | int error_reported; |
| 374 | |
| 375 | /* |
| 376 | * Description of ram blocks used throughout the code. |
| 377 | */ |
| 378 | RDMALocalBlocks local_ram_blocks; |
| 379 | RDMARemoteBlock *block; |
| 380 | |
| 381 | /* |
| 382 | * Migration on *destination* started. |
| 383 | * Then use coroutine yield function. |
| 384 | * Source runs in a thread, so we don't care. |
| 385 | */ |
| 386 | int migration_started_on_destination; |
| 387 | |
| 388 | int total_registrations; |
| 389 | int total_writes; |
| 390 | |
| 391 | int unregister_current, unregister_next; |
| 392 | uint64_t unregistrations[RDMA_SIGNALED_SEND_MAX]; |
| 393 | |
| 394 | GHashTable *blockmap; |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 395 | bool ipv6; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 396 | } RDMAContext; |
| 397 | |
| 398 | /* |
| 399 | * Interface to the rest of the migration call stack. |
| 400 | */ |
| 401 | typedef struct QEMUFileRDMA { |
| 402 | RDMAContext *rdma; |
| 403 | size_t len; |
| 404 | void *file; |
| 405 | } QEMUFileRDMA; |
| 406 | |
| 407 | /* |
| 408 | * Main structure for IB Send/Recv control messages. |
| 409 | * This gets prepended at the beginning of every Send/Recv. |
| 410 | */ |
| 411 | typedef struct QEMU_PACKED { |
| 412 | uint32_t len; /* Total length of data portion */ |
| 413 | uint32_t type; /* which control command to perform */ |
| 414 | uint32_t repeat; /* number of commands in data portion of same type */ |
| 415 | uint32_t padding; |
| 416 | } RDMAControlHeader; |
| 417 | |
| 418 | static void control_to_network(RDMAControlHeader *control) |
| 419 | { |
| 420 | control->type = htonl(control->type); |
| 421 | control->len = htonl(control->len); |
| 422 | control->repeat = htonl(control->repeat); |
| 423 | } |
| 424 | |
| 425 | static void network_to_control(RDMAControlHeader *control) |
| 426 | { |
| 427 | control->type = ntohl(control->type); |
| 428 | control->len = ntohl(control->len); |
| 429 | control->repeat = ntohl(control->repeat); |
| 430 | } |
| 431 | |
| 432 | /* |
| 433 | * Register a single Chunk. |
| 434 | * Information sent by the source VM to inform the dest |
| 435 | * to register an single chunk of memory before we can perform |
| 436 | * the actual RDMA operation. |
| 437 | */ |
| 438 | typedef struct QEMU_PACKED { |
| 439 | union QEMU_PACKED { |
| 440 | uint64_t current_addr; /* offset into the ramblock of the chunk */ |
| 441 | uint64_t chunk; /* chunk to lookup if unregistering */ |
| 442 | } key; |
| 443 | uint32_t current_index; /* which ramblock the chunk belongs to */ |
| 444 | uint32_t padding; |
| 445 | uint64_t chunks; /* how many sequential chunks to register */ |
| 446 | } RDMARegister; |
| 447 | |
| 448 | static void register_to_network(RDMARegister *reg) |
| 449 | { |
| 450 | reg->key.current_addr = htonll(reg->key.current_addr); |
| 451 | reg->current_index = htonl(reg->current_index); |
| 452 | reg->chunks = htonll(reg->chunks); |
| 453 | } |
| 454 | |
| 455 | static void network_to_register(RDMARegister *reg) |
| 456 | { |
| 457 | reg->key.current_addr = ntohll(reg->key.current_addr); |
| 458 | reg->current_index = ntohl(reg->current_index); |
| 459 | reg->chunks = ntohll(reg->chunks); |
| 460 | } |
| 461 | |
| 462 | typedef struct QEMU_PACKED { |
| 463 | uint32_t value; /* if zero, we will madvise() */ |
| 464 | uint32_t block_idx; /* which ram block index */ |
| 465 | uint64_t offset; /* where in the remote ramblock this chunk */ |
| 466 | uint64_t length; /* length of the chunk */ |
| 467 | } RDMACompress; |
| 468 | |
| 469 | static void compress_to_network(RDMACompress *comp) |
| 470 | { |
| 471 | comp->value = htonl(comp->value); |
| 472 | comp->block_idx = htonl(comp->block_idx); |
| 473 | comp->offset = htonll(comp->offset); |
| 474 | comp->length = htonll(comp->length); |
| 475 | } |
| 476 | |
| 477 | static void network_to_compress(RDMACompress *comp) |
| 478 | { |
| 479 | comp->value = ntohl(comp->value); |
| 480 | comp->block_idx = ntohl(comp->block_idx); |
| 481 | comp->offset = ntohll(comp->offset); |
| 482 | comp->length = ntohll(comp->length); |
| 483 | } |
| 484 | |
| 485 | /* |
| 486 | * The result of the dest's memory registration produces an "rkey" |
| 487 | * which the source VM must reference in order to perform |
| 488 | * the RDMA operation. |
| 489 | */ |
| 490 | typedef struct QEMU_PACKED { |
| 491 | uint32_t rkey; |
| 492 | uint32_t padding; |
| 493 | uint64_t host_addr; |
| 494 | } RDMARegisterResult; |
| 495 | |
| 496 | static void result_to_network(RDMARegisterResult *result) |
| 497 | { |
| 498 | result->rkey = htonl(result->rkey); |
| 499 | result->host_addr = htonll(result->host_addr); |
| 500 | }; |
| 501 | |
| 502 | static void network_to_result(RDMARegisterResult *result) |
| 503 | { |
| 504 | result->rkey = ntohl(result->rkey); |
| 505 | result->host_addr = ntohll(result->host_addr); |
| 506 | }; |
| 507 | |
| 508 | const char *print_wrid(int wrid); |
| 509 | static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, |
| 510 | uint8_t *data, RDMAControlHeader *resp, |
| 511 | int *resp_idx, |
| 512 | int (*callback)(RDMAContext *rdma)); |
| 513 | |
| 514 | static inline uint64_t ram_chunk_index(uint8_t *start, uint8_t *host) |
| 515 | { |
| 516 | return ((uintptr_t) host - (uintptr_t) start) >> RDMA_REG_CHUNK_SHIFT; |
| 517 | } |
| 518 | |
| 519 | static inline uint8_t *ram_chunk_start(RDMALocalBlock *rdma_ram_block, |
| 520 | uint64_t i) |
| 521 | { |
| 522 | return (uint8_t *) (((uintptr_t) rdma_ram_block->local_host_addr) |
| 523 | + (i << RDMA_REG_CHUNK_SHIFT)); |
| 524 | } |
| 525 | |
| 526 | static inline uint8_t *ram_chunk_end(RDMALocalBlock *rdma_ram_block, uint64_t i) |
| 527 | { |
| 528 | uint8_t *result = ram_chunk_start(rdma_ram_block, i) + |
| 529 | (1UL << RDMA_REG_CHUNK_SHIFT); |
| 530 | |
| 531 | if (result > (rdma_ram_block->local_host_addr + rdma_ram_block->length)) { |
| 532 | result = rdma_ram_block->local_host_addr + rdma_ram_block->length; |
| 533 | } |
| 534 | |
| 535 | return result; |
| 536 | } |
| 537 | |
| 538 | static int __qemu_rdma_add_block(RDMAContext *rdma, void *host_addr, |
| 539 | ram_addr_t block_offset, uint64_t length) |
| 540 | { |
| 541 | RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 542 | RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, |
| 543 | (void *) block_offset); |
| 544 | RDMALocalBlock *old = local->block; |
| 545 | |
| 546 | assert(block == NULL); |
| 547 | |
| 548 | local->block = g_malloc0(sizeof(RDMALocalBlock) * (local->nb_blocks + 1)); |
| 549 | |
| 550 | if (local->nb_blocks) { |
| 551 | int x; |
| 552 | |
| 553 | for (x = 0; x < local->nb_blocks; x++) { |
| 554 | g_hash_table_remove(rdma->blockmap, (void *)old[x].offset); |
| 555 | g_hash_table_insert(rdma->blockmap, (void *)old[x].offset, |
| 556 | &local->block[x]); |
| 557 | } |
| 558 | memcpy(local->block, old, sizeof(RDMALocalBlock) * local->nb_blocks); |
| 559 | g_free(old); |
| 560 | } |
| 561 | |
| 562 | block = &local->block[local->nb_blocks]; |
| 563 | |
| 564 | block->local_host_addr = host_addr; |
| 565 | block->offset = block_offset; |
| 566 | block->length = length; |
| 567 | block->index = local->nb_blocks; |
| 568 | block->nb_chunks = ram_chunk_index(host_addr, host_addr + length) + 1UL; |
| 569 | block->transit_bitmap = bitmap_new(block->nb_chunks); |
| 570 | bitmap_clear(block->transit_bitmap, 0, block->nb_chunks); |
| 571 | block->unregister_bitmap = bitmap_new(block->nb_chunks); |
| 572 | bitmap_clear(block->unregister_bitmap, 0, block->nb_chunks); |
| 573 | block->remote_keys = g_malloc0(block->nb_chunks * sizeof(uint32_t)); |
| 574 | |
| 575 | block->is_ram_block = local->init ? false : true; |
| 576 | |
| 577 | g_hash_table_insert(rdma->blockmap, (void *) block_offset, block); |
| 578 | |
| 579 | DDPRINTF("Added Block: %d, addr: %" PRIu64 ", offset: %" PRIu64 |
| 580 | " length: %" PRIu64 " end: %" PRIu64 " bits %" PRIu64 " chunks %d\n", |
| 581 | local->nb_blocks, (uint64_t) block->local_host_addr, block->offset, |
| 582 | block->length, (uint64_t) (block->local_host_addr + block->length), |
| 583 | BITS_TO_LONGS(block->nb_chunks) * |
| 584 | sizeof(unsigned long) * 8, block->nb_chunks); |
| 585 | |
| 586 | local->nb_blocks++; |
| 587 | |
| 588 | return 0; |
| 589 | } |
| 590 | |
| 591 | /* |
| 592 | * Memory regions need to be registered with the device and queue pairs setup |
| 593 | * in advanced before the migration starts. This tells us where the RAM blocks |
| 594 | * are so that we can register them individually. |
| 595 | */ |
| 596 | static void qemu_rdma_init_one_block(void *host_addr, |
| 597 | ram_addr_t block_offset, ram_addr_t length, void *opaque) |
| 598 | { |
| 599 | __qemu_rdma_add_block(opaque, host_addr, block_offset, length); |
| 600 | } |
| 601 | |
| 602 | /* |
| 603 | * Identify the RAMBlocks and their quantity. They will be references to |
| 604 | * identify chunk boundaries inside each RAMBlock and also be referenced |
| 605 | * during dynamic page registration. |
| 606 | */ |
| 607 | static int qemu_rdma_init_ram_blocks(RDMAContext *rdma) |
| 608 | { |
| 609 | RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 610 | |
| 611 | assert(rdma->blockmap == NULL); |
| 612 | rdma->blockmap = g_hash_table_new(g_direct_hash, g_direct_equal); |
| 613 | memset(local, 0, sizeof *local); |
| 614 | qemu_ram_foreach_block(qemu_rdma_init_one_block, rdma); |
| 615 | DPRINTF("Allocated %d local ram block structures\n", local->nb_blocks); |
| 616 | rdma->block = (RDMARemoteBlock *) g_malloc0(sizeof(RDMARemoteBlock) * |
| 617 | rdma->local_ram_blocks.nb_blocks); |
| 618 | local->init = true; |
| 619 | return 0; |
| 620 | } |
| 621 | |
| 622 | static int __qemu_rdma_delete_block(RDMAContext *rdma, ram_addr_t block_offset) |
| 623 | { |
| 624 | RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 625 | RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, |
| 626 | (void *) block_offset); |
| 627 | RDMALocalBlock *old = local->block; |
| 628 | int x; |
| 629 | |
| 630 | assert(block); |
| 631 | |
| 632 | if (block->pmr) { |
| 633 | int j; |
| 634 | |
| 635 | for (j = 0; j < block->nb_chunks; j++) { |
| 636 | if (!block->pmr[j]) { |
| 637 | continue; |
| 638 | } |
| 639 | ibv_dereg_mr(block->pmr[j]); |
| 640 | rdma->total_registrations--; |
| 641 | } |
| 642 | g_free(block->pmr); |
| 643 | block->pmr = NULL; |
| 644 | } |
| 645 | |
| 646 | if (block->mr) { |
| 647 | ibv_dereg_mr(block->mr); |
| 648 | rdma->total_registrations--; |
| 649 | block->mr = NULL; |
| 650 | } |
| 651 | |
| 652 | g_free(block->transit_bitmap); |
| 653 | block->transit_bitmap = NULL; |
| 654 | |
| 655 | g_free(block->unregister_bitmap); |
| 656 | block->unregister_bitmap = NULL; |
| 657 | |
| 658 | g_free(block->remote_keys); |
| 659 | block->remote_keys = NULL; |
| 660 | |
| 661 | for (x = 0; x < local->nb_blocks; x++) { |
| 662 | g_hash_table_remove(rdma->blockmap, (void *)old[x].offset); |
| 663 | } |
| 664 | |
| 665 | if (local->nb_blocks > 1) { |
| 666 | |
| 667 | local->block = g_malloc0(sizeof(RDMALocalBlock) * |
| 668 | (local->nb_blocks - 1)); |
| 669 | |
| 670 | if (block->index) { |
| 671 | memcpy(local->block, old, sizeof(RDMALocalBlock) * block->index); |
| 672 | } |
| 673 | |
| 674 | if (block->index < (local->nb_blocks - 1)) { |
| 675 | memcpy(local->block + block->index, old + (block->index + 1), |
| 676 | sizeof(RDMALocalBlock) * |
| 677 | (local->nb_blocks - (block->index + 1))); |
| 678 | } |
| 679 | } else { |
| 680 | assert(block == local->block); |
| 681 | local->block = NULL; |
| 682 | } |
| 683 | |
| 684 | DDPRINTF("Deleted Block: %d, addr: %" PRIu64 ", offset: %" PRIu64 |
| 685 | " length: %" PRIu64 " end: %" PRIu64 " bits %" PRIu64 " chunks %d\n", |
| 686 | local->nb_blocks, (uint64_t) block->local_host_addr, block->offset, |
| 687 | block->length, (uint64_t) (block->local_host_addr + block->length), |
| 688 | BITS_TO_LONGS(block->nb_chunks) * |
| 689 | sizeof(unsigned long) * 8, block->nb_chunks); |
| 690 | |
| 691 | g_free(old); |
| 692 | |
| 693 | local->nb_blocks--; |
| 694 | |
| 695 | if (local->nb_blocks) { |
| 696 | for (x = 0; x < local->nb_blocks; x++) { |
| 697 | g_hash_table_insert(rdma->blockmap, (void *)local->block[x].offset, |
| 698 | &local->block[x]); |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | return 0; |
| 703 | } |
| 704 | |
| 705 | /* |
| 706 | * Put in the log file which RDMA device was opened and the details |
| 707 | * associated with that device. |
| 708 | */ |
| 709 | static void qemu_rdma_dump_id(const char *who, struct ibv_context *verbs) |
| 710 | { |
| 711 | printf("%s RDMA Device opened: kernel name %s " |
| 712 | "uverbs device name %s, " |
| 713 | "infiniband_verbs class device path %s," |
| 714 | " infiniband class device path %s\n", |
| 715 | who, |
| 716 | verbs->device->name, |
| 717 | verbs->device->dev_name, |
| 718 | verbs->device->dev_path, |
| 719 | verbs->device->ibdev_path); |
| 720 | } |
| 721 | |
| 722 | /* |
| 723 | * Put in the log file the RDMA gid addressing information, |
| 724 | * useful for folks who have trouble understanding the |
| 725 | * RDMA device hierarchy in the kernel. |
| 726 | */ |
| 727 | static void qemu_rdma_dump_gid(const char *who, struct rdma_cm_id *id) |
| 728 | { |
| 729 | char sgid[33]; |
| 730 | char dgid[33]; |
| 731 | inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.sgid, sgid, sizeof sgid); |
| 732 | inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.dgid, dgid, sizeof dgid); |
| 733 | DPRINTF("%s Source GID: %s, Dest GID: %s\n", who, sgid, dgid); |
| 734 | } |
| 735 | |
| 736 | /* |
| 737 | * Figure out which RDMA device corresponds to the requested IP hostname |
| 738 | * Also create the initial connection manager identifiers for opening |
| 739 | * the connection. |
| 740 | */ |
| 741 | static int qemu_rdma_resolve_host(RDMAContext *rdma, Error **errp) |
| 742 | { |
| 743 | int ret; |
| 744 | struct addrinfo *res; |
| 745 | char port_str[16]; |
| 746 | struct rdma_cm_event *cm_event; |
| 747 | char ip[40] = "unknown"; |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 748 | int af = rdma->ipv6 ? PF_INET6 : PF_INET; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 749 | |
| 750 | if (rdma->host == NULL || !strcmp(rdma->host, "")) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 751 | ERROR(errp, "RDMA hostname has not been set"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 752 | return -1; |
| 753 | } |
| 754 | |
| 755 | /* create CM channel */ |
| 756 | rdma->channel = rdma_create_event_channel(); |
| 757 | if (!rdma->channel) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 758 | ERROR(errp, "could not create CM channel"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 759 | return -1; |
| 760 | } |
| 761 | |
| 762 | /* create CM id */ |
| 763 | ret = rdma_create_id(rdma->channel, &rdma->cm_id, NULL, RDMA_PS_TCP); |
| 764 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 765 | ERROR(errp, "could not create channel id"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 766 | goto err_resolve_create_id; |
| 767 | } |
| 768 | |
| 769 | snprintf(port_str, 16, "%d", rdma->port); |
| 770 | port_str[15] = '\0'; |
| 771 | |
| 772 | ret = getaddrinfo(rdma->host, port_str, NULL, &res); |
| 773 | if (ret < 0) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 774 | ERROR(errp, "could not getaddrinfo address %s", rdma->host); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 775 | goto err_resolve_get_addr; |
| 776 | } |
| 777 | |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 778 | inet_ntop(af, &((struct sockaddr_in *) res->ai_addr)->sin_addr, |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 779 | ip, sizeof ip); |
| 780 | DPRINTF("%s => %s\n", rdma->host, ip); |
| 781 | |
| 782 | /* resolve the first address */ |
| 783 | ret = rdma_resolve_addr(rdma->cm_id, NULL, res->ai_addr, |
| 784 | RDMA_RESOLVE_TIMEOUT_MS); |
| 785 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 786 | ERROR(errp, "could not resolve address %s", rdma->host); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 787 | goto err_resolve_get_addr; |
| 788 | } |
| 789 | |
| 790 | qemu_rdma_dump_gid("source_resolve_addr", rdma->cm_id); |
| 791 | |
| 792 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 793 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 794 | ERROR(errp, "could not perform event_addr_resolved"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 795 | goto err_resolve_get_addr; |
| 796 | } |
| 797 | |
| 798 | if (cm_event->event != RDMA_CM_EVENT_ADDR_RESOLVED) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 799 | ERROR(errp, "result not equal to event_addr_resolved %s", |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 800 | rdma_event_str(cm_event->event)); |
| 801 | perror("rdma_resolve_addr"); |
| 802 | goto err_resolve_get_addr; |
| 803 | } |
| 804 | rdma_ack_cm_event(cm_event); |
| 805 | |
| 806 | /* resolve route */ |
| 807 | ret = rdma_resolve_route(rdma->cm_id, RDMA_RESOLVE_TIMEOUT_MS); |
| 808 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 809 | ERROR(errp, "could not resolve rdma route"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 810 | goto err_resolve_get_addr; |
| 811 | } |
| 812 | |
| 813 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 814 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 815 | ERROR(errp, "could not perform event_route_resolved"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 816 | goto err_resolve_get_addr; |
| 817 | } |
| 818 | if (cm_event->event != RDMA_CM_EVENT_ROUTE_RESOLVED) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 819 | ERROR(errp, "result not equal to event_route_resolved: %s", |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 820 | rdma_event_str(cm_event->event)); |
| 821 | rdma_ack_cm_event(cm_event); |
| 822 | goto err_resolve_get_addr; |
| 823 | } |
| 824 | rdma_ack_cm_event(cm_event); |
| 825 | rdma->verbs = rdma->cm_id->verbs; |
| 826 | qemu_rdma_dump_id("source_resolve_host", rdma->cm_id->verbs); |
| 827 | qemu_rdma_dump_gid("source_resolve_host", rdma->cm_id); |
| 828 | return 0; |
| 829 | |
| 830 | err_resolve_get_addr: |
| 831 | rdma_destroy_id(rdma->cm_id); |
| 832 | rdma->cm_id = NULL; |
| 833 | err_resolve_create_id: |
| 834 | rdma_destroy_event_channel(rdma->channel); |
| 835 | rdma->channel = NULL; |
| 836 | |
| 837 | return -1; |
| 838 | } |
| 839 | |
| 840 | /* |
| 841 | * Create protection domain and completion queues |
| 842 | */ |
| 843 | static int qemu_rdma_alloc_pd_cq(RDMAContext *rdma) |
| 844 | { |
| 845 | /* allocate pd */ |
| 846 | rdma->pd = ibv_alloc_pd(rdma->verbs); |
| 847 | if (!rdma->pd) { |
| 848 | fprintf(stderr, "failed to allocate protection domain\n"); |
| 849 | return -1; |
| 850 | } |
| 851 | |
| 852 | /* create completion channel */ |
| 853 | rdma->comp_channel = ibv_create_comp_channel(rdma->verbs); |
| 854 | if (!rdma->comp_channel) { |
| 855 | fprintf(stderr, "failed to allocate completion channel\n"); |
| 856 | goto err_alloc_pd_cq; |
| 857 | } |
| 858 | |
| 859 | /* |
| 860 | * Completion queue can be filled by both read and write work requests, |
| 861 | * so must reflect the sum of both possible queue sizes. |
| 862 | */ |
| 863 | rdma->cq = ibv_create_cq(rdma->verbs, (RDMA_SIGNALED_SEND_MAX * 3), |
| 864 | NULL, rdma->comp_channel, 0); |
| 865 | if (!rdma->cq) { |
| 866 | fprintf(stderr, "failed to allocate completion queue\n"); |
| 867 | goto err_alloc_pd_cq; |
| 868 | } |
| 869 | |
| 870 | return 0; |
| 871 | |
| 872 | err_alloc_pd_cq: |
| 873 | if (rdma->pd) { |
| 874 | ibv_dealloc_pd(rdma->pd); |
| 875 | } |
| 876 | if (rdma->comp_channel) { |
| 877 | ibv_destroy_comp_channel(rdma->comp_channel); |
| 878 | } |
| 879 | rdma->pd = NULL; |
| 880 | rdma->comp_channel = NULL; |
| 881 | return -1; |
| 882 | |
| 883 | } |
| 884 | |
| 885 | /* |
| 886 | * Create queue pairs. |
| 887 | */ |
| 888 | static int qemu_rdma_alloc_qp(RDMAContext *rdma) |
| 889 | { |
| 890 | struct ibv_qp_init_attr attr = { 0 }; |
| 891 | int ret; |
| 892 | |
| 893 | attr.cap.max_send_wr = RDMA_SIGNALED_SEND_MAX; |
| 894 | attr.cap.max_recv_wr = 3; |
| 895 | attr.cap.max_send_sge = 1; |
| 896 | attr.cap.max_recv_sge = 1; |
| 897 | attr.send_cq = rdma->cq; |
| 898 | attr.recv_cq = rdma->cq; |
| 899 | attr.qp_type = IBV_QPT_RC; |
| 900 | |
| 901 | ret = rdma_create_qp(rdma->cm_id, rdma->pd, &attr); |
| 902 | if (ret) { |
| 903 | return -1; |
| 904 | } |
| 905 | |
| 906 | rdma->qp = rdma->cm_id->qp; |
| 907 | return 0; |
| 908 | } |
| 909 | |
| 910 | static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) |
| 911 | { |
| 912 | int i; |
| 913 | RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 914 | |
| 915 | for (i = 0; i < local->nb_blocks; i++) { |
| 916 | local->block[i].mr = |
| 917 | ibv_reg_mr(rdma->pd, |
| 918 | local->block[i].local_host_addr, |
| 919 | local->block[i].length, |
| 920 | IBV_ACCESS_LOCAL_WRITE | |
| 921 | IBV_ACCESS_REMOTE_WRITE |
| 922 | ); |
| 923 | if (!local->block[i].mr) { |
| 924 | perror("Failed to register local dest ram block!\n"); |
| 925 | break; |
| 926 | } |
| 927 | rdma->total_registrations++; |
| 928 | } |
| 929 | |
| 930 | if (i >= local->nb_blocks) { |
| 931 | return 0; |
| 932 | } |
| 933 | |
| 934 | for (i--; i >= 0; i--) { |
| 935 | ibv_dereg_mr(local->block[i].mr); |
| 936 | rdma->total_registrations--; |
| 937 | } |
| 938 | |
| 939 | return -1; |
| 940 | |
| 941 | } |
| 942 | |
| 943 | /* |
| 944 | * Find the ram block that corresponds to the page requested to be |
| 945 | * transmitted by QEMU. |
| 946 | * |
| 947 | * Once the block is found, also identify which 'chunk' within that |
| 948 | * block that the page belongs to. |
| 949 | * |
| 950 | * This search cannot fail or the migration will fail. |
| 951 | */ |
| 952 | static int qemu_rdma_search_ram_block(RDMAContext *rdma, |
| 953 | uint64_t block_offset, |
| 954 | uint64_t offset, |
| 955 | uint64_t length, |
| 956 | uint64_t *block_index, |
| 957 | uint64_t *chunk_index) |
| 958 | { |
| 959 | uint64_t current_addr = block_offset + offset; |
| 960 | RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, |
| 961 | (void *) block_offset); |
| 962 | assert(block); |
| 963 | assert(current_addr >= block->offset); |
| 964 | assert((current_addr + length) <= (block->offset + block->length)); |
| 965 | |
| 966 | *block_index = block->index; |
| 967 | *chunk_index = ram_chunk_index(block->local_host_addr, |
| 968 | block->local_host_addr + (current_addr - block->offset)); |
| 969 | |
| 970 | return 0; |
| 971 | } |
| 972 | |
| 973 | /* |
| 974 | * Register a chunk with IB. If the chunk was already registered |
| 975 | * previously, then skip. |
| 976 | * |
| 977 | * Also return the keys associated with the registration needed |
| 978 | * to perform the actual RDMA operation. |
| 979 | */ |
| 980 | static int qemu_rdma_register_and_get_keys(RDMAContext *rdma, |
| 981 | RDMALocalBlock *block, uint8_t *host_addr, |
| 982 | uint32_t *lkey, uint32_t *rkey, int chunk, |
| 983 | uint8_t *chunk_start, uint8_t *chunk_end) |
| 984 | { |
| 985 | if (block->mr) { |
| 986 | if (lkey) { |
| 987 | *lkey = block->mr->lkey; |
| 988 | } |
| 989 | if (rkey) { |
| 990 | *rkey = block->mr->rkey; |
| 991 | } |
| 992 | return 0; |
| 993 | } |
| 994 | |
| 995 | /* allocate memory to store chunk MRs */ |
| 996 | if (!block->pmr) { |
| 997 | block->pmr = g_malloc0(block->nb_chunks * sizeof(struct ibv_mr *)); |
| 998 | if (!block->pmr) { |
| 999 | return -1; |
| 1000 | } |
| 1001 | } |
| 1002 | |
| 1003 | /* |
| 1004 | * If 'rkey', then we're the destination, so grant access to the source. |
| 1005 | * |
| 1006 | * If 'lkey', then we're the source VM, so grant access only to ourselves. |
| 1007 | */ |
| 1008 | if (!block->pmr[chunk]) { |
| 1009 | uint64_t len = chunk_end - chunk_start; |
| 1010 | |
| 1011 | DDPRINTF("Registering %" PRIu64 " bytes @ %p\n", |
| 1012 | len, chunk_start); |
| 1013 | |
| 1014 | block->pmr[chunk] = ibv_reg_mr(rdma->pd, |
| 1015 | chunk_start, len, |
| 1016 | (rkey ? (IBV_ACCESS_LOCAL_WRITE | |
| 1017 | IBV_ACCESS_REMOTE_WRITE) : 0)); |
| 1018 | |
| 1019 | if (!block->pmr[chunk]) { |
| 1020 | perror("Failed to register chunk!"); |
| 1021 | fprintf(stderr, "Chunk details: block: %d chunk index %d" |
| 1022 | " start %" PRIu64 " end %" PRIu64 " host %" PRIu64 |
| 1023 | " local %" PRIu64 " registrations: %d\n", |
| 1024 | block->index, chunk, (uint64_t) chunk_start, |
| 1025 | (uint64_t) chunk_end, (uint64_t) host_addr, |
| 1026 | (uint64_t) block->local_host_addr, |
| 1027 | rdma->total_registrations); |
| 1028 | return -1; |
| 1029 | } |
| 1030 | rdma->total_registrations++; |
| 1031 | } |
| 1032 | |
| 1033 | if (lkey) { |
| 1034 | *lkey = block->pmr[chunk]->lkey; |
| 1035 | } |
| 1036 | if (rkey) { |
| 1037 | *rkey = block->pmr[chunk]->rkey; |
| 1038 | } |
| 1039 | return 0; |
| 1040 | } |
| 1041 | |
| 1042 | /* |
| 1043 | * Register (at connection time) the memory used for control |
| 1044 | * channel messages. |
| 1045 | */ |
| 1046 | static int qemu_rdma_reg_control(RDMAContext *rdma, int idx) |
| 1047 | { |
| 1048 | rdma->wr_data[idx].control_mr = ibv_reg_mr(rdma->pd, |
| 1049 | rdma->wr_data[idx].control, RDMA_CONTROL_MAX_BUFFER, |
| 1050 | IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); |
| 1051 | if (rdma->wr_data[idx].control_mr) { |
| 1052 | rdma->total_registrations++; |
| 1053 | return 0; |
| 1054 | } |
| 1055 | fprintf(stderr, "qemu_rdma_reg_control failed!\n"); |
| 1056 | return -1; |
| 1057 | } |
| 1058 | |
| 1059 | const char *print_wrid(int wrid) |
| 1060 | { |
| 1061 | if (wrid >= RDMA_WRID_RECV_CONTROL) { |
| 1062 | return wrid_desc[RDMA_WRID_RECV_CONTROL]; |
| 1063 | } |
| 1064 | return wrid_desc[wrid]; |
| 1065 | } |
| 1066 | |
| 1067 | /* |
| 1068 | * RDMA requires memory registration (mlock/pinning), but this is not good for |
| 1069 | * overcommitment. |
| 1070 | * |
| 1071 | * In preparation for the future where LRU information or workload-specific |
| 1072 | * writable writable working set memory access behavior is available to QEMU |
| 1073 | * it would be nice to have in place the ability to UN-register/UN-pin |
| 1074 | * particular memory regions from the RDMA hardware when it is determine that |
| 1075 | * those regions of memory will likely not be accessed again in the near future. |
| 1076 | * |
| 1077 | * While we do not yet have such information right now, the following |
| 1078 | * compile-time option allows us to perform a non-optimized version of this |
| 1079 | * behavior. |
| 1080 | * |
| 1081 | * By uncommenting this option, you will cause *all* RDMA transfers to be |
| 1082 | * unregistered immediately after the transfer completes on both sides of the |
| 1083 | * connection. This has no effect in 'rdma-pin-all' mode, only regular mode. |
| 1084 | * |
| 1085 | * This will have a terrible impact on migration performance, so until future |
| 1086 | * workload information or LRU information is available, do not attempt to use |
| 1087 | * this feature except for basic testing. |
| 1088 | */ |
| 1089 | //#define RDMA_UNREGISTRATION_EXAMPLE |
| 1090 | |
| 1091 | /* |
| 1092 | * Perform a non-optimized memory unregistration after every transfer |
| 1093 | * for demonsration purposes, only if pin-all is not requested. |
| 1094 | * |
| 1095 | * Potential optimizations: |
| 1096 | * 1. Start a new thread to run this function continuously |
| 1097 | - for bit clearing |
| 1098 | - and for receipt of unregister messages |
| 1099 | * 2. Use an LRU. |
| 1100 | * 3. Use workload hints. |
| 1101 | */ |
| 1102 | static int qemu_rdma_unregister_waiting(RDMAContext *rdma) |
| 1103 | { |
| 1104 | while (rdma->unregistrations[rdma->unregister_current]) { |
| 1105 | int ret; |
| 1106 | uint64_t wr_id = rdma->unregistrations[rdma->unregister_current]; |
| 1107 | uint64_t chunk = |
| 1108 | (wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; |
| 1109 | uint64_t index = |
| 1110 | (wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; |
| 1111 | RDMALocalBlock *block = |
| 1112 | &(rdma->local_ram_blocks.block[index]); |
| 1113 | RDMARegister reg = { .current_index = index }; |
| 1114 | RDMAControlHeader resp = { .type = RDMA_CONTROL_UNREGISTER_FINISHED, |
| 1115 | }; |
| 1116 | RDMAControlHeader head = { .len = sizeof(RDMARegister), |
| 1117 | .type = RDMA_CONTROL_UNREGISTER_REQUEST, |
| 1118 | .repeat = 1, |
| 1119 | }; |
| 1120 | |
| 1121 | DDPRINTF("Processing unregister for chunk: %" PRIu64 |
| 1122 | " at position %d\n", chunk, rdma->unregister_current); |
| 1123 | |
| 1124 | rdma->unregistrations[rdma->unregister_current] = 0; |
| 1125 | rdma->unregister_current++; |
| 1126 | |
| 1127 | if (rdma->unregister_current == RDMA_SIGNALED_SEND_MAX) { |
| 1128 | rdma->unregister_current = 0; |
| 1129 | } |
| 1130 | |
| 1131 | |
| 1132 | /* |
| 1133 | * Unregistration is speculative (because migration is single-threaded |
| 1134 | * and we cannot break the protocol's inifinband message ordering). |
| 1135 | * Thus, if the memory is currently being used for transmission, |
| 1136 | * then abort the attempt to unregister and try again |
| 1137 | * later the next time a completion is received for this memory. |
| 1138 | */ |
| 1139 | clear_bit(chunk, block->unregister_bitmap); |
| 1140 | |
| 1141 | if (test_bit(chunk, block->transit_bitmap)) { |
| 1142 | DDPRINTF("Cannot unregister inflight chunk: %" PRIu64 "\n", chunk); |
| 1143 | continue; |
| 1144 | } |
| 1145 | |
| 1146 | DDPRINTF("Sending unregister for chunk: %" PRIu64 "\n", chunk); |
| 1147 | |
| 1148 | ret = ibv_dereg_mr(block->pmr[chunk]); |
| 1149 | block->pmr[chunk] = NULL; |
| 1150 | block->remote_keys[chunk] = 0; |
| 1151 | |
| 1152 | if (ret != 0) { |
| 1153 | perror("unregistration chunk failed"); |
| 1154 | return -ret; |
| 1155 | } |
| 1156 | rdma->total_registrations--; |
| 1157 | |
| 1158 | reg.key.chunk = chunk; |
| 1159 | register_to_network(®); |
| 1160 | ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
| 1161 | &resp, NULL, NULL); |
| 1162 | if (ret < 0) { |
| 1163 | return ret; |
| 1164 | } |
| 1165 | |
| 1166 | DDPRINTF("Unregister for chunk: %" PRIu64 " complete.\n", chunk); |
| 1167 | } |
| 1168 | |
| 1169 | return 0; |
| 1170 | } |
| 1171 | |
| 1172 | static uint64_t qemu_rdma_make_wrid(uint64_t wr_id, uint64_t index, |
| 1173 | uint64_t chunk) |
| 1174 | { |
| 1175 | uint64_t result = wr_id & RDMA_WRID_TYPE_MASK; |
| 1176 | |
| 1177 | result |= (index << RDMA_WRID_BLOCK_SHIFT); |
| 1178 | result |= (chunk << RDMA_WRID_CHUNK_SHIFT); |
| 1179 | |
| 1180 | return result; |
| 1181 | } |
| 1182 | |
| 1183 | /* |
| 1184 | * Set bit for unregistration in the next iteration. |
| 1185 | * We cannot transmit right here, but will unpin later. |
| 1186 | */ |
| 1187 | static void qemu_rdma_signal_unregister(RDMAContext *rdma, uint64_t index, |
| 1188 | uint64_t chunk, uint64_t wr_id) |
| 1189 | { |
| 1190 | if (rdma->unregistrations[rdma->unregister_next] != 0) { |
| 1191 | fprintf(stderr, "rdma migration: queue is full!\n"); |
| 1192 | } else { |
| 1193 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); |
| 1194 | |
| 1195 | if (!test_and_set_bit(chunk, block->unregister_bitmap)) { |
| 1196 | DDPRINTF("Appending unregister chunk %" PRIu64 |
| 1197 | " at position %d\n", chunk, rdma->unregister_next); |
| 1198 | |
| 1199 | rdma->unregistrations[rdma->unregister_next++] = |
| 1200 | qemu_rdma_make_wrid(wr_id, index, chunk); |
| 1201 | |
| 1202 | if (rdma->unregister_next == RDMA_SIGNALED_SEND_MAX) { |
| 1203 | rdma->unregister_next = 0; |
| 1204 | } |
| 1205 | } else { |
| 1206 | DDPRINTF("Unregister chunk %" PRIu64 " already in queue.\n", |
| 1207 | chunk); |
| 1208 | } |
| 1209 | } |
| 1210 | } |
| 1211 | |
| 1212 | /* |
| 1213 | * Consult the connection manager to see a work request |
| 1214 | * (of any kind) has completed. |
| 1215 | * Return the work request ID that completed. |
| 1216 | */ |
| 1217 | static uint64_t qemu_rdma_poll(RDMAContext *rdma, uint64_t *wr_id_out) |
| 1218 | { |
| 1219 | int ret; |
| 1220 | struct ibv_wc wc; |
| 1221 | uint64_t wr_id; |
| 1222 | |
| 1223 | ret = ibv_poll_cq(rdma->cq, 1, &wc); |
| 1224 | |
| 1225 | if (!ret) { |
| 1226 | *wr_id_out = RDMA_WRID_NONE; |
| 1227 | return 0; |
| 1228 | } |
| 1229 | |
| 1230 | if (ret < 0) { |
| 1231 | fprintf(stderr, "ibv_poll_cq return %d!\n", ret); |
| 1232 | return ret; |
| 1233 | } |
| 1234 | |
| 1235 | wr_id = wc.wr_id & RDMA_WRID_TYPE_MASK; |
| 1236 | |
| 1237 | if (wc.status != IBV_WC_SUCCESS) { |
| 1238 | fprintf(stderr, "ibv_poll_cq wc.status=%d %s!\n", |
| 1239 | wc.status, ibv_wc_status_str(wc.status)); |
| 1240 | fprintf(stderr, "ibv_poll_cq wrid=%s!\n", wrid_desc[wr_id]); |
| 1241 | |
| 1242 | return -1; |
| 1243 | } |
| 1244 | |
| 1245 | if (rdma->control_ready_expected && |
| 1246 | (wr_id >= RDMA_WRID_RECV_CONTROL)) { |
| 1247 | DDDPRINTF("completion %s #%" PRId64 " received (%" PRId64 ")" |
| 1248 | " left %d\n", wrid_desc[RDMA_WRID_RECV_CONTROL], |
| 1249 | wr_id - RDMA_WRID_RECV_CONTROL, wr_id, rdma->nb_sent); |
| 1250 | rdma->control_ready_expected = 0; |
| 1251 | } |
| 1252 | |
| 1253 | if (wr_id == RDMA_WRID_RDMA_WRITE) { |
| 1254 | uint64_t chunk = |
| 1255 | (wc.wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; |
| 1256 | uint64_t index = |
| 1257 | (wc.wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; |
| 1258 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); |
| 1259 | |
| 1260 | DDDPRINTF("completions %s (%" PRId64 ") left %d, " |
| 1261 | "block %" PRIu64 ", chunk: %" PRIu64 " %p %p\n", |
| 1262 | print_wrid(wr_id), wr_id, rdma->nb_sent, index, chunk, |
| 1263 | block->local_host_addr, (void *)block->remote_host_addr); |
| 1264 | |
| 1265 | clear_bit(chunk, block->transit_bitmap); |
| 1266 | |
| 1267 | if (rdma->nb_sent > 0) { |
| 1268 | rdma->nb_sent--; |
| 1269 | } |
| 1270 | |
| 1271 | if (!rdma->pin_all) { |
| 1272 | /* |
| 1273 | * FYI: If one wanted to signal a specific chunk to be unregistered |
| 1274 | * using LRU or workload-specific information, this is the function |
| 1275 | * you would call to do so. That chunk would then get asynchronously |
| 1276 | * unregistered later. |
| 1277 | */ |
| 1278 | #ifdef RDMA_UNREGISTRATION_EXAMPLE |
| 1279 | qemu_rdma_signal_unregister(rdma, index, chunk, wc.wr_id); |
| 1280 | #endif |
| 1281 | } |
| 1282 | } else { |
| 1283 | DDDPRINTF("other completion %s (%" PRId64 ") received left %d\n", |
| 1284 | print_wrid(wr_id), wr_id, rdma->nb_sent); |
| 1285 | } |
| 1286 | |
| 1287 | *wr_id_out = wc.wr_id; |
| 1288 | |
| 1289 | return 0; |
| 1290 | } |
| 1291 | |
| 1292 | /* |
| 1293 | * Block until the next work request has completed. |
| 1294 | * |
| 1295 | * First poll to see if a work request has already completed, |
| 1296 | * otherwise block. |
| 1297 | * |
| 1298 | * If we encounter completed work requests for IDs other than |
| 1299 | * the one we're interested in, then that's generally an error. |
| 1300 | * |
| 1301 | * The only exception is actual RDMA Write completions. These |
| 1302 | * completions only need to be recorded, but do not actually |
| 1303 | * need further processing. |
| 1304 | */ |
| 1305 | static int qemu_rdma_block_for_wrid(RDMAContext *rdma, int wrid_requested) |
| 1306 | { |
| 1307 | int num_cq_events = 0, ret = 0; |
| 1308 | struct ibv_cq *cq; |
| 1309 | void *cq_ctx; |
| 1310 | uint64_t wr_id = RDMA_WRID_NONE, wr_id_in; |
| 1311 | |
| 1312 | if (ibv_req_notify_cq(rdma->cq, 0)) { |
| 1313 | return -1; |
| 1314 | } |
| 1315 | /* poll cq first */ |
| 1316 | while (wr_id != wrid_requested) { |
| 1317 | ret = qemu_rdma_poll(rdma, &wr_id_in); |
| 1318 | if (ret < 0) { |
| 1319 | return ret; |
| 1320 | } |
| 1321 | |
| 1322 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; |
| 1323 | |
| 1324 | if (wr_id == RDMA_WRID_NONE) { |
| 1325 | break; |
| 1326 | } |
| 1327 | if (wr_id != wrid_requested) { |
| 1328 | DDDPRINTF("A Wanted wrid %s (%d) but got %s (%" PRIu64 ")\n", |
| 1329 | print_wrid(wrid_requested), |
| 1330 | wrid_requested, print_wrid(wr_id), wr_id); |
| 1331 | } |
| 1332 | } |
| 1333 | |
| 1334 | if (wr_id == wrid_requested) { |
| 1335 | return 0; |
| 1336 | } |
| 1337 | |
| 1338 | while (1) { |
| 1339 | /* |
| 1340 | * Coroutine doesn't start until process_incoming_migration() |
| 1341 | * so don't yield unless we know we're running inside of a coroutine. |
| 1342 | */ |
| 1343 | if (rdma->migration_started_on_destination) { |
| 1344 | yield_until_fd_readable(rdma->comp_channel->fd); |
| 1345 | } |
| 1346 | |
| 1347 | if (ibv_get_cq_event(rdma->comp_channel, &cq, &cq_ctx)) { |
| 1348 | perror("ibv_get_cq_event"); |
| 1349 | goto err_block_for_wrid; |
| 1350 | } |
| 1351 | |
| 1352 | num_cq_events++; |
| 1353 | |
| 1354 | if (ibv_req_notify_cq(cq, 0)) { |
| 1355 | goto err_block_for_wrid; |
| 1356 | } |
| 1357 | |
| 1358 | while (wr_id != wrid_requested) { |
| 1359 | ret = qemu_rdma_poll(rdma, &wr_id_in); |
| 1360 | if (ret < 0) { |
| 1361 | goto err_block_for_wrid; |
| 1362 | } |
| 1363 | |
| 1364 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; |
| 1365 | |
| 1366 | if (wr_id == RDMA_WRID_NONE) { |
| 1367 | break; |
| 1368 | } |
| 1369 | if (wr_id != wrid_requested) { |
| 1370 | DDDPRINTF("B Wanted wrid %s (%d) but got %s (%" PRIu64 ")\n", |
| 1371 | print_wrid(wrid_requested), wrid_requested, |
| 1372 | print_wrid(wr_id), wr_id); |
| 1373 | } |
| 1374 | } |
| 1375 | |
| 1376 | if (wr_id == wrid_requested) { |
| 1377 | goto success_block_for_wrid; |
| 1378 | } |
| 1379 | } |
| 1380 | |
| 1381 | success_block_for_wrid: |
| 1382 | if (num_cq_events) { |
| 1383 | ibv_ack_cq_events(cq, num_cq_events); |
| 1384 | } |
| 1385 | return 0; |
| 1386 | |
| 1387 | err_block_for_wrid: |
| 1388 | if (num_cq_events) { |
| 1389 | ibv_ack_cq_events(cq, num_cq_events); |
| 1390 | } |
| 1391 | return ret; |
| 1392 | } |
| 1393 | |
| 1394 | /* |
| 1395 | * Post a SEND message work request for the control channel |
| 1396 | * containing some data and block until the post completes. |
| 1397 | */ |
| 1398 | static int qemu_rdma_post_send_control(RDMAContext *rdma, uint8_t *buf, |
| 1399 | RDMAControlHeader *head) |
| 1400 | { |
| 1401 | int ret = 0; |
Isaku Yamahata | 1f22364 | 2013-08-03 22:54:52 -0400 | [diff] [blame] | 1402 | RDMAWorkRequestData *wr = &rdma->wr_data[RDMA_WRID_CONTROL]; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 1403 | struct ibv_send_wr *bad_wr; |
| 1404 | struct ibv_sge sge = { |
| 1405 | .addr = (uint64_t)(wr->control), |
| 1406 | .length = head->len + sizeof(RDMAControlHeader), |
| 1407 | .lkey = wr->control_mr->lkey, |
| 1408 | }; |
| 1409 | struct ibv_send_wr send_wr = { |
| 1410 | .wr_id = RDMA_WRID_SEND_CONTROL, |
| 1411 | .opcode = IBV_WR_SEND, |
| 1412 | .send_flags = IBV_SEND_SIGNALED, |
| 1413 | .sg_list = &sge, |
| 1414 | .num_sge = 1, |
| 1415 | }; |
| 1416 | |
| 1417 | DDDPRINTF("CONTROL: sending %s..\n", control_desc[head->type]); |
| 1418 | |
| 1419 | /* |
| 1420 | * We don't actually need to do a memcpy() in here if we used |
| 1421 | * the "sge" properly, but since we're only sending control messages |
| 1422 | * (not RAM in a performance-critical path), then its OK for now. |
| 1423 | * |
| 1424 | * The copy makes the RDMAControlHeader simpler to manipulate |
| 1425 | * for the time being. |
| 1426 | */ |
| 1427 | memcpy(wr->control, head, sizeof(RDMAControlHeader)); |
| 1428 | control_to_network((void *) wr->control); |
| 1429 | |
| 1430 | if (buf) { |
| 1431 | memcpy(wr->control + sizeof(RDMAControlHeader), buf, head->len); |
| 1432 | } |
| 1433 | |
| 1434 | |
| 1435 | if (ibv_post_send(rdma->qp, &send_wr, &bad_wr)) { |
| 1436 | return -1; |
| 1437 | } |
| 1438 | |
| 1439 | if (ret < 0) { |
| 1440 | fprintf(stderr, "Failed to use post IB SEND for control!\n"); |
| 1441 | return ret; |
| 1442 | } |
| 1443 | |
| 1444 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_SEND_CONTROL); |
| 1445 | if (ret < 0) { |
| 1446 | fprintf(stderr, "rdma migration: send polling control error!\n"); |
| 1447 | } |
| 1448 | |
| 1449 | return ret; |
| 1450 | } |
| 1451 | |
| 1452 | /* |
| 1453 | * Post a RECV work request in anticipation of some future receipt |
| 1454 | * of data on the control channel. |
| 1455 | */ |
| 1456 | static int qemu_rdma_post_recv_control(RDMAContext *rdma, int idx) |
| 1457 | { |
| 1458 | struct ibv_recv_wr *bad_wr; |
| 1459 | struct ibv_sge sge = { |
| 1460 | .addr = (uint64_t)(rdma->wr_data[idx].control), |
| 1461 | .length = RDMA_CONTROL_MAX_BUFFER, |
| 1462 | .lkey = rdma->wr_data[idx].control_mr->lkey, |
| 1463 | }; |
| 1464 | |
| 1465 | struct ibv_recv_wr recv_wr = { |
| 1466 | .wr_id = RDMA_WRID_RECV_CONTROL + idx, |
| 1467 | .sg_list = &sge, |
| 1468 | .num_sge = 1, |
| 1469 | }; |
| 1470 | |
| 1471 | |
| 1472 | if (ibv_post_recv(rdma->qp, &recv_wr, &bad_wr)) { |
| 1473 | return -1; |
| 1474 | } |
| 1475 | |
| 1476 | return 0; |
| 1477 | } |
| 1478 | |
| 1479 | /* |
| 1480 | * Block and wait for a RECV control channel message to arrive. |
| 1481 | */ |
| 1482 | static int qemu_rdma_exchange_get_response(RDMAContext *rdma, |
| 1483 | RDMAControlHeader *head, int expecting, int idx) |
| 1484 | { |
| 1485 | int ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RECV_CONTROL + idx); |
| 1486 | |
| 1487 | if (ret < 0) { |
| 1488 | fprintf(stderr, "rdma migration: recv polling control error!\n"); |
| 1489 | return ret; |
| 1490 | } |
| 1491 | |
| 1492 | network_to_control((void *) rdma->wr_data[idx].control); |
| 1493 | memcpy(head, rdma->wr_data[idx].control, sizeof(RDMAControlHeader)); |
| 1494 | |
| 1495 | DDDPRINTF("CONTROL: %s receiving...\n", control_desc[expecting]); |
| 1496 | |
| 1497 | if (expecting == RDMA_CONTROL_NONE) { |
| 1498 | DDDPRINTF("Surprise: got %s (%d)\n", |
| 1499 | control_desc[head->type], head->type); |
| 1500 | } else if (head->type != expecting || head->type == RDMA_CONTROL_ERROR) { |
| 1501 | fprintf(stderr, "Was expecting a %s (%d) control message" |
| 1502 | ", but got: %s (%d), length: %d\n", |
| 1503 | control_desc[expecting], expecting, |
| 1504 | control_desc[head->type], head->type, head->len); |
| 1505 | return -EIO; |
| 1506 | } |
| 1507 | |
| 1508 | return 0; |
| 1509 | } |
| 1510 | |
| 1511 | /* |
| 1512 | * When a RECV work request has completed, the work request's |
| 1513 | * buffer is pointed at the header. |
| 1514 | * |
| 1515 | * This will advance the pointer to the data portion |
| 1516 | * of the control message of the work request's buffer that |
| 1517 | * was populated after the work request finished. |
| 1518 | */ |
| 1519 | static void qemu_rdma_move_header(RDMAContext *rdma, int idx, |
| 1520 | RDMAControlHeader *head) |
| 1521 | { |
| 1522 | rdma->wr_data[idx].control_len = head->len; |
| 1523 | rdma->wr_data[idx].control_curr = |
| 1524 | rdma->wr_data[idx].control + sizeof(RDMAControlHeader); |
| 1525 | } |
| 1526 | |
| 1527 | /* |
| 1528 | * This is an 'atomic' high-level operation to deliver a single, unified |
| 1529 | * control-channel message. |
| 1530 | * |
| 1531 | * Additionally, if the user is expecting some kind of reply to this message, |
| 1532 | * they can request a 'resp' response message be filled in by posting an |
| 1533 | * additional work request on behalf of the user and waiting for an additional |
| 1534 | * completion. |
| 1535 | * |
| 1536 | * The extra (optional) response is used during registration to us from having |
| 1537 | * to perform an *additional* exchange of message just to provide a response by |
| 1538 | * instead piggy-backing on the acknowledgement. |
| 1539 | */ |
| 1540 | static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, |
| 1541 | uint8_t *data, RDMAControlHeader *resp, |
| 1542 | int *resp_idx, |
| 1543 | int (*callback)(RDMAContext *rdma)) |
| 1544 | { |
| 1545 | int ret = 0; |
| 1546 | |
| 1547 | /* |
| 1548 | * Wait until the dest is ready before attempting to deliver the message |
| 1549 | * by waiting for a READY message. |
| 1550 | */ |
| 1551 | if (rdma->control_ready_expected) { |
| 1552 | RDMAControlHeader resp; |
| 1553 | ret = qemu_rdma_exchange_get_response(rdma, |
| 1554 | &resp, RDMA_CONTROL_READY, RDMA_WRID_READY); |
| 1555 | if (ret < 0) { |
| 1556 | return ret; |
| 1557 | } |
| 1558 | } |
| 1559 | |
| 1560 | /* |
| 1561 | * If the user is expecting a response, post a WR in anticipation of it. |
| 1562 | */ |
| 1563 | if (resp) { |
| 1564 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_DATA); |
| 1565 | if (ret) { |
| 1566 | fprintf(stderr, "rdma migration: error posting" |
| 1567 | " extra control recv for anticipated result!"); |
| 1568 | return ret; |
| 1569 | } |
| 1570 | } |
| 1571 | |
| 1572 | /* |
| 1573 | * Post a WR to replace the one we just consumed for the READY message. |
| 1574 | */ |
| 1575 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
| 1576 | if (ret) { |
| 1577 | fprintf(stderr, "rdma migration: error posting first control recv!"); |
| 1578 | return ret; |
| 1579 | } |
| 1580 | |
| 1581 | /* |
| 1582 | * Deliver the control message that was requested. |
| 1583 | */ |
| 1584 | ret = qemu_rdma_post_send_control(rdma, data, head); |
| 1585 | |
| 1586 | if (ret < 0) { |
| 1587 | fprintf(stderr, "Failed to send control buffer!\n"); |
| 1588 | return ret; |
| 1589 | } |
| 1590 | |
| 1591 | /* |
| 1592 | * If we're expecting a response, block and wait for it. |
| 1593 | */ |
| 1594 | if (resp) { |
| 1595 | if (callback) { |
| 1596 | DDPRINTF("Issuing callback before receiving response...\n"); |
| 1597 | ret = callback(rdma); |
| 1598 | if (ret < 0) { |
| 1599 | return ret; |
| 1600 | } |
| 1601 | } |
| 1602 | |
| 1603 | DDPRINTF("Waiting for response %s\n", control_desc[resp->type]); |
| 1604 | ret = qemu_rdma_exchange_get_response(rdma, resp, |
| 1605 | resp->type, RDMA_WRID_DATA); |
| 1606 | |
| 1607 | if (ret < 0) { |
| 1608 | return ret; |
| 1609 | } |
| 1610 | |
| 1611 | qemu_rdma_move_header(rdma, RDMA_WRID_DATA, resp); |
| 1612 | if (resp_idx) { |
| 1613 | *resp_idx = RDMA_WRID_DATA; |
| 1614 | } |
| 1615 | DDPRINTF("Response %s received.\n", control_desc[resp->type]); |
| 1616 | } |
| 1617 | |
| 1618 | rdma->control_ready_expected = 1; |
| 1619 | |
| 1620 | return 0; |
| 1621 | } |
| 1622 | |
| 1623 | /* |
| 1624 | * This is an 'atomic' high-level operation to receive a single, unified |
| 1625 | * control-channel message. |
| 1626 | */ |
| 1627 | static int qemu_rdma_exchange_recv(RDMAContext *rdma, RDMAControlHeader *head, |
| 1628 | int expecting) |
| 1629 | { |
| 1630 | RDMAControlHeader ready = { |
| 1631 | .len = 0, |
| 1632 | .type = RDMA_CONTROL_READY, |
| 1633 | .repeat = 1, |
| 1634 | }; |
| 1635 | int ret; |
| 1636 | |
| 1637 | /* |
| 1638 | * Inform the source that we're ready to receive a message. |
| 1639 | */ |
| 1640 | ret = qemu_rdma_post_send_control(rdma, NULL, &ready); |
| 1641 | |
| 1642 | if (ret < 0) { |
| 1643 | fprintf(stderr, "Failed to send control buffer!\n"); |
| 1644 | return ret; |
| 1645 | } |
| 1646 | |
| 1647 | /* |
| 1648 | * Block and wait for the message. |
| 1649 | */ |
| 1650 | ret = qemu_rdma_exchange_get_response(rdma, head, |
| 1651 | expecting, RDMA_WRID_READY); |
| 1652 | |
| 1653 | if (ret < 0) { |
| 1654 | return ret; |
| 1655 | } |
| 1656 | |
| 1657 | qemu_rdma_move_header(rdma, RDMA_WRID_READY, head); |
| 1658 | |
| 1659 | /* |
| 1660 | * Post a new RECV work request to replace the one we just consumed. |
| 1661 | */ |
| 1662 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
| 1663 | if (ret) { |
| 1664 | fprintf(stderr, "rdma migration: error posting second control recv!"); |
| 1665 | return ret; |
| 1666 | } |
| 1667 | |
| 1668 | return 0; |
| 1669 | } |
| 1670 | |
| 1671 | /* |
| 1672 | * Write an actual chunk of memory using RDMA. |
| 1673 | * |
| 1674 | * If we're using dynamic registration on the dest-side, we have to |
| 1675 | * send a registration command first. |
| 1676 | */ |
| 1677 | static int qemu_rdma_write_one(QEMUFile *f, RDMAContext *rdma, |
| 1678 | int current_index, uint64_t current_addr, |
| 1679 | uint64_t length) |
| 1680 | { |
| 1681 | struct ibv_sge sge; |
| 1682 | struct ibv_send_wr send_wr = { 0 }; |
| 1683 | struct ibv_send_wr *bad_wr; |
| 1684 | int reg_result_idx, ret, count = 0; |
| 1685 | uint64_t chunk, chunks; |
| 1686 | uint8_t *chunk_start, *chunk_end; |
| 1687 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[current_index]); |
| 1688 | RDMARegister reg; |
| 1689 | RDMARegisterResult *reg_result; |
| 1690 | RDMAControlHeader resp = { .type = RDMA_CONTROL_REGISTER_RESULT }; |
| 1691 | RDMAControlHeader head = { .len = sizeof(RDMARegister), |
| 1692 | .type = RDMA_CONTROL_REGISTER_REQUEST, |
| 1693 | .repeat = 1, |
| 1694 | }; |
| 1695 | |
| 1696 | retry: |
| 1697 | sge.addr = (uint64_t)(block->local_host_addr + |
| 1698 | (current_addr - block->offset)); |
| 1699 | sge.length = length; |
| 1700 | |
| 1701 | chunk = ram_chunk_index(block->local_host_addr, (uint8_t *) sge.addr); |
| 1702 | chunk_start = ram_chunk_start(block, chunk); |
| 1703 | |
| 1704 | if (block->is_ram_block) { |
| 1705 | chunks = length / (1UL << RDMA_REG_CHUNK_SHIFT); |
| 1706 | |
| 1707 | if (chunks && ((length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { |
| 1708 | chunks--; |
| 1709 | } |
| 1710 | } else { |
| 1711 | chunks = block->length / (1UL << RDMA_REG_CHUNK_SHIFT); |
| 1712 | |
| 1713 | if (chunks && ((block->length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { |
| 1714 | chunks--; |
| 1715 | } |
| 1716 | } |
| 1717 | |
| 1718 | DDPRINTF("Writing %" PRIu64 " chunks, (%" PRIu64 " MB)\n", |
| 1719 | chunks + 1, (chunks + 1) * (1UL << RDMA_REG_CHUNK_SHIFT) / 1024 / 1024); |
| 1720 | |
| 1721 | chunk_end = ram_chunk_end(block, chunk + chunks); |
| 1722 | |
| 1723 | if (!rdma->pin_all) { |
| 1724 | #ifdef RDMA_UNREGISTRATION_EXAMPLE |
| 1725 | qemu_rdma_unregister_waiting(rdma); |
| 1726 | #endif |
| 1727 | } |
| 1728 | |
| 1729 | while (test_bit(chunk, block->transit_bitmap)) { |
| 1730 | (void)count; |
| 1731 | DDPRINTF("(%d) Not clobbering: block: %d chunk %" PRIu64 |
| 1732 | " current %" PRIu64 " len %" PRIu64 " %d %d\n", |
| 1733 | count++, current_index, chunk, |
| 1734 | sge.addr, length, rdma->nb_sent, block->nb_chunks); |
| 1735 | |
| 1736 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE); |
| 1737 | |
| 1738 | if (ret < 0) { |
| 1739 | fprintf(stderr, "Failed to Wait for previous write to complete " |
| 1740 | "block %d chunk %" PRIu64 |
| 1741 | " current %" PRIu64 " len %" PRIu64 " %d\n", |
| 1742 | current_index, chunk, sge.addr, length, rdma->nb_sent); |
| 1743 | return ret; |
| 1744 | } |
| 1745 | } |
| 1746 | |
| 1747 | if (!rdma->pin_all || !block->is_ram_block) { |
| 1748 | if (!block->remote_keys[chunk]) { |
| 1749 | /* |
| 1750 | * This chunk has not yet been registered, so first check to see |
| 1751 | * if the entire chunk is zero. If so, tell the other size to |
| 1752 | * memset() + madvise() the entire chunk without RDMA. |
| 1753 | */ |
| 1754 | |
| 1755 | if (can_use_buffer_find_nonzero_offset((void *)sge.addr, length) |
| 1756 | && buffer_find_nonzero_offset((void *)sge.addr, |
| 1757 | length) == length) { |
| 1758 | RDMACompress comp = { |
| 1759 | .offset = current_addr, |
| 1760 | .value = 0, |
| 1761 | .block_idx = current_index, |
| 1762 | .length = length, |
| 1763 | }; |
| 1764 | |
| 1765 | head.len = sizeof(comp); |
| 1766 | head.type = RDMA_CONTROL_COMPRESS; |
| 1767 | |
| 1768 | DDPRINTF("Entire chunk is zero, sending compress: %" |
| 1769 | PRIu64 " for %d " |
| 1770 | "bytes, index: %d, offset: %" PRId64 "...\n", |
| 1771 | chunk, sge.length, current_index, current_addr); |
| 1772 | |
| 1773 | compress_to_network(&comp); |
| 1774 | ret = qemu_rdma_exchange_send(rdma, &head, |
| 1775 | (uint8_t *) &comp, NULL, NULL, NULL); |
| 1776 | |
| 1777 | if (ret < 0) { |
| 1778 | return -EIO; |
| 1779 | } |
| 1780 | |
| 1781 | acct_update_position(f, sge.length, true); |
| 1782 | |
| 1783 | return 1; |
| 1784 | } |
| 1785 | |
| 1786 | /* |
| 1787 | * Otherwise, tell other side to register. |
| 1788 | */ |
| 1789 | reg.current_index = current_index; |
| 1790 | if (block->is_ram_block) { |
| 1791 | reg.key.current_addr = current_addr; |
| 1792 | } else { |
| 1793 | reg.key.chunk = chunk; |
| 1794 | } |
| 1795 | reg.chunks = chunks; |
| 1796 | |
| 1797 | DDPRINTF("Sending registration request chunk %" PRIu64 " for %d " |
| 1798 | "bytes, index: %d, offset: %" PRId64 "...\n", |
| 1799 | chunk, sge.length, current_index, current_addr); |
| 1800 | |
| 1801 | register_to_network(®); |
| 1802 | ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
| 1803 | &resp, ®_result_idx, NULL); |
| 1804 | if (ret < 0) { |
| 1805 | return ret; |
| 1806 | } |
| 1807 | |
| 1808 | /* try to overlap this single registration with the one we sent. */ |
| 1809 | if (qemu_rdma_register_and_get_keys(rdma, block, |
| 1810 | (uint8_t *) sge.addr, |
| 1811 | &sge.lkey, NULL, chunk, |
| 1812 | chunk_start, chunk_end)) { |
| 1813 | fprintf(stderr, "cannot get lkey!\n"); |
| 1814 | return -EINVAL; |
| 1815 | } |
| 1816 | |
| 1817 | reg_result = (RDMARegisterResult *) |
| 1818 | rdma->wr_data[reg_result_idx].control_curr; |
| 1819 | |
| 1820 | network_to_result(reg_result); |
| 1821 | |
| 1822 | DDPRINTF("Received registration result:" |
| 1823 | " my key: %x their key %x, chunk %" PRIu64 "\n", |
| 1824 | block->remote_keys[chunk], reg_result->rkey, chunk); |
| 1825 | |
| 1826 | block->remote_keys[chunk] = reg_result->rkey; |
| 1827 | block->remote_host_addr = reg_result->host_addr; |
| 1828 | } else { |
| 1829 | /* already registered before */ |
| 1830 | if (qemu_rdma_register_and_get_keys(rdma, block, |
| 1831 | (uint8_t *)sge.addr, |
| 1832 | &sge.lkey, NULL, chunk, |
| 1833 | chunk_start, chunk_end)) { |
| 1834 | fprintf(stderr, "cannot get lkey!\n"); |
| 1835 | return -EINVAL; |
| 1836 | } |
| 1837 | } |
| 1838 | |
| 1839 | send_wr.wr.rdma.rkey = block->remote_keys[chunk]; |
| 1840 | } else { |
| 1841 | send_wr.wr.rdma.rkey = block->remote_rkey; |
| 1842 | |
| 1843 | if (qemu_rdma_register_and_get_keys(rdma, block, (uint8_t *)sge.addr, |
| 1844 | &sge.lkey, NULL, chunk, |
| 1845 | chunk_start, chunk_end)) { |
| 1846 | fprintf(stderr, "cannot get lkey!\n"); |
| 1847 | return -EINVAL; |
| 1848 | } |
| 1849 | } |
| 1850 | |
| 1851 | /* |
| 1852 | * Encode the ram block index and chunk within this wrid. |
| 1853 | * We will use this information at the time of completion |
| 1854 | * to figure out which bitmap to check against and then which |
| 1855 | * chunk in the bitmap to look for. |
| 1856 | */ |
| 1857 | send_wr.wr_id = qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE, |
| 1858 | current_index, chunk); |
| 1859 | |
| 1860 | send_wr.opcode = IBV_WR_RDMA_WRITE; |
| 1861 | send_wr.send_flags = IBV_SEND_SIGNALED; |
| 1862 | send_wr.sg_list = &sge; |
| 1863 | send_wr.num_sge = 1; |
| 1864 | send_wr.wr.rdma.remote_addr = block->remote_host_addr + |
| 1865 | (current_addr - block->offset); |
| 1866 | |
| 1867 | DDDPRINTF("Posting chunk: %" PRIu64 ", addr: %lx" |
| 1868 | " remote: %lx, bytes %" PRIu32 "\n", |
| 1869 | chunk, sge.addr, send_wr.wr.rdma.remote_addr, |
| 1870 | sge.length); |
| 1871 | |
| 1872 | /* |
| 1873 | * ibv_post_send() does not return negative error numbers, |
| 1874 | * per the specification they are positive - no idea why. |
| 1875 | */ |
| 1876 | ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr); |
| 1877 | |
| 1878 | if (ret == ENOMEM) { |
| 1879 | DDPRINTF("send queue is full. wait a little....\n"); |
| 1880 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE); |
| 1881 | if (ret < 0) { |
| 1882 | fprintf(stderr, "rdma migration: failed to make " |
| 1883 | "room in full send queue! %d\n", ret); |
| 1884 | return ret; |
| 1885 | } |
| 1886 | |
| 1887 | goto retry; |
| 1888 | |
| 1889 | } else if (ret > 0) { |
| 1890 | perror("rdma migration: post rdma write failed"); |
| 1891 | return -ret; |
| 1892 | } |
| 1893 | |
| 1894 | set_bit(chunk, block->transit_bitmap); |
| 1895 | acct_update_position(f, sge.length, false); |
| 1896 | rdma->total_writes++; |
| 1897 | |
| 1898 | return 0; |
| 1899 | } |
| 1900 | |
| 1901 | /* |
| 1902 | * Push out any unwritten RDMA operations. |
| 1903 | * |
| 1904 | * We support sending out multiple chunks at the same time. |
| 1905 | * Not all of them need to get signaled in the completion queue. |
| 1906 | */ |
| 1907 | static int qemu_rdma_write_flush(QEMUFile *f, RDMAContext *rdma) |
| 1908 | { |
| 1909 | int ret; |
| 1910 | |
| 1911 | if (!rdma->current_length) { |
| 1912 | return 0; |
| 1913 | } |
| 1914 | |
| 1915 | ret = qemu_rdma_write_one(f, rdma, |
| 1916 | rdma->current_index, rdma->current_addr, rdma->current_length); |
| 1917 | |
| 1918 | if (ret < 0) { |
| 1919 | return ret; |
| 1920 | } |
| 1921 | |
| 1922 | if (ret == 0) { |
| 1923 | rdma->nb_sent++; |
| 1924 | DDDPRINTF("sent total: %d\n", rdma->nb_sent); |
| 1925 | } |
| 1926 | |
| 1927 | rdma->current_length = 0; |
| 1928 | rdma->current_addr = 0; |
| 1929 | |
| 1930 | return 0; |
| 1931 | } |
| 1932 | |
| 1933 | static inline int qemu_rdma_buffer_mergable(RDMAContext *rdma, |
| 1934 | uint64_t offset, uint64_t len) |
| 1935 | { |
Isaku Yamahata | 44b5949 | 2013-08-03 22:54:51 -0400 | [diff] [blame] | 1936 | RDMALocalBlock *block; |
| 1937 | uint8_t *host_addr; |
| 1938 | uint8_t *chunk_end; |
| 1939 | |
| 1940 | if (rdma->current_index < 0) { |
| 1941 | return 0; |
| 1942 | } |
| 1943 | |
| 1944 | if (rdma->current_chunk < 0) { |
| 1945 | return 0; |
| 1946 | } |
| 1947 | |
| 1948 | block = &(rdma->local_ram_blocks.block[rdma->current_index]); |
| 1949 | host_addr = block->local_host_addr + (offset - block->offset); |
| 1950 | chunk_end = ram_chunk_end(block, rdma->current_chunk); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 1951 | |
| 1952 | if (rdma->current_length == 0) { |
| 1953 | return 0; |
| 1954 | } |
| 1955 | |
| 1956 | /* |
| 1957 | * Only merge into chunk sequentially. |
| 1958 | */ |
| 1959 | if (offset != (rdma->current_addr + rdma->current_length)) { |
| 1960 | return 0; |
| 1961 | } |
| 1962 | |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 1963 | if (offset < block->offset) { |
| 1964 | return 0; |
| 1965 | } |
| 1966 | |
| 1967 | if ((offset + len) > (block->offset + block->length)) { |
| 1968 | return 0; |
| 1969 | } |
| 1970 | |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 1971 | if ((host_addr + len) > chunk_end) { |
| 1972 | return 0; |
| 1973 | } |
| 1974 | |
| 1975 | return 1; |
| 1976 | } |
| 1977 | |
| 1978 | /* |
| 1979 | * We're not actually writing here, but doing three things: |
| 1980 | * |
| 1981 | * 1. Identify the chunk the buffer belongs to. |
| 1982 | * 2. If the chunk is full or the buffer doesn't belong to the current |
| 1983 | * chunk, then start a new chunk and flush() the old chunk. |
| 1984 | * 3. To keep the hardware busy, we also group chunks into batches |
| 1985 | * and only require that a batch gets acknowledged in the completion |
| 1986 | * qeueue instead of each individual chunk. |
| 1987 | */ |
| 1988 | static int qemu_rdma_write(QEMUFile *f, RDMAContext *rdma, |
| 1989 | uint64_t block_offset, uint64_t offset, |
| 1990 | uint64_t len) |
| 1991 | { |
| 1992 | uint64_t current_addr = block_offset + offset; |
| 1993 | uint64_t index = rdma->current_index; |
| 1994 | uint64_t chunk = rdma->current_chunk; |
| 1995 | int ret; |
| 1996 | |
| 1997 | /* If we cannot merge it, we flush the current buffer first. */ |
| 1998 | if (!qemu_rdma_buffer_mergable(rdma, current_addr, len)) { |
| 1999 | ret = qemu_rdma_write_flush(f, rdma); |
| 2000 | if (ret) { |
| 2001 | return ret; |
| 2002 | } |
| 2003 | rdma->current_length = 0; |
| 2004 | rdma->current_addr = current_addr; |
| 2005 | |
| 2006 | ret = qemu_rdma_search_ram_block(rdma, block_offset, |
| 2007 | offset, len, &index, &chunk); |
| 2008 | if (ret) { |
| 2009 | fprintf(stderr, "ram block search failed\n"); |
| 2010 | return ret; |
| 2011 | } |
| 2012 | rdma->current_index = index; |
| 2013 | rdma->current_chunk = chunk; |
| 2014 | } |
| 2015 | |
| 2016 | /* merge it */ |
| 2017 | rdma->current_length += len; |
| 2018 | |
| 2019 | /* flush it if buffer is too large */ |
| 2020 | if (rdma->current_length >= RDMA_MERGE_MAX) { |
| 2021 | return qemu_rdma_write_flush(f, rdma); |
| 2022 | } |
| 2023 | |
| 2024 | return 0; |
| 2025 | } |
| 2026 | |
| 2027 | static void qemu_rdma_cleanup(RDMAContext *rdma) |
| 2028 | { |
| 2029 | struct rdma_cm_event *cm_event; |
| 2030 | int ret, idx; |
| 2031 | |
| 2032 | if (rdma->cm_id) { |
| 2033 | if (rdma->error_state) { |
| 2034 | RDMAControlHeader head = { .len = 0, |
| 2035 | .type = RDMA_CONTROL_ERROR, |
| 2036 | .repeat = 1, |
| 2037 | }; |
| 2038 | fprintf(stderr, "Early error. Sending error.\n"); |
| 2039 | qemu_rdma_post_send_control(rdma, NULL, &head); |
| 2040 | } |
| 2041 | |
| 2042 | ret = rdma_disconnect(rdma->cm_id); |
| 2043 | if (!ret) { |
| 2044 | DDPRINTF("waiting for disconnect\n"); |
| 2045 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2046 | if (!ret) { |
| 2047 | rdma_ack_cm_event(cm_event); |
| 2048 | } |
| 2049 | } |
| 2050 | DDPRINTF("Disconnected.\n"); |
| 2051 | rdma->cm_id = NULL; |
| 2052 | } |
| 2053 | |
| 2054 | g_free(rdma->block); |
| 2055 | rdma->block = NULL; |
| 2056 | |
Isaku Yamahata | 1f22364 | 2013-08-03 22:54:52 -0400 | [diff] [blame] | 2057 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2058 | if (rdma->wr_data[idx].control_mr) { |
| 2059 | rdma->total_registrations--; |
| 2060 | ibv_dereg_mr(rdma->wr_data[idx].control_mr); |
| 2061 | } |
| 2062 | rdma->wr_data[idx].control_mr = NULL; |
| 2063 | } |
| 2064 | |
| 2065 | if (rdma->local_ram_blocks.block) { |
| 2066 | while (rdma->local_ram_blocks.nb_blocks) { |
| 2067 | __qemu_rdma_delete_block(rdma, |
| 2068 | rdma->local_ram_blocks.block->offset); |
| 2069 | } |
| 2070 | } |
| 2071 | |
| 2072 | if (rdma->qp) { |
| 2073 | ibv_destroy_qp(rdma->qp); |
| 2074 | rdma->qp = NULL; |
| 2075 | } |
| 2076 | if (rdma->cq) { |
| 2077 | ibv_destroy_cq(rdma->cq); |
| 2078 | rdma->cq = NULL; |
| 2079 | } |
| 2080 | if (rdma->comp_channel) { |
| 2081 | ibv_destroy_comp_channel(rdma->comp_channel); |
| 2082 | rdma->comp_channel = NULL; |
| 2083 | } |
| 2084 | if (rdma->pd) { |
| 2085 | ibv_dealloc_pd(rdma->pd); |
| 2086 | rdma->pd = NULL; |
| 2087 | } |
| 2088 | if (rdma->listen_id) { |
| 2089 | rdma_destroy_id(rdma->listen_id); |
| 2090 | rdma->listen_id = NULL; |
| 2091 | } |
| 2092 | if (rdma->cm_id) { |
| 2093 | rdma_destroy_id(rdma->cm_id); |
| 2094 | rdma->cm_id = NULL; |
| 2095 | } |
| 2096 | if (rdma->channel) { |
| 2097 | rdma_destroy_event_channel(rdma->channel); |
| 2098 | rdma->channel = NULL; |
| 2099 | } |
Isaku Yamahata | e1d0fb3 | 2013-08-03 22:54:54 -0400 | [diff] [blame^] | 2100 | g_free(rdma->host); |
| 2101 | rdma->host = NULL; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2102 | } |
| 2103 | |
| 2104 | |
| 2105 | static int qemu_rdma_source_init(RDMAContext *rdma, Error **errp, bool pin_all) |
| 2106 | { |
| 2107 | int ret, idx; |
| 2108 | Error *local_err = NULL, **temp = &local_err; |
| 2109 | |
| 2110 | /* |
| 2111 | * Will be validated against destination's actual capabilities |
| 2112 | * after the connect() completes. |
| 2113 | */ |
| 2114 | rdma->pin_all = pin_all; |
| 2115 | |
| 2116 | ret = qemu_rdma_resolve_host(rdma, temp); |
| 2117 | if (ret) { |
| 2118 | goto err_rdma_source_init; |
| 2119 | } |
| 2120 | |
| 2121 | ret = qemu_rdma_alloc_pd_cq(rdma); |
| 2122 | if (ret) { |
| 2123 | ERROR(temp, "rdma migration: error allocating pd and cq! Your mlock()" |
| 2124 | " limits may be too low. Please check $ ulimit -a # and " |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2125 | "search for 'ulimit -l' in the output"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2126 | goto err_rdma_source_init; |
| 2127 | } |
| 2128 | |
| 2129 | ret = qemu_rdma_alloc_qp(rdma); |
| 2130 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2131 | ERROR(temp, "rdma migration: error allocating qp!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2132 | goto err_rdma_source_init; |
| 2133 | } |
| 2134 | |
| 2135 | ret = qemu_rdma_init_ram_blocks(rdma); |
| 2136 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2137 | ERROR(temp, "rdma migration: error initializing ram blocks!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2138 | goto err_rdma_source_init; |
| 2139 | } |
| 2140 | |
Isaku Yamahata | 1f22364 | 2013-08-03 22:54:52 -0400 | [diff] [blame] | 2141 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2142 | ret = qemu_rdma_reg_control(rdma, idx); |
| 2143 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2144 | ERROR(temp, "rdma migration: error registering %d control!", |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2145 | idx); |
| 2146 | goto err_rdma_source_init; |
| 2147 | } |
| 2148 | } |
| 2149 | |
| 2150 | return 0; |
| 2151 | |
| 2152 | err_rdma_source_init: |
| 2153 | error_propagate(errp, local_err); |
| 2154 | qemu_rdma_cleanup(rdma); |
| 2155 | return -1; |
| 2156 | } |
| 2157 | |
| 2158 | static int qemu_rdma_connect(RDMAContext *rdma, Error **errp) |
| 2159 | { |
| 2160 | RDMACapabilities cap = { |
| 2161 | .version = RDMA_CONTROL_VERSION_CURRENT, |
| 2162 | .flags = 0, |
| 2163 | }; |
| 2164 | struct rdma_conn_param conn_param = { .initiator_depth = 2, |
| 2165 | .retry_count = 5, |
| 2166 | .private_data = &cap, |
| 2167 | .private_data_len = sizeof(cap), |
| 2168 | }; |
| 2169 | struct rdma_cm_event *cm_event; |
| 2170 | int ret; |
| 2171 | |
| 2172 | /* |
| 2173 | * Only negotiate the capability with destination if the user |
| 2174 | * on the source first requested the capability. |
| 2175 | */ |
| 2176 | if (rdma->pin_all) { |
| 2177 | DPRINTF("Server pin-all memory requested.\n"); |
| 2178 | cap.flags |= RDMA_CAPABILITY_PIN_ALL; |
| 2179 | } |
| 2180 | |
| 2181 | caps_to_network(&cap); |
| 2182 | |
| 2183 | ret = rdma_connect(rdma->cm_id, &conn_param); |
| 2184 | if (ret) { |
| 2185 | perror("rdma_connect"); |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2186 | ERROR(errp, "connecting to destination!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2187 | rdma_destroy_id(rdma->cm_id); |
| 2188 | rdma->cm_id = NULL; |
| 2189 | goto err_rdma_source_connect; |
| 2190 | } |
| 2191 | |
| 2192 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2193 | if (ret) { |
| 2194 | perror("rdma_get_cm_event after rdma_connect"); |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2195 | ERROR(errp, "connecting to destination!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2196 | rdma_ack_cm_event(cm_event); |
| 2197 | rdma_destroy_id(rdma->cm_id); |
| 2198 | rdma->cm_id = NULL; |
| 2199 | goto err_rdma_source_connect; |
| 2200 | } |
| 2201 | |
| 2202 | if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) { |
| 2203 | perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect"); |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2204 | ERROR(errp, "connecting to destination!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2205 | rdma_ack_cm_event(cm_event); |
| 2206 | rdma_destroy_id(rdma->cm_id); |
| 2207 | rdma->cm_id = NULL; |
| 2208 | goto err_rdma_source_connect; |
| 2209 | } |
| 2210 | |
| 2211 | memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap)); |
| 2212 | network_to_caps(&cap); |
| 2213 | |
| 2214 | /* |
| 2215 | * Verify that the *requested* capabilities are supported by the destination |
| 2216 | * and disable them otherwise. |
| 2217 | */ |
| 2218 | if (rdma->pin_all && !(cap.flags & RDMA_CAPABILITY_PIN_ALL)) { |
| 2219 | ERROR(errp, "Server cannot support pinning all memory. " |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2220 | "Will register memory dynamically."); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2221 | rdma->pin_all = false; |
| 2222 | } |
| 2223 | |
| 2224 | DPRINTF("Pin all memory: %s\n", rdma->pin_all ? "enabled" : "disabled"); |
| 2225 | |
| 2226 | rdma_ack_cm_event(cm_event); |
| 2227 | |
Isaku Yamahata | 8777263 | 2013-08-03 22:54:53 -0400 | [diff] [blame] | 2228 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2229 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2230 | ERROR(errp, "posting second control recv!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2231 | goto err_rdma_source_connect; |
| 2232 | } |
| 2233 | |
| 2234 | rdma->control_ready_expected = 1; |
| 2235 | rdma->nb_sent = 0; |
| 2236 | return 0; |
| 2237 | |
| 2238 | err_rdma_source_connect: |
| 2239 | qemu_rdma_cleanup(rdma); |
| 2240 | return -1; |
| 2241 | } |
| 2242 | |
| 2243 | static int qemu_rdma_dest_init(RDMAContext *rdma, Error **errp) |
| 2244 | { |
| 2245 | int ret = -EINVAL, idx; |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2246 | int af = rdma->ipv6 ? PF_INET6 : PF_INET; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2247 | struct sockaddr_in sin; |
| 2248 | struct rdma_cm_id *listen_id; |
| 2249 | char ip[40] = "unknown"; |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2250 | struct addrinfo *res; |
| 2251 | char port_str[16]; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2252 | |
Isaku Yamahata | 1f22364 | 2013-08-03 22:54:52 -0400 | [diff] [blame] | 2253 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2254 | rdma->wr_data[idx].control_len = 0; |
| 2255 | rdma->wr_data[idx].control_curr = NULL; |
| 2256 | } |
| 2257 | |
| 2258 | if (rdma->host == NULL) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2259 | ERROR(errp, "RDMA host is not set!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2260 | rdma->error_state = -EINVAL; |
| 2261 | return -1; |
| 2262 | } |
| 2263 | /* create CM channel */ |
| 2264 | rdma->channel = rdma_create_event_channel(); |
| 2265 | if (!rdma->channel) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2266 | ERROR(errp, "could not create rdma event channel"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2267 | rdma->error_state = -EINVAL; |
| 2268 | return -1; |
| 2269 | } |
| 2270 | |
| 2271 | /* create CM id */ |
| 2272 | ret = rdma_create_id(rdma->channel, &listen_id, NULL, RDMA_PS_TCP); |
| 2273 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2274 | ERROR(errp, "could not create cm_id!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2275 | goto err_dest_init_create_listen_id; |
| 2276 | } |
| 2277 | |
| 2278 | memset(&sin, 0, sizeof(sin)); |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2279 | sin.sin_family = af; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2280 | sin.sin_port = htons(rdma->port); |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2281 | snprintf(port_str, 16, "%d", rdma->port); |
| 2282 | port_str[15] = '\0'; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2283 | |
| 2284 | if (rdma->host && strcmp("", rdma->host)) { |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2285 | ret = getaddrinfo(rdma->host, port_str, NULL, &res); |
| 2286 | if (ret < 0) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2287 | ERROR(errp, "could not getaddrinfo address %s", rdma->host); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2288 | goto err_dest_init_bind_addr; |
| 2289 | } |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2290 | |
| 2291 | |
| 2292 | inet_ntop(af, &((struct sockaddr_in *) res->ai_addr)->sin_addr, |
| 2293 | ip, sizeof ip); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2294 | } else { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2295 | ERROR(errp, "migration host and port not specified!"); |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2296 | ret = -EINVAL; |
| 2297 | goto err_dest_init_bind_addr; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2298 | } |
| 2299 | |
| 2300 | DPRINTF("%s => %s\n", rdma->host, ip); |
| 2301 | |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2302 | ret = rdma_bind_addr(listen_id, res->ai_addr); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2303 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 2304 | ERROR(errp, "Error: could not rdma_bind_addr!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2305 | goto err_dest_init_bind_addr; |
| 2306 | } |
| 2307 | |
| 2308 | rdma->listen_id = listen_id; |
| 2309 | qemu_rdma_dump_gid("dest_init", listen_id); |
| 2310 | return 0; |
| 2311 | |
| 2312 | err_dest_init_bind_addr: |
| 2313 | rdma_destroy_id(listen_id); |
| 2314 | err_dest_init_create_listen_id: |
| 2315 | rdma_destroy_event_channel(rdma->channel); |
| 2316 | rdma->channel = NULL; |
| 2317 | rdma->error_state = ret; |
| 2318 | return ret; |
| 2319 | |
| 2320 | } |
| 2321 | |
| 2322 | static void *qemu_rdma_data_init(const char *host_port, Error **errp) |
| 2323 | { |
| 2324 | RDMAContext *rdma = NULL; |
| 2325 | InetSocketAddress *addr; |
| 2326 | |
| 2327 | if (host_port) { |
| 2328 | rdma = g_malloc0(sizeof(RDMAContext)); |
| 2329 | memset(rdma, 0, sizeof(RDMAContext)); |
| 2330 | rdma->current_index = -1; |
| 2331 | rdma->current_chunk = -1; |
| 2332 | |
| 2333 | addr = inet_parse(host_port, NULL); |
| 2334 | if (addr != NULL) { |
| 2335 | rdma->port = atoi(addr->port); |
| 2336 | rdma->host = g_strdup(addr->host); |
Michael R. Hines | b58c855 | 2013-08-03 22:54:48 -0400 | [diff] [blame] | 2337 | rdma->ipv6 = addr->ipv6; |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2338 | } else { |
| 2339 | ERROR(errp, "bad RDMA migration address '%s'", host_port); |
| 2340 | g_free(rdma); |
| 2341 | return NULL; |
| 2342 | } |
| 2343 | } |
| 2344 | |
| 2345 | return rdma; |
| 2346 | } |
| 2347 | |
| 2348 | /* |
| 2349 | * QEMUFile interface to the control channel. |
| 2350 | * SEND messages for control only. |
| 2351 | * pc.ram is handled with regular RDMA messages. |
| 2352 | */ |
| 2353 | static int qemu_rdma_put_buffer(void *opaque, const uint8_t *buf, |
| 2354 | int64_t pos, int size) |
| 2355 | { |
| 2356 | QEMUFileRDMA *r = opaque; |
| 2357 | QEMUFile *f = r->file; |
| 2358 | RDMAContext *rdma = r->rdma; |
| 2359 | size_t remaining = size; |
| 2360 | uint8_t * data = (void *) buf; |
| 2361 | int ret; |
| 2362 | |
| 2363 | CHECK_ERROR_STATE(); |
| 2364 | |
| 2365 | /* |
| 2366 | * Push out any writes that |
| 2367 | * we're queued up for pc.ram. |
| 2368 | */ |
| 2369 | ret = qemu_rdma_write_flush(f, rdma); |
| 2370 | if (ret < 0) { |
| 2371 | rdma->error_state = ret; |
| 2372 | return ret; |
| 2373 | } |
| 2374 | |
| 2375 | while (remaining) { |
| 2376 | RDMAControlHeader head; |
| 2377 | |
| 2378 | r->len = MIN(remaining, RDMA_SEND_INCREMENT); |
| 2379 | remaining -= r->len; |
| 2380 | |
| 2381 | head.len = r->len; |
| 2382 | head.type = RDMA_CONTROL_QEMU_FILE; |
| 2383 | |
| 2384 | ret = qemu_rdma_exchange_send(rdma, &head, data, NULL, NULL, NULL); |
| 2385 | |
| 2386 | if (ret < 0) { |
| 2387 | rdma->error_state = ret; |
| 2388 | return ret; |
| 2389 | } |
| 2390 | |
| 2391 | data += r->len; |
| 2392 | } |
| 2393 | |
| 2394 | return size; |
| 2395 | } |
| 2396 | |
| 2397 | static size_t qemu_rdma_fill(RDMAContext *rdma, uint8_t *buf, |
| 2398 | int size, int idx) |
| 2399 | { |
| 2400 | size_t len = 0; |
| 2401 | |
| 2402 | if (rdma->wr_data[idx].control_len) { |
| 2403 | DDDPRINTF("RDMA %" PRId64 " of %d bytes already in buffer\n", |
| 2404 | rdma->wr_data[idx].control_len, size); |
| 2405 | |
| 2406 | len = MIN(size, rdma->wr_data[idx].control_len); |
| 2407 | memcpy(buf, rdma->wr_data[idx].control_curr, len); |
| 2408 | rdma->wr_data[idx].control_curr += len; |
| 2409 | rdma->wr_data[idx].control_len -= len; |
| 2410 | } |
| 2411 | |
| 2412 | return len; |
| 2413 | } |
| 2414 | |
| 2415 | /* |
| 2416 | * QEMUFile interface to the control channel. |
| 2417 | * RDMA links don't use bytestreams, so we have to |
| 2418 | * return bytes to QEMUFile opportunistically. |
| 2419 | */ |
| 2420 | static int qemu_rdma_get_buffer(void *opaque, uint8_t *buf, |
| 2421 | int64_t pos, int size) |
| 2422 | { |
| 2423 | QEMUFileRDMA *r = opaque; |
| 2424 | RDMAContext *rdma = r->rdma; |
| 2425 | RDMAControlHeader head; |
| 2426 | int ret = 0; |
| 2427 | |
| 2428 | CHECK_ERROR_STATE(); |
| 2429 | |
| 2430 | /* |
| 2431 | * First, we hold on to the last SEND message we |
| 2432 | * were given and dish out the bytes until we run |
| 2433 | * out of bytes. |
| 2434 | */ |
| 2435 | r->len = qemu_rdma_fill(r->rdma, buf, size, 0); |
| 2436 | if (r->len) { |
| 2437 | return r->len; |
| 2438 | } |
| 2439 | |
| 2440 | /* |
| 2441 | * Once we run out, we block and wait for another |
| 2442 | * SEND message to arrive. |
| 2443 | */ |
| 2444 | ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_QEMU_FILE); |
| 2445 | |
| 2446 | if (ret < 0) { |
| 2447 | rdma->error_state = ret; |
| 2448 | return ret; |
| 2449 | } |
| 2450 | |
| 2451 | /* |
| 2452 | * SEND was received with new bytes, now try again. |
| 2453 | */ |
| 2454 | return qemu_rdma_fill(r->rdma, buf, size, 0); |
| 2455 | } |
| 2456 | |
| 2457 | /* |
| 2458 | * Block until all the outstanding chunks have been delivered by the hardware. |
| 2459 | */ |
| 2460 | static int qemu_rdma_drain_cq(QEMUFile *f, RDMAContext *rdma) |
| 2461 | { |
| 2462 | int ret; |
| 2463 | |
| 2464 | if (qemu_rdma_write_flush(f, rdma) < 0) { |
| 2465 | return -EIO; |
| 2466 | } |
| 2467 | |
| 2468 | while (rdma->nb_sent) { |
| 2469 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE); |
| 2470 | if (ret < 0) { |
| 2471 | fprintf(stderr, "rdma migration: complete polling error!\n"); |
| 2472 | return -EIO; |
| 2473 | } |
| 2474 | } |
| 2475 | |
| 2476 | qemu_rdma_unregister_waiting(rdma); |
| 2477 | |
| 2478 | return 0; |
| 2479 | } |
| 2480 | |
| 2481 | static int qemu_rdma_close(void *opaque) |
| 2482 | { |
| 2483 | DPRINTF("Shutting down connection.\n"); |
| 2484 | QEMUFileRDMA *r = opaque; |
| 2485 | if (r->rdma) { |
| 2486 | qemu_rdma_cleanup(r->rdma); |
| 2487 | g_free(r->rdma); |
| 2488 | } |
| 2489 | g_free(r); |
| 2490 | return 0; |
| 2491 | } |
| 2492 | |
| 2493 | /* |
| 2494 | * Parameters: |
| 2495 | * @offset == 0 : |
| 2496 | * This means that 'block_offset' is a full virtual address that does not |
| 2497 | * belong to a RAMBlock of the virtual machine and instead |
| 2498 | * represents a private malloc'd memory area that the caller wishes to |
| 2499 | * transfer. |
| 2500 | * |
| 2501 | * @offset != 0 : |
| 2502 | * Offset is an offset to be added to block_offset and used |
| 2503 | * to also lookup the corresponding RAMBlock. |
| 2504 | * |
| 2505 | * @size > 0 : |
| 2506 | * Initiate an transfer this size. |
| 2507 | * |
| 2508 | * @size == 0 : |
| 2509 | * A 'hint' or 'advice' that means that we wish to speculatively |
| 2510 | * and asynchronously unregister this memory. In this case, there is no |
Stefan Weil | 52f3502 | 2013-07-24 19:48:56 +0200 | [diff] [blame] | 2511 | * guarantee that the unregister will actually happen, for example, |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2512 | * if the memory is being actively transmitted. Additionally, the memory |
| 2513 | * may be re-registered at any future time if a write within the same |
| 2514 | * chunk was requested again, even if you attempted to unregister it |
| 2515 | * here. |
| 2516 | * |
| 2517 | * @size < 0 : TODO, not yet supported |
| 2518 | * Unregister the memory NOW. This means that the caller does not |
| 2519 | * expect there to be any future RDMA transfers and we just want to clean |
| 2520 | * things up. This is used in case the upper layer owns the memory and |
| 2521 | * cannot wait for qemu_fclose() to occur. |
| 2522 | * |
| 2523 | * @bytes_sent : User-specificed pointer to indicate how many bytes were |
| 2524 | * sent. Usually, this will not be more than a few bytes of |
| 2525 | * the protocol because most transfers are sent asynchronously. |
| 2526 | */ |
| 2527 | static size_t qemu_rdma_save_page(QEMUFile *f, void *opaque, |
| 2528 | ram_addr_t block_offset, ram_addr_t offset, |
| 2529 | size_t size, int *bytes_sent) |
| 2530 | { |
| 2531 | QEMUFileRDMA *rfile = opaque; |
| 2532 | RDMAContext *rdma = rfile->rdma; |
| 2533 | int ret; |
| 2534 | |
| 2535 | CHECK_ERROR_STATE(); |
| 2536 | |
| 2537 | qemu_fflush(f); |
| 2538 | |
| 2539 | if (size > 0) { |
| 2540 | /* |
| 2541 | * Add this page to the current 'chunk'. If the chunk |
| 2542 | * is full, or the page doen't belong to the current chunk, |
| 2543 | * an actual RDMA write will occur and a new chunk will be formed. |
| 2544 | */ |
| 2545 | ret = qemu_rdma_write(f, rdma, block_offset, offset, size); |
| 2546 | if (ret < 0) { |
| 2547 | fprintf(stderr, "rdma migration: write error! %d\n", ret); |
| 2548 | goto err; |
| 2549 | } |
| 2550 | |
| 2551 | /* |
| 2552 | * We always return 1 bytes because the RDMA |
| 2553 | * protocol is completely asynchronous. We do not yet know |
| 2554 | * whether an identified chunk is zero or not because we're |
| 2555 | * waiting for other pages to potentially be merged with |
| 2556 | * the current chunk. So, we have to call qemu_update_position() |
| 2557 | * later on when the actual write occurs. |
| 2558 | */ |
| 2559 | if (bytes_sent) { |
| 2560 | *bytes_sent = 1; |
| 2561 | } |
| 2562 | } else { |
| 2563 | uint64_t index, chunk; |
| 2564 | |
| 2565 | /* TODO: Change QEMUFileOps prototype to be signed: size_t => long |
| 2566 | if (size < 0) { |
| 2567 | ret = qemu_rdma_drain_cq(f, rdma); |
| 2568 | if (ret < 0) { |
| 2569 | fprintf(stderr, "rdma: failed to synchronously drain" |
| 2570 | " completion queue before unregistration.\n"); |
| 2571 | goto err; |
| 2572 | } |
| 2573 | } |
| 2574 | */ |
| 2575 | |
| 2576 | ret = qemu_rdma_search_ram_block(rdma, block_offset, |
| 2577 | offset, size, &index, &chunk); |
| 2578 | |
| 2579 | if (ret) { |
| 2580 | fprintf(stderr, "ram block search failed\n"); |
| 2581 | goto err; |
| 2582 | } |
| 2583 | |
| 2584 | qemu_rdma_signal_unregister(rdma, index, chunk, 0); |
| 2585 | |
| 2586 | /* |
Stefan Weil | 52f3502 | 2013-07-24 19:48:56 +0200 | [diff] [blame] | 2587 | * TODO: Synchronous, guaranteed unregistration (should not occur during |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2588 | * fast-path). Otherwise, unregisters will process on the next call to |
| 2589 | * qemu_rdma_drain_cq() |
| 2590 | if (size < 0) { |
| 2591 | qemu_rdma_unregister_waiting(rdma); |
| 2592 | } |
| 2593 | */ |
| 2594 | } |
| 2595 | |
| 2596 | /* |
| 2597 | * Drain the Completion Queue if possible, but do not block, |
| 2598 | * just poll. |
| 2599 | * |
| 2600 | * If nothing to poll, the end of the iteration will do this |
| 2601 | * again to make sure we don't overflow the request queue. |
| 2602 | */ |
| 2603 | while (1) { |
| 2604 | uint64_t wr_id, wr_id_in; |
| 2605 | int ret = qemu_rdma_poll(rdma, &wr_id_in); |
| 2606 | if (ret < 0) { |
| 2607 | fprintf(stderr, "rdma migration: polling error! %d\n", ret); |
| 2608 | goto err; |
| 2609 | } |
| 2610 | |
| 2611 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; |
| 2612 | |
| 2613 | if (wr_id == RDMA_WRID_NONE) { |
| 2614 | break; |
| 2615 | } |
| 2616 | } |
| 2617 | |
| 2618 | return RAM_SAVE_CONTROL_DELAYED; |
| 2619 | err: |
| 2620 | rdma->error_state = ret; |
| 2621 | return ret; |
| 2622 | } |
| 2623 | |
| 2624 | static int qemu_rdma_accept(RDMAContext *rdma) |
| 2625 | { |
| 2626 | RDMACapabilities cap; |
| 2627 | struct rdma_conn_param conn_param = { |
| 2628 | .responder_resources = 2, |
| 2629 | .private_data = &cap, |
| 2630 | .private_data_len = sizeof(cap), |
| 2631 | }; |
| 2632 | struct rdma_cm_event *cm_event; |
| 2633 | struct ibv_context *verbs; |
| 2634 | int ret = -EINVAL; |
| 2635 | int idx; |
| 2636 | |
| 2637 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2638 | if (ret) { |
| 2639 | goto err_rdma_dest_wait; |
| 2640 | } |
| 2641 | |
| 2642 | if (cm_event->event != RDMA_CM_EVENT_CONNECT_REQUEST) { |
| 2643 | rdma_ack_cm_event(cm_event); |
| 2644 | goto err_rdma_dest_wait; |
| 2645 | } |
| 2646 | |
| 2647 | memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap)); |
| 2648 | |
| 2649 | network_to_caps(&cap); |
| 2650 | |
| 2651 | if (cap.version < 1 || cap.version > RDMA_CONTROL_VERSION_CURRENT) { |
| 2652 | fprintf(stderr, "Unknown source RDMA version: %d, bailing...\n", |
| 2653 | cap.version); |
| 2654 | rdma_ack_cm_event(cm_event); |
| 2655 | goto err_rdma_dest_wait; |
| 2656 | } |
| 2657 | |
| 2658 | /* |
| 2659 | * Respond with only the capabilities this version of QEMU knows about. |
| 2660 | */ |
| 2661 | cap.flags &= known_capabilities; |
| 2662 | |
| 2663 | /* |
| 2664 | * Enable the ones that we do know about. |
| 2665 | * Add other checks here as new ones are introduced. |
| 2666 | */ |
| 2667 | if (cap.flags & RDMA_CAPABILITY_PIN_ALL) { |
| 2668 | rdma->pin_all = true; |
| 2669 | } |
| 2670 | |
| 2671 | rdma->cm_id = cm_event->id; |
| 2672 | verbs = cm_event->id->verbs; |
| 2673 | |
| 2674 | rdma_ack_cm_event(cm_event); |
| 2675 | |
| 2676 | DPRINTF("Memory pin all: %s\n", rdma->pin_all ? "enabled" : "disabled"); |
| 2677 | |
| 2678 | caps_to_network(&cap); |
| 2679 | |
| 2680 | DPRINTF("verbs context after listen: %p\n", verbs); |
| 2681 | |
| 2682 | if (!rdma->verbs) { |
| 2683 | rdma->verbs = verbs; |
| 2684 | } else if (rdma->verbs != verbs) { |
| 2685 | fprintf(stderr, "ibv context not matching %p, %p!\n", |
| 2686 | rdma->verbs, verbs); |
| 2687 | goto err_rdma_dest_wait; |
| 2688 | } |
| 2689 | |
| 2690 | qemu_rdma_dump_id("dest_init", verbs); |
| 2691 | |
| 2692 | ret = qemu_rdma_alloc_pd_cq(rdma); |
| 2693 | if (ret) { |
| 2694 | fprintf(stderr, "rdma migration: error allocating pd and cq!\n"); |
| 2695 | goto err_rdma_dest_wait; |
| 2696 | } |
| 2697 | |
| 2698 | ret = qemu_rdma_alloc_qp(rdma); |
| 2699 | if (ret) { |
| 2700 | fprintf(stderr, "rdma migration: error allocating qp!\n"); |
| 2701 | goto err_rdma_dest_wait; |
| 2702 | } |
| 2703 | |
| 2704 | ret = qemu_rdma_init_ram_blocks(rdma); |
| 2705 | if (ret) { |
| 2706 | fprintf(stderr, "rdma migration: error initializing ram blocks!\n"); |
| 2707 | goto err_rdma_dest_wait; |
| 2708 | } |
| 2709 | |
Isaku Yamahata | 1f22364 | 2013-08-03 22:54:52 -0400 | [diff] [blame] | 2710 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2711 | ret = qemu_rdma_reg_control(rdma, idx); |
| 2712 | if (ret) { |
| 2713 | fprintf(stderr, "rdma: error registering %d control!\n", idx); |
| 2714 | goto err_rdma_dest_wait; |
| 2715 | } |
| 2716 | } |
| 2717 | |
| 2718 | qemu_set_fd_handler2(rdma->channel->fd, NULL, NULL, NULL, NULL); |
| 2719 | |
| 2720 | ret = rdma_accept(rdma->cm_id, &conn_param); |
| 2721 | if (ret) { |
| 2722 | fprintf(stderr, "rdma_accept returns %d!\n", ret); |
| 2723 | goto err_rdma_dest_wait; |
| 2724 | } |
| 2725 | |
| 2726 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2727 | if (ret) { |
| 2728 | fprintf(stderr, "rdma_accept get_cm_event failed %d!\n", ret); |
| 2729 | goto err_rdma_dest_wait; |
| 2730 | } |
| 2731 | |
| 2732 | if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) { |
| 2733 | fprintf(stderr, "rdma_accept not event established!\n"); |
| 2734 | rdma_ack_cm_event(cm_event); |
| 2735 | goto err_rdma_dest_wait; |
| 2736 | } |
| 2737 | |
| 2738 | rdma_ack_cm_event(cm_event); |
| 2739 | |
Isaku Yamahata | 8777263 | 2013-08-03 22:54:53 -0400 | [diff] [blame] | 2740 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 2741 | if (ret) { |
| 2742 | fprintf(stderr, "rdma migration: error posting second control recv!\n"); |
| 2743 | goto err_rdma_dest_wait; |
| 2744 | } |
| 2745 | |
| 2746 | qemu_rdma_dump_gid("dest_connect", rdma->cm_id); |
| 2747 | |
| 2748 | return 0; |
| 2749 | |
| 2750 | err_rdma_dest_wait: |
| 2751 | rdma->error_state = ret; |
| 2752 | qemu_rdma_cleanup(rdma); |
| 2753 | return ret; |
| 2754 | } |
| 2755 | |
| 2756 | /* |
| 2757 | * During each iteration of the migration, we listen for instructions |
| 2758 | * by the source VM to perform dynamic page registrations before they |
| 2759 | * can perform RDMA operations. |
| 2760 | * |
| 2761 | * We respond with the 'rkey'. |
| 2762 | * |
| 2763 | * Keep doing this until the source tells us to stop. |
| 2764 | */ |
| 2765 | static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque, |
| 2766 | uint64_t flags) |
| 2767 | { |
| 2768 | RDMAControlHeader reg_resp = { .len = sizeof(RDMARegisterResult), |
| 2769 | .type = RDMA_CONTROL_REGISTER_RESULT, |
| 2770 | .repeat = 0, |
| 2771 | }; |
| 2772 | RDMAControlHeader unreg_resp = { .len = 0, |
| 2773 | .type = RDMA_CONTROL_UNREGISTER_FINISHED, |
| 2774 | .repeat = 0, |
| 2775 | }; |
| 2776 | RDMAControlHeader blocks = { .type = RDMA_CONTROL_RAM_BLOCKS_RESULT, |
| 2777 | .repeat = 1 }; |
| 2778 | QEMUFileRDMA *rfile = opaque; |
| 2779 | RDMAContext *rdma = rfile->rdma; |
| 2780 | RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 2781 | RDMAControlHeader head; |
| 2782 | RDMARegister *reg, *registers; |
| 2783 | RDMACompress *comp; |
| 2784 | RDMARegisterResult *reg_result; |
| 2785 | static RDMARegisterResult results[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE]; |
| 2786 | RDMALocalBlock *block; |
| 2787 | void *host_addr; |
| 2788 | int ret = 0; |
| 2789 | int idx = 0; |
| 2790 | int count = 0; |
| 2791 | int i = 0; |
| 2792 | |
| 2793 | CHECK_ERROR_STATE(); |
| 2794 | |
| 2795 | do { |
| 2796 | DDDPRINTF("Waiting for next request %" PRIu64 "...\n", flags); |
| 2797 | |
| 2798 | ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_NONE); |
| 2799 | |
| 2800 | if (ret < 0) { |
| 2801 | break; |
| 2802 | } |
| 2803 | |
| 2804 | if (head.repeat > RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE) { |
| 2805 | fprintf(stderr, "rdma: Too many requests in this message (%d)." |
| 2806 | "Bailing.\n", head.repeat); |
| 2807 | ret = -EIO; |
| 2808 | break; |
| 2809 | } |
| 2810 | |
| 2811 | switch (head.type) { |
| 2812 | case RDMA_CONTROL_COMPRESS: |
| 2813 | comp = (RDMACompress *) rdma->wr_data[idx].control_curr; |
| 2814 | network_to_compress(comp); |
| 2815 | |
| 2816 | DDPRINTF("Zapping zero chunk: %" PRId64 |
| 2817 | " bytes, index %d, offset %" PRId64 "\n", |
| 2818 | comp->length, comp->block_idx, comp->offset); |
| 2819 | block = &(rdma->local_ram_blocks.block[comp->block_idx]); |
| 2820 | |
| 2821 | host_addr = block->local_host_addr + |
| 2822 | (comp->offset - block->offset); |
| 2823 | |
| 2824 | ram_handle_compressed(host_addr, comp->value, comp->length); |
| 2825 | break; |
| 2826 | |
| 2827 | case RDMA_CONTROL_REGISTER_FINISHED: |
| 2828 | DDDPRINTF("Current registrations complete.\n"); |
| 2829 | goto out; |
| 2830 | |
| 2831 | case RDMA_CONTROL_RAM_BLOCKS_REQUEST: |
| 2832 | DPRINTF("Initial setup info requested.\n"); |
| 2833 | |
| 2834 | if (rdma->pin_all) { |
| 2835 | ret = qemu_rdma_reg_whole_ram_blocks(rdma); |
| 2836 | if (ret) { |
| 2837 | fprintf(stderr, "rdma migration: error dest " |
| 2838 | "registering ram blocks!\n"); |
| 2839 | goto out; |
| 2840 | } |
| 2841 | } |
| 2842 | |
| 2843 | /* |
| 2844 | * Dest uses this to prepare to transmit the RAMBlock descriptions |
| 2845 | * to the source VM after connection setup. |
| 2846 | * Both sides use the "remote" structure to communicate and update |
| 2847 | * their "local" descriptions with what was sent. |
| 2848 | */ |
| 2849 | for (i = 0; i < local->nb_blocks; i++) { |
| 2850 | rdma->block[i].remote_host_addr = |
| 2851 | (uint64_t)(local->block[i].local_host_addr); |
| 2852 | |
| 2853 | if (rdma->pin_all) { |
| 2854 | rdma->block[i].remote_rkey = local->block[i].mr->rkey; |
| 2855 | } |
| 2856 | |
| 2857 | rdma->block[i].offset = local->block[i].offset; |
| 2858 | rdma->block[i].length = local->block[i].length; |
| 2859 | |
| 2860 | remote_block_to_network(&rdma->block[i]); |
| 2861 | } |
| 2862 | |
| 2863 | blocks.len = rdma->local_ram_blocks.nb_blocks |
| 2864 | * sizeof(RDMARemoteBlock); |
| 2865 | |
| 2866 | |
| 2867 | ret = qemu_rdma_post_send_control(rdma, |
| 2868 | (uint8_t *) rdma->block, &blocks); |
| 2869 | |
| 2870 | if (ret < 0) { |
| 2871 | fprintf(stderr, "rdma migration: error sending remote info!\n"); |
| 2872 | goto out; |
| 2873 | } |
| 2874 | |
| 2875 | break; |
| 2876 | case RDMA_CONTROL_REGISTER_REQUEST: |
| 2877 | DDPRINTF("There are %d registration requests\n", head.repeat); |
| 2878 | |
| 2879 | reg_resp.repeat = head.repeat; |
| 2880 | registers = (RDMARegister *) rdma->wr_data[idx].control_curr; |
| 2881 | |
| 2882 | for (count = 0; count < head.repeat; count++) { |
| 2883 | uint64_t chunk; |
| 2884 | uint8_t *chunk_start, *chunk_end; |
| 2885 | |
| 2886 | reg = ®isters[count]; |
| 2887 | network_to_register(reg); |
| 2888 | |
| 2889 | reg_result = &results[count]; |
| 2890 | |
| 2891 | DDPRINTF("Registration request (%d): index %d, current_addr %" |
| 2892 | PRIu64 " chunks: %" PRIu64 "\n", count, |
| 2893 | reg->current_index, reg->key.current_addr, reg->chunks); |
| 2894 | |
| 2895 | block = &(rdma->local_ram_blocks.block[reg->current_index]); |
| 2896 | if (block->is_ram_block) { |
| 2897 | host_addr = (block->local_host_addr + |
| 2898 | (reg->key.current_addr - block->offset)); |
| 2899 | chunk = ram_chunk_index(block->local_host_addr, |
| 2900 | (uint8_t *) host_addr); |
| 2901 | } else { |
| 2902 | chunk = reg->key.chunk; |
| 2903 | host_addr = block->local_host_addr + |
| 2904 | (reg->key.chunk * (1UL << RDMA_REG_CHUNK_SHIFT)); |
| 2905 | } |
| 2906 | chunk_start = ram_chunk_start(block, chunk); |
| 2907 | chunk_end = ram_chunk_end(block, chunk + reg->chunks); |
| 2908 | if (qemu_rdma_register_and_get_keys(rdma, block, |
| 2909 | (uint8_t *)host_addr, NULL, ®_result->rkey, |
| 2910 | chunk, chunk_start, chunk_end)) { |
| 2911 | fprintf(stderr, "cannot get rkey!\n"); |
| 2912 | ret = -EINVAL; |
| 2913 | goto out; |
| 2914 | } |
| 2915 | |
| 2916 | reg_result->host_addr = (uint64_t) block->local_host_addr; |
| 2917 | |
| 2918 | DDPRINTF("Registered rkey for this request: %x\n", |
| 2919 | reg_result->rkey); |
| 2920 | |
| 2921 | result_to_network(reg_result); |
| 2922 | } |
| 2923 | |
| 2924 | ret = qemu_rdma_post_send_control(rdma, |
| 2925 | (uint8_t *) results, ®_resp); |
| 2926 | |
| 2927 | if (ret < 0) { |
| 2928 | fprintf(stderr, "Failed to send control buffer!\n"); |
| 2929 | goto out; |
| 2930 | } |
| 2931 | break; |
| 2932 | case RDMA_CONTROL_UNREGISTER_REQUEST: |
| 2933 | DDPRINTF("There are %d unregistration requests\n", head.repeat); |
| 2934 | unreg_resp.repeat = head.repeat; |
| 2935 | registers = (RDMARegister *) rdma->wr_data[idx].control_curr; |
| 2936 | |
| 2937 | for (count = 0; count < head.repeat; count++) { |
| 2938 | reg = ®isters[count]; |
| 2939 | network_to_register(reg); |
| 2940 | |
| 2941 | DDPRINTF("Unregistration request (%d): " |
| 2942 | " index %d, chunk %" PRIu64 "\n", |
| 2943 | count, reg->current_index, reg->key.chunk); |
| 2944 | |
| 2945 | block = &(rdma->local_ram_blocks.block[reg->current_index]); |
| 2946 | |
| 2947 | ret = ibv_dereg_mr(block->pmr[reg->key.chunk]); |
| 2948 | block->pmr[reg->key.chunk] = NULL; |
| 2949 | |
| 2950 | if (ret != 0) { |
| 2951 | perror("rdma unregistration chunk failed"); |
| 2952 | ret = -ret; |
| 2953 | goto out; |
| 2954 | } |
| 2955 | |
| 2956 | rdma->total_registrations--; |
| 2957 | |
| 2958 | DDPRINTF("Unregistered chunk %" PRIu64 " successfully.\n", |
| 2959 | reg->key.chunk); |
| 2960 | } |
| 2961 | |
| 2962 | ret = qemu_rdma_post_send_control(rdma, NULL, &unreg_resp); |
| 2963 | |
| 2964 | if (ret < 0) { |
| 2965 | fprintf(stderr, "Failed to send control buffer!\n"); |
| 2966 | goto out; |
| 2967 | } |
| 2968 | break; |
| 2969 | case RDMA_CONTROL_REGISTER_RESULT: |
| 2970 | fprintf(stderr, "Invalid RESULT message at dest.\n"); |
| 2971 | ret = -EIO; |
| 2972 | goto out; |
| 2973 | default: |
| 2974 | fprintf(stderr, "Unknown control message %s\n", |
| 2975 | control_desc[head.type]); |
| 2976 | ret = -EIO; |
| 2977 | goto out; |
| 2978 | } |
| 2979 | } while (1); |
| 2980 | out: |
| 2981 | if (ret < 0) { |
| 2982 | rdma->error_state = ret; |
| 2983 | } |
| 2984 | return ret; |
| 2985 | } |
| 2986 | |
| 2987 | static int qemu_rdma_registration_start(QEMUFile *f, void *opaque, |
| 2988 | uint64_t flags) |
| 2989 | { |
| 2990 | QEMUFileRDMA *rfile = opaque; |
| 2991 | RDMAContext *rdma = rfile->rdma; |
| 2992 | |
| 2993 | CHECK_ERROR_STATE(); |
| 2994 | |
| 2995 | DDDPRINTF("start section: %" PRIu64 "\n", flags); |
| 2996 | qemu_put_be64(f, RAM_SAVE_FLAG_HOOK); |
| 2997 | qemu_fflush(f); |
| 2998 | |
| 2999 | return 0; |
| 3000 | } |
| 3001 | |
| 3002 | /* |
| 3003 | * Inform dest that dynamic registrations are done for now. |
| 3004 | * First, flush writes, if any. |
| 3005 | */ |
| 3006 | static int qemu_rdma_registration_stop(QEMUFile *f, void *opaque, |
| 3007 | uint64_t flags) |
| 3008 | { |
| 3009 | Error *local_err = NULL, **errp = &local_err; |
| 3010 | QEMUFileRDMA *rfile = opaque; |
| 3011 | RDMAContext *rdma = rfile->rdma; |
| 3012 | RDMAControlHeader head = { .len = 0, .repeat = 1 }; |
| 3013 | int ret = 0; |
| 3014 | |
| 3015 | CHECK_ERROR_STATE(); |
| 3016 | |
| 3017 | qemu_fflush(f); |
| 3018 | ret = qemu_rdma_drain_cq(f, rdma); |
| 3019 | |
| 3020 | if (ret < 0) { |
| 3021 | goto err; |
| 3022 | } |
| 3023 | |
| 3024 | if (flags == RAM_CONTROL_SETUP) { |
| 3025 | RDMAControlHeader resp = {.type = RDMA_CONTROL_RAM_BLOCKS_RESULT }; |
| 3026 | RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 3027 | int reg_result_idx, i, j, nb_remote_blocks; |
| 3028 | |
| 3029 | head.type = RDMA_CONTROL_RAM_BLOCKS_REQUEST; |
| 3030 | DPRINTF("Sending registration setup for ram blocks...\n"); |
| 3031 | |
| 3032 | /* |
| 3033 | * Make sure that we parallelize the pinning on both sides. |
| 3034 | * For very large guests, doing this serially takes a really |
| 3035 | * long time, so we have to 'interleave' the pinning locally |
| 3036 | * with the control messages by performing the pinning on this |
| 3037 | * side before we receive the control response from the other |
| 3038 | * side that the pinning has completed. |
| 3039 | */ |
| 3040 | ret = qemu_rdma_exchange_send(rdma, &head, NULL, &resp, |
| 3041 | ®_result_idx, rdma->pin_all ? |
| 3042 | qemu_rdma_reg_whole_ram_blocks : NULL); |
| 3043 | if (ret < 0) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3044 | ERROR(errp, "receiving remote info!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3045 | return ret; |
| 3046 | } |
| 3047 | |
| 3048 | qemu_rdma_move_header(rdma, reg_result_idx, &resp); |
| 3049 | memcpy(rdma->block, |
| 3050 | rdma->wr_data[reg_result_idx].control_curr, resp.len); |
| 3051 | |
| 3052 | nb_remote_blocks = resp.len / sizeof(RDMARemoteBlock); |
| 3053 | |
| 3054 | /* |
| 3055 | * The protocol uses two different sets of rkeys (mutually exclusive): |
| 3056 | * 1. One key to represent the virtual address of the entire ram block. |
| 3057 | * (dynamic chunk registration disabled - pin everything with one rkey.) |
| 3058 | * 2. One to represent individual chunks within a ram block. |
| 3059 | * (dynamic chunk registration enabled - pin individual chunks.) |
| 3060 | * |
| 3061 | * Once the capability is successfully negotiated, the destination transmits |
| 3062 | * the keys to use (or sends them later) including the virtual addresses |
| 3063 | * and then propagates the remote ram block descriptions to his local copy. |
| 3064 | */ |
| 3065 | |
| 3066 | if (local->nb_blocks != nb_remote_blocks) { |
| 3067 | ERROR(errp, "ram blocks mismatch #1! " |
| 3068 | "Your QEMU command line parameters are probably " |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3069 | "not identical on both the source and destination."); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3070 | return -EINVAL; |
| 3071 | } |
| 3072 | |
| 3073 | for (i = 0; i < nb_remote_blocks; i++) { |
| 3074 | network_to_remote_block(&rdma->block[i]); |
| 3075 | |
| 3076 | /* search local ram blocks */ |
| 3077 | for (j = 0; j < local->nb_blocks; j++) { |
| 3078 | if (rdma->block[i].offset != local->block[j].offset) { |
| 3079 | continue; |
| 3080 | } |
| 3081 | |
| 3082 | if (rdma->block[i].length != local->block[j].length) { |
| 3083 | ERROR(errp, "ram blocks mismatch #2! " |
| 3084 | "Your QEMU command line parameters are probably " |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3085 | "not identical on both the source and destination."); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3086 | return -EINVAL; |
| 3087 | } |
| 3088 | local->block[j].remote_host_addr = |
| 3089 | rdma->block[i].remote_host_addr; |
| 3090 | local->block[j].remote_rkey = rdma->block[i].remote_rkey; |
| 3091 | break; |
| 3092 | } |
| 3093 | |
| 3094 | if (j >= local->nb_blocks) { |
| 3095 | ERROR(errp, "ram blocks mismatch #3! " |
| 3096 | "Your QEMU command line parameters are probably " |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3097 | "not identical on both the source and destination."); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3098 | return -EINVAL; |
| 3099 | } |
| 3100 | } |
| 3101 | } |
| 3102 | |
| 3103 | DDDPRINTF("Sending registration finish %" PRIu64 "...\n", flags); |
| 3104 | |
| 3105 | head.type = RDMA_CONTROL_REGISTER_FINISHED; |
| 3106 | ret = qemu_rdma_exchange_send(rdma, &head, NULL, NULL, NULL, NULL); |
| 3107 | |
| 3108 | if (ret < 0) { |
| 3109 | goto err; |
| 3110 | } |
| 3111 | |
| 3112 | return 0; |
| 3113 | err: |
| 3114 | rdma->error_state = ret; |
| 3115 | return ret; |
| 3116 | } |
| 3117 | |
| 3118 | static int qemu_rdma_get_fd(void *opaque) |
| 3119 | { |
| 3120 | QEMUFileRDMA *rfile = opaque; |
| 3121 | RDMAContext *rdma = rfile->rdma; |
| 3122 | |
| 3123 | return rdma->comp_channel->fd; |
| 3124 | } |
| 3125 | |
| 3126 | const QEMUFileOps rdma_read_ops = { |
| 3127 | .get_buffer = qemu_rdma_get_buffer, |
| 3128 | .get_fd = qemu_rdma_get_fd, |
| 3129 | .close = qemu_rdma_close, |
| 3130 | .hook_ram_load = qemu_rdma_registration_handle, |
| 3131 | }; |
| 3132 | |
| 3133 | const QEMUFileOps rdma_write_ops = { |
| 3134 | .put_buffer = qemu_rdma_put_buffer, |
| 3135 | .close = qemu_rdma_close, |
| 3136 | .before_ram_iterate = qemu_rdma_registration_start, |
| 3137 | .after_ram_iterate = qemu_rdma_registration_stop, |
| 3138 | .save_page = qemu_rdma_save_page, |
| 3139 | }; |
| 3140 | |
| 3141 | static void *qemu_fopen_rdma(RDMAContext *rdma, const char *mode) |
| 3142 | { |
| 3143 | QEMUFileRDMA *r = g_malloc0(sizeof(QEMUFileRDMA)); |
| 3144 | |
| 3145 | if (qemu_file_mode_is_not_valid(mode)) { |
| 3146 | return NULL; |
| 3147 | } |
| 3148 | |
| 3149 | r->rdma = rdma; |
| 3150 | |
| 3151 | if (mode[0] == 'w') { |
| 3152 | r->file = qemu_fopen_ops(r, &rdma_write_ops); |
| 3153 | } else { |
| 3154 | r->file = qemu_fopen_ops(r, &rdma_read_ops); |
| 3155 | } |
| 3156 | |
| 3157 | return r->file; |
| 3158 | } |
| 3159 | |
| 3160 | static void rdma_accept_incoming_migration(void *opaque) |
| 3161 | { |
| 3162 | RDMAContext *rdma = opaque; |
| 3163 | int ret; |
| 3164 | QEMUFile *f; |
| 3165 | Error *local_err = NULL, **errp = &local_err; |
| 3166 | |
| 3167 | DPRINTF("Accepting rdma connection...\n"); |
| 3168 | ret = qemu_rdma_accept(rdma); |
| 3169 | |
| 3170 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3171 | ERROR(errp, "RDMA Migration initialization failed!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3172 | return; |
| 3173 | } |
| 3174 | |
| 3175 | DPRINTF("Accepted migration\n"); |
| 3176 | |
| 3177 | f = qemu_fopen_rdma(rdma, "rb"); |
| 3178 | if (f == NULL) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3179 | ERROR(errp, "could not qemu_fopen_rdma!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3180 | qemu_rdma_cleanup(rdma); |
| 3181 | return; |
| 3182 | } |
| 3183 | |
| 3184 | rdma->migration_started_on_destination = 1; |
| 3185 | process_incoming_migration(f); |
| 3186 | } |
| 3187 | |
| 3188 | void rdma_start_incoming_migration(const char *host_port, Error **errp) |
| 3189 | { |
| 3190 | int ret; |
| 3191 | RDMAContext *rdma; |
| 3192 | Error *local_err = NULL; |
| 3193 | |
| 3194 | DPRINTF("Starting RDMA-based incoming migration\n"); |
| 3195 | rdma = qemu_rdma_data_init(host_port, &local_err); |
| 3196 | |
| 3197 | if (rdma == NULL) { |
| 3198 | goto err; |
| 3199 | } |
| 3200 | |
| 3201 | ret = qemu_rdma_dest_init(rdma, &local_err); |
| 3202 | |
| 3203 | if (ret) { |
| 3204 | goto err; |
| 3205 | } |
| 3206 | |
| 3207 | DPRINTF("qemu_rdma_dest_init success\n"); |
| 3208 | |
| 3209 | ret = rdma_listen(rdma->listen_id, 5); |
| 3210 | |
| 3211 | if (ret) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3212 | ERROR(errp, "listening on socket!"); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3213 | goto err; |
| 3214 | } |
| 3215 | |
| 3216 | DPRINTF("rdma_listen success\n"); |
| 3217 | |
| 3218 | qemu_set_fd_handler2(rdma->channel->fd, NULL, |
| 3219 | rdma_accept_incoming_migration, NULL, |
| 3220 | (void *)(intptr_t) rdma); |
| 3221 | return; |
| 3222 | err: |
| 3223 | error_propagate(errp, local_err); |
| 3224 | g_free(rdma); |
| 3225 | } |
| 3226 | |
| 3227 | void rdma_start_outgoing_migration(void *opaque, |
| 3228 | const char *host_port, Error **errp) |
| 3229 | { |
| 3230 | MigrationState *s = opaque; |
| 3231 | Error *local_err = NULL, **temp = &local_err; |
| 3232 | RDMAContext *rdma = qemu_rdma_data_init(host_port, &local_err); |
| 3233 | int ret = 0; |
| 3234 | |
| 3235 | if (rdma == NULL) { |
Michael R. Hines | 6698894 | 2013-08-03 22:54:50 -0400 | [diff] [blame] | 3236 | ERROR(temp, "Failed to initialize RDMA data structures! %d", ret); |
Michael R. Hines | 2da776d | 2013-07-22 10:01:54 -0400 | [diff] [blame] | 3237 | goto err; |
| 3238 | } |
| 3239 | |
| 3240 | ret = qemu_rdma_source_init(rdma, &local_err, |
| 3241 | s->enabled_capabilities[MIGRATION_CAPABILITY_X_RDMA_PIN_ALL]); |
| 3242 | |
| 3243 | if (ret) { |
| 3244 | goto err; |
| 3245 | } |
| 3246 | |
| 3247 | DPRINTF("qemu_rdma_source_init success\n"); |
| 3248 | ret = qemu_rdma_connect(rdma, &local_err); |
| 3249 | |
| 3250 | if (ret) { |
| 3251 | goto err; |
| 3252 | } |
| 3253 | |
| 3254 | DPRINTF("qemu_rdma_source_connect success\n"); |
| 3255 | |
| 3256 | s->file = qemu_fopen_rdma(rdma, "wb"); |
| 3257 | migrate_fd_connect(s); |
| 3258 | return; |
| 3259 | err: |
| 3260 | error_propagate(errp, local_err); |
| 3261 | g_free(rdma); |
| 3262 | migrate_fd_error(s); |
| 3263 | } |