2 * RDMA protocol and interfaces
4 * Copyright IBM, Corp. 2010-2013
7 * Michael R. Hines <mrhines@us.ibm.com>
8 * Jiuxing Liu <jl@us.ibm.com>
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.
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"
23 #include <sys/types.h>
24 #include <sys/socket.h>
26 #include <arpa/inet.h>
28 #include <rdma/rdma_cma.h>
32 * Print and error on both the Monitor and the Log file.
34 #define ERROR(errp, fmt, ...) \
36 fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \
37 if (errp && (*(errp) == NULL)) { \
38 error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \
42 #define RDMA_RESOLVE_TIMEOUT_MS 10000
44 /* Do not merge data if larger than this. */
45 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
46 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
48 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
51 * This is only for non-live state being migrated.
52 * Instead of RDMA_WRITE messages, we use RDMA_SEND
53 * messages for that state, which requires a different
54 * delivery design than main memory.
56 #define RDMA_SEND_INCREMENT 32768
59 * Maximum size infiniband SEND message
61 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
62 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
64 #define RDMA_CONTROL_VERSION_CURRENT 1
66 * Capabilities for negotiation.
68 #define RDMA_CAPABILITY_PIN_ALL 0x01
71 * Add the other flags above to this list of known capabilities
72 * as they are introduced.
74 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
76 #define CHECK_ERROR_STATE() \
78 if (rdma->error_state) { \
79 if (!rdma->error_reported) { \
80 error_report("RDMA is in an error state waiting migration" \
82 rdma->error_reported = 1; \
84 return rdma->error_state; \
89 * A work request ID is 64-bits and we split up these bits
92 * bits 0-15 : type of control message, 2^16
93 * bits 16-29: ram block index, 2^14
94 * bits 30-63: ram block chunk number, 2^34
96 * The last two bit ranges are only used for RDMA writes,
97 * in order to track their completion and potentially
98 * also track unregistration status of the message.
100 #define RDMA_WRID_TYPE_SHIFT 0UL
101 #define RDMA_WRID_BLOCK_SHIFT 16UL
102 #define RDMA_WRID_CHUNK_SHIFT 30UL
104 #define RDMA_WRID_TYPE_MASK \
105 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
107 #define RDMA_WRID_BLOCK_MASK \
108 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
110 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
113 * RDMA migration protocol:
114 * 1. RDMA Writes (data messages, i.e. RAM)
115 * 2. IB Send/Recv (control channel messages)
119 RDMA_WRID_RDMA_WRITE
= 1,
120 RDMA_WRID_SEND_CONTROL
= 2000,
121 RDMA_WRID_RECV_CONTROL
= 4000,
124 static const char *wrid_desc
[] = {
125 [RDMA_WRID_NONE
] = "NONE",
126 [RDMA_WRID_RDMA_WRITE
] = "WRITE RDMA",
127 [RDMA_WRID_SEND_CONTROL
] = "CONTROL SEND",
128 [RDMA_WRID_RECV_CONTROL
] = "CONTROL RECV",
132 * Work request IDs for IB SEND messages only (not RDMA writes).
133 * This is used by the migration protocol to transmit
134 * control messages (such as device state and registration commands)
136 * We could use more WRs, but we have enough for now.
146 * SEND/RECV IB Control Messages.
149 RDMA_CONTROL_NONE
= 0,
151 RDMA_CONTROL_READY
, /* ready to receive */
152 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
153 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
154 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
155 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
156 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
157 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
158 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
159 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
160 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
163 static const char *control_desc
[] = {
164 [RDMA_CONTROL_NONE
] = "NONE",
165 [RDMA_CONTROL_ERROR
] = "ERROR",
166 [RDMA_CONTROL_READY
] = "READY",
167 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
168 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
169 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
170 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
171 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
172 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
173 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
174 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
175 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
179 * Memory and MR structures used to represent an IB Send/Recv work request.
180 * This is *not* used for RDMA writes, only IB Send/Recv.
183 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
184 struct ibv_mr
*control_mr
; /* registration metadata */
185 size_t control_len
; /* length of the message */
186 uint8_t *control_curr
; /* start of unconsumed bytes */
187 } RDMAWorkRequestData
;
190 * Negotiate RDMA capabilities during connection-setup time.
197 static void caps_to_network(RDMACapabilities
*cap
)
199 cap
->version
= htonl(cap
->version
);
200 cap
->flags
= htonl(cap
->flags
);
203 static void network_to_caps(RDMACapabilities
*cap
)
205 cap
->version
= ntohl(cap
->version
);
206 cap
->flags
= ntohl(cap
->flags
);
210 * Representation of a RAMBlock from an RDMA perspective.
211 * This is not transmitted, only local.
212 * This and subsequent structures cannot be linked lists
213 * because we're using a single IB message to transmit
214 * the information. It's small anyway, so a list is overkill.
216 typedef struct RDMALocalBlock
{
217 uint8_t *local_host_addr
; /* local virtual address */
218 uint64_t remote_host_addr
; /* remote virtual address */
221 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
222 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
223 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
224 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
225 int index
; /* which block are we */
228 unsigned long *transit_bitmap
;
229 unsigned long *unregister_bitmap
;
233 * Also represents a RAMblock, but only on the dest.
234 * This gets transmitted by the dest during connection-time
235 * to the source VM and then is used to populate the
236 * corresponding RDMALocalBlock with
237 * the information needed to perform the actual RDMA.
239 typedef struct QEMU_PACKED RDMADestBlock
{
240 uint64_t remote_host_addr
;
243 uint32_t remote_rkey
;
247 static uint64_t htonll(uint64_t v
)
249 union { uint32_t lv
[2]; uint64_t llv
; } u
;
250 u
.lv
[0] = htonl(v
>> 32);
251 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
255 static uint64_t ntohll(uint64_t v
) {
256 union { uint32_t lv
[2]; uint64_t llv
; } u
;
258 return ((uint64_t)ntohl(u
.lv
[0]) << 32) | (uint64_t) ntohl(u
.lv
[1]);
261 static void dest_block_to_network(RDMADestBlock
*db
)
263 db
->remote_host_addr
= htonll(db
->remote_host_addr
);
264 db
->offset
= htonll(db
->offset
);
265 db
->length
= htonll(db
->length
);
266 db
->remote_rkey
= htonl(db
->remote_rkey
);
269 static void network_to_dest_block(RDMADestBlock
*db
)
271 db
->remote_host_addr
= ntohll(db
->remote_host_addr
);
272 db
->offset
= ntohll(db
->offset
);
273 db
->length
= ntohll(db
->length
);
274 db
->remote_rkey
= ntohl(db
->remote_rkey
);
278 * Virtual address of the above structures used for transmitting
279 * the RAMBlock descriptions at connection-time.
280 * This structure is *not* transmitted.
282 typedef struct RDMALocalBlocks
{
284 bool init
; /* main memory init complete */
285 RDMALocalBlock
*block
;
289 * Main data structure for RDMA state.
290 * While there is only one copy of this structure being allocated right now,
291 * this is the place where one would start if you wanted to consider
292 * having more than one RDMA connection open at the same time.
294 typedef struct RDMAContext
{
298 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
301 * This is used by *_exchange_send() to figure out whether or not
302 * the initial "READY" message has already been received or not.
303 * This is because other functions may potentially poll() and detect
304 * the READY message before send() does, in which case we need to
305 * know if it completed.
307 int control_ready_expected
;
309 /* number of outstanding writes */
312 /* store info about current buffer so that we can
313 merge it with future sends */
314 uint64_t current_addr
;
315 uint64_t current_length
;
316 /* index of ram block the current buffer belongs to */
318 /* index of the chunk in the current ram block */
324 * infiniband-specific variables for opening the device
325 * and maintaining connection state and so forth.
327 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
328 * cm_id->verbs, cm_id->channel, and cm_id->qp.
330 struct rdma_cm_id
*cm_id
; /* connection manager ID */
331 struct rdma_cm_id
*listen_id
;
334 struct ibv_context
*verbs
;
335 struct rdma_event_channel
*channel
;
336 struct ibv_qp
*qp
; /* queue pair */
337 struct ibv_comp_channel
*comp_channel
; /* completion channel */
338 struct ibv_pd
*pd
; /* protection domain */
339 struct ibv_cq
*cq
; /* completion queue */
342 * If a previous write failed (perhaps because of a failed
343 * memory registration, then do not attempt any future work
344 * and remember the error state.
350 * Description of ram blocks used throughout the code.
352 RDMALocalBlocks local_ram_blocks
;
353 RDMADestBlock
*dest_blocks
;
356 * Migration on *destination* started.
357 * Then use coroutine yield function.
358 * Source runs in a thread, so we don't care.
360 int migration_started_on_destination
;
362 int total_registrations
;
365 int unregister_current
, unregister_next
;
366 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
368 GHashTable
*blockmap
;
372 * Interface to the rest of the migration call stack.
374 typedef struct QEMUFileRDMA
{
381 * Main structure for IB Send/Recv control messages.
382 * This gets prepended at the beginning of every Send/Recv.
384 typedef struct QEMU_PACKED
{
385 uint32_t len
; /* Total length of data portion */
386 uint32_t type
; /* which control command to perform */
387 uint32_t repeat
; /* number of commands in data portion of same type */
391 static void control_to_network(RDMAControlHeader
*control
)
393 control
->type
= htonl(control
->type
);
394 control
->len
= htonl(control
->len
);
395 control
->repeat
= htonl(control
->repeat
);
398 static void network_to_control(RDMAControlHeader
*control
)
400 control
->type
= ntohl(control
->type
);
401 control
->len
= ntohl(control
->len
);
402 control
->repeat
= ntohl(control
->repeat
);
406 * Register a single Chunk.
407 * Information sent by the source VM to inform the dest
408 * to register an single chunk of memory before we can perform
409 * the actual RDMA operation.
411 typedef struct QEMU_PACKED
{
413 uint64_t current_addr
; /* offset into the ramblock of the chunk */
414 uint64_t chunk
; /* chunk to lookup if unregistering */
416 uint32_t current_index
; /* which ramblock the chunk belongs to */
418 uint64_t chunks
; /* how many sequential chunks to register */
421 static void register_to_network(RDMARegister
*reg
)
423 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
424 reg
->current_index
= htonl(reg
->current_index
);
425 reg
->chunks
= htonll(reg
->chunks
);
428 static void network_to_register(RDMARegister
*reg
)
430 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
431 reg
->current_index
= ntohl(reg
->current_index
);
432 reg
->chunks
= ntohll(reg
->chunks
);
435 typedef struct QEMU_PACKED
{
436 uint32_t value
; /* if zero, we will madvise() */
437 uint32_t block_idx
; /* which ram block index */
438 uint64_t offset
; /* where in the remote ramblock this chunk */
439 uint64_t length
; /* length of the chunk */
442 static void compress_to_network(RDMACompress
*comp
)
444 comp
->value
= htonl(comp
->value
);
445 comp
->block_idx
= htonl(comp
->block_idx
);
446 comp
->offset
= htonll(comp
->offset
);
447 comp
->length
= htonll(comp
->length
);
450 static void network_to_compress(RDMACompress
*comp
)
452 comp
->value
= ntohl(comp
->value
);
453 comp
->block_idx
= ntohl(comp
->block_idx
);
454 comp
->offset
= ntohll(comp
->offset
);
455 comp
->length
= ntohll(comp
->length
);
459 * The result of the dest's memory registration produces an "rkey"
460 * which the source VM must reference in order to perform
461 * the RDMA operation.
463 typedef struct QEMU_PACKED
{
467 } RDMARegisterResult
;
469 static void result_to_network(RDMARegisterResult
*result
)
471 result
->rkey
= htonl(result
->rkey
);
472 result
->host_addr
= htonll(result
->host_addr
);
475 static void network_to_result(RDMARegisterResult
*result
)
477 result
->rkey
= ntohl(result
->rkey
);
478 result
->host_addr
= ntohll(result
->host_addr
);
481 const char *print_wrid(int wrid
);
482 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
483 uint8_t *data
, RDMAControlHeader
*resp
,
485 int (*callback
)(RDMAContext
*rdma
));
487 static inline uint64_t ram_chunk_index(const uint8_t *start
,
490 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
493 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
496 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
497 (i
<< RDMA_REG_CHUNK_SHIFT
));
500 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
503 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
504 (1UL << RDMA_REG_CHUNK_SHIFT
);
506 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
507 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
513 static int rdma_add_block(RDMAContext
*rdma
, void *host_addr
,
514 ram_addr_t block_offset
, uint64_t length
)
516 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
517 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
518 (void *)(uintptr_t)block_offset
);
519 RDMALocalBlock
*old
= local
->block
;
521 assert(block
== NULL
);
523 local
->block
= g_malloc0(sizeof(RDMALocalBlock
) * (local
->nb_blocks
+ 1));
525 if (local
->nb_blocks
) {
528 for (x
= 0; x
< local
->nb_blocks
; x
++) {
529 g_hash_table_remove(rdma
->blockmap
,
530 (void *)(uintptr_t)old
[x
].offset
);
531 g_hash_table_insert(rdma
->blockmap
,
532 (void *)(uintptr_t)old
[x
].offset
,
535 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
539 block
= &local
->block
[local
->nb_blocks
];
541 block
->local_host_addr
= host_addr
;
542 block
->offset
= block_offset
;
543 block
->length
= length
;
544 block
->index
= local
->nb_blocks
;
545 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
546 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
547 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
548 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
549 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
550 block
->remote_keys
= g_malloc0(block
->nb_chunks
* sizeof(uint32_t));
552 block
->is_ram_block
= local
->init
? false : true;
554 g_hash_table_insert(rdma
->blockmap
, (void *) block_offset
, block
);
556 trace_rdma_add_block(local
->nb_blocks
, (uintptr_t) block
->local_host_addr
,
557 block
->offset
, block
->length
,
558 (uintptr_t) (block
->local_host_addr
+ block
->length
),
559 BITS_TO_LONGS(block
->nb_chunks
) *
560 sizeof(unsigned long) * 8,
569 * Memory regions need to be registered with the device and queue pairs setup
570 * in advanced before the migration starts. This tells us where the RAM blocks
571 * are so that we can register them individually.
573 static int qemu_rdma_init_one_block(const char *block_name
, void *host_addr
,
574 ram_addr_t block_offset
, ram_addr_t length
, void *opaque
)
576 return rdma_add_block(opaque
, host_addr
, block_offset
, length
);
580 * Identify the RAMBlocks and their quantity. They will be references to
581 * identify chunk boundaries inside each RAMBlock and also be referenced
582 * during dynamic page registration.
584 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
586 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
588 assert(rdma
->blockmap
== NULL
);
589 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
590 memset(local
, 0, sizeof *local
);
591 qemu_ram_foreach_block(qemu_rdma_init_one_block
, rdma
);
592 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
593 rdma
->dest_blocks
= (RDMADestBlock
*) g_malloc0(sizeof(RDMADestBlock
) *
594 rdma
->local_ram_blocks
.nb_blocks
);
599 static int rdma_delete_block(RDMAContext
*rdma
, ram_addr_t block_offset
)
601 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
602 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
603 (void *) block_offset
);
604 RDMALocalBlock
*old
= local
->block
;
612 for (j
= 0; j
< block
->nb_chunks
; j
++) {
613 if (!block
->pmr
[j
]) {
616 ibv_dereg_mr(block
->pmr
[j
]);
617 rdma
->total_registrations
--;
624 ibv_dereg_mr(block
->mr
);
625 rdma
->total_registrations
--;
629 g_free(block
->transit_bitmap
);
630 block
->transit_bitmap
= NULL
;
632 g_free(block
->unregister_bitmap
);
633 block
->unregister_bitmap
= NULL
;
635 g_free(block
->remote_keys
);
636 block
->remote_keys
= NULL
;
638 for (x
= 0; x
< local
->nb_blocks
; x
++) {
639 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)old
[x
].offset
);
642 if (local
->nb_blocks
> 1) {
644 local
->block
= g_malloc0(sizeof(RDMALocalBlock
) *
645 (local
->nb_blocks
- 1));
648 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
651 if (block
->index
< (local
->nb_blocks
- 1)) {
652 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
653 sizeof(RDMALocalBlock
) *
654 (local
->nb_blocks
- (block
->index
+ 1)));
657 assert(block
== local
->block
);
661 trace_rdma_delete_block(local
->nb_blocks
,
662 (uintptr_t)block
->local_host_addr
,
663 block
->offset
, block
->length
,
664 (uintptr_t)(block
->local_host_addr
+ block
->length
),
665 BITS_TO_LONGS(block
->nb_chunks
) *
666 sizeof(unsigned long) * 8, block
->nb_chunks
);
672 if (local
->nb_blocks
) {
673 for (x
= 0; x
< local
->nb_blocks
; x
++) {
674 g_hash_table_insert(rdma
->blockmap
,
675 (void *)(uintptr_t)local
->block
[x
].offset
,
684 * Put in the log file which RDMA device was opened and the details
685 * associated with that device.
687 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
689 struct ibv_port_attr port
;
691 if (ibv_query_port(verbs
, 1, &port
)) {
692 error_report("Failed to query port information");
696 printf("%s RDMA Device opened: kernel name %s "
697 "uverbs device name %s, "
698 "infiniband_verbs class device path %s, "
699 "infiniband class device path %s, "
700 "transport: (%d) %s\n",
703 verbs
->device
->dev_name
,
704 verbs
->device
->dev_path
,
705 verbs
->device
->ibdev_path
,
707 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
708 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
709 ? "Ethernet" : "Unknown"));
713 * Put in the log file the RDMA gid addressing information,
714 * useful for folks who have trouble understanding the
715 * RDMA device hierarchy in the kernel.
717 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
721 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
722 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
723 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
727 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
728 * We will try the next addrinfo struct, and fail if there are
729 * no other valid addresses to bind against.
731 * If user is listening on '[::]', then we will not have a opened a device
732 * yet and have no way of verifying if the device is RoCE or not.
734 * In this case, the source VM will throw an error for ALL types of
735 * connections (both IPv4 and IPv6) if the destination machine does not have
736 * a regular infiniband network available for use.
738 * The only way to guarantee that an error is thrown for broken kernels is
739 * for the management software to choose a *specific* interface at bind time
740 * and validate what time of hardware it is.
742 * Unfortunately, this puts the user in a fix:
744 * If the source VM connects with an IPv4 address without knowing that the
745 * destination has bound to '[::]' the migration will unconditionally fail
746 * unless the management software is explicitly listening on the the IPv4
747 * address while using a RoCE-based device.
749 * If the source VM connects with an IPv6 address, then we're OK because we can
750 * throw an error on the source (and similarly on the destination).
752 * But in mixed environments, this will be broken for a while until it is fixed
755 * We do provide a *tiny* bit of help in this function: We can list all of the
756 * devices in the system and check to see if all the devices are RoCE or
759 * If we detect that we have a *pure* RoCE environment, then we can safely
760 * thrown an error even if the management software has specified '[::]' as the
763 * However, if there is are multiple hetergeneous devices, then we cannot make
764 * this assumption and the user just has to be sure they know what they are
767 * Patches are being reviewed on linux-rdma.
769 static int qemu_rdma_broken_ipv6_kernel(Error
**errp
, struct ibv_context
*verbs
)
771 struct ibv_port_attr port_attr
;
773 /* This bug only exists in linux, to our knowledge. */
777 * Verbs are only NULL if management has bound to '[::]'.
779 * Let's iterate through all the devices and see if there any pure IB
780 * devices (non-ethernet).
782 * If not, then we can safely proceed with the migration.
783 * Otherwise, there are no guarantees until the bug is fixed in linux.
787 struct ibv_device
** dev_list
= ibv_get_device_list(&num_devices
);
788 bool roce_found
= false;
789 bool ib_found
= false;
791 for (x
= 0; x
< num_devices
; x
++) {
792 verbs
= ibv_open_device(dev_list
[x
]);
794 if (errno
== EPERM
) {
801 if (ibv_query_port(verbs
, 1, &port_attr
)) {
802 ibv_close_device(verbs
);
803 ERROR(errp
, "Could not query initial IB port");
807 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
809 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
813 ibv_close_device(verbs
);
819 fprintf(stderr
, "WARN: migrations may fail:"
820 " IPv6 over RoCE / iWARP in linux"
821 " is broken. But since you appear to have a"
822 " mixed RoCE / IB environment, be sure to only"
823 " migrate over the IB fabric until the kernel "
824 " fixes the bug.\n");
826 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
827 " and your management software has specified '[::]'"
828 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
837 * If we have a verbs context, that means that some other than '[::]' was
838 * used by the management software for binding. In which case we can
839 * actually warn the user about a potentially broken kernel.
842 /* IB ports start with 1, not 0 */
843 if (ibv_query_port(verbs
, 1, &port_attr
)) {
844 ERROR(errp
, "Could not query initial IB port");
848 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
849 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
850 "(but patches on linux-rdma in progress)");
860 * Figure out which RDMA device corresponds to the requested IP hostname
861 * Also create the initial connection manager identifiers for opening
864 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
867 struct rdma_addrinfo
*res
;
869 struct rdma_cm_event
*cm_event
;
870 char ip
[40] = "unknown";
871 struct rdma_addrinfo
*e
;
873 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
874 ERROR(errp
, "RDMA hostname has not been set");
878 /* create CM channel */
879 rdma
->channel
= rdma_create_event_channel();
880 if (!rdma
->channel
) {
881 ERROR(errp
, "could not create CM channel");
886 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
888 ERROR(errp
, "could not create channel id");
889 goto err_resolve_create_id
;
892 snprintf(port_str
, 16, "%d", rdma
->port
);
895 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
897 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
898 goto err_resolve_get_addr
;
901 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
902 inet_ntop(e
->ai_family
,
903 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
904 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
906 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
907 RDMA_RESOLVE_TIMEOUT_MS
);
909 if (e
->ai_family
== AF_INET6
) {
910 ret
= qemu_rdma_broken_ipv6_kernel(errp
, rdma
->cm_id
->verbs
);
919 ERROR(errp
, "could not resolve address %s", rdma
->host
);
920 goto err_resolve_get_addr
;
923 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
925 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
927 ERROR(errp
, "could not perform event_addr_resolved");
928 goto err_resolve_get_addr
;
931 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
932 ERROR(errp
, "result not equal to event_addr_resolved %s",
933 rdma_event_str(cm_event
->event
));
934 perror("rdma_resolve_addr");
935 rdma_ack_cm_event(cm_event
);
937 goto err_resolve_get_addr
;
939 rdma_ack_cm_event(cm_event
);
942 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
944 ERROR(errp
, "could not resolve rdma route");
945 goto err_resolve_get_addr
;
948 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
950 ERROR(errp
, "could not perform event_route_resolved");
951 goto err_resolve_get_addr
;
953 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
954 ERROR(errp
, "result not equal to event_route_resolved: %s",
955 rdma_event_str(cm_event
->event
));
956 rdma_ack_cm_event(cm_event
);
958 goto err_resolve_get_addr
;
960 rdma_ack_cm_event(cm_event
);
961 rdma
->verbs
= rdma
->cm_id
->verbs
;
962 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
963 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
966 err_resolve_get_addr
:
967 rdma_destroy_id(rdma
->cm_id
);
969 err_resolve_create_id
:
970 rdma_destroy_event_channel(rdma
->channel
);
971 rdma
->channel
= NULL
;
976 * Create protection domain and completion queues
978 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
981 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
983 error_report("failed to allocate protection domain");
987 /* create completion channel */
988 rdma
->comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
989 if (!rdma
->comp_channel
) {
990 error_report("failed to allocate completion channel");
991 goto err_alloc_pd_cq
;
995 * Completion queue can be filled by both read and write work requests,
996 * so must reflect the sum of both possible queue sizes.
998 rdma
->cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
999 NULL
, rdma
->comp_channel
, 0);
1001 error_report("failed to allocate completion queue");
1002 goto err_alloc_pd_cq
;
1009 ibv_dealloc_pd(rdma
->pd
);
1011 if (rdma
->comp_channel
) {
1012 ibv_destroy_comp_channel(rdma
->comp_channel
);
1015 rdma
->comp_channel
= NULL
;
1021 * Create queue pairs.
1023 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1025 struct ibv_qp_init_attr attr
= { 0 };
1028 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1029 attr
.cap
.max_recv_wr
= 3;
1030 attr
.cap
.max_send_sge
= 1;
1031 attr
.cap
.max_recv_sge
= 1;
1032 attr
.send_cq
= rdma
->cq
;
1033 attr
.recv_cq
= rdma
->cq
;
1034 attr
.qp_type
= IBV_QPT_RC
;
1036 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1041 rdma
->qp
= rdma
->cm_id
->qp
;
1045 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1048 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1050 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1051 local
->block
[i
].mr
=
1052 ibv_reg_mr(rdma
->pd
,
1053 local
->block
[i
].local_host_addr
,
1054 local
->block
[i
].length
,
1055 IBV_ACCESS_LOCAL_WRITE
|
1056 IBV_ACCESS_REMOTE_WRITE
1058 if (!local
->block
[i
].mr
) {
1059 perror("Failed to register local dest ram block!\n");
1062 rdma
->total_registrations
++;
1065 if (i
>= local
->nb_blocks
) {
1069 for (i
--; i
>= 0; i
--) {
1070 ibv_dereg_mr(local
->block
[i
].mr
);
1071 rdma
->total_registrations
--;
1079 * Find the ram block that corresponds to the page requested to be
1080 * transmitted by QEMU.
1082 * Once the block is found, also identify which 'chunk' within that
1083 * block that the page belongs to.
1085 * This search cannot fail or the migration will fail.
1087 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1088 uintptr_t block_offset
,
1091 uint64_t *block_index
,
1092 uint64_t *chunk_index
)
1094 uint64_t current_addr
= block_offset
+ offset
;
1095 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1096 (void *) block_offset
);
1098 assert(current_addr
>= block
->offset
);
1099 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1101 *block_index
= block
->index
;
1102 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1103 block
->local_host_addr
+ (current_addr
- block
->offset
));
1109 * Register a chunk with IB. If the chunk was already registered
1110 * previously, then skip.
1112 * Also return the keys associated with the registration needed
1113 * to perform the actual RDMA operation.
1115 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1116 RDMALocalBlock
*block
, uintptr_t host_addr
,
1117 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1118 uint8_t *chunk_start
, uint8_t *chunk_end
)
1122 *lkey
= block
->mr
->lkey
;
1125 *rkey
= block
->mr
->rkey
;
1130 /* allocate memory to store chunk MRs */
1132 block
->pmr
= g_malloc0(block
->nb_chunks
* sizeof(struct ibv_mr
*));
1136 * If 'rkey', then we're the destination, so grant access to the source.
1138 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1140 if (!block
->pmr
[chunk
]) {
1141 uint64_t len
= chunk_end
- chunk_start
;
1143 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1145 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
,
1147 (rkey
? (IBV_ACCESS_LOCAL_WRITE
|
1148 IBV_ACCESS_REMOTE_WRITE
) : 0));
1150 if (!block
->pmr
[chunk
]) {
1151 perror("Failed to register chunk!");
1152 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1153 " start %" PRIuPTR
" end %" PRIuPTR
1155 " local %" PRIuPTR
" registrations: %d\n",
1156 block
->index
, chunk
, (uintptr_t)chunk_start
,
1157 (uintptr_t)chunk_end
, host_addr
,
1158 (uintptr_t)block
->local_host_addr
,
1159 rdma
->total_registrations
);
1162 rdma
->total_registrations
++;
1166 *lkey
= block
->pmr
[chunk
]->lkey
;
1169 *rkey
= block
->pmr
[chunk
]->rkey
;
1175 * Register (at connection time) the memory used for control
1178 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1180 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1181 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1182 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1183 if (rdma
->wr_data
[idx
].control_mr
) {
1184 rdma
->total_registrations
++;
1187 error_report("qemu_rdma_reg_control failed");
1191 const char *print_wrid(int wrid
)
1193 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1194 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1196 return wrid_desc
[wrid
];
1200 * RDMA requires memory registration (mlock/pinning), but this is not good for
1203 * In preparation for the future where LRU information or workload-specific
1204 * writable writable working set memory access behavior is available to QEMU
1205 * it would be nice to have in place the ability to UN-register/UN-pin
1206 * particular memory regions from the RDMA hardware when it is determine that
1207 * those regions of memory will likely not be accessed again in the near future.
1209 * While we do not yet have such information right now, the following
1210 * compile-time option allows us to perform a non-optimized version of this
1213 * By uncommenting this option, you will cause *all* RDMA transfers to be
1214 * unregistered immediately after the transfer completes on both sides of the
1215 * connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
1217 * This will have a terrible impact on migration performance, so until future
1218 * workload information or LRU information is available, do not attempt to use
1219 * this feature except for basic testing.
1221 //#define RDMA_UNREGISTRATION_EXAMPLE
1224 * Perform a non-optimized memory unregistration after every transfer
1225 * for demonsration purposes, only if pin-all is not requested.
1227 * Potential optimizations:
1228 * 1. Start a new thread to run this function continuously
1230 - and for receipt of unregister messages
1232 * 3. Use workload hints.
1234 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1236 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1238 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1240 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1242 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1243 RDMALocalBlock
*block
=
1244 &(rdma
->local_ram_blocks
.block
[index
]);
1245 RDMARegister reg
= { .current_index
= index
};
1246 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1248 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1249 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1253 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1254 rdma
->unregister_current
);
1256 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1257 rdma
->unregister_current
++;
1259 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1260 rdma
->unregister_current
= 0;
1265 * Unregistration is speculative (because migration is single-threaded
1266 * and we cannot break the protocol's inifinband message ordering).
1267 * Thus, if the memory is currently being used for transmission,
1268 * then abort the attempt to unregister and try again
1269 * later the next time a completion is received for this memory.
1271 clear_bit(chunk
, block
->unregister_bitmap
);
1273 if (test_bit(chunk
, block
->transit_bitmap
)) {
1274 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1278 trace_qemu_rdma_unregister_waiting_send(chunk
);
1280 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1281 block
->pmr
[chunk
] = NULL
;
1282 block
->remote_keys
[chunk
] = 0;
1285 perror("unregistration chunk failed");
1288 rdma
->total_registrations
--;
1290 reg
.key
.chunk
= chunk
;
1291 register_to_network(®
);
1292 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1298 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1304 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1307 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1309 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1310 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1316 * Set bit for unregistration in the next iteration.
1317 * We cannot transmit right here, but will unpin later.
1319 static void qemu_rdma_signal_unregister(RDMAContext
*rdma
, uint64_t index
,
1320 uint64_t chunk
, uint64_t wr_id
)
1322 if (rdma
->unregistrations
[rdma
->unregister_next
] != 0) {
1323 error_report("rdma migration: queue is full");
1325 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1327 if (!test_and_set_bit(chunk
, block
->unregister_bitmap
)) {
1328 trace_qemu_rdma_signal_unregister_append(chunk
,
1329 rdma
->unregister_next
);
1331 rdma
->unregistrations
[rdma
->unregister_next
++] =
1332 qemu_rdma_make_wrid(wr_id
, index
, chunk
);
1334 if (rdma
->unregister_next
== RDMA_SIGNALED_SEND_MAX
) {
1335 rdma
->unregister_next
= 0;
1338 trace_qemu_rdma_signal_unregister_already(chunk
);
1344 * Consult the connection manager to see a work request
1345 * (of any kind) has completed.
1346 * Return the work request ID that completed.
1348 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, uint64_t *wr_id_out
,
1355 ret
= ibv_poll_cq(rdma
->cq
, 1, &wc
);
1358 *wr_id_out
= RDMA_WRID_NONE
;
1363 error_report("ibv_poll_cq return %d", ret
);
1367 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1369 if (wc
.status
!= IBV_WC_SUCCESS
) {
1370 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1371 wc
.status
, ibv_wc_status_str(wc
.status
));
1372 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1377 if (rdma
->control_ready_expected
&&
1378 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1379 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1380 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1381 rdma
->control_ready_expected
= 0;
1384 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1386 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1388 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1389 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1391 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1392 index
, chunk
, block
->local_host_addr
,
1393 (void *)(uintptr_t)block
->remote_host_addr
);
1395 clear_bit(chunk
, block
->transit_bitmap
);
1397 if (rdma
->nb_sent
> 0) {
1401 if (!rdma
->pin_all
) {
1403 * FYI: If one wanted to signal a specific chunk to be unregistered
1404 * using LRU or workload-specific information, this is the function
1405 * you would call to do so. That chunk would then get asynchronously
1406 * unregistered later.
1408 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1409 qemu_rdma_signal_unregister(rdma
, index
, chunk
, wc
.wr_id
);
1413 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1416 *wr_id_out
= wc
.wr_id
;
1418 *byte_len
= wc
.byte_len
;
1425 * Block until the next work request has completed.
1427 * First poll to see if a work request has already completed,
1430 * If we encounter completed work requests for IDs other than
1431 * the one we're interested in, then that's generally an error.
1433 * The only exception is actual RDMA Write completions. These
1434 * completions only need to be recorded, but do not actually
1435 * need further processing.
1437 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1440 int num_cq_events
= 0, ret
= 0;
1443 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1445 if (ibv_req_notify_cq(rdma
->cq
, 0)) {
1449 while (wr_id
!= wrid_requested
) {
1450 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1455 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1457 if (wr_id
== RDMA_WRID_NONE
) {
1460 if (wr_id
!= wrid_requested
) {
1461 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1462 wrid_requested
, print_wrid(wr_id
), wr_id
);
1466 if (wr_id
== wrid_requested
) {
1472 * Coroutine doesn't start until process_incoming_migration()
1473 * so don't yield unless we know we're running inside of a coroutine.
1475 if (rdma
->migration_started_on_destination
) {
1476 yield_until_fd_readable(rdma
->comp_channel
->fd
);
1479 if (ibv_get_cq_event(rdma
->comp_channel
, &cq
, &cq_ctx
)) {
1480 perror("ibv_get_cq_event");
1481 goto err_block_for_wrid
;
1486 if (ibv_req_notify_cq(cq
, 0)) {
1487 goto err_block_for_wrid
;
1490 while (wr_id
!= wrid_requested
) {
1491 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1493 goto err_block_for_wrid
;
1496 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1498 if (wr_id
== RDMA_WRID_NONE
) {
1501 if (wr_id
!= wrid_requested
) {
1502 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1503 wrid_requested
, print_wrid(wr_id
), wr_id
);
1507 if (wr_id
== wrid_requested
) {
1508 goto success_block_for_wrid
;
1512 success_block_for_wrid
:
1513 if (num_cq_events
) {
1514 ibv_ack_cq_events(cq
, num_cq_events
);
1519 if (num_cq_events
) {
1520 ibv_ack_cq_events(cq
, num_cq_events
);
1526 * Post a SEND message work request for the control channel
1527 * containing some data and block until the post completes.
1529 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1530 RDMAControlHeader
*head
)
1533 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1534 struct ibv_send_wr
*bad_wr
;
1535 struct ibv_sge sge
= {
1536 .addr
= (uintptr_t)(wr
->control
),
1537 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1538 .lkey
= wr
->control_mr
->lkey
,
1540 struct ibv_send_wr send_wr
= {
1541 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1542 .opcode
= IBV_WR_SEND
,
1543 .send_flags
= IBV_SEND_SIGNALED
,
1548 trace_qemu_rdma_post_send_control(control_desc
[head
->type
]);
1551 * We don't actually need to do a memcpy() in here if we used
1552 * the "sge" properly, but since we're only sending control messages
1553 * (not RAM in a performance-critical path), then its OK for now.
1555 * The copy makes the RDMAControlHeader simpler to manipulate
1556 * for the time being.
1558 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1559 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1560 control_to_network((void *) wr
->control
);
1563 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1567 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1570 error_report("Failed to use post IB SEND for control");
1574 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1576 error_report("rdma migration: send polling control error");
1583 * Post a RECV work request in anticipation of some future receipt
1584 * of data on the control channel.
1586 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1588 struct ibv_recv_wr
*bad_wr
;
1589 struct ibv_sge sge
= {
1590 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1591 .length
= RDMA_CONTROL_MAX_BUFFER
,
1592 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1595 struct ibv_recv_wr recv_wr
= {
1596 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1602 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1610 * Block and wait for a RECV control channel message to arrive.
1612 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1613 RDMAControlHeader
*head
, int expecting
, int idx
)
1616 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1620 error_report("rdma migration: recv polling control error!");
1624 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1625 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1627 trace_qemu_rdma_exchange_get_response_start(control_desc
[expecting
]);
1629 if (expecting
== RDMA_CONTROL_NONE
) {
1630 trace_qemu_rdma_exchange_get_response_none(control_desc
[head
->type
],
1632 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1633 error_report("Was expecting a %s (%d) control message"
1634 ", but got: %s (%d), length: %d",
1635 control_desc
[expecting
], expecting
,
1636 control_desc
[head
->type
], head
->type
, head
->len
);
1639 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1640 error_report("too long length: %d", head
->len
);
1643 if (sizeof(*head
) + head
->len
!= byte_len
) {
1644 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1652 * When a RECV work request has completed, the work request's
1653 * buffer is pointed at the header.
1655 * This will advance the pointer to the data portion
1656 * of the control message of the work request's buffer that
1657 * was populated after the work request finished.
1659 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1660 RDMAControlHeader
*head
)
1662 rdma
->wr_data
[idx
].control_len
= head
->len
;
1663 rdma
->wr_data
[idx
].control_curr
=
1664 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1668 * This is an 'atomic' high-level operation to deliver a single, unified
1669 * control-channel message.
1671 * Additionally, if the user is expecting some kind of reply to this message,
1672 * they can request a 'resp' response message be filled in by posting an
1673 * additional work request on behalf of the user and waiting for an additional
1676 * The extra (optional) response is used during registration to us from having
1677 * to perform an *additional* exchange of message just to provide a response by
1678 * instead piggy-backing on the acknowledgement.
1680 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1681 uint8_t *data
, RDMAControlHeader
*resp
,
1683 int (*callback
)(RDMAContext
*rdma
))
1688 * Wait until the dest is ready before attempting to deliver the message
1689 * by waiting for a READY message.
1691 if (rdma
->control_ready_expected
) {
1692 RDMAControlHeader resp
;
1693 ret
= qemu_rdma_exchange_get_response(rdma
,
1694 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1701 * If the user is expecting a response, post a WR in anticipation of it.
1704 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1706 error_report("rdma migration: error posting"
1707 " extra control recv for anticipated result!");
1713 * Post a WR to replace the one we just consumed for the READY message.
1715 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1717 error_report("rdma migration: error posting first control recv!");
1722 * Deliver the control message that was requested.
1724 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1727 error_report("Failed to send control buffer!");
1732 * If we're expecting a response, block and wait for it.
1736 trace_qemu_rdma_exchange_send_issue_callback();
1737 ret
= callback(rdma
);
1743 trace_qemu_rdma_exchange_send_waiting(control_desc
[resp
->type
]);
1744 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1745 resp
->type
, RDMA_WRID_DATA
);
1751 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1753 *resp_idx
= RDMA_WRID_DATA
;
1755 trace_qemu_rdma_exchange_send_received(control_desc
[resp
->type
]);
1758 rdma
->control_ready_expected
= 1;
1764 * This is an 'atomic' high-level operation to receive a single, unified
1765 * control-channel message.
1767 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1770 RDMAControlHeader ready
= {
1772 .type
= RDMA_CONTROL_READY
,
1778 * Inform the source that we're ready to receive a message.
1780 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1783 error_report("Failed to send control buffer!");
1788 * Block and wait for the message.
1790 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1791 expecting
, RDMA_WRID_READY
);
1797 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1800 * Post a new RECV work request to replace the one we just consumed.
1802 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1804 error_report("rdma migration: error posting second control recv!");
1812 * Write an actual chunk of memory using RDMA.
1814 * If we're using dynamic registration on the dest-side, we have to
1815 * send a registration command first.
1817 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
1818 int current_index
, uint64_t current_addr
,
1822 struct ibv_send_wr send_wr
= { 0 };
1823 struct ibv_send_wr
*bad_wr
;
1824 int reg_result_idx
, ret
, count
= 0;
1825 uint64_t chunk
, chunks
;
1826 uint8_t *chunk_start
, *chunk_end
;
1827 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
1829 RDMARegisterResult
*reg_result
;
1830 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
1831 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1832 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
1837 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
1838 (current_addr
- block
->offset
));
1839 sge
.length
= length
;
1841 chunk
= ram_chunk_index(block
->local_host_addr
,
1842 (uint8_t *)(uintptr_t)sge
.addr
);
1843 chunk_start
= ram_chunk_start(block
, chunk
);
1845 if (block
->is_ram_block
) {
1846 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
1848 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
1852 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
1854 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
1859 trace_qemu_rdma_write_one_top(chunks
+ 1,
1861 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
1863 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
1865 if (!rdma
->pin_all
) {
1866 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1867 qemu_rdma_unregister_waiting(rdma
);
1871 while (test_bit(chunk
, block
->transit_bitmap
)) {
1873 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
1874 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
1876 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
1879 error_report("Failed to Wait for previous write to complete "
1880 "block %d chunk %" PRIu64
1881 " current %" PRIu64
" len %" PRIu64
" %d",
1882 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
1887 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
1888 if (!block
->remote_keys
[chunk
]) {
1890 * This chunk has not yet been registered, so first check to see
1891 * if the entire chunk is zero. If so, tell the other size to
1892 * memset() + madvise() the entire chunk without RDMA.
1895 if (can_use_buffer_find_nonzero_offset((void *)(uintptr_t)sge
.addr
,
1897 && buffer_find_nonzero_offset((void *)(uintptr_t)sge
.addr
,
1898 length
) == length
) {
1899 RDMACompress comp
= {
1900 .offset
= current_addr
,
1902 .block_idx
= current_index
,
1906 head
.len
= sizeof(comp
);
1907 head
.type
= RDMA_CONTROL_COMPRESS
;
1909 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
1910 current_index
, current_addr
);
1912 compress_to_network(&comp
);
1913 ret
= qemu_rdma_exchange_send(rdma
, &head
,
1914 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
1920 acct_update_position(f
, sge
.length
, true);
1926 * Otherwise, tell other side to register.
1928 reg
.current_index
= current_index
;
1929 if (block
->is_ram_block
) {
1930 reg
.key
.current_addr
= current_addr
;
1932 reg
.key
.chunk
= chunk
;
1934 reg
.chunks
= chunks
;
1936 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
1939 register_to_network(®
);
1940 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1941 &resp
, ®_result_idx
, NULL
);
1946 /* try to overlap this single registration with the one we sent. */
1947 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
1948 &sge
.lkey
, NULL
, chunk
,
1949 chunk_start
, chunk_end
)) {
1950 error_report("cannot get lkey");
1954 reg_result
= (RDMARegisterResult
*)
1955 rdma
->wr_data
[reg_result_idx
].control_curr
;
1957 network_to_result(reg_result
);
1959 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
1960 reg_result
->rkey
, chunk
);
1962 block
->remote_keys
[chunk
] = reg_result
->rkey
;
1963 block
->remote_host_addr
= reg_result
->host_addr
;
1965 /* already registered before */
1966 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
1967 &sge
.lkey
, NULL
, chunk
,
1968 chunk_start
, chunk_end
)) {
1969 error_report("cannot get lkey!");
1974 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
1976 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
1978 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
1979 &sge
.lkey
, NULL
, chunk
,
1980 chunk_start
, chunk_end
)) {
1981 error_report("cannot get lkey!");
1987 * Encode the ram block index and chunk within this wrid.
1988 * We will use this information at the time of completion
1989 * to figure out which bitmap to check against and then which
1990 * chunk in the bitmap to look for.
1992 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
1993 current_index
, chunk
);
1995 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
1996 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
1997 send_wr
.sg_list
= &sge
;
1998 send_wr
.num_sge
= 1;
1999 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2000 (current_addr
- block
->offset
);
2002 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2006 * ibv_post_send() does not return negative error numbers,
2007 * per the specification they are positive - no idea why.
2009 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2011 if (ret
== ENOMEM
) {
2012 trace_qemu_rdma_write_one_queue_full();
2013 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2015 error_report("rdma migration: failed to make "
2016 "room in full send queue! %d", ret
);
2022 } else if (ret
> 0) {
2023 perror("rdma migration: post rdma write failed");
2027 set_bit(chunk
, block
->transit_bitmap
);
2028 acct_update_position(f
, sge
.length
, false);
2029 rdma
->total_writes
++;
2035 * Push out any unwritten RDMA operations.
2037 * We support sending out multiple chunks at the same time.
2038 * Not all of them need to get signaled in the completion queue.
2040 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2044 if (!rdma
->current_length
) {
2048 ret
= qemu_rdma_write_one(f
, rdma
,
2049 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2057 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2060 rdma
->current_length
= 0;
2061 rdma
->current_addr
= 0;
2066 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2067 uint64_t offset
, uint64_t len
)
2069 RDMALocalBlock
*block
;
2073 if (rdma
->current_index
< 0) {
2077 if (rdma
->current_chunk
< 0) {
2081 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2082 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2083 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2085 if (rdma
->current_length
== 0) {
2090 * Only merge into chunk sequentially.
2092 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2096 if (offset
< block
->offset
) {
2100 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2104 if ((host_addr
+ len
) > chunk_end
) {
2112 * We're not actually writing here, but doing three things:
2114 * 1. Identify the chunk the buffer belongs to.
2115 * 2. If the chunk is full or the buffer doesn't belong to the current
2116 * chunk, then start a new chunk and flush() the old chunk.
2117 * 3. To keep the hardware busy, we also group chunks into batches
2118 * and only require that a batch gets acknowledged in the completion
2119 * qeueue instead of each individual chunk.
2121 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2122 uint64_t block_offset
, uint64_t offset
,
2125 uint64_t current_addr
= block_offset
+ offset
;
2126 uint64_t index
= rdma
->current_index
;
2127 uint64_t chunk
= rdma
->current_chunk
;
2130 /* If we cannot merge it, we flush the current buffer first. */
2131 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2132 ret
= qemu_rdma_write_flush(f
, rdma
);
2136 rdma
->current_length
= 0;
2137 rdma
->current_addr
= current_addr
;
2139 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2140 offset
, len
, &index
, &chunk
);
2142 error_report("ram block search failed");
2145 rdma
->current_index
= index
;
2146 rdma
->current_chunk
= chunk
;
2150 rdma
->current_length
+= len
;
2152 /* flush it if buffer is too large */
2153 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2154 return qemu_rdma_write_flush(f
, rdma
);
2160 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2162 struct rdma_cm_event
*cm_event
;
2165 if (rdma
->cm_id
&& rdma
->connected
) {
2166 if (rdma
->error_state
) {
2167 RDMAControlHeader head
= { .len
= 0,
2168 .type
= RDMA_CONTROL_ERROR
,
2171 error_report("Early error. Sending error.");
2172 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2175 ret
= rdma_disconnect(rdma
->cm_id
);
2177 trace_qemu_rdma_cleanup_waiting_for_disconnect();
2178 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2180 rdma_ack_cm_event(cm_event
);
2183 trace_qemu_rdma_cleanup_disconnect();
2184 rdma
->connected
= false;
2187 g_free(rdma
->dest_blocks
);
2188 rdma
->dest_blocks
= NULL
;
2190 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2191 if (rdma
->wr_data
[idx
].control_mr
) {
2192 rdma
->total_registrations
--;
2193 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2195 rdma
->wr_data
[idx
].control_mr
= NULL
;
2198 if (rdma
->local_ram_blocks
.block
) {
2199 while (rdma
->local_ram_blocks
.nb_blocks
) {
2200 rdma_delete_block(rdma
, rdma
->local_ram_blocks
.block
->offset
);
2205 rdma_destroy_qp(rdma
->cm_id
);
2209 ibv_destroy_cq(rdma
->cq
);
2212 if (rdma
->comp_channel
) {
2213 ibv_destroy_comp_channel(rdma
->comp_channel
);
2214 rdma
->comp_channel
= NULL
;
2217 ibv_dealloc_pd(rdma
->pd
);
2221 rdma_destroy_id(rdma
->cm_id
);
2224 if (rdma
->listen_id
) {
2225 rdma_destroy_id(rdma
->listen_id
);
2226 rdma
->listen_id
= NULL
;
2228 if (rdma
->channel
) {
2229 rdma_destroy_event_channel(rdma
->channel
);
2230 rdma
->channel
= NULL
;
2237 static int qemu_rdma_source_init(RDMAContext
*rdma
, Error
**errp
, bool pin_all
)
2240 Error
*local_err
= NULL
, **temp
= &local_err
;
2243 * Will be validated against destination's actual capabilities
2244 * after the connect() completes.
2246 rdma
->pin_all
= pin_all
;
2248 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2250 goto err_rdma_source_init
;
2253 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2255 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2256 " limits may be too low. Please check $ ulimit -a # and "
2257 "search for 'ulimit -l' in the output");
2258 goto err_rdma_source_init
;
2261 ret
= qemu_rdma_alloc_qp(rdma
);
2263 ERROR(temp
, "rdma migration: error allocating qp!");
2264 goto err_rdma_source_init
;
2267 ret
= qemu_rdma_init_ram_blocks(rdma
);
2269 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2270 goto err_rdma_source_init
;
2273 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2274 ret
= qemu_rdma_reg_control(rdma
, idx
);
2276 ERROR(temp
, "rdma migration: error registering %d control!",
2278 goto err_rdma_source_init
;
2284 err_rdma_source_init
:
2285 error_propagate(errp
, local_err
);
2286 qemu_rdma_cleanup(rdma
);
2290 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
)
2292 RDMACapabilities cap
= {
2293 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2296 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2298 .private_data
= &cap
,
2299 .private_data_len
= sizeof(cap
),
2301 struct rdma_cm_event
*cm_event
;
2305 * Only negotiate the capability with destination if the user
2306 * on the source first requested the capability.
2308 if (rdma
->pin_all
) {
2309 trace_qemu_rdma_connect_pin_all_requested();
2310 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2313 caps_to_network(&cap
);
2315 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2317 perror("rdma_connect");
2318 ERROR(errp
, "connecting to destination!");
2319 goto err_rdma_source_connect
;
2322 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2324 perror("rdma_get_cm_event after rdma_connect");
2325 ERROR(errp
, "connecting to destination!");
2326 rdma_ack_cm_event(cm_event
);
2327 goto err_rdma_source_connect
;
2330 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2331 perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2332 ERROR(errp
, "connecting to destination!");
2333 rdma_ack_cm_event(cm_event
);
2334 goto err_rdma_source_connect
;
2336 rdma
->connected
= true;
2338 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2339 network_to_caps(&cap
);
2342 * Verify that the *requested* capabilities are supported by the destination
2343 * and disable them otherwise.
2345 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2346 ERROR(errp
, "Server cannot support pinning all memory. "
2347 "Will register memory dynamically.");
2348 rdma
->pin_all
= false;
2351 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2353 rdma_ack_cm_event(cm_event
);
2355 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2357 ERROR(errp
, "posting second control recv!");
2358 goto err_rdma_source_connect
;
2361 rdma
->control_ready_expected
= 1;
2365 err_rdma_source_connect
:
2366 qemu_rdma_cleanup(rdma
);
2370 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2373 struct rdma_cm_id
*listen_id
;
2374 char ip
[40] = "unknown";
2375 struct rdma_addrinfo
*res
, *e
;
2378 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2379 rdma
->wr_data
[idx
].control_len
= 0;
2380 rdma
->wr_data
[idx
].control_curr
= NULL
;
2383 if (!rdma
->host
|| !rdma
->host
[0]) {
2384 ERROR(errp
, "RDMA host is not set!");
2385 rdma
->error_state
= -EINVAL
;
2388 /* create CM channel */
2389 rdma
->channel
= rdma_create_event_channel();
2390 if (!rdma
->channel
) {
2391 ERROR(errp
, "could not create rdma event channel");
2392 rdma
->error_state
= -EINVAL
;
2397 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2399 ERROR(errp
, "could not create cm_id!");
2400 goto err_dest_init_create_listen_id
;
2403 snprintf(port_str
, 16, "%d", rdma
->port
);
2404 port_str
[15] = '\0';
2406 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2408 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2409 goto err_dest_init_bind_addr
;
2412 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2413 inet_ntop(e
->ai_family
,
2414 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2415 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2416 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2420 if (e
->ai_family
== AF_INET6
) {
2421 ret
= qemu_rdma_broken_ipv6_kernel(errp
, listen_id
->verbs
);
2430 ERROR(errp
, "Error: could not rdma_bind_addr!");
2431 goto err_dest_init_bind_addr
;
2434 rdma
->listen_id
= listen_id
;
2435 qemu_rdma_dump_gid("dest_init", listen_id
);
2438 err_dest_init_bind_addr
:
2439 rdma_destroy_id(listen_id
);
2440 err_dest_init_create_listen_id
:
2441 rdma_destroy_event_channel(rdma
->channel
);
2442 rdma
->channel
= NULL
;
2443 rdma
->error_state
= ret
;
2448 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2450 RDMAContext
*rdma
= NULL
;
2451 InetSocketAddress
*addr
;
2454 rdma
= g_malloc0(sizeof(RDMAContext
));
2455 rdma
->current_index
= -1;
2456 rdma
->current_chunk
= -1;
2458 addr
= inet_parse(host_port
, NULL
);
2460 rdma
->port
= atoi(addr
->port
);
2461 rdma
->host
= g_strdup(addr
->host
);
2463 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2468 qapi_free_InetSocketAddress(addr
);
2475 * QEMUFile interface to the control channel.
2476 * SEND messages for control only.
2477 * VM's ram is handled with regular RDMA messages.
2479 static int qemu_rdma_put_buffer(void *opaque
, const uint8_t *buf
,
2480 int64_t pos
, int size
)
2482 QEMUFileRDMA
*r
= opaque
;
2483 QEMUFile
*f
= r
->file
;
2484 RDMAContext
*rdma
= r
->rdma
;
2485 size_t remaining
= size
;
2486 uint8_t * data
= (void *) buf
;
2489 CHECK_ERROR_STATE();
2492 * Push out any writes that
2493 * we're queued up for VM's ram.
2495 ret
= qemu_rdma_write_flush(f
, rdma
);
2497 rdma
->error_state
= ret
;
2502 RDMAControlHeader head
;
2504 r
->len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2505 remaining
-= r
->len
;
2508 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2510 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2513 rdma
->error_state
= ret
;
2523 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2528 if (rdma
->wr_data
[idx
].control_len
) {
2529 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2531 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2532 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2533 rdma
->wr_data
[idx
].control_curr
+= len
;
2534 rdma
->wr_data
[idx
].control_len
-= len
;
2541 * QEMUFile interface to the control channel.
2542 * RDMA links don't use bytestreams, so we have to
2543 * return bytes to QEMUFile opportunistically.
2545 static int qemu_rdma_get_buffer(void *opaque
, uint8_t *buf
,
2546 int64_t pos
, int size
)
2548 QEMUFileRDMA
*r
= opaque
;
2549 RDMAContext
*rdma
= r
->rdma
;
2550 RDMAControlHeader head
;
2553 CHECK_ERROR_STATE();
2556 * First, we hold on to the last SEND message we
2557 * were given and dish out the bytes until we run
2560 r
->len
= qemu_rdma_fill(r
->rdma
, buf
, size
, 0);
2566 * Once we run out, we block and wait for another
2567 * SEND message to arrive.
2569 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2572 rdma
->error_state
= ret
;
2577 * SEND was received with new bytes, now try again.
2579 return qemu_rdma_fill(r
->rdma
, buf
, size
, 0);
2583 * Block until all the outstanding chunks have been delivered by the hardware.
2585 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2589 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2593 while (rdma
->nb_sent
) {
2594 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2596 error_report("rdma migration: complete polling error!");
2601 qemu_rdma_unregister_waiting(rdma
);
2606 static int qemu_rdma_close(void *opaque
)
2608 trace_qemu_rdma_close();
2609 QEMUFileRDMA
*r
= opaque
;
2611 qemu_rdma_cleanup(r
->rdma
);
2621 * This means that 'block_offset' is a full virtual address that does not
2622 * belong to a RAMBlock of the virtual machine and instead
2623 * represents a private malloc'd memory area that the caller wishes to
2627 * Offset is an offset to be added to block_offset and used
2628 * to also lookup the corresponding RAMBlock.
2631 * Initiate an transfer this size.
2634 * A 'hint' or 'advice' that means that we wish to speculatively
2635 * and asynchronously unregister this memory. In this case, there is no
2636 * guarantee that the unregister will actually happen, for example,
2637 * if the memory is being actively transmitted. Additionally, the memory
2638 * may be re-registered at any future time if a write within the same
2639 * chunk was requested again, even if you attempted to unregister it
2642 * @size < 0 : TODO, not yet supported
2643 * Unregister the memory NOW. This means that the caller does not
2644 * expect there to be any future RDMA transfers and we just want to clean
2645 * things up. This is used in case the upper layer owns the memory and
2646 * cannot wait for qemu_fclose() to occur.
2648 * @bytes_sent : User-specificed pointer to indicate how many bytes were
2649 * sent. Usually, this will not be more than a few bytes of
2650 * the protocol because most transfers are sent asynchronously.
2652 static size_t qemu_rdma_save_page(QEMUFile
*f
, void *opaque
,
2653 ram_addr_t block_offset
, ram_addr_t offset
,
2654 size_t size
, uint64_t *bytes_sent
)
2656 QEMUFileRDMA
*rfile
= opaque
;
2657 RDMAContext
*rdma
= rfile
->rdma
;
2660 CHECK_ERROR_STATE();
2666 * Add this page to the current 'chunk'. If the chunk
2667 * is full, or the page doen't belong to the current chunk,
2668 * an actual RDMA write will occur and a new chunk will be formed.
2670 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
2672 error_report("rdma migration: write error! %d", ret
);
2677 * We always return 1 bytes because the RDMA
2678 * protocol is completely asynchronous. We do not yet know
2679 * whether an identified chunk is zero or not because we're
2680 * waiting for other pages to potentially be merged with
2681 * the current chunk. So, we have to call qemu_update_position()
2682 * later on when the actual write occurs.
2688 uint64_t index
, chunk
;
2690 /* TODO: Change QEMUFileOps prototype to be signed: size_t => long
2692 ret = qemu_rdma_drain_cq(f, rdma);
2694 fprintf(stderr, "rdma: failed to synchronously drain"
2695 " completion queue before unregistration.\n");
2701 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2702 offset
, size
, &index
, &chunk
);
2705 error_report("ram block search failed");
2709 qemu_rdma_signal_unregister(rdma
, index
, chunk
, 0);
2712 * TODO: Synchronous, guaranteed unregistration (should not occur during
2713 * fast-path). Otherwise, unregisters will process on the next call to
2714 * qemu_rdma_drain_cq()
2716 qemu_rdma_unregister_waiting(rdma);
2722 * Drain the Completion Queue if possible, but do not block,
2725 * If nothing to poll, the end of the iteration will do this
2726 * again to make sure we don't overflow the request queue.
2729 uint64_t wr_id
, wr_id_in
;
2730 int ret
= qemu_rdma_poll(rdma
, &wr_id_in
, NULL
);
2732 error_report("rdma migration: polling error! %d", ret
);
2736 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
2738 if (wr_id
== RDMA_WRID_NONE
) {
2743 return RAM_SAVE_CONTROL_DELAYED
;
2745 rdma
->error_state
= ret
;
2749 static int qemu_rdma_accept(RDMAContext
*rdma
)
2751 RDMACapabilities cap
;
2752 struct rdma_conn_param conn_param
= {
2753 .responder_resources
= 2,
2754 .private_data
= &cap
,
2755 .private_data_len
= sizeof(cap
),
2757 struct rdma_cm_event
*cm_event
;
2758 struct ibv_context
*verbs
;
2762 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2764 goto err_rdma_dest_wait
;
2767 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
2768 rdma_ack_cm_event(cm_event
);
2769 goto err_rdma_dest_wait
;
2772 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2774 network_to_caps(&cap
);
2776 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
2777 error_report("Unknown source RDMA version: %d, bailing...",
2779 rdma_ack_cm_event(cm_event
);
2780 goto err_rdma_dest_wait
;
2784 * Respond with only the capabilities this version of QEMU knows about.
2786 cap
.flags
&= known_capabilities
;
2789 * Enable the ones that we do know about.
2790 * Add other checks here as new ones are introduced.
2792 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
2793 rdma
->pin_all
= true;
2796 rdma
->cm_id
= cm_event
->id
;
2797 verbs
= cm_event
->id
->verbs
;
2799 rdma_ack_cm_event(cm_event
);
2801 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
2803 caps_to_network(&cap
);
2805 trace_qemu_rdma_accept_pin_verbsc(verbs
);
2808 rdma
->verbs
= verbs
;
2809 } else if (rdma
->verbs
!= verbs
) {
2810 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
2812 goto err_rdma_dest_wait
;
2815 qemu_rdma_dump_id("dest_init", verbs
);
2817 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2819 error_report("rdma migration: error allocating pd and cq!");
2820 goto err_rdma_dest_wait
;
2823 ret
= qemu_rdma_alloc_qp(rdma
);
2825 error_report("rdma migration: error allocating qp!");
2826 goto err_rdma_dest_wait
;
2829 ret
= qemu_rdma_init_ram_blocks(rdma
);
2831 error_report("rdma migration: error initializing ram blocks!");
2832 goto err_rdma_dest_wait
;
2835 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2836 ret
= qemu_rdma_reg_control(rdma
, idx
);
2838 error_report("rdma: error registering %d control", idx
);
2839 goto err_rdma_dest_wait
;
2843 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2845 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
2847 error_report("rdma_accept returns %d", ret
);
2848 goto err_rdma_dest_wait
;
2851 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2853 error_report("rdma_accept get_cm_event failed %d", ret
);
2854 goto err_rdma_dest_wait
;
2857 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2858 error_report("rdma_accept not event established");
2859 rdma_ack_cm_event(cm_event
);
2860 goto err_rdma_dest_wait
;
2863 rdma_ack_cm_event(cm_event
);
2864 rdma
->connected
= true;
2866 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2868 error_report("rdma migration: error posting second control recv");
2869 goto err_rdma_dest_wait
;
2872 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
2877 rdma
->error_state
= ret
;
2878 qemu_rdma_cleanup(rdma
);
2883 * During each iteration of the migration, we listen for instructions
2884 * by the source VM to perform dynamic page registrations before they
2885 * can perform RDMA operations.
2887 * We respond with the 'rkey'.
2889 * Keep doing this until the source tells us to stop.
2891 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
,
2894 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
2895 .type
= RDMA_CONTROL_REGISTER_RESULT
,
2898 RDMAControlHeader unreg_resp
= { .len
= 0,
2899 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
2902 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
2904 QEMUFileRDMA
*rfile
= opaque
;
2905 RDMAContext
*rdma
= rfile
->rdma
;
2906 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
2907 RDMAControlHeader head
;
2908 RDMARegister
*reg
, *registers
;
2910 RDMARegisterResult
*reg_result
;
2911 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
2912 RDMALocalBlock
*block
;
2919 CHECK_ERROR_STATE();
2922 trace_qemu_rdma_registration_handle_wait(flags
);
2924 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
2930 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
2931 error_report("rdma: Too many requests in this message (%d)."
2932 "Bailing.", head
.repeat
);
2937 switch (head
.type
) {
2938 case RDMA_CONTROL_COMPRESS
:
2939 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
2940 network_to_compress(comp
);
2942 trace_qemu_rdma_registration_handle_compress(comp
->length
,
2945 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
2947 host_addr
= block
->local_host_addr
+
2948 (comp
->offset
- block
->offset
);
2950 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
2953 case RDMA_CONTROL_REGISTER_FINISHED
:
2954 trace_qemu_rdma_registration_handle_finished();
2957 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
2958 trace_qemu_rdma_registration_handle_ram_blocks();
2960 if (rdma
->pin_all
) {
2961 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
2963 error_report("rdma migration: error dest "
2964 "registering ram blocks");
2970 * Dest uses this to prepare to transmit the RAMBlock descriptions
2971 * to the source VM after connection setup.
2972 * Both sides use the "remote" structure to communicate and update
2973 * their "local" descriptions with what was sent.
2975 for (i
= 0; i
< local
->nb_blocks
; i
++) {
2976 rdma
->dest_blocks
[i
].remote_host_addr
=
2977 (uintptr_t)(local
->block
[i
].local_host_addr
);
2979 if (rdma
->pin_all
) {
2980 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
2983 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
2984 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
2986 dest_block_to_network(&rdma
->dest_blocks
[i
]);
2989 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
2990 * sizeof(RDMADestBlock
);
2993 ret
= qemu_rdma_post_send_control(rdma
,
2994 (uint8_t *) rdma
->dest_blocks
, &blocks
);
2997 error_report("rdma migration: error sending remote info");
3002 case RDMA_CONTROL_REGISTER_REQUEST
:
3003 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3005 reg_resp
.repeat
= head
.repeat
;
3006 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3008 for (count
= 0; count
< head
.repeat
; count
++) {
3010 uint8_t *chunk_start
, *chunk_end
;
3012 reg
= ®isters
[count
];
3013 network_to_register(reg
);
3015 reg_result
= &results
[count
];
3017 trace_qemu_rdma_registration_handle_register_loop(count
,
3018 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3020 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3021 if (block
->is_ram_block
) {
3022 host_addr
= (block
->local_host_addr
+
3023 (reg
->key
.current_addr
- block
->offset
));
3024 chunk
= ram_chunk_index(block
->local_host_addr
,
3025 (uint8_t *) host_addr
);
3027 chunk
= reg
->key
.chunk
;
3028 host_addr
= block
->local_host_addr
+
3029 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3031 chunk_start
= ram_chunk_start(block
, chunk
);
3032 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3033 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3034 (uintptr_t)host_addr
, NULL
, ®_result
->rkey
,
3035 chunk
, chunk_start
, chunk_end
)) {
3036 error_report("cannot get rkey");
3041 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3043 trace_qemu_rdma_registration_handle_register_rkey(
3046 result_to_network(reg_result
);
3049 ret
= qemu_rdma_post_send_control(rdma
,
3050 (uint8_t *) results
, ®_resp
);
3053 error_report("Failed to send control buffer");
3057 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3058 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3059 unreg_resp
.repeat
= head
.repeat
;
3060 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3062 for (count
= 0; count
< head
.repeat
; count
++) {
3063 reg
= ®isters
[count
];
3064 network_to_register(reg
);
3066 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3067 reg
->current_index
, reg
->key
.chunk
);
3069 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3071 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3072 block
->pmr
[reg
->key
.chunk
] = NULL
;
3075 perror("rdma unregistration chunk failed");
3080 rdma
->total_registrations
--;
3082 trace_qemu_rdma_registration_handle_unregister_success(
3086 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3089 error_report("Failed to send control buffer");
3093 case RDMA_CONTROL_REGISTER_RESULT
:
3094 error_report("Invalid RESULT message at dest.");
3098 error_report("Unknown control message %s", control_desc
[head
.type
]);
3105 rdma
->error_state
= ret
;
3110 static int qemu_rdma_registration_start(QEMUFile
*f
, void *opaque
,
3113 QEMUFileRDMA
*rfile
= opaque
;
3114 RDMAContext
*rdma
= rfile
->rdma
;
3116 CHECK_ERROR_STATE();
3118 trace_qemu_rdma_registration_start(flags
);
3119 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3126 * Inform dest that dynamic registrations are done for now.
3127 * First, flush writes, if any.
3129 static int qemu_rdma_registration_stop(QEMUFile
*f
, void *opaque
,
3132 Error
*local_err
= NULL
, **errp
= &local_err
;
3133 QEMUFileRDMA
*rfile
= opaque
;
3134 RDMAContext
*rdma
= rfile
->rdma
;
3135 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3138 CHECK_ERROR_STATE();
3141 ret
= qemu_rdma_drain_cq(f
, rdma
);
3147 if (flags
== RAM_CONTROL_SETUP
) {
3148 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3149 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3150 int reg_result_idx
, i
, j
, nb_dest_blocks
;
3152 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3153 trace_qemu_rdma_registration_stop_ram();
3156 * Make sure that we parallelize the pinning on both sides.
3157 * For very large guests, doing this serially takes a really
3158 * long time, so we have to 'interleave' the pinning locally
3159 * with the control messages by performing the pinning on this
3160 * side before we receive the control response from the other
3161 * side that the pinning has completed.
3163 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3164 ®_result_idx
, rdma
->pin_all
?
3165 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3167 ERROR(errp
, "receiving remote info!");
3171 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3174 * The protocol uses two different sets of rkeys (mutually exclusive):
3175 * 1. One key to represent the virtual address of the entire ram block.
3176 * (dynamic chunk registration disabled - pin everything with one rkey.)
3177 * 2. One to represent individual chunks within a ram block.
3178 * (dynamic chunk registration enabled - pin individual chunks.)
3180 * Once the capability is successfully negotiated, the destination transmits
3181 * the keys to use (or sends them later) including the virtual addresses
3182 * and then propagates the remote ram block descriptions to his local copy.
3185 if (local
->nb_blocks
!= nb_dest_blocks
) {
3186 ERROR(errp
, "ram blocks mismatch #1! "
3187 "Your QEMU command line parameters are probably "
3188 "not identical on both the source and destination.");
3192 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3193 memcpy(rdma
->dest_blocks
,
3194 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3195 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3196 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3198 /* search local ram blocks */
3199 for (j
= 0; j
< local
->nb_blocks
; j
++) {
3200 if (rdma
->dest_blocks
[i
].offset
!= local
->block
[j
].offset
) {
3204 if (rdma
->dest_blocks
[i
].length
!= local
->block
[j
].length
) {
3205 ERROR(errp
, "ram blocks mismatch #2! "
3206 "Your QEMU command line parameters are probably "
3207 "not identical on both the source and destination.");
3210 local
->block
[j
].remote_host_addr
=
3211 rdma
->dest_blocks
[i
].remote_host_addr
;
3212 local
->block
[j
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3216 if (j
>= local
->nb_blocks
) {
3217 ERROR(errp
, "ram blocks mismatch #3! "
3218 "Your QEMU command line parameters are probably "
3219 "not identical on both the source and destination.");
3225 trace_qemu_rdma_registration_stop(flags
);
3227 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3228 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3236 rdma
->error_state
= ret
;
3240 static int qemu_rdma_get_fd(void *opaque
)
3242 QEMUFileRDMA
*rfile
= opaque
;
3243 RDMAContext
*rdma
= rfile
->rdma
;
3245 return rdma
->comp_channel
->fd
;
3248 static const QEMUFileOps rdma_read_ops
= {
3249 .get_buffer
= qemu_rdma_get_buffer
,
3250 .get_fd
= qemu_rdma_get_fd
,
3251 .close
= qemu_rdma_close
,
3252 .hook_ram_load
= qemu_rdma_registration_handle
,
3255 static const QEMUFileOps rdma_write_ops
= {
3256 .put_buffer
= qemu_rdma_put_buffer
,
3257 .close
= qemu_rdma_close
,
3258 .before_ram_iterate
= qemu_rdma_registration_start
,
3259 .after_ram_iterate
= qemu_rdma_registration_stop
,
3260 .save_page
= qemu_rdma_save_page
,
3263 static void *qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
3265 QEMUFileRDMA
*r
= g_malloc0(sizeof(QEMUFileRDMA
));
3267 if (qemu_file_mode_is_not_valid(mode
)) {
3273 if (mode
[0] == 'w') {
3274 r
->file
= qemu_fopen_ops(r
, &rdma_write_ops
);
3276 r
->file
= qemu_fopen_ops(r
, &rdma_read_ops
);
3282 static void rdma_accept_incoming_migration(void *opaque
)
3284 RDMAContext
*rdma
= opaque
;
3287 Error
*local_err
= NULL
, **errp
= &local_err
;
3289 trace_qemu_dma_accept_incoming_migration();
3290 ret
= qemu_rdma_accept(rdma
);
3293 ERROR(errp
, "RDMA Migration initialization failed!");
3297 trace_qemu_dma_accept_incoming_migration_accepted();
3299 f
= qemu_fopen_rdma(rdma
, "rb");
3301 ERROR(errp
, "could not qemu_fopen_rdma!");
3302 qemu_rdma_cleanup(rdma
);
3306 rdma
->migration_started_on_destination
= 1;
3307 process_incoming_migration(f
);
3310 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
3314 Error
*local_err
= NULL
;
3316 trace_rdma_start_incoming_migration();
3317 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
3323 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
3329 trace_rdma_start_incoming_migration_after_dest_init();
3331 ret
= rdma_listen(rdma
->listen_id
, 5);
3334 ERROR(errp
, "listening on socket!");
3338 trace_rdma_start_incoming_migration_after_rdma_listen();
3340 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3341 NULL
, (void *)(intptr_t)rdma
);
3344 error_propagate(errp
, local_err
);
3348 void rdma_start_outgoing_migration(void *opaque
,
3349 const char *host_port
, Error
**errp
)
3351 MigrationState
*s
= opaque
;
3352 Error
*local_err
= NULL
, **temp
= &local_err
;
3353 RDMAContext
*rdma
= qemu_rdma_data_init(host_port
, &local_err
);
3357 ERROR(temp
, "Failed to initialize RDMA data structures! %d", ret
);
3361 ret
= qemu_rdma_source_init(rdma
, &local_err
,
3362 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
]);
3368 trace_rdma_start_outgoing_migration_after_rdma_source_init();
3369 ret
= qemu_rdma_connect(rdma
, &local_err
);
3375 trace_rdma_start_outgoing_migration_after_rdma_connect();
3377 s
->file
= qemu_fopen_rdma(rdma
, "wb");
3378 migrate_fd_connect(s
);
3381 error_propagate(errp
, local_err
);
3383 migrate_fd_error(s
);