root / migration-rdma.c @ 5d12aa63
History | View | Annotate | Download (104.9 kB)
| 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 |
|
| 30 |
//#define DEBUG_RDMA
|
| 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 { \
|
| 63 |
fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \ |
| 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 |
|
| 325 |
RDMAWorkRequestData wr_data[RDMA_WRID_MAX]; |
| 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 |
bool connected;
|
| 360 |
|
| 361 |
struct ibv_context *verbs;
|
| 362 |
struct rdma_event_channel *channel;
|
| 363 |
struct ibv_qp *qp; /* queue pair */ |
| 364 |
struct ibv_comp_channel *comp_channel; /* completion channel */ |
| 365 |
struct ibv_pd *pd; /* protection domain */ |
| 366 |
struct ibv_cq *cq; /* completion queue */ |
| 367 |
|
| 368 |
/*
|
| 369 |
* If a previous write failed (perhaps because of a failed
|
| 370 |
* memory registration, then do not attempt any future work
|
| 371 |
* and remember the error state.
|
| 372 |
*/
|
| 373 |
int error_state;
|
| 374 |
int error_reported;
|
| 375 |
|
| 376 |
/*
|
| 377 |
* Description of ram blocks used throughout the code.
|
| 378 |
*/
|
| 379 |
RDMALocalBlocks local_ram_blocks; |
| 380 |
RDMARemoteBlock *block; |
| 381 |
|
| 382 |
/*
|
| 383 |
* Migration on *destination* started.
|
| 384 |
* Then use coroutine yield function.
|
| 385 |
* Source runs in a thread, so we don't care.
|
| 386 |
*/
|
| 387 |
int migration_started_on_destination;
|
| 388 |
|
| 389 |
int total_registrations;
|
| 390 |
int total_writes;
|
| 391 |
|
| 392 |
int unregister_current, unregister_next;
|
| 393 |
uint64_t unregistrations[RDMA_SIGNALED_SEND_MAX]; |
| 394 |
|
| 395 |
GHashTable *blockmap; |
| 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(const uint8_t *start, |
| 515 |
const uint8_t *host)
|
| 516 |
{
|
| 517 |
return ((uintptr_t) host - (uintptr_t) start) >> RDMA_REG_CHUNK_SHIFT;
|
| 518 |
} |
| 519 |
|
| 520 |
static inline uint8_t *ram_chunk_start(const RDMALocalBlock *rdma_ram_block, |
| 521 |
uint64_t i) |
| 522 |
{
|
| 523 |
return (uint8_t *) (((uintptr_t) rdma_ram_block->local_host_addr)
|
| 524 |
+ (i << RDMA_REG_CHUNK_SHIFT)); |
| 525 |
} |
| 526 |
|
| 527 |
static inline uint8_t *ram_chunk_end(const RDMALocalBlock *rdma_ram_block, |
| 528 |
uint64_t i) |
| 529 |
{
|
| 530 |
uint8_t *result = ram_chunk_start(rdma_ram_block, i) + |
| 531 |
(1UL << RDMA_REG_CHUNK_SHIFT);
|
| 532 |
|
| 533 |
if (result > (rdma_ram_block->local_host_addr + rdma_ram_block->length)) {
|
| 534 |
result = rdma_ram_block->local_host_addr + rdma_ram_block->length; |
| 535 |
} |
| 536 |
|
| 537 |
return result;
|
| 538 |
} |
| 539 |
|
| 540 |
static int __qemu_rdma_add_block(RDMAContext *rdma, void *host_addr, |
| 541 |
ram_addr_t block_offset, uint64_t length) |
| 542 |
{
|
| 543 |
RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 544 |
RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, |
| 545 |
(void *) block_offset);
|
| 546 |
RDMALocalBlock *old = local->block; |
| 547 |
|
| 548 |
assert(block == NULL);
|
| 549 |
|
| 550 |
local->block = g_malloc0(sizeof(RDMALocalBlock) * (local->nb_blocks + 1)); |
| 551 |
|
| 552 |
if (local->nb_blocks) {
|
| 553 |
int x;
|
| 554 |
|
| 555 |
for (x = 0; x < local->nb_blocks; x++) { |
| 556 |
g_hash_table_remove(rdma->blockmap, (void *)old[x].offset);
|
| 557 |
g_hash_table_insert(rdma->blockmap, (void *)old[x].offset,
|
| 558 |
&local->block[x]); |
| 559 |
} |
| 560 |
memcpy(local->block, old, sizeof(RDMALocalBlock) * local->nb_blocks);
|
| 561 |
g_free(old); |
| 562 |
} |
| 563 |
|
| 564 |
block = &local->block[local->nb_blocks]; |
| 565 |
|
| 566 |
block->local_host_addr = host_addr; |
| 567 |
block->offset = block_offset; |
| 568 |
block->length = length; |
| 569 |
block->index = local->nb_blocks; |
| 570 |
block->nb_chunks = ram_chunk_index(host_addr, host_addr + length) + 1UL;
|
| 571 |
block->transit_bitmap = bitmap_new(block->nb_chunks); |
| 572 |
bitmap_clear(block->transit_bitmap, 0, block->nb_chunks);
|
| 573 |
block->unregister_bitmap = bitmap_new(block->nb_chunks); |
| 574 |
bitmap_clear(block->unregister_bitmap, 0, block->nb_chunks);
|
| 575 |
block->remote_keys = g_malloc0(block->nb_chunks * sizeof(uint32_t));
|
| 576 |
|
| 577 |
block->is_ram_block = local->init ? false : true; |
| 578 |
|
| 579 |
g_hash_table_insert(rdma->blockmap, (void *) block_offset, block);
|
| 580 |
|
| 581 |
DDPRINTF("Added Block: %d, addr: %" PRIu64 ", offset: %" PRIu64 |
| 582 |
" length: %" PRIu64 " end: %" PRIu64 " bits %" PRIu64 " chunks %d\n", |
| 583 |
local->nb_blocks, (uint64_t) block->local_host_addr, block->offset, |
| 584 |
block->length, (uint64_t) (block->local_host_addr + block->length), |
| 585 |
BITS_TO_LONGS(block->nb_chunks) * |
| 586 |
sizeof(unsigned long) * 8, block->nb_chunks); |
| 587 |
|
| 588 |
local->nb_blocks++; |
| 589 |
|
| 590 |
return 0; |
| 591 |
} |
| 592 |
|
| 593 |
/*
|
| 594 |
* Memory regions need to be registered with the device and queue pairs setup
|
| 595 |
* in advanced before the migration starts. This tells us where the RAM blocks
|
| 596 |
* are so that we can register them individually.
|
| 597 |
*/
|
| 598 |
static void qemu_rdma_init_one_block(void *host_addr, |
| 599 |
ram_addr_t block_offset, ram_addr_t length, void *opaque)
|
| 600 |
{
|
| 601 |
__qemu_rdma_add_block(opaque, host_addr, block_offset, length); |
| 602 |
} |
| 603 |
|
| 604 |
/*
|
| 605 |
* Identify the RAMBlocks and their quantity. They will be references to
|
| 606 |
* identify chunk boundaries inside each RAMBlock and also be referenced
|
| 607 |
* during dynamic page registration.
|
| 608 |
*/
|
| 609 |
static int qemu_rdma_init_ram_blocks(RDMAContext *rdma) |
| 610 |
{
|
| 611 |
RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 612 |
|
| 613 |
assert(rdma->blockmap == NULL);
|
| 614 |
rdma->blockmap = g_hash_table_new(g_direct_hash, g_direct_equal); |
| 615 |
memset(local, 0, sizeof *local); |
| 616 |
qemu_ram_foreach_block(qemu_rdma_init_one_block, rdma); |
| 617 |
DPRINTF("Allocated %d local ram block structures\n", local->nb_blocks);
|
| 618 |
rdma->block = (RDMARemoteBlock *) g_malloc0(sizeof(RDMARemoteBlock) *
|
| 619 |
rdma->local_ram_blocks.nb_blocks); |
| 620 |
local->init = true;
|
| 621 |
return 0; |
| 622 |
} |
| 623 |
|
| 624 |
static int __qemu_rdma_delete_block(RDMAContext *rdma, ram_addr_t block_offset) |
| 625 |
{
|
| 626 |
RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 627 |
RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, |
| 628 |
(void *) block_offset);
|
| 629 |
RDMALocalBlock *old = local->block; |
| 630 |
int x;
|
| 631 |
|
| 632 |
assert(block); |
| 633 |
|
| 634 |
if (block->pmr) {
|
| 635 |
int j;
|
| 636 |
|
| 637 |
for (j = 0; j < block->nb_chunks; j++) { |
| 638 |
if (!block->pmr[j]) {
|
| 639 |
continue;
|
| 640 |
} |
| 641 |
ibv_dereg_mr(block->pmr[j]); |
| 642 |
rdma->total_registrations--; |
| 643 |
} |
| 644 |
g_free(block->pmr); |
| 645 |
block->pmr = NULL;
|
| 646 |
} |
| 647 |
|
| 648 |
if (block->mr) {
|
| 649 |
ibv_dereg_mr(block->mr); |
| 650 |
rdma->total_registrations--; |
| 651 |
block->mr = NULL;
|
| 652 |
} |
| 653 |
|
| 654 |
g_free(block->transit_bitmap); |
| 655 |
block->transit_bitmap = NULL;
|
| 656 |
|
| 657 |
g_free(block->unregister_bitmap); |
| 658 |
block->unregister_bitmap = NULL;
|
| 659 |
|
| 660 |
g_free(block->remote_keys); |
| 661 |
block->remote_keys = NULL;
|
| 662 |
|
| 663 |
for (x = 0; x < local->nb_blocks; x++) { |
| 664 |
g_hash_table_remove(rdma->blockmap, (void *)old[x].offset);
|
| 665 |
} |
| 666 |
|
| 667 |
if (local->nb_blocks > 1) { |
| 668 |
|
| 669 |
local->block = g_malloc0(sizeof(RDMALocalBlock) *
|
| 670 |
(local->nb_blocks - 1));
|
| 671 |
|
| 672 |
if (block->index) {
|
| 673 |
memcpy(local->block, old, sizeof(RDMALocalBlock) * block->index);
|
| 674 |
} |
| 675 |
|
| 676 |
if (block->index < (local->nb_blocks - 1)) { |
| 677 |
memcpy(local->block + block->index, old + (block->index + 1),
|
| 678 |
sizeof(RDMALocalBlock) *
|
| 679 |
(local->nb_blocks - (block->index + 1)));
|
| 680 |
} |
| 681 |
} else {
|
| 682 |
assert(block == local->block); |
| 683 |
local->block = NULL;
|
| 684 |
} |
| 685 |
|
| 686 |
DDPRINTF("Deleted Block: %d, addr: %" PRIu64 ", offset: %" PRIu64 |
| 687 |
" length: %" PRIu64 " end: %" PRIu64 " bits %" PRIu64 " chunks %d\n", |
| 688 |
local->nb_blocks, (uint64_t) block->local_host_addr, block->offset, |
| 689 |
block->length, (uint64_t) (block->local_host_addr + block->length), |
| 690 |
BITS_TO_LONGS(block->nb_chunks) * |
| 691 |
sizeof(unsigned long) * 8, block->nb_chunks); |
| 692 |
|
| 693 |
g_free(old); |
| 694 |
|
| 695 |
local->nb_blocks--; |
| 696 |
|
| 697 |
if (local->nb_blocks) {
|
| 698 |
for (x = 0; x < local->nb_blocks; x++) { |
| 699 |
g_hash_table_insert(rdma->blockmap, (void *)local->block[x].offset,
|
| 700 |
&local->block[x]); |
| 701 |
} |
| 702 |
} |
| 703 |
|
| 704 |
return 0; |
| 705 |
} |
| 706 |
|
| 707 |
/*
|
| 708 |
* Put in the log file which RDMA device was opened and the details
|
| 709 |
* associated with that device.
|
| 710 |
*/
|
| 711 |
static void qemu_rdma_dump_id(const char *who, struct ibv_context *verbs) |
| 712 |
{
|
| 713 |
struct ibv_port_attr port;
|
| 714 |
|
| 715 |
if (ibv_query_port(verbs, 1, &port)) { |
| 716 |
fprintf(stderr, "FAILED TO QUERY PORT INFORMATION!\n");
|
| 717 |
return;
|
| 718 |
} |
| 719 |
|
| 720 |
printf("%s RDMA Device opened: kernel name %s "
|
| 721 |
"uverbs device name %s, "
|
| 722 |
"infiniband_verbs class device path %s, "
|
| 723 |
"infiniband class device path %s, "
|
| 724 |
"transport: (%d) %s\n",
|
| 725 |
who, |
| 726 |
verbs->device->name, |
| 727 |
verbs->device->dev_name, |
| 728 |
verbs->device->dev_path, |
| 729 |
verbs->device->ibdev_path, |
| 730 |
port.link_layer, |
| 731 |
(port.link_layer == IBV_LINK_LAYER_INFINIBAND) ? "Infiniband" :
|
| 732 |
((port.link_layer == IBV_LINK_LAYER_ETHERNET) |
| 733 |
? "Ethernet" : "Unknown")); |
| 734 |
} |
| 735 |
|
| 736 |
/*
|
| 737 |
* Put in the log file the RDMA gid addressing information,
|
| 738 |
* useful for folks who have trouble understanding the
|
| 739 |
* RDMA device hierarchy in the kernel.
|
| 740 |
*/
|
| 741 |
static void qemu_rdma_dump_gid(const char *who, struct rdma_cm_id *id) |
| 742 |
{
|
| 743 |
char sgid[33]; |
| 744 |
char dgid[33]; |
| 745 |
inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.sgid, sgid, sizeof sgid);
|
| 746 |
inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.dgid, dgid, sizeof dgid);
|
| 747 |
DPRINTF("%s Source GID: %s, Dest GID: %s\n", who, sgid, dgid);
|
| 748 |
} |
| 749 |
|
| 750 |
/*
|
| 751 |
* As of now, IPv6 over RoCE / iWARP is not supported by linux.
|
| 752 |
* We will try the next addrinfo struct, and fail if there are
|
| 753 |
* no other valid addresses to bind against.
|
| 754 |
*
|
| 755 |
* If user is listening on '[::]', then we will not have a opened a device
|
| 756 |
* yet and have no way of verifying if the device is RoCE or not.
|
| 757 |
*
|
| 758 |
* In this case, the source VM will throw an error for ALL types of
|
| 759 |
* connections (both IPv4 and IPv6) if the destination machine does not have
|
| 760 |
* a regular infiniband network available for use.
|
| 761 |
*
|
| 762 |
* The only way to guarantee that an error is thrown for broken kernels is
|
| 763 |
* for the management software to choose a *specific* interface at bind time
|
| 764 |
* and validate what time of hardware it is.
|
| 765 |
*
|
| 766 |
* Unfortunately, this puts the user in a fix:
|
| 767 |
*
|
| 768 |
* If the source VM connects with an IPv4 address without knowing that the
|
| 769 |
* destination has bound to '[::]' the migration will unconditionally fail
|
| 770 |
* unless the management software is explicitly listening on the the IPv4
|
| 771 |
* address while using a RoCE-based device.
|
| 772 |
*
|
| 773 |
* If the source VM connects with an IPv6 address, then we're OK because we can
|
| 774 |
* throw an error on the source (and similarly on the destination).
|
| 775 |
*
|
| 776 |
* But in mixed environments, this will be broken for a while until it is fixed
|
| 777 |
* inside linux.
|
| 778 |
*
|
| 779 |
* We do provide a *tiny* bit of help in this function: We can list all of the
|
| 780 |
* devices in the system and check to see if all the devices are RoCE or
|
| 781 |
* Infiniband.
|
| 782 |
*
|
| 783 |
* If we detect that we have a *pure* RoCE environment, then we can safely
|
| 784 |
* thrown an error even if the management software has specified '[::]' as the
|
| 785 |
* bind address.
|
| 786 |
*
|
| 787 |
* However, if there is are multiple hetergeneous devices, then we cannot make
|
| 788 |
* this assumption and the user just has to be sure they know what they are
|
| 789 |
* doing.
|
| 790 |
*
|
| 791 |
* Patches are being reviewed on linux-rdma.
|
| 792 |
*/
|
| 793 |
static int qemu_rdma_broken_ipv6_kernel(Error **errp, struct ibv_context *verbs) |
| 794 |
{
|
| 795 |
struct ibv_port_attr port_attr;
|
| 796 |
|
| 797 |
/* This bug only exists in linux, to our knowledge. */
|
| 798 |
#ifdef CONFIG_LINUX
|
| 799 |
|
| 800 |
/*
|
| 801 |
* Verbs are only NULL if management has bound to '[::]'.
|
| 802 |
*
|
| 803 |
* Let's iterate through all the devices and see if there any pure IB
|
| 804 |
* devices (non-ethernet).
|
| 805 |
*
|
| 806 |
* If not, then we can safely proceed with the migration.
|
| 807 |
* Otherwise, there are no guarantees until the bug is fixed in linux.
|
| 808 |
*/
|
| 809 |
if (!verbs) {
|
| 810 |
int num_devices, x;
|
| 811 |
struct ibv_device ** dev_list = ibv_get_device_list(&num_devices);
|
| 812 |
bool roce_found = false; |
| 813 |
bool ib_found = false; |
| 814 |
|
| 815 |
for (x = 0; x < num_devices; x++) { |
| 816 |
verbs = ibv_open_device(dev_list[x]); |
| 817 |
|
| 818 |
if (ibv_query_port(verbs, 1, &port_attr)) { |
| 819 |
ibv_close_device(verbs); |
| 820 |
ERROR(errp, "Could not query initial IB port");
|
| 821 |
return -EINVAL;
|
| 822 |
} |
| 823 |
|
| 824 |
if (port_attr.link_layer == IBV_LINK_LAYER_INFINIBAND) {
|
| 825 |
ib_found = true;
|
| 826 |
} else if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) { |
| 827 |
roce_found = true;
|
| 828 |
} |
| 829 |
|
| 830 |
ibv_close_device(verbs); |
| 831 |
|
| 832 |
} |
| 833 |
|
| 834 |
if (roce_found) {
|
| 835 |
if (ib_found) {
|
| 836 |
fprintf(stderr, "WARN: migrations may fail:"
|
| 837 |
" IPv6 over RoCE / iWARP in linux"
|
| 838 |
" is broken. But since you appear to have a"
|
| 839 |
" mixed RoCE / IB environment, be sure to only"
|
| 840 |
" migrate over the IB fabric until the kernel "
|
| 841 |
" fixes the bug.\n");
|
| 842 |
} else {
|
| 843 |
ERROR(errp, "You only have RoCE / iWARP devices in your systems"
|
| 844 |
" and your management software has specified '[::]'"
|
| 845 |
", but IPv6 over RoCE / iWARP is not supported in Linux.");
|
| 846 |
return -ENONET;
|
| 847 |
} |
| 848 |
} |
| 849 |
|
| 850 |
return 0; |
| 851 |
} |
| 852 |
|
| 853 |
/*
|
| 854 |
* If we have a verbs context, that means that some other than '[::]' was
|
| 855 |
* used by the management software for binding. In which case we can actually
|
| 856 |
* warn the user about a potential broken kernel;
|
| 857 |
*/
|
| 858 |
|
| 859 |
/* IB ports start with 1, not 0 */
|
| 860 |
if (ibv_query_port(verbs, 1, &port_attr)) { |
| 861 |
ERROR(errp, "Could not query initial IB port");
|
| 862 |
return -EINVAL;
|
| 863 |
} |
| 864 |
|
| 865 |
if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) {
|
| 866 |
ERROR(errp, "Linux kernel's RoCE / iWARP does not support IPv6 "
|
| 867 |
"(but patches on linux-rdma in progress)");
|
| 868 |
return -ENONET;
|
| 869 |
} |
| 870 |
|
| 871 |
#endif
|
| 872 |
|
| 873 |
return 0; |
| 874 |
} |
| 875 |
|
| 876 |
/*
|
| 877 |
* Figure out which RDMA device corresponds to the requested IP hostname
|
| 878 |
* Also create the initial connection manager identifiers for opening
|
| 879 |
* the connection.
|
| 880 |
*/
|
| 881 |
static int qemu_rdma_resolve_host(RDMAContext *rdma, Error **errp) |
| 882 |
{
|
| 883 |
int ret;
|
| 884 |
struct rdma_addrinfo *res;
|
| 885 |
char port_str[16]; |
| 886 |
struct rdma_cm_event *cm_event;
|
| 887 |
char ip[40] = "unknown"; |
| 888 |
struct rdma_addrinfo *e;
|
| 889 |
|
| 890 |
if (rdma->host == NULL || !strcmp(rdma->host, "")) { |
| 891 |
ERROR(errp, "RDMA hostname has not been set");
|
| 892 |
return -EINVAL;
|
| 893 |
} |
| 894 |
|
| 895 |
/* create CM channel */
|
| 896 |
rdma->channel = rdma_create_event_channel(); |
| 897 |
if (!rdma->channel) {
|
| 898 |
ERROR(errp, "could not create CM channel");
|
| 899 |
return -EINVAL;
|
| 900 |
} |
| 901 |
|
| 902 |
/* create CM id */
|
| 903 |
ret = rdma_create_id(rdma->channel, &rdma->cm_id, NULL, RDMA_PS_TCP);
|
| 904 |
if (ret) {
|
| 905 |
ERROR(errp, "could not create channel id");
|
| 906 |
goto err_resolve_create_id;
|
| 907 |
} |
| 908 |
|
| 909 |
snprintf(port_str, 16, "%d", rdma->port); |
| 910 |
port_str[15] = '\0'; |
| 911 |
|
| 912 |
ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res);
|
| 913 |
if (ret < 0) { |
| 914 |
ERROR(errp, "could not rdma_getaddrinfo address %s", rdma->host);
|
| 915 |
goto err_resolve_get_addr;
|
| 916 |
} |
| 917 |
|
| 918 |
for (e = res; e != NULL; e = e->ai_next) { |
| 919 |
inet_ntop(e->ai_family, |
| 920 |
&((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip); |
| 921 |
DPRINTF("Trying %s => %s\n", rdma->host, ip);
|
| 922 |
|
| 923 |
ret = rdma_resolve_addr(rdma->cm_id, NULL, e->ai_dst_addr,
|
| 924 |
RDMA_RESOLVE_TIMEOUT_MS); |
| 925 |
if (!ret) {
|
| 926 |
if (e->ai_family == AF_INET6) {
|
| 927 |
ret = qemu_rdma_broken_ipv6_kernel(errp, rdma->cm_id->verbs); |
| 928 |
if (ret) {
|
| 929 |
continue;
|
| 930 |
} |
| 931 |
} |
| 932 |
goto route;
|
| 933 |
} |
| 934 |
} |
| 935 |
|
| 936 |
ERROR(errp, "could not resolve address %s", rdma->host);
|
| 937 |
goto err_resolve_get_addr;
|
| 938 |
|
| 939 |
route:
|
| 940 |
qemu_rdma_dump_gid("source_resolve_addr", rdma->cm_id);
|
| 941 |
|
| 942 |
ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 943 |
if (ret) {
|
| 944 |
ERROR(errp, "could not perform event_addr_resolved");
|
| 945 |
goto err_resolve_get_addr;
|
| 946 |
} |
| 947 |
|
| 948 |
if (cm_event->event != RDMA_CM_EVENT_ADDR_RESOLVED) {
|
| 949 |
ERROR(errp, "result not equal to event_addr_resolved %s",
|
| 950 |
rdma_event_str(cm_event->event)); |
| 951 |
perror("rdma_resolve_addr");
|
| 952 |
ret = -EINVAL; |
| 953 |
goto err_resolve_get_addr;
|
| 954 |
} |
| 955 |
rdma_ack_cm_event(cm_event); |
| 956 |
|
| 957 |
/* resolve route */
|
| 958 |
ret = rdma_resolve_route(rdma->cm_id, RDMA_RESOLVE_TIMEOUT_MS); |
| 959 |
if (ret) {
|
| 960 |
ERROR(errp, "could not resolve rdma route");
|
| 961 |
goto err_resolve_get_addr;
|
| 962 |
} |
| 963 |
|
| 964 |
ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 965 |
if (ret) {
|
| 966 |
ERROR(errp, "could not perform event_route_resolved");
|
| 967 |
goto err_resolve_get_addr;
|
| 968 |
} |
| 969 |
if (cm_event->event != RDMA_CM_EVENT_ROUTE_RESOLVED) {
|
| 970 |
ERROR(errp, "result not equal to event_route_resolved: %s",
|
| 971 |
rdma_event_str(cm_event->event)); |
| 972 |
rdma_ack_cm_event(cm_event); |
| 973 |
ret = -EINVAL; |
| 974 |
goto err_resolve_get_addr;
|
| 975 |
} |
| 976 |
rdma_ack_cm_event(cm_event); |
| 977 |
rdma->verbs = rdma->cm_id->verbs; |
| 978 |
qemu_rdma_dump_id("source_resolve_host", rdma->cm_id->verbs);
|
| 979 |
qemu_rdma_dump_gid("source_resolve_host", rdma->cm_id);
|
| 980 |
return 0; |
| 981 |
|
| 982 |
err_resolve_get_addr:
|
| 983 |
rdma_destroy_id(rdma->cm_id); |
| 984 |
rdma->cm_id = NULL;
|
| 985 |
err_resolve_create_id:
|
| 986 |
rdma_destroy_event_channel(rdma->channel); |
| 987 |
rdma->channel = NULL;
|
| 988 |
return ret;
|
| 989 |
} |
| 990 |
|
| 991 |
/*
|
| 992 |
* Create protection domain and completion queues
|
| 993 |
*/
|
| 994 |
static int qemu_rdma_alloc_pd_cq(RDMAContext *rdma) |
| 995 |
{
|
| 996 |
/* allocate pd */
|
| 997 |
rdma->pd = ibv_alloc_pd(rdma->verbs); |
| 998 |
if (!rdma->pd) {
|
| 999 |
fprintf(stderr, "failed to allocate protection domain\n");
|
| 1000 |
return -1; |
| 1001 |
} |
| 1002 |
|
| 1003 |
/* create completion channel */
|
| 1004 |
rdma->comp_channel = ibv_create_comp_channel(rdma->verbs); |
| 1005 |
if (!rdma->comp_channel) {
|
| 1006 |
fprintf(stderr, "failed to allocate completion channel\n");
|
| 1007 |
goto err_alloc_pd_cq;
|
| 1008 |
} |
| 1009 |
|
| 1010 |
/*
|
| 1011 |
* Completion queue can be filled by both read and write work requests,
|
| 1012 |
* so must reflect the sum of both possible queue sizes.
|
| 1013 |
*/
|
| 1014 |
rdma->cq = ibv_create_cq(rdma->verbs, (RDMA_SIGNALED_SEND_MAX * 3),
|
| 1015 |
NULL, rdma->comp_channel, 0); |
| 1016 |
if (!rdma->cq) {
|
| 1017 |
fprintf(stderr, "failed to allocate completion queue\n");
|
| 1018 |
goto err_alloc_pd_cq;
|
| 1019 |
} |
| 1020 |
|
| 1021 |
return 0; |
| 1022 |
|
| 1023 |
err_alloc_pd_cq:
|
| 1024 |
if (rdma->pd) {
|
| 1025 |
ibv_dealloc_pd(rdma->pd); |
| 1026 |
} |
| 1027 |
if (rdma->comp_channel) {
|
| 1028 |
ibv_destroy_comp_channel(rdma->comp_channel); |
| 1029 |
} |
| 1030 |
rdma->pd = NULL;
|
| 1031 |
rdma->comp_channel = NULL;
|
| 1032 |
return -1; |
| 1033 |
|
| 1034 |
} |
| 1035 |
|
| 1036 |
/*
|
| 1037 |
* Create queue pairs.
|
| 1038 |
*/
|
| 1039 |
static int qemu_rdma_alloc_qp(RDMAContext *rdma) |
| 1040 |
{
|
| 1041 |
struct ibv_qp_init_attr attr = { 0 }; |
| 1042 |
int ret;
|
| 1043 |
|
| 1044 |
attr.cap.max_send_wr = RDMA_SIGNALED_SEND_MAX; |
| 1045 |
attr.cap.max_recv_wr = 3;
|
| 1046 |
attr.cap.max_send_sge = 1;
|
| 1047 |
attr.cap.max_recv_sge = 1;
|
| 1048 |
attr.send_cq = rdma->cq; |
| 1049 |
attr.recv_cq = rdma->cq; |
| 1050 |
attr.qp_type = IBV_QPT_RC; |
| 1051 |
|
| 1052 |
ret = rdma_create_qp(rdma->cm_id, rdma->pd, &attr); |
| 1053 |
if (ret) {
|
| 1054 |
return -1; |
| 1055 |
} |
| 1056 |
|
| 1057 |
rdma->qp = rdma->cm_id->qp; |
| 1058 |
return 0; |
| 1059 |
} |
| 1060 |
|
| 1061 |
static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) |
| 1062 |
{
|
| 1063 |
int i;
|
| 1064 |
RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 1065 |
|
| 1066 |
for (i = 0; i < local->nb_blocks; i++) { |
| 1067 |
local->block[i].mr = |
| 1068 |
ibv_reg_mr(rdma->pd, |
| 1069 |
local->block[i].local_host_addr, |
| 1070 |
local->block[i].length, |
| 1071 |
IBV_ACCESS_LOCAL_WRITE | |
| 1072 |
IBV_ACCESS_REMOTE_WRITE |
| 1073 |
); |
| 1074 |
if (!local->block[i].mr) {
|
| 1075 |
perror("Failed to register local dest ram block!\n");
|
| 1076 |
break;
|
| 1077 |
} |
| 1078 |
rdma->total_registrations++; |
| 1079 |
} |
| 1080 |
|
| 1081 |
if (i >= local->nb_blocks) {
|
| 1082 |
return 0; |
| 1083 |
} |
| 1084 |
|
| 1085 |
for (i--; i >= 0; i--) { |
| 1086 |
ibv_dereg_mr(local->block[i].mr); |
| 1087 |
rdma->total_registrations--; |
| 1088 |
} |
| 1089 |
|
| 1090 |
return -1; |
| 1091 |
|
| 1092 |
} |
| 1093 |
|
| 1094 |
/*
|
| 1095 |
* Find the ram block that corresponds to the page requested to be
|
| 1096 |
* transmitted by QEMU.
|
| 1097 |
*
|
| 1098 |
* Once the block is found, also identify which 'chunk' within that
|
| 1099 |
* block that the page belongs to.
|
| 1100 |
*
|
| 1101 |
* This search cannot fail or the migration will fail.
|
| 1102 |
*/
|
| 1103 |
static int qemu_rdma_search_ram_block(RDMAContext *rdma, |
| 1104 |
uint64_t block_offset, |
| 1105 |
uint64_t offset, |
| 1106 |
uint64_t length, |
| 1107 |
uint64_t *block_index, |
| 1108 |
uint64_t *chunk_index) |
| 1109 |
{
|
| 1110 |
uint64_t current_addr = block_offset + offset; |
| 1111 |
RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, |
| 1112 |
(void *) block_offset);
|
| 1113 |
assert(block); |
| 1114 |
assert(current_addr >= block->offset); |
| 1115 |
assert((current_addr + length) <= (block->offset + block->length)); |
| 1116 |
|
| 1117 |
*block_index = block->index; |
| 1118 |
*chunk_index = ram_chunk_index(block->local_host_addr, |
| 1119 |
block->local_host_addr + (current_addr - block->offset)); |
| 1120 |
|
| 1121 |
return 0; |
| 1122 |
} |
| 1123 |
|
| 1124 |
/*
|
| 1125 |
* Register a chunk with IB. If the chunk was already registered
|
| 1126 |
* previously, then skip.
|
| 1127 |
*
|
| 1128 |
* Also return the keys associated with the registration needed
|
| 1129 |
* to perform the actual RDMA operation.
|
| 1130 |
*/
|
| 1131 |
static int qemu_rdma_register_and_get_keys(RDMAContext *rdma, |
| 1132 |
RDMALocalBlock *block, uint8_t *host_addr, |
| 1133 |
uint32_t *lkey, uint32_t *rkey, int chunk,
|
| 1134 |
uint8_t *chunk_start, uint8_t *chunk_end) |
| 1135 |
{
|
| 1136 |
if (block->mr) {
|
| 1137 |
if (lkey) {
|
| 1138 |
*lkey = block->mr->lkey; |
| 1139 |
} |
| 1140 |
if (rkey) {
|
| 1141 |
*rkey = block->mr->rkey; |
| 1142 |
} |
| 1143 |
return 0; |
| 1144 |
} |
| 1145 |
|
| 1146 |
/* allocate memory to store chunk MRs */
|
| 1147 |
if (!block->pmr) {
|
| 1148 |
block->pmr = g_malloc0(block->nb_chunks * sizeof(struct ibv_mr *)); |
| 1149 |
if (!block->pmr) {
|
| 1150 |
return -1; |
| 1151 |
} |
| 1152 |
} |
| 1153 |
|
| 1154 |
/*
|
| 1155 |
* If 'rkey', then we're the destination, so grant access to the source.
|
| 1156 |
*
|
| 1157 |
* If 'lkey', then we're the source VM, so grant access only to ourselves.
|
| 1158 |
*/
|
| 1159 |
if (!block->pmr[chunk]) {
|
| 1160 |
uint64_t len = chunk_end - chunk_start; |
| 1161 |
|
| 1162 |
DDPRINTF("Registering %" PRIu64 " bytes @ %p\n", |
| 1163 |
len, chunk_start); |
| 1164 |
|
| 1165 |
block->pmr[chunk] = ibv_reg_mr(rdma->pd, |
| 1166 |
chunk_start, len, |
| 1167 |
(rkey ? (IBV_ACCESS_LOCAL_WRITE | |
| 1168 |
IBV_ACCESS_REMOTE_WRITE) : 0));
|
| 1169 |
|
| 1170 |
if (!block->pmr[chunk]) {
|
| 1171 |
perror("Failed to register chunk!");
|
| 1172 |
fprintf(stderr, "Chunk details: block: %d chunk index %d"
|
| 1173 |
" start %" PRIu64 " end %" PRIu64 " host %" PRIu64 |
| 1174 |
" local %" PRIu64 " registrations: %d\n", |
| 1175 |
block->index, chunk, (uint64_t) chunk_start, |
| 1176 |
(uint64_t) chunk_end, (uint64_t) host_addr, |
| 1177 |
(uint64_t) block->local_host_addr, |
| 1178 |
rdma->total_registrations); |
| 1179 |
return -1; |
| 1180 |
} |
| 1181 |
rdma->total_registrations++; |
| 1182 |
} |
| 1183 |
|
| 1184 |
if (lkey) {
|
| 1185 |
*lkey = block->pmr[chunk]->lkey; |
| 1186 |
} |
| 1187 |
if (rkey) {
|
| 1188 |
*rkey = block->pmr[chunk]->rkey; |
| 1189 |
} |
| 1190 |
return 0; |
| 1191 |
} |
| 1192 |
|
| 1193 |
/*
|
| 1194 |
* Register (at connection time) the memory used for control
|
| 1195 |
* channel messages.
|
| 1196 |
*/
|
| 1197 |
static int qemu_rdma_reg_control(RDMAContext *rdma, int idx) |
| 1198 |
{
|
| 1199 |
rdma->wr_data[idx].control_mr = ibv_reg_mr(rdma->pd, |
| 1200 |
rdma->wr_data[idx].control, RDMA_CONTROL_MAX_BUFFER, |
| 1201 |
IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); |
| 1202 |
if (rdma->wr_data[idx].control_mr) {
|
| 1203 |
rdma->total_registrations++; |
| 1204 |
return 0; |
| 1205 |
} |
| 1206 |
fprintf(stderr, "qemu_rdma_reg_control failed!\n");
|
| 1207 |
return -1; |
| 1208 |
} |
| 1209 |
|
| 1210 |
const char *print_wrid(int wrid) |
| 1211 |
{
|
| 1212 |
if (wrid >= RDMA_WRID_RECV_CONTROL) {
|
| 1213 |
return wrid_desc[RDMA_WRID_RECV_CONTROL];
|
| 1214 |
} |
| 1215 |
return wrid_desc[wrid];
|
| 1216 |
} |
| 1217 |
|
| 1218 |
/*
|
| 1219 |
* RDMA requires memory registration (mlock/pinning), but this is not good for
|
| 1220 |
* overcommitment.
|
| 1221 |
*
|
| 1222 |
* In preparation for the future where LRU information or workload-specific
|
| 1223 |
* writable writable working set memory access behavior is available to QEMU
|
| 1224 |
* it would be nice to have in place the ability to UN-register/UN-pin
|
| 1225 |
* particular memory regions from the RDMA hardware when it is determine that
|
| 1226 |
* those regions of memory will likely not be accessed again in the near future.
|
| 1227 |
*
|
| 1228 |
* While we do not yet have such information right now, the following
|
| 1229 |
* compile-time option allows us to perform a non-optimized version of this
|
| 1230 |
* behavior.
|
| 1231 |
*
|
| 1232 |
* By uncommenting this option, you will cause *all* RDMA transfers to be
|
| 1233 |
* unregistered immediately after the transfer completes on both sides of the
|
| 1234 |
* connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
|
| 1235 |
*
|
| 1236 |
* This will have a terrible impact on migration performance, so until future
|
| 1237 |
* workload information or LRU information is available, do not attempt to use
|
| 1238 |
* this feature except for basic testing.
|
| 1239 |
*/
|
| 1240 |
//#define RDMA_UNREGISTRATION_EXAMPLE
|
| 1241 |
|
| 1242 |
/*
|
| 1243 |
* Perform a non-optimized memory unregistration after every transfer
|
| 1244 |
* for demonsration purposes, only if pin-all is not requested.
|
| 1245 |
*
|
| 1246 |
* Potential optimizations:
|
| 1247 |
* 1. Start a new thread to run this function continuously
|
| 1248 |
- for bit clearing
|
| 1249 |
- and for receipt of unregister messages
|
| 1250 |
* 2. Use an LRU.
|
| 1251 |
* 3. Use workload hints.
|
| 1252 |
*/
|
| 1253 |
static int qemu_rdma_unregister_waiting(RDMAContext *rdma) |
| 1254 |
{
|
| 1255 |
while (rdma->unregistrations[rdma->unregister_current]) {
|
| 1256 |
int ret;
|
| 1257 |
uint64_t wr_id = rdma->unregistrations[rdma->unregister_current]; |
| 1258 |
uint64_t chunk = |
| 1259 |
(wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; |
| 1260 |
uint64_t index = |
| 1261 |
(wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; |
| 1262 |
RDMALocalBlock *block = |
| 1263 |
&(rdma->local_ram_blocks.block[index]); |
| 1264 |
RDMARegister reg = { .current_index = index };
|
| 1265 |
RDMAControlHeader resp = { .type = RDMA_CONTROL_UNREGISTER_FINISHED,
|
| 1266 |
}; |
| 1267 |
RDMAControlHeader head = { .len = sizeof(RDMARegister),
|
| 1268 |
.type = RDMA_CONTROL_UNREGISTER_REQUEST, |
| 1269 |
.repeat = 1,
|
| 1270 |
}; |
| 1271 |
|
| 1272 |
DDPRINTF("Processing unregister for chunk: %" PRIu64
|
| 1273 |
" at position %d\n", chunk, rdma->unregister_current);
|
| 1274 |
|
| 1275 |
rdma->unregistrations[rdma->unregister_current] = 0;
|
| 1276 |
rdma->unregister_current++; |
| 1277 |
|
| 1278 |
if (rdma->unregister_current == RDMA_SIGNALED_SEND_MAX) {
|
| 1279 |
rdma->unregister_current = 0;
|
| 1280 |
} |
| 1281 |
|
| 1282 |
|
| 1283 |
/*
|
| 1284 |
* Unregistration is speculative (because migration is single-threaded
|
| 1285 |
* and we cannot break the protocol's inifinband message ordering).
|
| 1286 |
* Thus, if the memory is currently being used for transmission,
|
| 1287 |
* then abort the attempt to unregister and try again
|
| 1288 |
* later the next time a completion is received for this memory.
|
| 1289 |
*/
|
| 1290 |
clear_bit(chunk, block->unregister_bitmap); |
| 1291 |
|
| 1292 |
if (test_bit(chunk, block->transit_bitmap)) {
|
| 1293 |
DDPRINTF("Cannot unregister inflight chunk: %" PRIu64 "\n", chunk); |
| 1294 |
continue;
|
| 1295 |
} |
| 1296 |
|
| 1297 |
DDPRINTF("Sending unregister for chunk: %" PRIu64 "\n", chunk); |
| 1298 |
|
| 1299 |
ret = ibv_dereg_mr(block->pmr[chunk]); |
| 1300 |
block->pmr[chunk] = NULL;
|
| 1301 |
block->remote_keys[chunk] = 0;
|
| 1302 |
|
| 1303 |
if (ret != 0) { |
| 1304 |
perror("unregistration chunk failed");
|
| 1305 |
return -ret;
|
| 1306 |
} |
| 1307 |
rdma->total_registrations--; |
| 1308 |
|
| 1309 |
reg.key.chunk = chunk; |
| 1310 |
register_to_network(®); |
| 1311 |
ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
| 1312 |
&resp, NULL, NULL); |
| 1313 |
if (ret < 0) { |
| 1314 |
return ret;
|
| 1315 |
} |
| 1316 |
|
| 1317 |
DDPRINTF("Unregister for chunk: %" PRIu64 " complete.\n", chunk); |
| 1318 |
} |
| 1319 |
|
| 1320 |
return 0; |
| 1321 |
} |
| 1322 |
|
| 1323 |
static uint64_t qemu_rdma_make_wrid(uint64_t wr_id, uint64_t index,
|
| 1324 |
uint64_t chunk) |
| 1325 |
{
|
| 1326 |
uint64_t result = wr_id & RDMA_WRID_TYPE_MASK; |
| 1327 |
|
| 1328 |
result |= (index << RDMA_WRID_BLOCK_SHIFT); |
| 1329 |
result |= (chunk << RDMA_WRID_CHUNK_SHIFT); |
| 1330 |
|
| 1331 |
return result;
|
| 1332 |
} |
| 1333 |
|
| 1334 |
/*
|
| 1335 |
* Set bit for unregistration in the next iteration.
|
| 1336 |
* We cannot transmit right here, but will unpin later.
|
| 1337 |
*/
|
| 1338 |
static void qemu_rdma_signal_unregister(RDMAContext *rdma, uint64_t index, |
| 1339 |
uint64_t chunk, uint64_t wr_id) |
| 1340 |
{
|
| 1341 |
if (rdma->unregistrations[rdma->unregister_next] != 0) { |
| 1342 |
fprintf(stderr, "rdma migration: queue is full!\n");
|
| 1343 |
} else {
|
| 1344 |
RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); |
| 1345 |
|
| 1346 |
if (!test_and_set_bit(chunk, block->unregister_bitmap)) {
|
| 1347 |
DDPRINTF("Appending unregister chunk %" PRIu64
|
| 1348 |
" at position %d\n", chunk, rdma->unregister_next);
|
| 1349 |
|
| 1350 |
rdma->unregistrations[rdma->unregister_next++] = |
| 1351 |
qemu_rdma_make_wrid(wr_id, index, chunk); |
| 1352 |
|
| 1353 |
if (rdma->unregister_next == RDMA_SIGNALED_SEND_MAX) {
|
| 1354 |
rdma->unregister_next = 0;
|
| 1355 |
} |
| 1356 |
} else {
|
| 1357 |
DDPRINTF("Unregister chunk %" PRIu64 " already in queue.\n", |
| 1358 |
chunk); |
| 1359 |
} |
| 1360 |
} |
| 1361 |
} |
| 1362 |
|
| 1363 |
/*
|
| 1364 |
* Consult the connection manager to see a work request
|
| 1365 |
* (of any kind) has completed.
|
| 1366 |
* Return the work request ID that completed.
|
| 1367 |
*/
|
| 1368 |
static uint64_t qemu_rdma_poll(RDMAContext *rdma, uint64_t *wr_id_out,
|
| 1369 |
uint32_t *byte_len) |
| 1370 |
{
|
| 1371 |
int ret;
|
| 1372 |
struct ibv_wc wc;
|
| 1373 |
uint64_t wr_id; |
| 1374 |
|
| 1375 |
ret = ibv_poll_cq(rdma->cq, 1, &wc);
|
| 1376 |
|
| 1377 |
if (!ret) {
|
| 1378 |
*wr_id_out = RDMA_WRID_NONE; |
| 1379 |
return 0; |
| 1380 |
} |
| 1381 |
|
| 1382 |
if (ret < 0) { |
| 1383 |
fprintf(stderr, "ibv_poll_cq return %d!\n", ret);
|
| 1384 |
return ret;
|
| 1385 |
} |
| 1386 |
|
| 1387 |
wr_id = wc.wr_id & RDMA_WRID_TYPE_MASK; |
| 1388 |
|
| 1389 |
if (wc.status != IBV_WC_SUCCESS) {
|
| 1390 |
fprintf(stderr, "ibv_poll_cq wc.status=%d %s!\n",
|
| 1391 |
wc.status, ibv_wc_status_str(wc.status)); |
| 1392 |
fprintf(stderr, "ibv_poll_cq wrid=%s!\n", wrid_desc[wr_id]);
|
| 1393 |
|
| 1394 |
return -1; |
| 1395 |
} |
| 1396 |
|
| 1397 |
if (rdma->control_ready_expected &&
|
| 1398 |
(wr_id >= RDMA_WRID_RECV_CONTROL)) {
|
| 1399 |
DDDPRINTF("completion %s #%" PRId64 " received (%" PRId64 ")" |
| 1400 |
" left %d\n", wrid_desc[RDMA_WRID_RECV_CONTROL],
|
| 1401 |
wr_id - RDMA_WRID_RECV_CONTROL, wr_id, rdma->nb_sent); |
| 1402 |
rdma->control_ready_expected = 0;
|
| 1403 |
} |
| 1404 |
|
| 1405 |
if (wr_id == RDMA_WRID_RDMA_WRITE) {
|
| 1406 |
uint64_t chunk = |
| 1407 |
(wc.wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; |
| 1408 |
uint64_t index = |
| 1409 |
(wc.wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; |
| 1410 |
RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); |
| 1411 |
|
| 1412 |
DDDPRINTF("completions %s (%" PRId64 ") left %d, " |
| 1413 |
"block %" PRIu64 ", chunk: %" PRIu64 " %p %p\n", |
| 1414 |
print_wrid(wr_id), wr_id, rdma->nb_sent, index, chunk, |
| 1415 |
block->local_host_addr, (void *)block->remote_host_addr);
|
| 1416 |
|
| 1417 |
clear_bit(chunk, block->transit_bitmap); |
| 1418 |
|
| 1419 |
if (rdma->nb_sent > 0) { |
| 1420 |
rdma->nb_sent--; |
| 1421 |
} |
| 1422 |
|
| 1423 |
if (!rdma->pin_all) {
|
| 1424 |
/*
|
| 1425 |
* FYI: If one wanted to signal a specific chunk to be unregistered
|
| 1426 |
* using LRU or workload-specific information, this is the function
|
| 1427 |
* you would call to do so. That chunk would then get asynchronously
|
| 1428 |
* unregistered later.
|
| 1429 |
*/
|
| 1430 |
#ifdef RDMA_UNREGISTRATION_EXAMPLE
|
| 1431 |
qemu_rdma_signal_unregister(rdma, index, chunk, wc.wr_id); |
| 1432 |
#endif
|
| 1433 |
} |
| 1434 |
} else {
|
| 1435 |
DDDPRINTF("other completion %s (%" PRId64 ") received left %d\n", |
| 1436 |
print_wrid(wr_id), wr_id, rdma->nb_sent); |
| 1437 |
} |
| 1438 |
|
| 1439 |
*wr_id_out = wc.wr_id; |
| 1440 |
if (byte_len) {
|
| 1441 |
*byte_len = wc.byte_len; |
| 1442 |
} |
| 1443 |
|
| 1444 |
return 0; |
| 1445 |
} |
| 1446 |
|
| 1447 |
/*
|
| 1448 |
* Block until the next work request has completed.
|
| 1449 |
*
|
| 1450 |
* First poll to see if a work request has already completed,
|
| 1451 |
* otherwise block.
|
| 1452 |
*
|
| 1453 |
* If we encounter completed work requests for IDs other than
|
| 1454 |
* the one we're interested in, then that's generally an error.
|
| 1455 |
*
|
| 1456 |
* The only exception is actual RDMA Write completions. These
|
| 1457 |
* completions only need to be recorded, but do not actually
|
| 1458 |
* need further processing.
|
| 1459 |
*/
|
| 1460 |
static int qemu_rdma_block_for_wrid(RDMAContext *rdma, int wrid_requested, |
| 1461 |
uint32_t *byte_len) |
| 1462 |
{
|
| 1463 |
int num_cq_events = 0, ret = 0; |
| 1464 |
struct ibv_cq *cq;
|
| 1465 |
void *cq_ctx;
|
| 1466 |
uint64_t wr_id = RDMA_WRID_NONE, wr_id_in; |
| 1467 |
|
| 1468 |
if (ibv_req_notify_cq(rdma->cq, 0)) { |
| 1469 |
return -1; |
| 1470 |
} |
| 1471 |
/* poll cq first */
|
| 1472 |
while (wr_id != wrid_requested) {
|
| 1473 |
ret = qemu_rdma_poll(rdma, &wr_id_in, byte_len); |
| 1474 |
if (ret < 0) { |
| 1475 |
return ret;
|
| 1476 |
} |
| 1477 |
|
| 1478 |
wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; |
| 1479 |
|
| 1480 |
if (wr_id == RDMA_WRID_NONE) {
|
| 1481 |
break;
|
| 1482 |
} |
| 1483 |
if (wr_id != wrid_requested) {
|
| 1484 |
DDDPRINTF("A Wanted wrid %s (%d) but got %s (%" PRIu64 ")\n", |
| 1485 |
print_wrid(wrid_requested), |
| 1486 |
wrid_requested, print_wrid(wr_id), wr_id); |
| 1487 |
} |
| 1488 |
} |
| 1489 |
|
| 1490 |
if (wr_id == wrid_requested) {
|
| 1491 |
return 0; |
| 1492 |
} |
| 1493 |
|
| 1494 |
while (1) { |
| 1495 |
/*
|
| 1496 |
* Coroutine doesn't start until process_incoming_migration()
|
| 1497 |
* so don't yield unless we know we're running inside of a coroutine.
|
| 1498 |
*/
|
| 1499 |
if (rdma->migration_started_on_destination) {
|
| 1500 |
yield_until_fd_readable(rdma->comp_channel->fd); |
| 1501 |
} |
| 1502 |
|
| 1503 |
if (ibv_get_cq_event(rdma->comp_channel, &cq, &cq_ctx)) {
|
| 1504 |
perror("ibv_get_cq_event");
|
| 1505 |
goto err_block_for_wrid;
|
| 1506 |
} |
| 1507 |
|
| 1508 |
num_cq_events++; |
| 1509 |
|
| 1510 |
if (ibv_req_notify_cq(cq, 0)) { |
| 1511 |
goto err_block_for_wrid;
|
| 1512 |
} |
| 1513 |
|
| 1514 |
while (wr_id != wrid_requested) {
|
| 1515 |
ret = qemu_rdma_poll(rdma, &wr_id_in, byte_len); |
| 1516 |
if (ret < 0) { |
| 1517 |
goto err_block_for_wrid;
|
| 1518 |
} |
| 1519 |
|
| 1520 |
wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; |
| 1521 |
|
| 1522 |
if (wr_id == RDMA_WRID_NONE) {
|
| 1523 |
break;
|
| 1524 |
} |
| 1525 |
if (wr_id != wrid_requested) {
|
| 1526 |
DDDPRINTF("B Wanted wrid %s (%d) but got %s (%" PRIu64 ")\n", |
| 1527 |
print_wrid(wrid_requested), wrid_requested, |
| 1528 |
print_wrid(wr_id), wr_id); |
| 1529 |
} |
| 1530 |
} |
| 1531 |
|
| 1532 |
if (wr_id == wrid_requested) {
|
| 1533 |
goto success_block_for_wrid;
|
| 1534 |
} |
| 1535 |
} |
| 1536 |
|
| 1537 |
success_block_for_wrid:
|
| 1538 |
if (num_cq_events) {
|
| 1539 |
ibv_ack_cq_events(cq, num_cq_events); |
| 1540 |
} |
| 1541 |
return 0; |
| 1542 |
|
| 1543 |
err_block_for_wrid:
|
| 1544 |
if (num_cq_events) {
|
| 1545 |
ibv_ack_cq_events(cq, num_cq_events); |
| 1546 |
} |
| 1547 |
return ret;
|
| 1548 |
} |
| 1549 |
|
| 1550 |
/*
|
| 1551 |
* Post a SEND message work request for the control channel
|
| 1552 |
* containing some data and block until the post completes.
|
| 1553 |
*/
|
| 1554 |
static int qemu_rdma_post_send_control(RDMAContext *rdma, uint8_t *buf, |
| 1555 |
RDMAControlHeader *head) |
| 1556 |
{
|
| 1557 |
int ret = 0; |
| 1558 |
RDMAWorkRequestData *wr = &rdma->wr_data[RDMA_WRID_CONTROL]; |
| 1559 |
struct ibv_send_wr *bad_wr;
|
| 1560 |
struct ibv_sge sge = {
|
| 1561 |
.addr = (uint64_t)(wr->control), |
| 1562 |
.length = head->len + sizeof(RDMAControlHeader),
|
| 1563 |
.lkey = wr->control_mr->lkey, |
| 1564 |
}; |
| 1565 |
struct ibv_send_wr send_wr = {
|
| 1566 |
.wr_id = RDMA_WRID_SEND_CONTROL, |
| 1567 |
.opcode = IBV_WR_SEND, |
| 1568 |
.send_flags = IBV_SEND_SIGNALED, |
| 1569 |
.sg_list = &sge, |
| 1570 |
.num_sge = 1,
|
| 1571 |
}; |
| 1572 |
|
| 1573 |
DDDPRINTF("CONTROL: sending %s..\n", control_desc[head->type]);
|
| 1574 |
|
| 1575 |
/*
|
| 1576 |
* We don't actually need to do a memcpy() in here if we used
|
| 1577 |
* the "sge" properly, but since we're only sending control messages
|
| 1578 |
* (not RAM in a performance-critical path), then its OK for now.
|
| 1579 |
*
|
| 1580 |
* The copy makes the RDMAControlHeader simpler to manipulate
|
| 1581 |
* for the time being.
|
| 1582 |
*/
|
| 1583 |
assert(head->len <= RDMA_CONTROL_MAX_BUFFER - sizeof(*head));
|
| 1584 |
memcpy(wr->control, head, sizeof(RDMAControlHeader));
|
| 1585 |
control_to_network((void *) wr->control);
|
| 1586 |
|
| 1587 |
if (buf) {
|
| 1588 |
memcpy(wr->control + sizeof(RDMAControlHeader), buf, head->len);
|
| 1589 |
} |
| 1590 |
|
| 1591 |
|
| 1592 |
if (ibv_post_send(rdma->qp, &send_wr, &bad_wr)) {
|
| 1593 |
return -1; |
| 1594 |
} |
| 1595 |
|
| 1596 |
if (ret < 0) { |
| 1597 |
fprintf(stderr, "Failed to use post IB SEND for control!\n");
|
| 1598 |
return ret;
|
| 1599 |
} |
| 1600 |
|
| 1601 |
ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_SEND_CONTROL, NULL);
|
| 1602 |
if (ret < 0) { |
| 1603 |
fprintf(stderr, "rdma migration: send polling control error!\n");
|
| 1604 |
} |
| 1605 |
|
| 1606 |
return ret;
|
| 1607 |
} |
| 1608 |
|
| 1609 |
/*
|
| 1610 |
* Post a RECV work request in anticipation of some future receipt
|
| 1611 |
* of data on the control channel.
|
| 1612 |
*/
|
| 1613 |
static int qemu_rdma_post_recv_control(RDMAContext *rdma, int idx) |
| 1614 |
{
|
| 1615 |
struct ibv_recv_wr *bad_wr;
|
| 1616 |
struct ibv_sge sge = {
|
| 1617 |
.addr = (uint64_t)(rdma->wr_data[idx].control), |
| 1618 |
.length = RDMA_CONTROL_MAX_BUFFER, |
| 1619 |
.lkey = rdma->wr_data[idx].control_mr->lkey, |
| 1620 |
}; |
| 1621 |
|
| 1622 |
struct ibv_recv_wr recv_wr = {
|
| 1623 |
.wr_id = RDMA_WRID_RECV_CONTROL + idx, |
| 1624 |
.sg_list = &sge, |
| 1625 |
.num_sge = 1,
|
| 1626 |
}; |
| 1627 |
|
| 1628 |
|
| 1629 |
if (ibv_post_recv(rdma->qp, &recv_wr, &bad_wr)) {
|
| 1630 |
return -1; |
| 1631 |
} |
| 1632 |
|
| 1633 |
return 0; |
| 1634 |
} |
| 1635 |
|
| 1636 |
/*
|
| 1637 |
* Block and wait for a RECV control channel message to arrive.
|
| 1638 |
*/
|
| 1639 |
static int qemu_rdma_exchange_get_response(RDMAContext *rdma, |
| 1640 |
RDMAControlHeader *head, int expecting, int idx) |
| 1641 |
{
|
| 1642 |
uint32_t byte_len; |
| 1643 |
int ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RECV_CONTROL + idx,
|
| 1644 |
&byte_len); |
| 1645 |
|
| 1646 |
if (ret < 0) { |
| 1647 |
fprintf(stderr, "rdma migration: recv polling control error!\n");
|
| 1648 |
return ret;
|
| 1649 |
} |
| 1650 |
|
| 1651 |
network_to_control((void *) rdma->wr_data[idx].control);
|
| 1652 |
memcpy(head, rdma->wr_data[idx].control, sizeof(RDMAControlHeader));
|
| 1653 |
|
| 1654 |
DDDPRINTF("CONTROL: %s receiving...\n", control_desc[expecting]);
|
| 1655 |
|
| 1656 |
if (expecting == RDMA_CONTROL_NONE) {
|
| 1657 |
DDDPRINTF("Surprise: got %s (%d)\n",
|
| 1658 |
control_desc[head->type], head->type); |
| 1659 |
} else if (head->type != expecting || head->type == RDMA_CONTROL_ERROR) { |
| 1660 |
fprintf(stderr, "Was expecting a %s (%d) control message"
|
| 1661 |
", but got: %s (%d), length: %d\n",
|
| 1662 |
control_desc[expecting], expecting, |
| 1663 |
control_desc[head->type], head->type, head->len); |
| 1664 |
return -EIO;
|
| 1665 |
} |
| 1666 |
if (head->len > RDMA_CONTROL_MAX_BUFFER - sizeof(*head)) { |
| 1667 |
fprintf(stderr, "too long length: %d\n", head->len);
|
| 1668 |
return -EINVAL;
|
| 1669 |
} |
| 1670 |
if (sizeof(*head) + head->len != byte_len) { |
| 1671 |
fprintf(stderr, "Malformed length: %d byte_len %d\n",
|
| 1672 |
head->len, byte_len); |
| 1673 |
return -EINVAL;
|
| 1674 |
} |
| 1675 |
|
| 1676 |
return 0; |
| 1677 |
} |
| 1678 |
|
| 1679 |
/*
|
| 1680 |
* When a RECV work request has completed, the work request's
|
| 1681 |
* buffer is pointed at the header.
|
| 1682 |
*
|
| 1683 |
* This will advance the pointer to the data portion
|
| 1684 |
* of the control message of the work request's buffer that
|
| 1685 |
* was populated after the work request finished.
|
| 1686 |
*/
|
| 1687 |
static void qemu_rdma_move_header(RDMAContext *rdma, int idx, |
| 1688 |
RDMAControlHeader *head) |
| 1689 |
{
|
| 1690 |
rdma->wr_data[idx].control_len = head->len; |
| 1691 |
rdma->wr_data[idx].control_curr = |
| 1692 |
rdma->wr_data[idx].control + sizeof(RDMAControlHeader);
|
| 1693 |
} |
| 1694 |
|
| 1695 |
/*
|
| 1696 |
* This is an 'atomic' high-level operation to deliver a single, unified
|
| 1697 |
* control-channel message.
|
| 1698 |
*
|
| 1699 |
* Additionally, if the user is expecting some kind of reply to this message,
|
| 1700 |
* they can request a 'resp' response message be filled in by posting an
|
| 1701 |
* additional work request on behalf of the user and waiting for an additional
|
| 1702 |
* completion.
|
| 1703 |
*
|
| 1704 |
* The extra (optional) response is used during registration to us from having
|
| 1705 |
* to perform an *additional* exchange of message just to provide a response by
|
| 1706 |
* instead piggy-backing on the acknowledgement.
|
| 1707 |
*/
|
| 1708 |
static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, |
| 1709 |
uint8_t *data, RDMAControlHeader *resp, |
| 1710 |
int *resp_idx,
|
| 1711 |
int (*callback)(RDMAContext *rdma))
|
| 1712 |
{
|
| 1713 |
int ret = 0; |
| 1714 |
|
| 1715 |
/*
|
| 1716 |
* Wait until the dest is ready before attempting to deliver the message
|
| 1717 |
* by waiting for a READY message.
|
| 1718 |
*/
|
| 1719 |
if (rdma->control_ready_expected) {
|
| 1720 |
RDMAControlHeader resp; |
| 1721 |
ret = qemu_rdma_exchange_get_response(rdma, |
| 1722 |
&resp, RDMA_CONTROL_READY, RDMA_WRID_READY); |
| 1723 |
if (ret < 0) { |
| 1724 |
return ret;
|
| 1725 |
} |
| 1726 |
} |
| 1727 |
|
| 1728 |
/*
|
| 1729 |
* If the user is expecting a response, post a WR in anticipation of it.
|
| 1730 |
*/
|
| 1731 |
if (resp) {
|
| 1732 |
ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_DATA); |
| 1733 |
if (ret) {
|
| 1734 |
fprintf(stderr, "rdma migration: error posting"
|
| 1735 |
" extra control recv for anticipated result!");
|
| 1736 |
return ret;
|
| 1737 |
} |
| 1738 |
} |
| 1739 |
|
| 1740 |
/*
|
| 1741 |
* Post a WR to replace the one we just consumed for the READY message.
|
| 1742 |
*/
|
| 1743 |
ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
| 1744 |
if (ret) {
|
| 1745 |
fprintf(stderr, "rdma migration: error posting first control recv!");
|
| 1746 |
return ret;
|
| 1747 |
} |
| 1748 |
|
| 1749 |
/*
|
| 1750 |
* Deliver the control message that was requested.
|
| 1751 |
*/
|
| 1752 |
ret = qemu_rdma_post_send_control(rdma, data, head); |
| 1753 |
|
| 1754 |
if (ret < 0) { |
| 1755 |
fprintf(stderr, "Failed to send control buffer!\n");
|
| 1756 |
return ret;
|
| 1757 |
} |
| 1758 |
|
| 1759 |
/*
|
| 1760 |
* If we're expecting a response, block and wait for it.
|
| 1761 |
*/
|
| 1762 |
if (resp) {
|
| 1763 |
if (callback) {
|
| 1764 |
DDPRINTF("Issuing callback before receiving response...\n");
|
| 1765 |
ret = callback(rdma); |
| 1766 |
if (ret < 0) { |
| 1767 |
return ret;
|
| 1768 |
} |
| 1769 |
} |
| 1770 |
|
| 1771 |
DDPRINTF("Waiting for response %s\n", control_desc[resp->type]);
|
| 1772 |
ret = qemu_rdma_exchange_get_response(rdma, resp, |
| 1773 |
resp->type, RDMA_WRID_DATA); |
| 1774 |
|
| 1775 |
if (ret < 0) { |
| 1776 |
return ret;
|
| 1777 |
} |
| 1778 |
|
| 1779 |
qemu_rdma_move_header(rdma, RDMA_WRID_DATA, resp); |
| 1780 |
if (resp_idx) {
|
| 1781 |
*resp_idx = RDMA_WRID_DATA; |
| 1782 |
} |
| 1783 |
DDPRINTF("Response %s received.\n", control_desc[resp->type]);
|
| 1784 |
} |
| 1785 |
|
| 1786 |
rdma->control_ready_expected = 1;
|
| 1787 |
|
| 1788 |
return 0; |
| 1789 |
} |
| 1790 |
|
| 1791 |
/*
|
| 1792 |
* This is an 'atomic' high-level operation to receive a single, unified
|
| 1793 |
* control-channel message.
|
| 1794 |
*/
|
| 1795 |
static int qemu_rdma_exchange_recv(RDMAContext *rdma, RDMAControlHeader *head, |
| 1796 |
int expecting)
|
| 1797 |
{
|
| 1798 |
RDMAControlHeader ready = {
|
| 1799 |
.len = 0,
|
| 1800 |
.type = RDMA_CONTROL_READY, |
| 1801 |
.repeat = 1,
|
| 1802 |
}; |
| 1803 |
int ret;
|
| 1804 |
|
| 1805 |
/*
|
| 1806 |
* Inform the source that we're ready to receive a message.
|
| 1807 |
*/
|
| 1808 |
ret = qemu_rdma_post_send_control(rdma, NULL, &ready);
|
| 1809 |
|
| 1810 |
if (ret < 0) { |
| 1811 |
fprintf(stderr, "Failed to send control buffer!\n");
|
| 1812 |
return ret;
|
| 1813 |
} |
| 1814 |
|
| 1815 |
/*
|
| 1816 |
* Block and wait for the message.
|
| 1817 |
*/
|
| 1818 |
ret = qemu_rdma_exchange_get_response(rdma, head, |
| 1819 |
expecting, RDMA_WRID_READY); |
| 1820 |
|
| 1821 |
if (ret < 0) { |
| 1822 |
return ret;
|
| 1823 |
} |
| 1824 |
|
| 1825 |
qemu_rdma_move_header(rdma, RDMA_WRID_READY, head); |
| 1826 |
|
| 1827 |
/*
|
| 1828 |
* Post a new RECV work request to replace the one we just consumed.
|
| 1829 |
*/
|
| 1830 |
ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
| 1831 |
if (ret) {
|
| 1832 |
fprintf(stderr, "rdma migration: error posting second control recv!");
|
| 1833 |
return ret;
|
| 1834 |
} |
| 1835 |
|
| 1836 |
return 0; |
| 1837 |
} |
| 1838 |
|
| 1839 |
/*
|
| 1840 |
* Write an actual chunk of memory using RDMA.
|
| 1841 |
*
|
| 1842 |
* If we're using dynamic registration on the dest-side, we have to
|
| 1843 |
* send a registration command first.
|
| 1844 |
*/
|
| 1845 |
static int qemu_rdma_write_one(QEMUFile *f, RDMAContext *rdma, |
| 1846 |
int current_index, uint64_t current_addr,
|
| 1847 |
uint64_t length) |
| 1848 |
{
|
| 1849 |
struct ibv_sge sge;
|
| 1850 |
struct ibv_send_wr send_wr = { 0 }; |
| 1851 |
struct ibv_send_wr *bad_wr;
|
| 1852 |
int reg_result_idx, ret, count = 0; |
| 1853 |
uint64_t chunk, chunks; |
| 1854 |
uint8_t *chunk_start, *chunk_end; |
| 1855 |
RDMALocalBlock *block = &(rdma->local_ram_blocks.block[current_index]); |
| 1856 |
RDMARegister reg; |
| 1857 |
RDMARegisterResult *reg_result; |
| 1858 |
RDMAControlHeader resp = { .type = RDMA_CONTROL_REGISTER_RESULT };
|
| 1859 |
RDMAControlHeader head = { .len = sizeof(RDMARegister),
|
| 1860 |
.type = RDMA_CONTROL_REGISTER_REQUEST, |
| 1861 |
.repeat = 1,
|
| 1862 |
}; |
| 1863 |
|
| 1864 |
retry:
|
| 1865 |
sge.addr = (uint64_t)(block->local_host_addr + |
| 1866 |
(current_addr - block->offset)); |
| 1867 |
sge.length = length; |
| 1868 |
|
| 1869 |
chunk = ram_chunk_index(block->local_host_addr, (uint8_t *) sge.addr); |
| 1870 |
chunk_start = ram_chunk_start(block, chunk); |
| 1871 |
|
| 1872 |
if (block->is_ram_block) {
|
| 1873 |
chunks = length / (1UL << RDMA_REG_CHUNK_SHIFT);
|
| 1874 |
|
| 1875 |
if (chunks && ((length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { |
| 1876 |
chunks--; |
| 1877 |
} |
| 1878 |
} else {
|
| 1879 |
chunks = block->length / (1UL << RDMA_REG_CHUNK_SHIFT);
|
| 1880 |
|
| 1881 |
if (chunks && ((block->length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { |
| 1882 |
chunks--; |
| 1883 |
} |
| 1884 |
} |
| 1885 |
|
| 1886 |
DDPRINTF("Writing %" PRIu64 " chunks, (%" PRIu64 " MB)\n", |
| 1887 |
chunks + 1, (chunks + 1) * (1UL << RDMA_REG_CHUNK_SHIFT) / 1024 / 1024); |
| 1888 |
|
| 1889 |
chunk_end = ram_chunk_end(block, chunk + chunks); |
| 1890 |
|
| 1891 |
if (!rdma->pin_all) {
|
| 1892 |
#ifdef RDMA_UNREGISTRATION_EXAMPLE
|
| 1893 |
qemu_rdma_unregister_waiting(rdma); |
| 1894 |
#endif
|
| 1895 |
} |
| 1896 |
|
| 1897 |
while (test_bit(chunk, block->transit_bitmap)) {
|
| 1898 |
(void)count;
|
| 1899 |
DDPRINTF("(%d) Not clobbering: block: %d chunk %" PRIu64
|
| 1900 |
" current %" PRIu64 " len %" PRIu64 " %d %d\n", |
| 1901 |
count++, current_index, chunk, |
| 1902 |
sge.addr, length, rdma->nb_sent, block->nb_chunks); |
| 1903 |
|
| 1904 |
ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL);
|
| 1905 |
|
| 1906 |
if (ret < 0) { |
| 1907 |
fprintf(stderr, "Failed to Wait for previous write to complete "
|
| 1908 |
"block %d chunk %" PRIu64
|
| 1909 |
" current %" PRIu64 " len %" PRIu64 " %d\n", |
| 1910 |
current_index, chunk, sge.addr, length, rdma->nb_sent); |
| 1911 |
return ret;
|
| 1912 |
} |
| 1913 |
} |
| 1914 |
|
| 1915 |
if (!rdma->pin_all || !block->is_ram_block) {
|
| 1916 |
if (!block->remote_keys[chunk]) {
|
| 1917 |
/*
|
| 1918 |
* This chunk has not yet been registered, so first check to see
|
| 1919 |
* if the entire chunk is zero. If so, tell the other size to
|
| 1920 |
* memset() + madvise() the entire chunk without RDMA.
|
| 1921 |
*/
|
| 1922 |
|
| 1923 |
if (can_use_buffer_find_nonzero_offset((void *)sge.addr, length) |
| 1924 |
&& buffer_find_nonzero_offset((void *)sge.addr,
|
| 1925 |
length) == length) {
|
| 1926 |
RDMACompress comp = {
|
| 1927 |
.offset = current_addr, |
| 1928 |
.value = 0,
|
| 1929 |
.block_idx = current_index, |
| 1930 |
.length = length, |
| 1931 |
}; |
| 1932 |
|
| 1933 |
head.len = sizeof(comp);
|
| 1934 |
head.type = RDMA_CONTROL_COMPRESS; |
| 1935 |
|
| 1936 |
DDPRINTF("Entire chunk is zero, sending compress: %"
|
| 1937 |
PRIu64 " for %d "
|
| 1938 |
"bytes, index: %d, offset: %" PRId64 "...\n", |
| 1939 |
chunk, sge.length, current_index, current_addr); |
| 1940 |
|
| 1941 |
compress_to_network(&comp); |
| 1942 |
ret = qemu_rdma_exchange_send(rdma, &head, |
| 1943 |
(uint8_t *) &comp, NULL, NULL, NULL); |
| 1944 |
|
| 1945 |
if (ret < 0) { |
| 1946 |
return -EIO;
|
| 1947 |
} |
| 1948 |
|
| 1949 |
acct_update_position(f, sge.length, true);
|
| 1950 |
|
| 1951 |
return 1; |
| 1952 |
} |
| 1953 |
|
| 1954 |
/*
|
| 1955 |
* Otherwise, tell other side to register.
|
| 1956 |
*/
|
| 1957 |
reg.current_index = current_index; |
| 1958 |
if (block->is_ram_block) {
|
| 1959 |
reg.key.current_addr = current_addr; |
| 1960 |
} else {
|
| 1961 |
reg.key.chunk = chunk; |
| 1962 |
} |
| 1963 |
reg.chunks = chunks; |
| 1964 |
|
| 1965 |
DDPRINTF("Sending registration request chunk %" PRIu64 " for %d " |
| 1966 |
"bytes, index: %d, offset: %" PRId64 "...\n", |
| 1967 |
chunk, sge.length, current_index, current_addr); |
| 1968 |
|
| 1969 |
register_to_network(®); |
| 1970 |
ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
| 1971 |
&resp, ®_result_idx, NULL);
|
| 1972 |
if (ret < 0) { |
| 1973 |
return ret;
|
| 1974 |
} |
| 1975 |
|
| 1976 |
/* try to overlap this single registration with the one we sent. */
|
| 1977 |
if (qemu_rdma_register_and_get_keys(rdma, block,
|
| 1978 |
(uint8_t *) sge.addr, |
| 1979 |
&sge.lkey, NULL, chunk,
|
| 1980 |
chunk_start, chunk_end)) {
|
| 1981 |
fprintf(stderr, "cannot get lkey!\n");
|
| 1982 |
return -EINVAL;
|
| 1983 |
} |
| 1984 |
|
| 1985 |
reg_result = (RDMARegisterResult *) |
| 1986 |
rdma->wr_data[reg_result_idx].control_curr; |
| 1987 |
|
| 1988 |
network_to_result(reg_result); |
| 1989 |
|
| 1990 |
DDPRINTF("Received registration result:"
|
| 1991 |
" my key: %x their key %x, chunk %" PRIu64 "\n", |
| 1992 |
block->remote_keys[chunk], reg_result->rkey, chunk); |
| 1993 |
|
| 1994 |
block->remote_keys[chunk] = reg_result->rkey; |
| 1995 |
block->remote_host_addr = reg_result->host_addr; |
| 1996 |
} else {
|
| 1997 |
/* already registered before */
|
| 1998 |
if (qemu_rdma_register_and_get_keys(rdma, block,
|
| 1999 |
(uint8_t *)sge.addr, |
| 2000 |
&sge.lkey, NULL, chunk,
|
| 2001 |
chunk_start, chunk_end)) {
|
| 2002 |
fprintf(stderr, "cannot get lkey!\n");
|
| 2003 |
return -EINVAL;
|
| 2004 |
} |
| 2005 |
} |
| 2006 |
|
| 2007 |
send_wr.wr.rdma.rkey = block->remote_keys[chunk]; |
| 2008 |
} else {
|
| 2009 |
send_wr.wr.rdma.rkey = block->remote_rkey; |
| 2010 |
|
| 2011 |
if (qemu_rdma_register_and_get_keys(rdma, block, (uint8_t *)sge.addr,
|
| 2012 |
&sge.lkey, NULL, chunk,
|
| 2013 |
chunk_start, chunk_end)) {
|
| 2014 |
fprintf(stderr, "cannot get lkey!\n");
|
| 2015 |
return -EINVAL;
|
| 2016 |
} |
| 2017 |
} |
| 2018 |
|
| 2019 |
/*
|
| 2020 |
* Encode the ram block index and chunk within this wrid.
|
| 2021 |
* We will use this information at the time of completion
|
| 2022 |
* to figure out which bitmap to check against and then which
|
| 2023 |
* chunk in the bitmap to look for.
|
| 2024 |
*/
|
| 2025 |
send_wr.wr_id = qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE, |
| 2026 |
current_index, chunk); |
| 2027 |
|
| 2028 |
send_wr.opcode = IBV_WR_RDMA_WRITE; |
| 2029 |
send_wr.send_flags = IBV_SEND_SIGNALED; |
| 2030 |
send_wr.sg_list = &sge; |
| 2031 |
send_wr.num_sge = 1;
|
| 2032 |
send_wr.wr.rdma.remote_addr = block->remote_host_addr + |
| 2033 |
(current_addr - block->offset); |
| 2034 |
|
| 2035 |
DDDPRINTF("Posting chunk: %" PRIu64 ", addr: %lx" |
| 2036 |
" remote: %lx, bytes %" PRIu32 "\n", |
| 2037 |
chunk, sge.addr, send_wr.wr.rdma.remote_addr, |
| 2038 |
sge.length); |
| 2039 |
|
| 2040 |
/*
|
| 2041 |
* ibv_post_send() does not return negative error numbers,
|
| 2042 |
* per the specification they are positive - no idea why.
|
| 2043 |
*/
|
| 2044 |
ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr); |
| 2045 |
|
| 2046 |
if (ret == ENOMEM) {
|
| 2047 |
DDPRINTF("send queue is full. wait a little....\n");
|
| 2048 |
ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL);
|
| 2049 |
if (ret < 0) { |
| 2050 |
fprintf(stderr, "rdma migration: failed to make "
|
| 2051 |
"room in full send queue! %d\n", ret);
|
| 2052 |
return ret;
|
| 2053 |
} |
| 2054 |
|
| 2055 |
goto retry;
|
| 2056 |
|
| 2057 |
} else if (ret > 0) { |
| 2058 |
perror("rdma migration: post rdma write failed");
|
| 2059 |
return -ret;
|
| 2060 |
} |
| 2061 |
|
| 2062 |
set_bit(chunk, block->transit_bitmap); |
| 2063 |
acct_update_position(f, sge.length, false);
|
| 2064 |
rdma->total_writes++; |
| 2065 |
|
| 2066 |
return 0; |
| 2067 |
} |
| 2068 |
|
| 2069 |
/*
|
| 2070 |
* Push out any unwritten RDMA operations.
|
| 2071 |
*
|
| 2072 |
* We support sending out multiple chunks at the same time.
|
| 2073 |
* Not all of them need to get signaled in the completion queue.
|
| 2074 |
*/
|
| 2075 |
static int qemu_rdma_write_flush(QEMUFile *f, RDMAContext *rdma) |
| 2076 |
{
|
| 2077 |
int ret;
|
| 2078 |
|
| 2079 |
if (!rdma->current_length) {
|
| 2080 |
return 0; |
| 2081 |
} |
| 2082 |
|
| 2083 |
ret = qemu_rdma_write_one(f, rdma, |
| 2084 |
rdma->current_index, rdma->current_addr, rdma->current_length); |
| 2085 |
|
| 2086 |
if (ret < 0) { |
| 2087 |
return ret;
|
| 2088 |
} |
| 2089 |
|
| 2090 |
if (ret == 0) { |
| 2091 |
rdma->nb_sent++; |
| 2092 |
DDDPRINTF("sent total: %d\n", rdma->nb_sent);
|
| 2093 |
} |
| 2094 |
|
| 2095 |
rdma->current_length = 0;
|
| 2096 |
rdma->current_addr = 0;
|
| 2097 |
|
| 2098 |
return 0; |
| 2099 |
} |
| 2100 |
|
| 2101 |
static inline int qemu_rdma_buffer_mergable(RDMAContext *rdma, |
| 2102 |
uint64_t offset, uint64_t len) |
| 2103 |
{
|
| 2104 |
RDMALocalBlock *block; |
| 2105 |
uint8_t *host_addr; |
| 2106 |
uint8_t *chunk_end; |
| 2107 |
|
| 2108 |
if (rdma->current_index < 0) { |
| 2109 |
return 0; |
| 2110 |
} |
| 2111 |
|
| 2112 |
if (rdma->current_chunk < 0) { |
| 2113 |
return 0; |
| 2114 |
} |
| 2115 |
|
| 2116 |
block = &(rdma->local_ram_blocks.block[rdma->current_index]); |
| 2117 |
host_addr = block->local_host_addr + (offset - block->offset); |
| 2118 |
chunk_end = ram_chunk_end(block, rdma->current_chunk); |
| 2119 |
|
| 2120 |
if (rdma->current_length == 0) { |
| 2121 |
return 0; |
| 2122 |
} |
| 2123 |
|
| 2124 |
/*
|
| 2125 |
* Only merge into chunk sequentially.
|
| 2126 |
*/
|
| 2127 |
if (offset != (rdma->current_addr + rdma->current_length)) {
|
| 2128 |
return 0; |
| 2129 |
} |
| 2130 |
|
| 2131 |
if (offset < block->offset) {
|
| 2132 |
return 0; |
| 2133 |
} |
| 2134 |
|
| 2135 |
if ((offset + len) > (block->offset + block->length)) {
|
| 2136 |
return 0; |
| 2137 |
} |
| 2138 |
|
| 2139 |
if ((host_addr + len) > chunk_end) {
|
| 2140 |
return 0; |
| 2141 |
} |
| 2142 |
|
| 2143 |
return 1; |
| 2144 |
} |
| 2145 |
|
| 2146 |
/*
|
| 2147 |
* We're not actually writing here, but doing three things:
|
| 2148 |
*
|
| 2149 |
* 1. Identify the chunk the buffer belongs to.
|
| 2150 |
* 2. If the chunk is full or the buffer doesn't belong to the current
|
| 2151 |
* chunk, then start a new chunk and flush() the old chunk.
|
| 2152 |
* 3. To keep the hardware busy, we also group chunks into batches
|
| 2153 |
* and only require that a batch gets acknowledged in the completion
|
| 2154 |
* qeueue instead of each individual chunk.
|
| 2155 |
*/
|
| 2156 |
static int qemu_rdma_write(QEMUFile *f, RDMAContext *rdma, |
| 2157 |
uint64_t block_offset, uint64_t offset, |
| 2158 |
uint64_t len) |
| 2159 |
{
|
| 2160 |
uint64_t current_addr = block_offset + offset; |
| 2161 |
uint64_t index = rdma->current_index; |
| 2162 |
uint64_t chunk = rdma->current_chunk; |
| 2163 |
int ret;
|
| 2164 |
|
| 2165 |
/* If we cannot merge it, we flush the current buffer first. */
|
| 2166 |
if (!qemu_rdma_buffer_mergable(rdma, current_addr, len)) {
|
| 2167 |
ret = qemu_rdma_write_flush(f, rdma); |
| 2168 |
if (ret) {
|
| 2169 |
return ret;
|
| 2170 |
} |
| 2171 |
rdma->current_length = 0;
|
| 2172 |
rdma->current_addr = current_addr; |
| 2173 |
|
| 2174 |
ret = qemu_rdma_search_ram_block(rdma, block_offset, |
| 2175 |
offset, len, &index, &chunk); |
| 2176 |
if (ret) {
|
| 2177 |
fprintf(stderr, "ram block search failed\n");
|
| 2178 |
return ret;
|
| 2179 |
} |
| 2180 |
rdma->current_index = index; |
| 2181 |
rdma->current_chunk = chunk; |
| 2182 |
} |
| 2183 |
|
| 2184 |
/* merge it */
|
| 2185 |
rdma->current_length += len; |
| 2186 |
|
| 2187 |
/* flush it if buffer is too large */
|
| 2188 |
if (rdma->current_length >= RDMA_MERGE_MAX) {
|
| 2189 |
return qemu_rdma_write_flush(f, rdma);
|
| 2190 |
} |
| 2191 |
|
| 2192 |
return 0; |
| 2193 |
} |
| 2194 |
|
| 2195 |
static void qemu_rdma_cleanup(RDMAContext *rdma) |
| 2196 |
{
|
| 2197 |
struct rdma_cm_event *cm_event;
|
| 2198 |
int ret, idx;
|
| 2199 |
|
| 2200 |
if (rdma->cm_id && rdma->connected) {
|
| 2201 |
if (rdma->error_state) {
|
| 2202 |
RDMAControlHeader head = { .len = 0,
|
| 2203 |
.type = RDMA_CONTROL_ERROR, |
| 2204 |
.repeat = 1,
|
| 2205 |
}; |
| 2206 |
fprintf(stderr, "Early error. Sending error.\n");
|
| 2207 |
qemu_rdma_post_send_control(rdma, NULL, &head);
|
| 2208 |
} |
| 2209 |
|
| 2210 |
ret = rdma_disconnect(rdma->cm_id); |
| 2211 |
if (!ret) {
|
| 2212 |
DDPRINTF("waiting for disconnect\n");
|
| 2213 |
ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2214 |
if (!ret) {
|
| 2215 |
rdma_ack_cm_event(cm_event); |
| 2216 |
} |
| 2217 |
} |
| 2218 |
DDPRINTF("Disconnected.\n");
|
| 2219 |
rdma->connected = false;
|
| 2220 |
} |
| 2221 |
|
| 2222 |
g_free(rdma->block); |
| 2223 |
rdma->block = NULL;
|
| 2224 |
|
| 2225 |
for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
| 2226 |
if (rdma->wr_data[idx].control_mr) {
|
| 2227 |
rdma->total_registrations--; |
| 2228 |
ibv_dereg_mr(rdma->wr_data[idx].control_mr); |
| 2229 |
} |
| 2230 |
rdma->wr_data[idx].control_mr = NULL;
|
| 2231 |
} |
| 2232 |
|
| 2233 |
if (rdma->local_ram_blocks.block) {
|
| 2234 |
while (rdma->local_ram_blocks.nb_blocks) {
|
| 2235 |
__qemu_rdma_delete_block(rdma, |
| 2236 |
rdma->local_ram_blocks.block->offset); |
| 2237 |
} |
| 2238 |
} |
| 2239 |
|
| 2240 |
if (rdma->qp) {
|
| 2241 |
rdma_destroy_qp(rdma->cm_id); |
| 2242 |
rdma->qp = NULL;
|
| 2243 |
} |
| 2244 |
if (rdma->cq) {
|
| 2245 |
ibv_destroy_cq(rdma->cq); |
| 2246 |
rdma->cq = NULL;
|
| 2247 |
} |
| 2248 |
if (rdma->comp_channel) {
|
| 2249 |
ibv_destroy_comp_channel(rdma->comp_channel); |
| 2250 |
rdma->comp_channel = NULL;
|
| 2251 |
} |
| 2252 |
if (rdma->pd) {
|
| 2253 |
ibv_dealloc_pd(rdma->pd); |
| 2254 |
rdma->pd = NULL;
|
| 2255 |
} |
| 2256 |
if (rdma->listen_id) {
|
| 2257 |
rdma_destroy_id(rdma->listen_id); |
| 2258 |
rdma->listen_id = NULL;
|
| 2259 |
} |
| 2260 |
if (rdma->cm_id) {
|
| 2261 |
rdma_destroy_id(rdma->cm_id); |
| 2262 |
rdma->cm_id = NULL;
|
| 2263 |
} |
| 2264 |
if (rdma->channel) {
|
| 2265 |
rdma_destroy_event_channel(rdma->channel); |
| 2266 |
rdma->channel = NULL;
|
| 2267 |
} |
| 2268 |
g_free(rdma->host); |
| 2269 |
rdma->host = NULL;
|
| 2270 |
} |
| 2271 |
|
| 2272 |
|
| 2273 |
static int qemu_rdma_source_init(RDMAContext *rdma, Error **errp, bool pin_all) |
| 2274 |
{
|
| 2275 |
int ret, idx;
|
| 2276 |
Error *local_err = NULL, **temp = &local_err;
|
| 2277 |
|
| 2278 |
/*
|
| 2279 |
* Will be validated against destination's actual capabilities
|
| 2280 |
* after the connect() completes.
|
| 2281 |
*/
|
| 2282 |
rdma->pin_all = pin_all; |
| 2283 |
|
| 2284 |
ret = qemu_rdma_resolve_host(rdma, temp); |
| 2285 |
if (ret) {
|
| 2286 |
goto err_rdma_source_init;
|
| 2287 |
} |
| 2288 |
|
| 2289 |
ret = qemu_rdma_alloc_pd_cq(rdma); |
| 2290 |
if (ret) {
|
| 2291 |
ERROR(temp, "rdma migration: error allocating pd and cq! Your mlock()"
|
| 2292 |
" limits may be too low. Please check $ ulimit -a # and "
|
| 2293 |
"search for 'ulimit -l' in the output");
|
| 2294 |
goto err_rdma_source_init;
|
| 2295 |
} |
| 2296 |
|
| 2297 |
ret = qemu_rdma_alloc_qp(rdma); |
| 2298 |
if (ret) {
|
| 2299 |
ERROR(temp, "rdma migration: error allocating qp!");
|
| 2300 |
goto err_rdma_source_init;
|
| 2301 |
} |
| 2302 |
|
| 2303 |
ret = qemu_rdma_init_ram_blocks(rdma); |
| 2304 |
if (ret) {
|
| 2305 |
ERROR(temp, "rdma migration: error initializing ram blocks!");
|
| 2306 |
goto err_rdma_source_init;
|
| 2307 |
} |
| 2308 |
|
| 2309 |
for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
| 2310 |
ret = qemu_rdma_reg_control(rdma, idx); |
| 2311 |
if (ret) {
|
| 2312 |
ERROR(temp, "rdma migration: error registering %d control!",
|
| 2313 |
idx); |
| 2314 |
goto err_rdma_source_init;
|
| 2315 |
} |
| 2316 |
} |
| 2317 |
|
| 2318 |
return 0; |
| 2319 |
|
| 2320 |
err_rdma_source_init:
|
| 2321 |
error_propagate(errp, local_err); |
| 2322 |
qemu_rdma_cleanup(rdma); |
| 2323 |
return -1; |
| 2324 |
} |
| 2325 |
|
| 2326 |
static int qemu_rdma_connect(RDMAContext *rdma, Error **errp) |
| 2327 |
{
|
| 2328 |
RDMACapabilities cap = {
|
| 2329 |
.version = RDMA_CONTROL_VERSION_CURRENT, |
| 2330 |
.flags = 0,
|
| 2331 |
}; |
| 2332 |
struct rdma_conn_param conn_param = { .initiator_depth = 2, |
| 2333 |
.retry_count = 5,
|
| 2334 |
.private_data = &cap, |
| 2335 |
.private_data_len = sizeof(cap),
|
| 2336 |
}; |
| 2337 |
struct rdma_cm_event *cm_event;
|
| 2338 |
int ret;
|
| 2339 |
|
| 2340 |
/*
|
| 2341 |
* Only negotiate the capability with destination if the user
|
| 2342 |
* on the source first requested the capability.
|
| 2343 |
*/
|
| 2344 |
if (rdma->pin_all) {
|
| 2345 |
DPRINTF("Server pin-all memory requested.\n");
|
| 2346 |
cap.flags |= RDMA_CAPABILITY_PIN_ALL; |
| 2347 |
} |
| 2348 |
|
| 2349 |
caps_to_network(&cap); |
| 2350 |
|
| 2351 |
ret = rdma_connect(rdma->cm_id, &conn_param); |
| 2352 |
if (ret) {
|
| 2353 |
perror("rdma_connect");
|
| 2354 |
ERROR(errp, "connecting to destination!");
|
| 2355 |
rdma_destroy_id(rdma->cm_id); |
| 2356 |
rdma->cm_id = NULL;
|
| 2357 |
goto err_rdma_source_connect;
|
| 2358 |
} |
| 2359 |
|
| 2360 |
ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2361 |
if (ret) {
|
| 2362 |
perror("rdma_get_cm_event after rdma_connect");
|
| 2363 |
ERROR(errp, "connecting to destination!");
|
| 2364 |
rdma_ack_cm_event(cm_event); |
| 2365 |
rdma_destroy_id(rdma->cm_id); |
| 2366 |
rdma->cm_id = NULL;
|
| 2367 |
goto err_rdma_source_connect;
|
| 2368 |
} |
| 2369 |
|
| 2370 |
if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) {
|
| 2371 |
perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
|
| 2372 |
ERROR(errp, "connecting to destination!");
|
| 2373 |
rdma_ack_cm_event(cm_event); |
| 2374 |
rdma_destroy_id(rdma->cm_id); |
| 2375 |
rdma->cm_id = NULL;
|
| 2376 |
goto err_rdma_source_connect;
|
| 2377 |
} |
| 2378 |
rdma->connected = true;
|
| 2379 |
|
| 2380 |
memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap));
|
| 2381 |
network_to_caps(&cap); |
| 2382 |
|
| 2383 |
/*
|
| 2384 |
* Verify that the *requested* capabilities are supported by the destination
|
| 2385 |
* and disable them otherwise.
|
| 2386 |
*/
|
| 2387 |
if (rdma->pin_all && !(cap.flags & RDMA_CAPABILITY_PIN_ALL)) {
|
| 2388 |
ERROR(errp, "Server cannot support pinning all memory. "
|
| 2389 |
"Will register memory dynamically.");
|
| 2390 |
rdma->pin_all = false;
|
| 2391 |
} |
| 2392 |
|
| 2393 |
DPRINTF("Pin all memory: %s\n", rdma->pin_all ? "enabled" : "disabled"); |
| 2394 |
|
| 2395 |
rdma_ack_cm_event(cm_event); |
| 2396 |
|
| 2397 |
ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
| 2398 |
if (ret) {
|
| 2399 |
ERROR(errp, "posting second control recv!");
|
| 2400 |
goto err_rdma_source_connect;
|
| 2401 |
} |
| 2402 |
|
| 2403 |
rdma->control_ready_expected = 1;
|
| 2404 |
rdma->nb_sent = 0;
|
| 2405 |
return 0; |
| 2406 |
|
| 2407 |
err_rdma_source_connect:
|
| 2408 |
qemu_rdma_cleanup(rdma); |
| 2409 |
return -1; |
| 2410 |
} |
| 2411 |
|
| 2412 |
static int qemu_rdma_dest_init(RDMAContext *rdma, Error **errp) |
| 2413 |
{
|
| 2414 |
int ret = -EINVAL, idx;
|
| 2415 |
struct rdma_cm_id *listen_id;
|
| 2416 |
char ip[40] = "unknown"; |
| 2417 |
struct rdma_addrinfo *res;
|
| 2418 |
char port_str[16]; |
| 2419 |
|
| 2420 |
for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
| 2421 |
rdma->wr_data[idx].control_len = 0;
|
| 2422 |
rdma->wr_data[idx].control_curr = NULL;
|
| 2423 |
} |
| 2424 |
|
| 2425 |
if (rdma->host == NULL) { |
| 2426 |
ERROR(errp, "RDMA host is not set!");
|
| 2427 |
rdma->error_state = -EINVAL; |
| 2428 |
return -1; |
| 2429 |
} |
| 2430 |
/* create CM channel */
|
| 2431 |
rdma->channel = rdma_create_event_channel(); |
| 2432 |
if (!rdma->channel) {
|
| 2433 |
ERROR(errp, "could not create rdma event channel");
|
| 2434 |
rdma->error_state = -EINVAL; |
| 2435 |
return -1; |
| 2436 |
} |
| 2437 |
|
| 2438 |
/* create CM id */
|
| 2439 |
ret = rdma_create_id(rdma->channel, &listen_id, NULL, RDMA_PS_TCP);
|
| 2440 |
if (ret) {
|
| 2441 |
ERROR(errp, "could not create cm_id!");
|
| 2442 |
goto err_dest_init_create_listen_id;
|
| 2443 |
} |
| 2444 |
|
| 2445 |
snprintf(port_str, 16, "%d", rdma->port); |
| 2446 |
port_str[15] = '\0'; |
| 2447 |
|
| 2448 |
if (rdma->host && strcmp("", rdma->host)) { |
| 2449 |
struct rdma_addrinfo *e;
|
| 2450 |
|
| 2451 |
ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res);
|
| 2452 |
if (ret < 0) { |
| 2453 |
ERROR(errp, "could not rdma_getaddrinfo address %s", rdma->host);
|
| 2454 |
goto err_dest_init_bind_addr;
|
| 2455 |
} |
| 2456 |
|
| 2457 |
for (e = res; e != NULL; e = e->ai_next) { |
| 2458 |
inet_ntop(e->ai_family, |
| 2459 |
&((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip); |
| 2460 |
DPRINTF("Trying %s => %s\n", rdma->host, ip);
|
| 2461 |
ret = rdma_bind_addr(listen_id, e->ai_dst_addr); |
| 2462 |
if (!ret) {
|
| 2463 |
if (e->ai_family == AF_INET6) {
|
| 2464 |
ret = qemu_rdma_broken_ipv6_kernel(errp, listen_id->verbs); |
| 2465 |
if (ret) {
|
| 2466 |
continue;
|
| 2467 |
} |
| 2468 |
} |
| 2469 |
|
| 2470 |
goto listen;
|
| 2471 |
} |
| 2472 |
} |
| 2473 |
|
| 2474 |
ERROR(errp, "Error: could not rdma_bind_addr!");
|
| 2475 |
goto err_dest_init_bind_addr;
|
| 2476 |
} else {
|
| 2477 |
ERROR(errp, "migration host and port not specified!");
|
| 2478 |
ret = -EINVAL; |
| 2479 |
goto err_dest_init_bind_addr;
|
| 2480 |
} |
| 2481 |
listen:
|
| 2482 |
|
| 2483 |
rdma->listen_id = listen_id; |
| 2484 |
qemu_rdma_dump_gid("dest_init", listen_id);
|
| 2485 |
return 0; |
| 2486 |
|
| 2487 |
err_dest_init_bind_addr:
|
| 2488 |
rdma_destroy_id(listen_id); |
| 2489 |
err_dest_init_create_listen_id:
|
| 2490 |
rdma_destroy_event_channel(rdma->channel); |
| 2491 |
rdma->channel = NULL;
|
| 2492 |
rdma->error_state = ret; |
| 2493 |
return ret;
|
| 2494 |
|
| 2495 |
} |
| 2496 |
|
| 2497 |
static void *qemu_rdma_data_init(const char *host_port, Error **errp) |
| 2498 |
{
|
| 2499 |
RDMAContext *rdma = NULL;
|
| 2500 |
InetSocketAddress *addr; |
| 2501 |
|
| 2502 |
if (host_port) {
|
| 2503 |
rdma = g_malloc0(sizeof(RDMAContext));
|
| 2504 |
memset(rdma, 0, sizeof(RDMAContext)); |
| 2505 |
rdma->current_index = -1;
|
| 2506 |
rdma->current_chunk = -1;
|
| 2507 |
|
| 2508 |
addr = inet_parse(host_port, NULL);
|
| 2509 |
if (addr != NULL) { |
| 2510 |
rdma->port = atoi(addr->port); |
| 2511 |
rdma->host = g_strdup(addr->host); |
| 2512 |
} else {
|
| 2513 |
ERROR(errp, "bad RDMA migration address '%s'", host_port);
|
| 2514 |
g_free(rdma); |
| 2515 |
return NULL; |
| 2516 |
} |
| 2517 |
} |
| 2518 |
|
| 2519 |
return rdma;
|
| 2520 |
} |
| 2521 |
|
| 2522 |
/*
|
| 2523 |
* QEMUFile interface to the control channel.
|
| 2524 |
* SEND messages for control only.
|
| 2525 |
* pc.ram is handled with regular RDMA messages.
|
| 2526 |
*/
|
| 2527 |
static int qemu_rdma_put_buffer(void *opaque, const uint8_t *buf, |
| 2528 |
int64_t pos, int size)
|
| 2529 |
{
|
| 2530 |
QEMUFileRDMA *r = opaque; |
| 2531 |
QEMUFile *f = r->file; |
| 2532 |
RDMAContext *rdma = r->rdma; |
| 2533 |
size_t remaining = size; |
| 2534 |
uint8_t * data = (void *) buf;
|
| 2535 |
int ret;
|
| 2536 |
|
| 2537 |
CHECK_ERROR_STATE(); |
| 2538 |
|
| 2539 |
/*
|
| 2540 |
* Push out any writes that
|
| 2541 |
* we're queued up for pc.ram.
|
| 2542 |
*/
|
| 2543 |
ret = qemu_rdma_write_flush(f, rdma); |
| 2544 |
if (ret < 0) { |
| 2545 |
rdma->error_state = ret; |
| 2546 |
return ret;
|
| 2547 |
} |
| 2548 |
|
| 2549 |
while (remaining) {
|
| 2550 |
RDMAControlHeader head; |
| 2551 |
|
| 2552 |
r->len = MIN(remaining, RDMA_SEND_INCREMENT); |
| 2553 |
remaining -= r->len; |
| 2554 |
|
| 2555 |
head.len = r->len; |
| 2556 |
head.type = RDMA_CONTROL_QEMU_FILE; |
| 2557 |
|
| 2558 |
ret = qemu_rdma_exchange_send(rdma, &head, data, NULL, NULL, NULL); |
| 2559 |
|
| 2560 |
if (ret < 0) { |
| 2561 |
rdma->error_state = ret; |
| 2562 |
return ret;
|
| 2563 |
} |
| 2564 |
|
| 2565 |
data += r->len; |
| 2566 |
} |
| 2567 |
|
| 2568 |
return size;
|
| 2569 |
} |
| 2570 |
|
| 2571 |
static size_t qemu_rdma_fill(RDMAContext *rdma, uint8_t *buf,
|
| 2572 |
int size, int idx) |
| 2573 |
{
|
| 2574 |
size_t len = 0;
|
| 2575 |
|
| 2576 |
if (rdma->wr_data[idx].control_len) {
|
| 2577 |
DDDPRINTF("RDMA %" PRId64 " of %d bytes already in buffer\n", |
| 2578 |
rdma->wr_data[idx].control_len, size); |
| 2579 |
|
| 2580 |
len = MIN(size, rdma->wr_data[idx].control_len); |
| 2581 |
memcpy(buf, rdma->wr_data[idx].control_curr, len); |
| 2582 |
rdma->wr_data[idx].control_curr += len; |
| 2583 |
rdma->wr_data[idx].control_len -= len; |
| 2584 |
} |
| 2585 |
|
| 2586 |
return len;
|
| 2587 |
} |
| 2588 |
|
| 2589 |
/*
|
| 2590 |
* QEMUFile interface to the control channel.
|
| 2591 |
* RDMA links don't use bytestreams, so we have to
|
| 2592 |
* return bytes to QEMUFile opportunistically.
|
| 2593 |
*/
|
| 2594 |
static int qemu_rdma_get_buffer(void *opaque, uint8_t *buf, |
| 2595 |
int64_t pos, int size)
|
| 2596 |
{
|
| 2597 |
QEMUFileRDMA *r = opaque; |
| 2598 |
RDMAContext *rdma = r->rdma; |
| 2599 |
RDMAControlHeader head; |
| 2600 |
int ret = 0; |
| 2601 |
|
| 2602 |
CHECK_ERROR_STATE(); |
| 2603 |
|
| 2604 |
/*
|
| 2605 |
* First, we hold on to the last SEND message we
|
| 2606 |
* were given and dish out the bytes until we run
|
| 2607 |
* out of bytes.
|
| 2608 |
*/
|
| 2609 |
r->len = qemu_rdma_fill(r->rdma, buf, size, 0);
|
| 2610 |
if (r->len) {
|
| 2611 |
return r->len;
|
| 2612 |
} |
| 2613 |
|
| 2614 |
/*
|
| 2615 |
* Once we run out, we block and wait for another
|
| 2616 |
* SEND message to arrive.
|
| 2617 |
*/
|
| 2618 |
ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_QEMU_FILE); |
| 2619 |
|
| 2620 |
if (ret < 0) { |
| 2621 |
rdma->error_state = ret; |
| 2622 |
return ret;
|
| 2623 |
} |
| 2624 |
|
| 2625 |
/*
|
| 2626 |
* SEND was received with new bytes, now try again.
|
| 2627 |
*/
|
| 2628 |
return qemu_rdma_fill(r->rdma, buf, size, 0); |
| 2629 |
} |
| 2630 |
|
| 2631 |
/*
|
| 2632 |
* Block until all the outstanding chunks have been delivered by the hardware.
|
| 2633 |
*/
|
| 2634 |
static int qemu_rdma_drain_cq(QEMUFile *f, RDMAContext *rdma) |
| 2635 |
{
|
| 2636 |
int ret;
|
| 2637 |
|
| 2638 |
if (qemu_rdma_write_flush(f, rdma) < 0) { |
| 2639 |
return -EIO;
|
| 2640 |
} |
| 2641 |
|
| 2642 |
while (rdma->nb_sent) {
|
| 2643 |
ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL);
|
| 2644 |
if (ret < 0) { |
| 2645 |
fprintf(stderr, "rdma migration: complete polling error!\n");
|
| 2646 |
return -EIO;
|
| 2647 |
} |
| 2648 |
} |
| 2649 |
|
| 2650 |
qemu_rdma_unregister_waiting(rdma); |
| 2651 |
|
| 2652 |
return 0; |
| 2653 |
} |
| 2654 |
|
| 2655 |
static int qemu_rdma_close(void *opaque) |
| 2656 |
{
|
| 2657 |
DPRINTF("Shutting down connection.\n");
|
| 2658 |
QEMUFileRDMA *r = opaque; |
| 2659 |
if (r->rdma) {
|
| 2660 |
qemu_rdma_cleanup(r->rdma); |
| 2661 |
g_free(r->rdma); |
| 2662 |
} |
| 2663 |
g_free(r); |
| 2664 |
return 0; |
| 2665 |
} |
| 2666 |
|
| 2667 |
/*
|
| 2668 |
* Parameters:
|
| 2669 |
* @offset == 0 :
|
| 2670 |
* This means that 'block_offset' is a full virtual address that does not
|
| 2671 |
* belong to a RAMBlock of the virtual machine and instead
|
| 2672 |
* represents a private malloc'd memory area that the caller wishes to
|
| 2673 |
* transfer.
|
| 2674 |
*
|
| 2675 |
* @offset != 0 :
|
| 2676 |
* Offset is an offset to be added to block_offset and used
|
| 2677 |
* to also lookup the corresponding RAMBlock.
|
| 2678 |
*
|
| 2679 |
* @size > 0 :
|
| 2680 |
* Initiate an transfer this size.
|
| 2681 |
*
|
| 2682 |
* @size == 0 :
|
| 2683 |
* A 'hint' or 'advice' that means that we wish to speculatively
|
| 2684 |
* and asynchronously unregister this memory. In this case, there is no
|
| 2685 |
* guarantee that the unregister will actually happen, for example,
|
| 2686 |
* if the memory is being actively transmitted. Additionally, the memory
|
| 2687 |
* may be re-registered at any future time if a write within the same
|
| 2688 |
* chunk was requested again, even if you attempted to unregister it
|
| 2689 |
* here.
|
| 2690 |
*
|
| 2691 |
* @size < 0 : TODO, not yet supported
|
| 2692 |
* Unregister the memory NOW. This means that the caller does not
|
| 2693 |
* expect there to be any future RDMA transfers and we just want to clean
|
| 2694 |
* things up. This is used in case the upper layer owns the memory and
|
| 2695 |
* cannot wait for qemu_fclose() to occur.
|
| 2696 |
*
|
| 2697 |
* @bytes_sent : User-specificed pointer to indicate how many bytes were
|
| 2698 |
* sent. Usually, this will not be more than a few bytes of
|
| 2699 |
* the protocol because most transfers are sent asynchronously.
|
| 2700 |
*/
|
| 2701 |
static size_t qemu_rdma_save_page(QEMUFile *f, void *opaque, |
| 2702 |
ram_addr_t block_offset, ram_addr_t offset, |
| 2703 |
size_t size, int *bytes_sent)
|
| 2704 |
{
|
| 2705 |
QEMUFileRDMA *rfile = opaque; |
| 2706 |
RDMAContext *rdma = rfile->rdma; |
| 2707 |
int ret;
|
| 2708 |
|
| 2709 |
CHECK_ERROR_STATE(); |
| 2710 |
|
| 2711 |
qemu_fflush(f); |
| 2712 |
|
| 2713 |
if (size > 0) { |
| 2714 |
/*
|
| 2715 |
* Add this page to the current 'chunk'. If the chunk
|
| 2716 |
* is full, or the page doen't belong to the current chunk,
|
| 2717 |
* an actual RDMA write will occur and a new chunk will be formed.
|
| 2718 |
*/
|
| 2719 |
ret = qemu_rdma_write(f, rdma, block_offset, offset, size); |
| 2720 |
if (ret < 0) { |
| 2721 |
fprintf(stderr, "rdma migration: write error! %d\n", ret);
|
| 2722 |
goto err;
|
| 2723 |
} |
| 2724 |
|
| 2725 |
/*
|
| 2726 |
* We always return 1 bytes because the RDMA
|
| 2727 |
* protocol is completely asynchronous. We do not yet know
|
| 2728 |
* whether an identified chunk is zero or not because we're
|
| 2729 |
* waiting for other pages to potentially be merged with
|
| 2730 |
* the current chunk. So, we have to call qemu_update_position()
|
| 2731 |
* later on when the actual write occurs.
|
| 2732 |
*/
|
| 2733 |
if (bytes_sent) {
|
| 2734 |
*bytes_sent = 1;
|
| 2735 |
} |
| 2736 |
} else {
|
| 2737 |
uint64_t index, chunk; |
| 2738 |
|
| 2739 |
/* TODO: Change QEMUFileOps prototype to be signed: size_t => long
|
| 2740 |
if (size < 0) {
|
| 2741 |
ret = qemu_rdma_drain_cq(f, rdma);
|
| 2742 |
if (ret < 0) {
|
| 2743 |
fprintf(stderr, "rdma: failed to synchronously drain"
|
| 2744 |
" completion queue before unregistration.\n");
|
| 2745 |
goto err;
|
| 2746 |
}
|
| 2747 |
}
|
| 2748 |
*/
|
| 2749 |
|
| 2750 |
ret = qemu_rdma_search_ram_block(rdma, block_offset, |
| 2751 |
offset, size, &index, &chunk); |
| 2752 |
|
| 2753 |
if (ret) {
|
| 2754 |
fprintf(stderr, "ram block search failed\n");
|
| 2755 |
goto err;
|
| 2756 |
} |
| 2757 |
|
| 2758 |
qemu_rdma_signal_unregister(rdma, index, chunk, 0);
|
| 2759 |
|
| 2760 |
/*
|
| 2761 |
* TODO: Synchronous, guaranteed unregistration (should not occur during
|
| 2762 |
* fast-path). Otherwise, unregisters will process on the next call to
|
| 2763 |
* qemu_rdma_drain_cq()
|
| 2764 |
if (size < 0) {
|
| 2765 |
qemu_rdma_unregister_waiting(rdma);
|
| 2766 |
}
|
| 2767 |
*/
|
| 2768 |
} |
| 2769 |
|
| 2770 |
/*
|
| 2771 |
* Drain the Completion Queue if possible, but do not block,
|
| 2772 |
* just poll.
|
| 2773 |
*
|
| 2774 |
* If nothing to poll, the end of the iteration will do this
|
| 2775 |
* again to make sure we don't overflow the request queue.
|
| 2776 |
*/
|
| 2777 |
while (1) { |
| 2778 |
uint64_t wr_id, wr_id_in; |
| 2779 |
int ret = qemu_rdma_poll(rdma, &wr_id_in, NULL); |
| 2780 |
if (ret < 0) { |
| 2781 |
fprintf(stderr, "rdma migration: polling error! %d\n", ret);
|
| 2782 |
goto err;
|
| 2783 |
} |
| 2784 |
|
| 2785 |
wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; |
| 2786 |
|
| 2787 |
if (wr_id == RDMA_WRID_NONE) {
|
| 2788 |
break;
|
| 2789 |
} |
| 2790 |
} |
| 2791 |
|
| 2792 |
return RAM_SAVE_CONTROL_DELAYED;
|
| 2793 |
err:
|
| 2794 |
rdma->error_state = ret; |
| 2795 |
return ret;
|
| 2796 |
} |
| 2797 |
|
| 2798 |
static int qemu_rdma_accept(RDMAContext *rdma) |
| 2799 |
{
|
| 2800 |
RDMACapabilities cap; |
| 2801 |
struct rdma_conn_param conn_param = {
|
| 2802 |
.responder_resources = 2,
|
| 2803 |
.private_data = &cap, |
| 2804 |
.private_data_len = sizeof(cap),
|
| 2805 |
}; |
| 2806 |
struct rdma_cm_event *cm_event;
|
| 2807 |
struct ibv_context *verbs;
|
| 2808 |
int ret = -EINVAL;
|
| 2809 |
int idx;
|
| 2810 |
|
| 2811 |
ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2812 |
if (ret) {
|
| 2813 |
goto err_rdma_dest_wait;
|
| 2814 |
} |
| 2815 |
|
| 2816 |
if (cm_event->event != RDMA_CM_EVENT_CONNECT_REQUEST) {
|
| 2817 |
rdma_ack_cm_event(cm_event); |
| 2818 |
goto err_rdma_dest_wait;
|
| 2819 |
} |
| 2820 |
|
| 2821 |
memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap));
|
| 2822 |
|
| 2823 |
network_to_caps(&cap); |
| 2824 |
|
| 2825 |
if (cap.version < 1 || cap.version > RDMA_CONTROL_VERSION_CURRENT) { |
| 2826 |
fprintf(stderr, "Unknown source RDMA version: %d, bailing...\n",
|
| 2827 |
cap.version); |
| 2828 |
rdma_ack_cm_event(cm_event); |
| 2829 |
goto err_rdma_dest_wait;
|
| 2830 |
} |
| 2831 |
|
| 2832 |
/*
|
| 2833 |
* Respond with only the capabilities this version of QEMU knows about.
|
| 2834 |
*/
|
| 2835 |
cap.flags &= known_capabilities; |
| 2836 |
|
| 2837 |
/*
|
| 2838 |
* Enable the ones that we do know about.
|
| 2839 |
* Add other checks here as new ones are introduced.
|
| 2840 |
*/
|
| 2841 |
if (cap.flags & RDMA_CAPABILITY_PIN_ALL) {
|
| 2842 |
rdma->pin_all = true;
|
| 2843 |
} |
| 2844 |
|
| 2845 |
rdma->cm_id = cm_event->id; |
| 2846 |
verbs = cm_event->id->verbs; |
| 2847 |
|
| 2848 |
rdma_ack_cm_event(cm_event); |
| 2849 |
|
| 2850 |
DPRINTF("Memory pin all: %s\n", rdma->pin_all ? "enabled" : "disabled"); |
| 2851 |
|
| 2852 |
caps_to_network(&cap); |
| 2853 |
|
| 2854 |
DPRINTF("verbs context after listen: %p\n", verbs);
|
| 2855 |
|
| 2856 |
if (!rdma->verbs) {
|
| 2857 |
rdma->verbs = verbs; |
| 2858 |
} else if (rdma->verbs != verbs) { |
| 2859 |
fprintf(stderr, "ibv context not matching %p, %p!\n",
|
| 2860 |
rdma->verbs, verbs); |
| 2861 |
goto err_rdma_dest_wait;
|
| 2862 |
} |
| 2863 |
|
| 2864 |
qemu_rdma_dump_id("dest_init", verbs);
|
| 2865 |
|
| 2866 |
ret = qemu_rdma_alloc_pd_cq(rdma); |
| 2867 |
if (ret) {
|
| 2868 |
fprintf(stderr, "rdma migration: error allocating pd and cq!\n");
|
| 2869 |
goto err_rdma_dest_wait;
|
| 2870 |
} |
| 2871 |
|
| 2872 |
ret = qemu_rdma_alloc_qp(rdma); |
| 2873 |
if (ret) {
|
| 2874 |
fprintf(stderr, "rdma migration: error allocating qp!\n");
|
| 2875 |
goto err_rdma_dest_wait;
|
| 2876 |
} |
| 2877 |
|
| 2878 |
ret = qemu_rdma_init_ram_blocks(rdma); |
| 2879 |
if (ret) {
|
| 2880 |
fprintf(stderr, "rdma migration: error initializing ram blocks!\n");
|
| 2881 |
goto err_rdma_dest_wait;
|
| 2882 |
} |
| 2883 |
|
| 2884 |
for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
| 2885 |
ret = qemu_rdma_reg_control(rdma, idx); |
| 2886 |
if (ret) {
|
| 2887 |
fprintf(stderr, "rdma: error registering %d control!\n", idx);
|
| 2888 |
goto err_rdma_dest_wait;
|
| 2889 |
} |
| 2890 |
} |
| 2891 |
|
| 2892 |
qemu_set_fd_handler2(rdma->channel->fd, NULL, NULL, NULL, NULL); |
| 2893 |
|
| 2894 |
ret = rdma_accept(rdma->cm_id, &conn_param); |
| 2895 |
if (ret) {
|
| 2896 |
fprintf(stderr, "rdma_accept returns %d!\n", ret);
|
| 2897 |
goto err_rdma_dest_wait;
|
| 2898 |
} |
| 2899 |
|
| 2900 |
ret = rdma_get_cm_event(rdma->channel, &cm_event); |
| 2901 |
if (ret) {
|
| 2902 |
fprintf(stderr, "rdma_accept get_cm_event failed %d!\n", ret);
|
| 2903 |
goto err_rdma_dest_wait;
|
| 2904 |
} |
| 2905 |
|
| 2906 |
if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) {
|
| 2907 |
fprintf(stderr, "rdma_accept not event established!\n");
|
| 2908 |
rdma_ack_cm_event(cm_event); |
| 2909 |
goto err_rdma_dest_wait;
|
| 2910 |
} |
| 2911 |
|
| 2912 |
rdma_ack_cm_event(cm_event); |
| 2913 |
rdma->connected = true;
|
| 2914 |
|
| 2915 |
ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
| 2916 |
if (ret) {
|
| 2917 |
fprintf(stderr, "rdma migration: error posting second control recv!\n");
|
| 2918 |
goto err_rdma_dest_wait;
|
| 2919 |
} |
| 2920 |
|
| 2921 |
qemu_rdma_dump_gid("dest_connect", rdma->cm_id);
|
| 2922 |
|
| 2923 |
return 0; |
| 2924 |
|
| 2925 |
err_rdma_dest_wait:
|
| 2926 |
rdma->error_state = ret; |
| 2927 |
qemu_rdma_cleanup(rdma); |
| 2928 |
return ret;
|
| 2929 |
} |
| 2930 |
|
| 2931 |
/*
|
| 2932 |
* During each iteration of the migration, we listen for instructions
|
| 2933 |
* by the source VM to perform dynamic page registrations before they
|
| 2934 |
* can perform RDMA operations.
|
| 2935 |
*
|
| 2936 |
* We respond with the 'rkey'.
|
| 2937 |
*
|
| 2938 |
* Keep doing this until the source tells us to stop.
|
| 2939 |
*/
|
| 2940 |
static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque, |
| 2941 |
uint64_t flags) |
| 2942 |
{
|
| 2943 |
RDMAControlHeader reg_resp = { .len = sizeof(RDMARegisterResult),
|
| 2944 |
.type = RDMA_CONTROL_REGISTER_RESULT, |
| 2945 |
.repeat = 0,
|
| 2946 |
}; |
| 2947 |
RDMAControlHeader unreg_resp = { .len = 0,
|
| 2948 |
.type = RDMA_CONTROL_UNREGISTER_FINISHED, |
| 2949 |
.repeat = 0,
|
| 2950 |
}; |
| 2951 |
RDMAControlHeader blocks = { .type = RDMA_CONTROL_RAM_BLOCKS_RESULT,
|
| 2952 |
.repeat = 1 };
|
| 2953 |
QEMUFileRDMA *rfile = opaque; |
| 2954 |
RDMAContext *rdma = rfile->rdma; |
| 2955 |
RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 2956 |
RDMAControlHeader head; |
| 2957 |
RDMARegister *reg, *registers; |
| 2958 |
RDMACompress *comp; |
| 2959 |
RDMARegisterResult *reg_result; |
| 2960 |
static RDMARegisterResult results[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE];
|
| 2961 |
RDMALocalBlock *block; |
| 2962 |
void *host_addr;
|
| 2963 |
int ret = 0; |
| 2964 |
int idx = 0; |
| 2965 |
int count = 0; |
| 2966 |
int i = 0; |
| 2967 |
|
| 2968 |
CHECK_ERROR_STATE(); |
| 2969 |
|
| 2970 |
do {
|
| 2971 |
DDDPRINTF("Waiting for next request %" PRIu64 "...\n", flags); |
| 2972 |
|
| 2973 |
ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_NONE); |
| 2974 |
|
| 2975 |
if (ret < 0) { |
| 2976 |
break;
|
| 2977 |
} |
| 2978 |
|
| 2979 |
if (head.repeat > RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE) {
|
| 2980 |
fprintf(stderr, "rdma: Too many requests in this message (%d)."
|
| 2981 |
"Bailing.\n", head.repeat);
|
| 2982 |
ret = -EIO; |
| 2983 |
break;
|
| 2984 |
} |
| 2985 |
|
| 2986 |
switch (head.type) {
|
| 2987 |
case RDMA_CONTROL_COMPRESS:
|
| 2988 |
comp = (RDMACompress *) rdma->wr_data[idx].control_curr; |
| 2989 |
network_to_compress(comp); |
| 2990 |
|
| 2991 |
DDPRINTF("Zapping zero chunk: %" PRId64
|
| 2992 |
" bytes, index %d, offset %" PRId64 "\n", |
| 2993 |
comp->length, comp->block_idx, comp->offset); |
| 2994 |
block = &(rdma->local_ram_blocks.block[comp->block_idx]); |
| 2995 |
|
| 2996 |
host_addr = block->local_host_addr + |
| 2997 |
(comp->offset - block->offset); |
| 2998 |
|
| 2999 |
ram_handle_compressed(host_addr, comp->value, comp->length); |
| 3000 |
break;
|
| 3001 |
|
| 3002 |
case RDMA_CONTROL_REGISTER_FINISHED:
|
| 3003 |
DDDPRINTF("Current registrations complete.\n");
|
| 3004 |
goto out;
|
| 3005 |
|
| 3006 |
case RDMA_CONTROL_RAM_BLOCKS_REQUEST:
|
| 3007 |
DPRINTF("Initial setup info requested.\n");
|
| 3008 |
|
| 3009 |
if (rdma->pin_all) {
|
| 3010 |
ret = qemu_rdma_reg_whole_ram_blocks(rdma); |
| 3011 |
if (ret) {
|
| 3012 |
fprintf(stderr, "rdma migration: error dest "
|
| 3013 |
"registering ram blocks!\n");
|
| 3014 |
goto out;
|
| 3015 |
} |
| 3016 |
} |
| 3017 |
|
| 3018 |
/*
|
| 3019 |
* Dest uses this to prepare to transmit the RAMBlock descriptions
|
| 3020 |
* to the source VM after connection setup.
|
| 3021 |
* Both sides use the "remote" structure to communicate and update
|
| 3022 |
* their "local" descriptions with what was sent.
|
| 3023 |
*/
|
| 3024 |
for (i = 0; i < local->nb_blocks; i++) { |
| 3025 |
rdma->block[i].remote_host_addr = |
| 3026 |
(uint64_t)(local->block[i].local_host_addr); |
| 3027 |
|
| 3028 |
if (rdma->pin_all) {
|
| 3029 |
rdma->block[i].remote_rkey = local->block[i].mr->rkey; |
| 3030 |
} |
| 3031 |
|
| 3032 |
rdma->block[i].offset = local->block[i].offset; |
| 3033 |
rdma->block[i].length = local->block[i].length; |
| 3034 |
|
| 3035 |
remote_block_to_network(&rdma->block[i]); |
| 3036 |
} |
| 3037 |
|
| 3038 |
blocks.len = rdma->local_ram_blocks.nb_blocks |
| 3039 |
* sizeof(RDMARemoteBlock);
|
| 3040 |
|
| 3041 |
|
| 3042 |
ret = qemu_rdma_post_send_control(rdma, |
| 3043 |
(uint8_t *) rdma->block, &blocks); |
| 3044 |
|
| 3045 |
if (ret < 0) { |
| 3046 |
fprintf(stderr, "rdma migration: error sending remote info!\n");
|
| 3047 |
goto out;
|
| 3048 |
} |
| 3049 |
|
| 3050 |
break;
|
| 3051 |
case RDMA_CONTROL_REGISTER_REQUEST:
|
| 3052 |
DDPRINTF("There are %d registration requests\n", head.repeat);
|
| 3053 |
|
| 3054 |
reg_resp.repeat = head.repeat; |
| 3055 |
registers = (RDMARegister *) rdma->wr_data[idx].control_curr; |
| 3056 |
|
| 3057 |
for (count = 0; count < head.repeat; count++) { |
| 3058 |
uint64_t chunk; |
| 3059 |
uint8_t *chunk_start, *chunk_end; |
| 3060 |
|
| 3061 |
reg = ®isters[count]; |
| 3062 |
network_to_register(reg); |
| 3063 |
|
| 3064 |
reg_result = &results[count]; |
| 3065 |
|
| 3066 |
DDPRINTF("Registration request (%d): index %d, current_addr %"
|
| 3067 |
PRIu64 " chunks: %" PRIu64 "\n", count, |
| 3068 |
reg->current_index, reg->key.current_addr, reg->chunks); |
| 3069 |
|
| 3070 |
block = &(rdma->local_ram_blocks.block[reg->current_index]); |
| 3071 |
if (block->is_ram_block) {
|
| 3072 |
host_addr = (block->local_host_addr + |
| 3073 |
(reg->key.current_addr - block->offset)); |
| 3074 |
chunk = ram_chunk_index(block->local_host_addr, |
| 3075 |
(uint8_t *) host_addr); |
| 3076 |
} else {
|
| 3077 |
chunk = reg->key.chunk; |
| 3078 |
host_addr = block->local_host_addr + |
| 3079 |
(reg->key.chunk * (1UL << RDMA_REG_CHUNK_SHIFT));
|
| 3080 |
} |
| 3081 |
chunk_start = ram_chunk_start(block, chunk); |
| 3082 |
chunk_end = ram_chunk_end(block, chunk + reg->chunks); |
| 3083 |
if (qemu_rdma_register_and_get_keys(rdma, block,
|
| 3084 |
(uint8_t *)host_addr, NULL, ®_result->rkey,
|
| 3085 |
chunk, chunk_start, chunk_end)) {
|
| 3086 |
fprintf(stderr, "cannot get rkey!\n");
|
| 3087 |
ret = -EINVAL; |
| 3088 |
goto out;
|
| 3089 |
} |
| 3090 |
|
| 3091 |
reg_result->host_addr = (uint64_t) block->local_host_addr; |
| 3092 |
|
| 3093 |
DDPRINTF("Registered rkey for this request: %x\n",
|
| 3094 |
reg_result->rkey); |
| 3095 |
|
| 3096 |
result_to_network(reg_result); |
| 3097 |
} |
| 3098 |
|
| 3099 |
ret = qemu_rdma_post_send_control(rdma, |
| 3100 |
(uint8_t *) results, ®_resp); |
| 3101 |
|
| 3102 |
if (ret < 0) { |
| 3103 |
fprintf(stderr, "Failed to send control buffer!\n");
|
| 3104 |
goto out;
|
| 3105 |
} |
| 3106 |
break;
|
| 3107 |
case RDMA_CONTROL_UNREGISTER_REQUEST:
|
| 3108 |
DDPRINTF("There are %d unregistration requests\n", head.repeat);
|
| 3109 |
unreg_resp.repeat = head.repeat; |
| 3110 |
registers = (RDMARegister *) rdma->wr_data[idx].control_curr; |
| 3111 |
|
| 3112 |
for (count = 0; count < head.repeat; count++) { |
| 3113 |
reg = ®isters[count]; |
| 3114 |
network_to_register(reg); |
| 3115 |
|
| 3116 |
DDPRINTF("Unregistration request (%d): "
|
| 3117 |
" index %d, chunk %" PRIu64 "\n", |
| 3118 |
count, reg->current_index, reg->key.chunk); |
| 3119 |
|
| 3120 |
block = &(rdma->local_ram_blocks.block[reg->current_index]); |
| 3121 |
|
| 3122 |
ret = ibv_dereg_mr(block->pmr[reg->key.chunk]); |
| 3123 |
block->pmr[reg->key.chunk] = NULL;
|
| 3124 |
|
| 3125 |
if (ret != 0) { |
| 3126 |
perror("rdma unregistration chunk failed");
|
| 3127 |
ret = -ret; |
| 3128 |
goto out;
|
| 3129 |
} |
| 3130 |
|
| 3131 |
rdma->total_registrations--; |
| 3132 |
|
| 3133 |
DDPRINTF("Unregistered chunk %" PRIu64 " successfully.\n", |
| 3134 |
reg->key.chunk); |
| 3135 |
} |
| 3136 |
|
| 3137 |
ret = qemu_rdma_post_send_control(rdma, NULL, &unreg_resp);
|
| 3138 |
|
| 3139 |
if (ret < 0) { |
| 3140 |
fprintf(stderr, "Failed to send control buffer!\n");
|
| 3141 |
goto out;
|
| 3142 |
} |
| 3143 |
break;
|
| 3144 |
case RDMA_CONTROL_REGISTER_RESULT:
|
| 3145 |
fprintf(stderr, "Invalid RESULT message at dest.\n");
|
| 3146 |
ret = -EIO; |
| 3147 |
goto out;
|
| 3148 |
default:
|
| 3149 |
fprintf(stderr, "Unknown control message %s\n",
|
| 3150 |
control_desc[head.type]); |
| 3151 |
ret = -EIO; |
| 3152 |
goto out;
|
| 3153 |
} |
| 3154 |
} while (1); |
| 3155 |
out:
|
| 3156 |
if (ret < 0) { |
| 3157 |
rdma->error_state = ret; |
| 3158 |
} |
| 3159 |
return ret;
|
| 3160 |
} |
| 3161 |
|
| 3162 |
static int qemu_rdma_registration_start(QEMUFile *f, void *opaque, |
| 3163 |
uint64_t flags) |
| 3164 |
{
|
| 3165 |
QEMUFileRDMA *rfile = opaque; |
| 3166 |
RDMAContext *rdma = rfile->rdma; |
| 3167 |
|
| 3168 |
CHECK_ERROR_STATE(); |
| 3169 |
|
| 3170 |
DDDPRINTF("start section: %" PRIu64 "\n", flags); |
| 3171 |
qemu_put_be64(f, RAM_SAVE_FLAG_HOOK); |
| 3172 |
qemu_fflush(f); |
| 3173 |
|
| 3174 |
return 0; |
| 3175 |
} |
| 3176 |
|
| 3177 |
/*
|
| 3178 |
* Inform dest that dynamic registrations are done for now.
|
| 3179 |
* First, flush writes, if any.
|
| 3180 |
*/
|
| 3181 |
static int qemu_rdma_registration_stop(QEMUFile *f, void *opaque, |
| 3182 |
uint64_t flags) |
| 3183 |
{
|
| 3184 |
Error *local_err = NULL, **errp = &local_err;
|
| 3185 |
QEMUFileRDMA *rfile = opaque; |
| 3186 |
RDMAContext *rdma = rfile->rdma; |
| 3187 |
RDMAControlHeader head = { .len = 0, .repeat = 1 };
|
| 3188 |
int ret = 0; |
| 3189 |
|
| 3190 |
CHECK_ERROR_STATE(); |
| 3191 |
|
| 3192 |
qemu_fflush(f); |
| 3193 |
ret = qemu_rdma_drain_cq(f, rdma); |
| 3194 |
|
| 3195 |
if (ret < 0) { |
| 3196 |
goto err;
|
| 3197 |
} |
| 3198 |
|
| 3199 |
if (flags == RAM_CONTROL_SETUP) {
|
| 3200 |
RDMAControlHeader resp = {.type = RDMA_CONTROL_RAM_BLOCKS_RESULT };
|
| 3201 |
RDMALocalBlocks *local = &rdma->local_ram_blocks; |
| 3202 |
int reg_result_idx, i, j, nb_remote_blocks;
|
| 3203 |
|
| 3204 |
head.type = RDMA_CONTROL_RAM_BLOCKS_REQUEST; |
| 3205 |
DPRINTF("Sending registration setup for ram blocks...\n");
|
| 3206 |
|
| 3207 |
/*
|
| 3208 |
* Make sure that we parallelize the pinning on both sides.
|
| 3209 |
* For very large guests, doing this serially takes a really
|
| 3210 |
* long time, so we have to 'interleave' the pinning locally
|
| 3211 |
* with the control messages by performing the pinning on this
|
| 3212 |
* side before we receive the control response from the other
|
| 3213 |
* side that the pinning has completed.
|
| 3214 |
*/
|
| 3215 |
ret = qemu_rdma_exchange_send(rdma, &head, NULL, &resp,
|
| 3216 |
®_result_idx, rdma->pin_all ? |
| 3217 |
qemu_rdma_reg_whole_ram_blocks : NULL);
|
| 3218 |
if (ret < 0) { |
| 3219 |
ERROR(errp, "receiving remote info!");
|
| 3220 |
return ret;
|
| 3221 |
} |
| 3222 |
|
| 3223 |
nb_remote_blocks = resp.len / sizeof(RDMARemoteBlock);
|
| 3224 |
|
| 3225 |
/*
|
| 3226 |
* The protocol uses two different sets of rkeys (mutually exclusive):
|
| 3227 |
* 1. One key to represent the virtual address of the entire ram block.
|
| 3228 |
* (dynamic chunk registration disabled - pin everything with one rkey.)
|
| 3229 |
* 2. One to represent individual chunks within a ram block.
|
| 3230 |
* (dynamic chunk registration enabled - pin individual chunks.)
|
| 3231 |
*
|
| 3232 |
* Once the capability is successfully negotiated, the destination transmits
|
| 3233 |
* the keys to use (or sends them later) including the virtual addresses
|
| 3234 |
* and then propagates the remote ram block descriptions to his local copy.
|
| 3235 |
*/
|
| 3236 |
|
| 3237 |
if (local->nb_blocks != nb_remote_blocks) {
|
| 3238 |
ERROR(errp, "ram blocks mismatch #1! "
|
| 3239 |
"Your QEMU command line parameters are probably "
|
| 3240 |
"not identical on both the source and destination.");
|
| 3241 |
return -EINVAL;
|
| 3242 |
} |
| 3243 |
|
| 3244 |
qemu_rdma_move_header(rdma, reg_result_idx, &resp); |
| 3245 |
memcpy(rdma->block, |
| 3246 |
rdma->wr_data[reg_result_idx].control_curr, resp.len); |
| 3247 |
for (i = 0; i < nb_remote_blocks; i++) { |
| 3248 |
network_to_remote_block(&rdma->block[i]); |
| 3249 |
|
| 3250 |
/* search local ram blocks */
|
| 3251 |
for (j = 0; j < local->nb_blocks; j++) { |
| 3252 |
if (rdma->block[i].offset != local->block[j].offset) {
|
| 3253 |
continue;
|
| 3254 |
} |
| 3255 |
|
| 3256 |
if (rdma->block[i].length != local->block[j].length) {
|
| 3257 |
ERROR(errp, "ram blocks mismatch #2! "
|
| 3258 |
"Your QEMU command line parameters are probably "
|
| 3259 |
"not identical on both the source and destination.");
|
| 3260 |
return -EINVAL;
|
| 3261 |
} |
| 3262 |
local->block[j].remote_host_addr = |
| 3263 |
rdma->block[i].remote_host_addr; |
| 3264 |
local->block[j].remote_rkey = rdma->block[i].remote_rkey; |
| 3265 |
break;
|
| 3266 |
} |
| 3267 |
|
| 3268 |
if (j >= local->nb_blocks) {
|
| 3269 |
ERROR(errp, "ram blocks mismatch #3! "
|
| 3270 |
"Your QEMU command line parameters are probably "
|
| 3271 |
"not identical on both the source and destination.");
|
| 3272 |
return -EINVAL;
|
| 3273 |
} |
| 3274 |
} |
| 3275 |
} |
| 3276 |
|
| 3277 |
DDDPRINTF("Sending registration finish %" PRIu64 "...\n", flags); |
| 3278 |
|
| 3279 |
head.type = RDMA_CONTROL_REGISTER_FINISHED; |
| 3280 |
ret = qemu_rdma_exchange_send(rdma, &head, NULL, NULL, NULL, NULL); |
| 3281 |
|
| 3282 |
if (ret < 0) { |
| 3283 |
goto err;
|
| 3284 |
} |
| 3285 |
|
| 3286 |
return 0; |
| 3287 |
err:
|
| 3288 |
rdma->error_state = ret; |
| 3289 |
return ret;
|
| 3290 |
} |
| 3291 |
|
| 3292 |
static int qemu_rdma_get_fd(void *opaque) |
| 3293 |
{
|
| 3294 |
QEMUFileRDMA *rfile = opaque; |
| 3295 |
RDMAContext *rdma = rfile->rdma; |
| 3296 |
|
| 3297 |
return rdma->comp_channel->fd;
|
| 3298 |
} |
| 3299 |
|
| 3300 |
const QEMUFileOps rdma_read_ops = {
|
| 3301 |
.get_buffer = qemu_rdma_get_buffer, |
| 3302 |
.get_fd = qemu_rdma_get_fd, |
| 3303 |
.close = qemu_rdma_close, |
| 3304 |
.hook_ram_load = qemu_rdma_registration_handle, |
| 3305 |
}; |
| 3306 |
|
| 3307 |
const QEMUFileOps rdma_write_ops = {
|
| 3308 |
.put_buffer = qemu_rdma_put_buffer, |
| 3309 |
.close = qemu_rdma_close, |
| 3310 |
.before_ram_iterate = qemu_rdma_registration_start, |
| 3311 |
.after_ram_iterate = qemu_rdma_registration_stop, |
| 3312 |
.save_page = qemu_rdma_save_page, |
| 3313 |
}; |
| 3314 |
|
| 3315 |
static void *qemu_fopen_rdma(RDMAContext *rdma, const char *mode) |
| 3316 |
{
|
| 3317 |
QEMUFileRDMA *r = g_malloc0(sizeof(QEMUFileRDMA));
|
| 3318 |
|
| 3319 |
if (qemu_file_mode_is_not_valid(mode)) {
|
| 3320 |
return NULL; |
| 3321 |
} |
| 3322 |
|
| 3323 |
r->rdma = rdma; |
| 3324 |
|
| 3325 |
if (mode[0] == 'w') { |
| 3326 |
r->file = qemu_fopen_ops(r, &rdma_write_ops); |
| 3327 |
} else {
|
| 3328 |
r->file = qemu_fopen_ops(r, &rdma_read_ops); |
| 3329 |
} |
| 3330 |
|
| 3331 |
return r->file;
|
| 3332 |
} |
| 3333 |
|
| 3334 |
static void rdma_accept_incoming_migration(void *opaque) |
| 3335 |
{
|
| 3336 |
RDMAContext *rdma = opaque; |
| 3337 |
int ret;
|
| 3338 |
QEMUFile *f; |
| 3339 |
Error *local_err = NULL, **errp = &local_err;
|
| 3340 |
|
| 3341 |
DPRINTF("Accepting rdma connection...\n");
|
| 3342 |
ret = qemu_rdma_accept(rdma); |
| 3343 |
|
| 3344 |
if (ret) {
|
| 3345 |
ERROR(errp, "RDMA Migration initialization failed!");
|
| 3346 |
return;
|
| 3347 |
} |
| 3348 |
|
| 3349 |
DPRINTF("Accepted migration\n");
|
| 3350 |
|
| 3351 |
f = qemu_fopen_rdma(rdma, "rb");
|
| 3352 |
if (f == NULL) { |
| 3353 |
ERROR(errp, "could not qemu_fopen_rdma!");
|
| 3354 |
qemu_rdma_cleanup(rdma); |
| 3355 |
return;
|
| 3356 |
} |
| 3357 |
|
| 3358 |
rdma->migration_started_on_destination = 1;
|
| 3359 |
process_incoming_migration(f); |
| 3360 |
} |
| 3361 |
|
| 3362 |
void rdma_start_incoming_migration(const char *host_port, Error **errp) |
| 3363 |
{
|
| 3364 |
int ret;
|
| 3365 |
RDMAContext *rdma; |
| 3366 |
Error *local_err = NULL;
|
| 3367 |
|
| 3368 |
DPRINTF("Starting RDMA-based incoming migration\n");
|
| 3369 |
rdma = qemu_rdma_data_init(host_port, &local_err); |
| 3370 |
|
| 3371 |
if (rdma == NULL) { |
| 3372 |
goto err;
|
| 3373 |
} |
| 3374 |
|
| 3375 |
ret = qemu_rdma_dest_init(rdma, &local_err); |
| 3376 |
|
| 3377 |
if (ret) {
|
| 3378 |
goto err;
|
| 3379 |
} |
| 3380 |
|
| 3381 |
DPRINTF("qemu_rdma_dest_init success\n");
|
| 3382 |
|
| 3383 |
ret = rdma_listen(rdma->listen_id, 5);
|
| 3384 |
|
| 3385 |
if (ret) {
|
| 3386 |
ERROR(errp, "listening on socket!");
|
| 3387 |
goto err;
|
| 3388 |
} |
| 3389 |
|
| 3390 |
DPRINTF("rdma_listen success\n");
|
| 3391 |
|
| 3392 |
qemu_set_fd_handler2(rdma->channel->fd, NULL,
|
| 3393 |
rdma_accept_incoming_migration, NULL,
|
| 3394 |
(void *)(intptr_t) rdma);
|
| 3395 |
return;
|
| 3396 |
err:
|
| 3397 |
error_propagate(errp, local_err); |
| 3398 |
g_free(rdma); |
| 3399 |
} |
| 3400 |
|
| 3401 |
void rdma_start_outgoing_migration(void *opaque, |
| 3402 |
const char *host_port, Error **errp) |
| 3403 |
{
|
| 3404 |
MigrationState *s = opaque; |
| 3405 |
Error *local_err = NULL, **temp = &local_err;
|
| 3406 |
RDMAContext *rdma = qemu_rdma_data_init(host_port, &local_err); |
| 3407 |
int ret = 0; |
| 3408 |
|
| 3409 |
if (rdma == NULL) { |
| 3410 |
ERROR(temp, "Failed to initialize RDMA data structures! %d", ret);
|
| 3411 |
goto err;
|
| 3412 |
} |
| 3413 |
|
| 3414 |
ret = qemu_rdma_source_init(rdma, &local_err, |
| 3415 |
s->enabled_capabilities[MIGRATION_CAPABILITY_X_RDMA_PIN_ALL]); |
| 3416 |
|
| 3417 |
if (ret) {
|
| 3418 |
goto err;
|
| 3419 |
} |
| 3420 |
|
| 3421 |
DPRINTF("qemu_rdma_source_init success\n");
|
| 3422 |
ret = qemu_rdma_connect(rdma, &local_err); |
| 3423 |
|
| 3424 |
if (ret) {
|
| 3425 |
goto err;
|
| 3426 |
} |
| 3427 |
|
| 3428 |
DPRINTF("qemu_rdma_source_connect success\n");
|
| 3429 |
|
| 3430 |
s->file = qemu_fopen_rdma(rdma, "wb");
|
| 3431 |
migrate_fd_connect(s); |
| 3432 |
return;
|
| 3433 |
err:
|
| 3434 |
error_propagate(errp, local_err); |
| 3435 |
g_free(rdma); |
| 3436 |
migrate_fd_error(s); |
| 3437 |
} |