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 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 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 RDMARemoteBlock
{
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 remote_block_to_network(RDMARemoteBlock
*rb
)
263 rb
->remote_host_addr
= htonll(rb
->remote_host_addr
);
264 rb
->offset
= htonll(rb
->offset
);
265 rb
->length
= htonll(rb
->length
);
266 rb
->remote_rkey
= htonl(rb
->remote_rkey
);
269 static void network_to_remote_block(RDMARemoteBlock
*rb
)
271 rb
->remote_host_addr
= ntohll(rb
->remote_host_addr
);
272 rb
->offset
= ntohll(rb
->offset
);
273 rb
->length
= ntohll(rb
->length
);
274 rb
->remote_rkey
= ntohl(rb
->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 RDMARemoteBlock
*block
;
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 __qemu_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 *) 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
, (void *)old
[x
].offset
);
530 g_hash_table_insert(rdma
->blockmap
, (void *)old
[x
].offset
,
533 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
537 block
= &local
->block
[local
->nb_blocks
];
539 block
->local_host_addr
= host_addr
;
540 block
->offset
= block_offset
;
541 block
->length
= length
;
542 block
->index
= local
->nb_blocks
;
543 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
544 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
545 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
546 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
547 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
548 block
->remote_keys
= g_malloc0(block
->nb_chunks
* sizeof(uint32_t));
550 block
->is_ram_block
= local
->init
? false : true;
552 g_hash_table_insert(rdma
->blockmap
, (void *) block_offset
, block
);
554 trace___qemu_rdma_add_block(local
->nb_blocks
,
555 (uint64_t) block
->local_host_addr
, block
->offset
,
557 (uint64_t) (block
->local_host_addr
+ block
->length
),
558 BITS_TO_LONGS(block
->nb_chunks
) *
559 sizeof(unsigned long) * 8,
568 * Memory regions need to be registered with the device and queue pairs setup
569 * in advanced before the migration starts. This tells us where the RAM blocks
570 * are so that we can register them individually.
572 static void qemu_rdma_init_one_block(void *host_addr
,
573 ram_addr_t block_offset
, ram_addr_t length
, void *opaque
)
575 __qemu_rdma_add_block(opaque
, host_addr
, block_offset
, length
);
579 * Identify the RAMBlocks and their quantity. They will be references to
580 * identify chunk boundaries inside each RAMBlock and also be referenced
581 * during dynamic page registration.
583 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
585 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
587 assert(rdma
->blockmap
== NULL
);
588 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
589 memset(local
, 0, sizeof *local
);
590 qemu_ram_foreach_block(qemu_rdma_init_one_block
, rdma
);
591 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
592 rdma
->block
= (RDMARemoteBlock
*) g_malloc0(sizeof(RDMARemoteBlock
) *
593 rdma
->local_ram_blocks
.nb_blocks
);
598 static int __qemu_rdma_delete_block(RDMAContext
*rdma
, ram_addr_t block_offset
)
600 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
601 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
602 (void *) block_offset
);
603 RDMALocalBlock
*old
= local
->block
;
611 for (j
= 0; j
< block
->nb_chunks
; j
++) {
612 if (!block
->pmr
[j
]) {
615 ibv_dereg_mr(block
->pmr
[j
]);
616 rdma
->total_registrations
--;
623 ibv_dereg_mr(block
->mr
);
624 rdma
->total_registrations
--;
628 g_free(block
->transit_bitmap
);
629 block
->transit_bitmap
= NULL
;
631 g_free(block
->unregister_bitmap
);
632 block
->unregister_bitmap
= NULL
;
634 g_free(block
->remote_keys
);
635 block
->remote_keys
= NULL
;
637 for (x
= 0; x
< local
->nb_blocks
; x
++) {
638 g_hash_table_remove(rdma
->blockmap
, (void *)old
[x
].offset
);
641 if (local
->nb_blocks
> 1) {
643 local
->block
= g_malloc0(sizeof(RDMALocalBlock
) *
644 (local
->nb_blocks
- 1));
647 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
650 if (block
->index
< (local
->nb_blocks
- 1)) {
651 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
652 sizeof(RDMALocalBlock
) *
653 (local
->nb_blocks
- (block
->index
+ 1)));
656 assert(block
== local
->block
);
660 trace___qemu_rdma_delete_block(local
->nb_blocks
,
661 (uint64_t)block
->local_host_addr
,
662 block
->offset
, block
->length
,
663 (uint64_t)(block
->local_host_addr
+ block
->length
),
664 BITS_TO_LONGS(block
->nb_chunks
) *
665 sizeof(unsigned long) * 8, block
->nb_chunks
);
671 if (local
->nb_blocks
) {
672 for (x
= 0; x
< local
->nb_blocks
; x
++) {
673 g_hash_table_insert(rdma
->blockmap
, (void *)local
->block
[x
].offset
,
682 * Put in the log file which RDMA device was opened and the details
683 * associated with that device.
685 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
687 struct ibv_port_attr port
;
689 if (ibv_query_port(verbs
, 1, &port
)) {
690 error_report("Failed to query port information");
694 printf("%s RDMA Device opened: kernel name %s "
695 "uverbs device name %s, "
696 "infiniband_verbs class device path %s, "
697 "infiniband class device path %s, "
698 "transport: (%d) %s\n",
701 verbs
->device
->dev_name
,
702 verbs
->device
->dev_path
,
703 verbs
->device
->ibdev_path
,
705 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
706 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
707 ? "Ethernet" : "Unknown"));
711 * Put in the log file the RDMA gid addressing information,
712 * useful for folks who have trouble understanding the
713 * RDMA device hierarchy in the kernel.
715 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
719 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
720 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
721 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
725 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
726 * We will try the next addrinfo struct, and fail if there are
727 * no other valid addresses to bind against.
729 * If user is listening on '[::]', then we will not have a opened a device
730 * yet and have no way of verifying if the device is RoCE or not.
732 * In this case, the source VM will throw an error for ALL types of
733 * connections (both IPv4 and IPv6) if the destination machine does not have
734 * a regular infiniband network available for use.
736 * The only way to guarantee that an error is thrown for broken kernels is
737 * for the management software to choose a *specific* interface at bind time
738 * and validate what time of hardware it is.
740 * Unfortunately, this puts the user in a fix:
742 * If the source VM connects with an IPv4 address without knowing that the
743 * destination has bound to '[::]' the migration will unconditionally fail
744 * unless the management software is explicitly listening on the the IPv4
745 * address while using a RoCE-based device.
747 * If the source VM connects with an IPv6 address, then we're OK because we can
748 * throw an error on the source (and similarly on the destination).
750 * But in mixed environments, this will be broken for a while until it is fixed
753 * We do provide a *tiny* bit of help in this function: We can list all of the
754 * devices in the system and check to see if all the devices are RoCE or
757 * If we detect that we have a *pure* RoCE environment, then we can safely
758 * thrown an error even if the management software has specified '[::]' as the
761 * However, if there is are multiple hetergeneous devices, then we cannot make
762 * this assumption and the user just has to be sure they know what they are
765 * Patches are being reviewed on linux-rdma.
767 static int qemu_rdma_broken_ipv6_kernel(Error
**errp
, struct ibv_context
*verbs
)
769 struct ibv_port_attr port_attr
;
771 /* This bug only exists in linux, to our knowledge. */
775 * Verbs are only NULL if management has bound to '[::]'.
777 * Let's iterate through all the devices and see if there any pure IB
778 * devices (non-ethernet).
780 * If not, then we can safely proceed with the migration.
781 * Otherwise, there are no guarantees until the bug is fixed in linux.
785 struct ibv_device
** dev_list
= ibv_get_device_list(&num_devices
);
786 bool roce_found
= false;
787 bool ib_found
= false;
789 for (x
= 0; x
< num_devices
; x
++) {
790 verbs
= ibv_open_device(dev_list
[x
]);
792 if (ibv_query_port(verbs
, 1, &port_attr
)) {
793 ibv_close_device(verbs
);
794 ERROR(errp
, "Could not query initial IB port");
798 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
800 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
804 ibv_close_device(verbs
);
810 fprintf(stderr
, "WARN: migrations may fail:"
811 " IPv6 over RoCE / iWARP in linux"
812 " is broken. But since you appear to have a"
813 " mixed RoCE / IB environment, be sure to only"
814 " migrate over the IB fabric until the kernel "
815 " fixes the bug.\n");
817 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
818 " and your management software has specified '[::]'"
819 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
828 * If we have a verbs context, that means that some other than '[::]' was
829 * used by the management software for binding. In which case we can actually
830 * warn the user about a potential broken kernel;
833 /* IB ports start with 1, not 0 */
834 if (ibv_query_port(verbs
, 1, &port_attr
)) {
835 ERROR(errp
, "Could not query initial IB port");
839 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
840 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
841 "(but patches on linux-rdma in progress)");
851 * Figure out which RDMA device corresponds to the requested IP hostname
852 * Also create the initial connection manager identifiers for opening
855 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
858 struct rdma_addrinfo
*res
;
860 struct rdma_cm_event
*cm_event
;
861 char ip
[40] = "unknown";
862 struct rdma_addrinfo
*e
;
864 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
865 ERROR(errp
, "RDMA hostname has not been set");
869 /* create CM channel */
870 rdma
->channel
= rdma_create_event_channel();
871 if (!rdma
->channel
) {
872 ERROR(errp
, "could not create CM channel");
877 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
879 ERROR(errp
, "could not create channel id");
880 goto err_resolve_create_id
;
883 snprintf(port_str
, 16, "%d", rdma
->port
);
886 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
888 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
889 goto err_resolve_get_addr
;
892 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
893 inet_ntop(e
->ai_family
,
894 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
895 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
897 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
898 RDMA_RESOLVE_TIMEOUT_MS
);
900 if (e
->ai_family
== AF_INET6
) {
901 ret
= qemu_rdma_broken_ipv6_kernel(errp
, rdma
->cm_id
->verbs
);
910 ERROR(errp
, "could not resolve address %s", rdma
->host
);
911 goto err_resolve_get_addr
;
914 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
916 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
918 ERROR(errp
, "could not perform event_addr_resolved");
919 goto err_resolve_get_addr
;
922 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
923 ERROR(errp
, "result not equal to event_addr_resolved %s",
924 rdma_event_str(cm_event
->event
));
925 perror("rdma_resolve_addr");
926 rdma_ack_cm_event(cm_event
);
928 goto err_resolve_get_addr
;
930 rdma_ack_cm_event(cm_event
);
933 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
935 ERROR(errp
, "could not resolve rdma route");
936 goto err_resolve_get_addr
;
939 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
941 ERROR(errp
, "could not perform event_route_resolved");
942 goto err_resolve_get_addr
;
944 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
945 ERROR(errp
, "result not equal to event_route_resolved: %s",
946 rdma_event_str(cm_event
->event
));
947 rdma_ack_cm_event(cm_event
);
949 goto err_resolve_get_addr
;
951 rdma_ack_cm_event(cm_event
);
952 rdma
->verbs
= rdma
->cm_id
->verbs
;
953 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
954 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
957 err_resolve_get_addr
:
958 rdma_destroy_id(rdma
->cm_id
);
960 err_resolve_create_id
:
961 rdma_destroy_event_channel(rdma
->channel
);
962 rdma
->channel
= NULL
;
967 * Create protection domain and completion queues
969 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
972 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
974 error_report("failed to allocate protection domain");
978 /* create completion channel */
979 rdma
->comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
980 if (!rdma
->comp_channel
) {
981 error_report("failed to allocate completion channel");
982 goto err_alloc_pd_cq
;
986 * Completion queue can be filled by both read and write work requests,
987 * so must reflect the sum of both possible queue sizes.
989 rdma
->cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
990 NULL
, rdma
->comp_channel
, 0);
992 error_report("failed to allocate completion queue");
993 goto err_alloc_pd_cq
;
1000 ibv_dealloc_pd(rdma
->pd
);
1002 if (rdma
->comp_channel
) {
1003 ibv_destroy_comp_channel(rdma
->comp_channel
);
1006 rdma
->comp_channel
= NULL
;
1012 * Create queue pairs.
1014 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1016 struct ibv_qp_init_attr attr
= { 0 };
1019 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1020 attr
.cap
.max_recv_wr
= 3;
1021 attr
.cap
.max_send_sge
= 1;
1022 attr
.cap
.max_recv_sge
= 1;
1023 attr
.send_cq
= rdma
->cq
;
1024 attr
.recv_cq
= rdma
->cq
;
1025 attr
.qp_type
= IBV_QPT_RC
;
1027 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1032 rdma
->qp
= rdma
->cm_id
->qp
;
1036 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1039 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1041 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1042 local
->block
[i
].mr
=
1043 ibv_reg_mr(rdma
->pd
,
1044 local
->block
[i
].local_host_addr
,
1045 local
->block
[i
].length
,
1046 IBV_ACCESS_LOCAL_WRITE
|
1047 IBV_ACCESS_REMOTE_WRITE
1049 if (!local
->block
[i
].mr
) {
1050 perror("Failed to register local dest ram block!\n");
1053 rdma
->total_registrations
++;
1056 if (i
>= local
->nb_blocks
) {
1060 for (i
--; i
>= 0; i
--) {
1061 ibv_dereg_mr(local
->block
[i
].mr
);
1062 rdma
->total_registrations
--;
1070 * Find the ram block that corresponds to the page requested to be
1071 * transmitted by QEMU.
1073 * Once the block is found, also identify which 'chunk' within that
1074 * block that the page belongs to.
1076 * This search cannot fail or the migration will fail.
1078 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1079 uint64_t block_offset
,
1082 uint64_t *block_index
,
1083 uint64_t *chunk_index
)
1085 uint64_t current_addr
= block_offset
+ offset
;
1086 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1087 (void *) block_offset
);
1089 assert(current_addr
>= block
->offset
);
1090 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1092 *block_index
= block
->index
;
1093 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1094 block
->local_host_addr
+ (current_addr
- block
->offset
));
1100 * Register a chunk with IB. If the chunk was already registered
1101 * previously, then skip.
1103 * Also return the keys associated with the registration needed
1104 * to perform the actual RDMA operation.
1106 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1107 RDMALocalBlock
*block
, uint8_t *host_addr
,
1108 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1109 uint8_t *chunk_start
, uint8_t *chunk_end
)
1113 *lkey
= block
->mr
->lkey
;
1116 *rkey
= block
->mr
->rkey
;
1121 /* allocate memory to store chunk MRs */
1123 block
->pmr
= g_malloc0(block
->nb_chunks
* sizeof(struct ibv_mr
*));
1130 * If 'rkey', then we're the destination, so grant access to the source.
1132 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1134 if (!block
->pmr
[chunk
]) {
1135 uint64_t len
= chunk_end
- chunk_start
;
1137 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1139 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
,
1141 (rkey
? (IBV_ACCESS_LOCAL_WRITE
|
1142 IBV_ACCESS_REMOTE_WRITE
) : 0));
1144 if (!block
->pmr
[chunk
]) {
1145 perror("Failed to register chunk!");
1146 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1147 " start %" PRIu64
" end %" PRIu64
" host %" PRIu64
1148 " local %" PRIu64
" registrations: %d\n",
1149 block
->index
, chunk
, (uint64_t) chunk_start
,
1150 (uint64_t) chunk_end
, (uint64_t) host_addr
,
1151 (uint64_t) block
->local_host_addr
,
1152 rdma
->total_registrations
);
1155 rdma
->total_registrations
++;
1159 *lkey
= block
->pmr
[chunk
]->lkey
;
1162 *rkey
= block
->pmr
[chunk
]->rkey
;
1168 * Register (at connection time) the memory used for control
1171 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1173 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1174 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1175 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1176 if (rdma
->wr_data
[idx
].control_mr
) {
1177 rdma
->total_registrations
++;
1180 error_report("qemu_rdma_reg_control failed");
1184 const char *print_wrid(int wrid
)
1186 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1187 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1189 return wrid_desc
[wrid
];
1193 * RDMA requires memory registration (mlock/pinning), but this is not good for
1196 * In preparation for the future where LRU information or workload-specific
1197 * writable writable working set memory access behavior is available to QEMU
1198 * it would be nice to have in place the ability to UN-register/UN-pin
1199 * particular memory regions from the RDMA hardware when it is determine that
1200 * those regions of memory will likely not be accessed again in the near future.
1202 * While we do not yet have such information right now, the following
1203 * compile-time option allows us to perform a non-optimized version of this
1206 * By uncommenting this option, you will cause *all* RDMA transfers to be
1207 * unregistered immediately after the transfer completes on both sides of the
1208 * connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
1210 * This will have a terrible impact on migration performance, so until future
1211 * workload information or LRU information is available, do not attempt to use
1212 * this feature except for basic testing.
1214 //#define RDMA_UNREGISTRATION_EXAMPLE
1217 * Perform a non-optimized memory unregistration after every transfer
1218 * for demonsration purposes, only if pin-all is not requested.
1220 * Potential optimizations:
1221 * 1. Start a new thread to run this function continuously
1223 - and for receipt of unregister messages
1225 * 3. Use workload hints.
1227 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1229 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1231 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1233 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1235 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1236 RDMALocalBlock
*block
=
1237 &(rdma
->local_ram_blocks
.block
[index
]);
1238 RDMARegister reg
= { .current_index
= index
};
1239 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1241 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1242 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1246 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1247 rdma
->unregister_current
);
1249 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1250 rdma
->unregister_current
++;
1252 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1253 rdma
->unregister_current
= 0;
1258 * Unregistration is speculative (because migration is single-threaded
1259 * and we cannot break the protocol's inifinband message ordering).
1260 * Thus, if the memory is currently being used for transmission,
1261 * then abort the attempt to unregister and try again
1262 * later the next time a completion is received for this memory.
1264 clear_bit(chunk
, block
->unregister_bitmap
);
1266 if (test_bit(chunk
, block
->transit_bitmap
)) {
1267 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1271 trace_qemu_rdma_unregister_waiting_send(chunk
);
1273 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1274 block
->pmr
[chunk
] = NULL
;
1275 block
->remote_keys
[chunk
] = 0;
1278 perror("unregistration chunk failed");
1281 rdma
->total_registrations
--;
1283 reg
.key
.chunk
= chunk
;
1284 register_to_network(®
);
1285 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1291 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1297 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1300 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1302 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1303 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1309 * Set bit for unregistration in the next iteration.
1310 * We cannot transmit right here, but will unpin later.
1312 static void qemu_rdma_signal_unregister(RDMAContext
*rdma
, uint64_t index
,
1313 uint64_t chunk
, uint64_t wr_id
)
1315 if (rdma
->unregistrations
[rdma
->unregister_next
] != 0) {
1316 error_report("rdma migration: queue is full");
1318 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1320 if (!test_and_set_bit(chunk
, block
->unregister_bitmap
)) {
1321 trace_qemu_rdma_signal_unregister_append(chunk
,
1322 rdma
->unregister_next
);
1324 rdma
->unregistrations
[rdma
->unregister_next
++] =
1325 qemu_rdma_make_wrid(wr_id
, index
, chunk
);
1327 if (rdma
->unregister_next
== RDMA_SIGNALED_SEND_MAX
) {
1328 rdma
->unregister_next
= 0;
1331 trace_qemu_rdma_signal_unregister_already(chunk
);
1337 * Consult the connection manager to see a work request
1338 * (of any kind) has completed.
1339 * Return the work request ID that completed.
1341 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, uint64_t *wr_id_out
,
1348 ret
= ibv_poll_cq(rdma
->cq
, 1, &wc
);
1351 *wr_id_out
= RDMA_WRID_NONE
;
1356 error_report("ibv_poll_cq return %d", ret
);
1360 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1362 if (wc
.status
!= IBV_WC_SUCCESS
) {
1363 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1364 wc
.status
, ibv_wc_status_str(wc
.status
));
1365 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1370 if (rdma
->control_ready_expected
&&
1371 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1372 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1373 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1374 rdma
->control_ready_expected
= 0;
1377 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1379 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1381 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1382 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1384 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1386 block
->local_host_addr
, (void *)block
->remote_host_addr
);
1388 clear_bit(chunk
, block
->transit_bitmap
);
1390 if (rdma
->nb_sent
> 0) {
1394 if (!rdma
->pin_all
) {
1396 * FYI: If one wanted to signal a specific chunk to be unregistered
1397 * using LRU or workload-specific information, this is the function
1398 * you would call to do so. That chunk would then get asynchronously
1399 * unregistered later.
1401 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1402 qemu_rdma_signal_unregister(rdma
, index
, chunk
, wc
.wr_id
);
1406 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1409 *wr_id_out
= wc
.wr_id
;
1411 *byte_len
= wc
.byte_len
;
1418 * Block until the next work request has completed.
1420 * First poll to see if a work request has already completed,
1423 * If we encounter completed work requests for IDs other than
1424 * the one we're interested in, then that's generally an error.
1426 * The only exception is actual RDMA Write completions. These
1427 * completions only need to be recorded, but do not actually
1428 * need further processing.
1430 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1433 int num_cq_events
= 0, ret
= 0;
1436 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1438 if (ibv_req_notify_cq(rdma
->cq
, 0)) {
1442 while (wr_id
!= wrid_requested
) {
1443 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1448 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1450 if (wr_id
== RDMA_WRID_NONE
) {
1453 if (wr_id
!= wrid_requested
) {
1454 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1455 wrid_requested
, print_wrid(wr_id
), wr_id
);
1459 if (wr_id
== wrid_requested
) {
1465 * Coroutine doesn't start until process_incoming_migration()
1466 * so don't yield unless we know we're running inside of a coroutine.
1468 if (rdma
->migration_started_on_destination
) {
1469 yield_until_fd_readable(rdma
->comp_channel
->fd
);
1472 if (ibv_get_cq_event(rdma
->comp_channel
, &cq
, &cq_ctx
)) {
1473 perror("ibv_get_cq_event");
1474 goto err_block_for_wrid
;
1479 if (ibv_req_notify_cq(cq
, 0)) {
1480 goto err_block_for_wrid
;
1483 while (wr_id
!= wrid_requested
) {
1484 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1486 goto err_block_for_wrid
;
1489 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1491 if (wr_id
== RDMA_WRID_NONE
) {
1494 if (wr_id
!= wrid_requested
) {
1495 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1496 wrid_requested
, print_wrid(wr_id
), wr_id
);
1500 if (wr_id
== wrid_requested
) {
1501 goto success_block_for_wrid
;
1505 success_block_for_wrid
:
1506 if (num_cq_events
) {
1507 ibv_ack_cq_events(cq
, num_cq_events
);
1512 if (num_cq_events
) {
1513 ibv_ack_cq_events(cq
, num_cq_events
);
1519 * Post a SEND message work request for the control channel
1520 * containing some data and block until the post completes.
1522 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1523 RDMAControlHeader
*head
)
1526 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1527 struct ibv_send_wr
*bad_wr
;
1528 struct ibv_sge sge
= {
1529 .addr
= (uint64_t)(wr
->control
),
1530 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1531 .lkey
= wr
->control_mr
->lkey
,
1533 struct ibv_send_wr send_wr
= {
1534 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1535 .opcode
= IBV_WR_SEND
,
1536 .send_flags
= IBV_SEND_SIGNALED
,
1541 trace_qemu_rdma_post_send_control(control_desc
[head
->type
]);
1544 * We don't actually need to do a memcpy() in here if we used
1545 * the "sge" properly, but since we're only sending control messages
1546 * (not RAM in a performance-critical path), then its OK for now.
1548 * The copy makes the RDMAControlHeader simpler to manipulate
1549 * for the time being.
1551 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1552 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1553 control_to_network((void *) wr
->control
);
1556 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1560 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1563 error_report("Failed to use post IB SEND for control");
1567 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1569 error_report("rdma migration: send polling control error");
1576 * Post a RECV work request in anticipation of some future receipt
1577 * of data on the control channel.
1579 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1581 struct ibv_recv_wr
*bad_wr
;
1582 struct ibv_sge sge
= {
1583 .addr
= (uint64_t)(rdma
->wr_data
[idx
].control
),
1584 .length
= RDMA_CONTROL_MAX_BUFFER
,
1585 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1588 struct ibv_recv_wr recv_wr
= {
1589 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1595 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1603 * Block and wait for a RECV control channel message to arrive.
1605 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1606 RDMAControlHeader
*head
, int expecting
, int idx
)
1609 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1613 error_report("rdma migration: recv polling control error!");
1617 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1618 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1620 trace_qemu_rdma_exchange_get_response_start(control_desc
[expecting
]);
1622 if (expecting
== RDMA_CONTROL_NONE
) {
1623 trace_qemu_rdma_exchange_get_response_none(control_desc
[head
->type
],
1625 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1626 error_report("Was expecting a %s (%d) control message"
1627 ", but got: %s (%d), length: %d",
1628 control_desc
[expecting
], expecting
,
1629 control_desc
[head
->type
], head
->type
, head
->len
);
1632 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1633 error_report("too long length: %d\n", head
->len
);
1636 if (sizeof(*head
) + head
->len
!= byte_len
) {
1637 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1645 * When a RECV work request has completed, the work request's
1646 * buffer is pointed at the header.
1648 * This will advance the pointer to the data portion
1649 * of the control message of the work request's buffer that
1650 * was populated after the work request finished.
1652 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1653 RDMAControlHeader
*head
)
1655 rdma
->wr_data
[idx
].control_len
= head
->len
;
1656 rdma
->wr_data
[idx
].control_curr
=
1657 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1661 * This is an 'atomic' high-level operation to deliver a single, unified
1662 * control-channel message.
1664 * Additionally, if the user is expecting some kind of reply to this message,
1665 * they can request a 'resp' response message be filled in by posting an
1666 * additional work request on behalf of the user and waiting for an additional
1669 * The extra (optional) response is used during registration to us from having
1670 * to perform an *additional* exchange of message just to provide a response by
1671 * instead piggy-backing on the acknowledgement.
1673 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1674 uint8_t *data
, RDMAControlHeader
*resp
,
1676 int (*callback
)(RDMAContext
*rdma
))
1681 * Wait until the dest is ready before attempting to deliver the message
1682 * by waiting for a READY message.
1684 if (rdma
->control_ready_expected
) {
1685 RDMAControlHeader resp
;
1686 ret
= qemu_rdma_exchange_get_response(rdma
,
1687 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1694 * If the user is expecting a response, post a WR in anticipation of it.
1697 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1699 error_report("rdma migration: error posting"
1700 " extra control recv for anticipated result!");
1706 * Post a WR to replace the one we just consumed for the READY message.
1708 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1710 error_report("rdma migration: error posting first control recv!");
1715 * Deliver the control message that was requested.
1717 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1720 error_report("Failed to send control buffer!");
1725 * If we're expecting a response, block and wait for it.
1729 trace_qemu_rdma_exchange_send_issue_callback();
1730 ret
= callback(rdma
);
1736 trace_qemu_rdma_exchange_send_waiting(control_desc
[resp
->type
]);
1737 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1738 resp
->type
, RDMA_WRID_DATA
);
1744 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1746 *resp_idx
= RDMA_WRID_DATA
;
1748 trace_qemu_rdma_exchange_send_received(control_desc
[resp
->type
]);
1751 rdma
->control_ready_expected
= 1;
1757 * This is an 'atomic' high-level operation to receive a single, unified
1758 * control-channel message.
1760 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1763 RDMAControlHeader ready
= {
1765 .type
= RDMA_CONTROL_READY
,
1771 * Inform the source that we're ready to receive a message.
1773 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1776 error_report("Failed to send control buffer!");
1781 * Block and wait for the message.
1783 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1784 expecting
, RDMA_WRID_READY
);
1790 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1793 * Post a new RECV work request to replace the one we just consumed.
1795 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1797 error_report("rdma migration: error posting second control recv!");
1805 * Write an actual chunk of memory using RDMA.
1807 * If we're using dynamic registration on the dest-side, we have to
1808 * send a registration command first.
1810 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
1811 int current_index
, uint64_t current_addr
,
1815 struct ibv_send_wr send_wr
= { 0 };
1816 struct ibv_send_wr
*bad_wr
;
1817 int reg_result_idx
, ret
, count
= 0;
1818 uint64_t chunk
, chunks
;
1819 uint8_t *chunk_start
, *chunk_end
;
1820 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
1822 RDMARegisterResult
*reg_result
;
1823 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
1824 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1825 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
1830 sge
.addr
= (uint64_t)(block
->local_host_addr
+
1831 (current_addr
- block
->offset
));
1832 sge
.length
= length
;
1834 chunk
= ram_chunk_index(block
->local_host_addr
, (uint8_t *) sge
.addr
);
1835 chunk_start
= ram_chunk_start(block
, chunk
);
1837 if (block
->is_ram_block
) {
1838 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
1840 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
1844 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
1846 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
1851 trace_qemu_rdma_write_one_top(chunks
+ 1,
1853 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
1855 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
1857 if (!rdma
->pin_all
) {
1858 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1859 qemu_rdma_unregister_waiting(rdma
);
1863 while (test_bit(chunk
, block
->transit_bitmap
)) {
1865 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
1866 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
1868 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
1871 error_report("Failed to Wait for previous write to complete "
1872 "block %d chunk %" PRIu64
1873 " current %" PRIu64
" len %" PRIu64
" %d",
1874 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
1879 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
1880 if (!block
->remote_keys
[chunk
]) {
1882 * This chunk has not yet been registered, so first check to see
1883 * if the entire chunk is zero. If so, tell the other size to
1884 * memset() + madvise() the entire chunk without RDMA.
1887 if (can_use_buffer_find_nonzero_offset((void *)sge
.addr
, length
)
1888 && buffer_find_nonzero_offset((void *)sge
.addr
,
1889 length
) == length
) {
1890 RDMACompress comp
= {
1891 .offset
= current_addr
,
1893 .block_idx
= current_index
,
1897 head
.len
= sizeof(comp
);
1898 head
.type
= RDMA_CONTROL_COMPRESS
;
1900 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
1901 current_index
, current_addr
);
1903 compress_to_network(&comp
);
1904 ret
= qemu_rdma_exchange_send(rdma
, &head
,
1905 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
1911 acct_update_position(f
, sge
.length
, true);
1917 * Otherwise, tell other side to register.
1919 reg
.current_index
= current_index
;
1920 if (block
->is_ram_block
) {
1921 reg
.key
.current_addr
= current_addr
;
1923 reg
.key
.chunk
= chunk
;
1925 reg
.chunks
= chunks
;
1927 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
1930 register_to_network(®
);
1931 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1932 &resp
, ®_result_idx
, NULL
);
1937 /* try to overlap this single registration with the one we sent. */
1938 if (qemu_rdma_register_and_get_keys(rdma
, block
,
1939 (uint8_t *) sge
.addr
,
1940 &sge
.lkey
, NULL
, chunk
,
1941 chunk_start
, chunk_end
)) {
1942 error_report("cannot get lkey");
1946 reg_result
= (RDMARegisterResult
*)
1947 rdma
->wr_data
[reg_result_idx
].control_curr
;
1949 network_to_result(reg_result
);
1951 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
1952 reg_result
->rkey
, chunk
);
1954 block
->remote_keys
[chunk
] = reg_result
->rkey
;
1955 block
->remote_host_addr
= reg_result
->host_addr
;
1957 /* already registered before */
1958 if (qemu_rdma_register_and_get_keys(rdma
, block
,
1959 (uint8_t *)sge
.addr
,
1960 &sge
.lkey
, NULL
, chunk
,
1961 chunk_start
, chunk_end
)) {
1962 error_report("cannot get lkey!");
1967 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
1969 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
1971 if (qemu_rdma_register_and_get_keys(rdma
, block
, (uint8_t *)sge
.addr
,
1972 &sge
.lkey
, NULL
, chunk
,
1973 chunk_start
, chunk_end
)) {
1974 error_report("cannot get lkey!");
1980 * Encode the ram block index and chunk within this wrid.
1981 * We will use this information at the time of completion
1982 * to figure out which bitmap to check against and then which
1983 * chunk in the bitmap to look for.
1985 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
1986 current_index
, chunk
);
1988 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
1989 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
1990 send_wr
.sg_list
= &sge
;
1991 send_wr
.num_sge
= 1;
1992 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
1993 (current_addr
- block
->offset
);
1995 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
1999 * ibv_post_send() does not return negative error numbers,
2000 * per the specification they are positive - no idea why.
2002 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2004 if (ret
== ENOMEM
) {
2005 trace_qemu_rdma_write_one_queue_full();
2006 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2008 error_report("rdma migration: failed to make "
2009 "room in full send queue! %d", ret
);
2015 } else if (ret
> 0) {
2016 perror("rdma migration: post rdma write failed");
2020 set_bit(chunk
, block
->transit_bitmap
);
2021 acct_update_position(f
, sge
.length
, false);
2022 rdma
->total_writes
++;
2028 * Push out any unwritten RDMA operations.
2030 * We support sending out multiple chunks at the same time.
2031 * Not all of them need to get signaled in the completion queue.
2033 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2037 if (!rdma
->current_length
) {
2041 ret
= qemu_rdma_write_one(f
, rdma
,
2042 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2050 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2053 rdma
->current_length
= 0;
2054 rdma
->current_addr
= 0;
2059 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2060 uint64_t offset
, uint64_t len
)
2062 RDMALocalBlock
*block
;
2066 if (rdma
->current_index
< 0) {
2070 if (rdma
->current_chunk
< 0) {
2074 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2075 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2076 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2078 if (rdma
->current_length
== 0) {
2083 * Only merge into chunk sequentially.
2085 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2089 if (offset
< block
->offset
) {
2093 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2097 if ((host_addr
+ len
) > chunk_end
) {
2105 * We're not actually writing here, but doing three things:
2107 * 1. Identify the chunk the buffer belongs to.
2108 * 2. If the chunk is full or the buffer doesn't belong to the current
2109 * chunk, then start a new chunk and flush() the old chunk.
2110 * 3. To keep the hardware busy, we also group chunks into batches
2111 * and only require that a batch gets acknowledged in the completion
2112 * qeueue instead of each individual chunk.
2114 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2115 uint64_t block_offset
, uint64_t offset
,
2118 uint64_t current_addr
= block_offset
+ offset
;
2119 uint64_t index
= rdma
->current_index
;
2120 uint64_t chunk
= rdma
->current_chunk
;
2123 /* If we cannot merge it, we flush the current buffer first. */
2124 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2125 ret
= qemu_rdma_write_flush(f
, rdma
);
2129 rdma
->current_length
= 0;
2130 rdma
->current_addr
= current_addr
;
2132 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2133 offset
, len
, &index
, &chunk
);
2135 error_report("ram block search failed");
2138 rdma
->current_index
= index
;
2139 rdma
->current_chunk
= chunk
;
2143 rdma
->current_length
+= len
;
2145 /* flush it if buffer is too large */
2146 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2147 return qemu_rdma_write_flush(f
, rdma
);
2153 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2155 struct rdma_cm_event
*cm_event
;
2158 if (rdma
->cm_id
&& rdma
->connected
) {
2159 if (rdma
->error_state
) {
2160 RDMAControlHeader head
= { .len
= 0,
2161 .type
= RDMA_CONTROL_ERROR
,
2164 error_report("Early error. Sending error.");
2165 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2168 ret
= rdma_disconnect(rdma
->cm_id
);
2170 trace_qemu_rdma_cleanup_waiting_for_disconnect();
2171 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2173 rdma_ack_cm_event(cm_event
);
2176 trace_qemu_rdma_cleanup_disconnect();
2177 rdma
->connected
= false;
2180 g_free(rdma
->block
);
2183 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2184 if (rdma
->wr_data
[idx
].control_mr
) {
2185 rdma
->total_registrations
--;
2186 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2188 rdma
->wr_data
[idx
].control_mr
= NULL
;
2191 if (rdma
->local_ram_blocks
.block
) {
2192 while (rdma
->local_ram_blocks
.nb_blocks
) {
2193 __qemu_rdma_delete_block(rdma
,
2194 rdma
->local_ram_blocks
.block
->offset
);
2199 ibv_destroy_cq(rdma
->cq
);
2202 if (rdma
->comp_channel
) {
2203 ibv_destroy_comp_channel(rdma
->comp_channel
);
2204 rdma
->comp_channel
= NULL
;
2207 ibv_dealloc_pd(rdma
->pd
);
2210 if (rdma
->listen_id
) {
2211 rdma_destroy_id(rdma
->listen_id
);
2212 rdma
->listen_id
= NULL
;
2216 rdma_destroy_qp(rdma
->cm_id
);
2219 rdma_destroy_id(rdma
->cm_id
);
2222 if (rdma
->channel
) {
2223 rdma_destroy_event_channel(rdma
->channel
);
2224 rdma
->channel
= NULL
;
2231 static int qemu_rdma_source_init(RDMAContext
*rdma
, Error
**errp
, bool pin_all
)
2234 Error
*local_err
= NULL
, **temp
= &local_err
;
2237 * Will be validated against destination's actual capabilities
2238 * after the connect() completes.
2240 rdma
->pin_all
= pin_all
;
2242 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2244 goto err_rdma_source_init
;
2247 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2249 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2250 " limits may be too low. Please check $ ulimit -a # and "
2251 "search for 'ulimit -l' in the output");
2252 goto err_rdma_source_init
;
2255 ret
= qemu_rdma_alloc_qp(rdma
);
2257 ERROR(temp
, "rdma migration: error allocating qp!");
2258 goto err_rdma_source_init
;
2261 ret
= qemu_rdma_init_ram_blocks(rdma
);
2263 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2264 goto err_rdma_source_init
;
2267 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2268 ret
= qemu_rdma_reg_control(rdma
, idx
);
2270 ERROR(temp
, "rdma migration: error registering %d control!",
2272 goto err_rdma_source_init
;
2278 err_rdma_source_init
:
2279 error_propagate(errp
, local_err
);
2280 qemu_rdma_cleanup(rdma
);
2284 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
)
2286 RDMACapabilities cap
= {
2287 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2290 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2292 .private_data
= &cap
,
2293 .private_data_len
= sizeof(cap
),
2295 struct rdma_cm_event
*cm_event
;
2299 * Only negotiate the capability with destination if the user
2300 * on the source first requested the capability.
2302 if (rdma
->pin_all
) {
2303 trace_qemu_rdma_connect_pin_all_requested();
2304 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2307 caps_to_network(&cap
);
2309 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2311 perror("rdma_connect");
2312 ERROR(errp
, "connecting to destination!");
2313 rdma_destroy_id(rdma
->cm_id
);
2315 goto err_rdma_source_connect
;
2318 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2320 perror("rdma_get_cm_event after rdma_connect");
2321 ERROR(errp
, "connecting to destination!");
2322 rdma_ack_cm_event(cm_event
);
2323 rdma_destroy_id(rdma
->cm_id
);
2325 goto err_rdma_source_connect
;
2328 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2329 perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2330 ERROR(errp
, "connecting to destination!");
2331 rdma_ack_cm_event(cm_event
);
2332 rdma_destroy_id(rdma
->cm_id
);
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
)
2372 int ret
= -EINVAL
, idx
;
2373 struct rdma_cm_id
*listen_id
;
2374 char ip
[40] = "unknown";
2375 struct rdma_addrinfo
*res
;
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
== NULL
) {
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 if (rdma
->host
&& strcmp("", rdma
->host
)) {
2407 struct rdma_addrinfo
*e
;
2409 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2411 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2412 goto err_dest_init_bind_addr
;
2415 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2416 inet_ntop(e
->ai_family
,
2417 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2418 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2419 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2421 if (e
->ai_family
== AF_INET6
) {
2422 ret
= qemu_rdma_broken_ipv6_kernel(errp
, listen_id
->verbs
);
2432 ERROR(errp
, "Error: could not rdma_bind_addr!");
2433 goto err_dest_init_bind_addr
;
2435 ERROR(errp
, "migration host and port not specified!");
2437 goto err_dest_init_bind_addr
;
2441 rdma
->listen_id
= listen_id
;
2442 qemu_rdma_dump_gid("dest_init", listen_id
);
2445 err_dest_init_bind_addr
:
2446 rdma_destroy_id(listen_id
);
2447 err_dest_init_create_listen_id
:
2448 rdma_destroy_event_channel(rdma
->channel
);
2449 rdma
->channel
= NULL
;
2450 rdma
->error_state
= ret
;
2455 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2457 RDMAContext
*rdma
= NULL
;
2458 InetSocketAddress
*addr
;
2461 rdma
= g_malloc0(sizeof(RDMAContext
));
2462 memset(rdma
, 0, sizeof(RDMAContext
));
2463 rdma
->current_index
= -1;
2464 rdma
->current_chunk
= -1;
2466 addr
= inet_parse(host_port
, NULL
);
2468 rdma
->port
= atoi(addr
->port
);
2469 rdma
->host
= g_strdup(addr
->host
);
2471 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2476 qapi_free_InetSocketAddress(addr
);
2483 * QEMUFile interface to the control channel.
2484 * SEND messages for control only.
2485 * VM's ram is handled with regular RDMA messages.
2487 static int qemu_rdma_put_buffer(void *opaque
, const uint8_t *buf
,
2488 int64_t pos
, int size
)
2490 QEMUFileRDMA
*r
= opaque
;
2491 QEMUFile
*f
= r
->file
;
2492 RDMAContext
*rdma
= r
->rdma
;
2493 size_t remaining
= size
;
2494 uint8_t * data
= (void *) buf
;
2497 CHECK_ERROR_STATE();
2500 * Push out any writes that
2501 * we're queued up for VM's ram.
2503 ret
= qemu_rdma_write_flush(f
, rdma
);
2505 rdma
->error_state
= ret
;
2510 RDMAControlHeader head
;
2512 r
->len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2513 remaining
-= r
->len
;
2516 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2518 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2521 rdma
->error_state
= ret
;
2531 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2536 if (rdma
->wr_data
[idx
].control_len
) {
2537 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2539 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2540 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2541 rdma
->wr_data
[idx
].control_curr
+= len
;
2542 rdma
->wr_data
[idx
].control_len
-= len
;
2549 * QEMUFile interface to the control channel.
2550 * RDMA links don't use bytestreams, so we have to
2551 * return bytes to QEMUFile opportunistically.
2553 static int qemu_rdma_get_buffer(void *opaque
, uint8_t *buf
,
2554 int64_t pos
, int size
)
2556 QEMUFileRDMA
*r
= opaque
;
2557 RDMAContext
*rdma
= r
->rdma
;
2558 RDMAControlHeader head
;
2561 CHECK_ERROR_STATE();
2564 * First, we hold on to the last SEND message we
2565 * were given and dish out the bytes until we run
2568 r
->len
= qemu_rdma_fill(r
->rdma
, buf
, size
, 0);
2574 * Once we run out, we block and wait for another
2575 * SEND message to arrive.
2577 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2580 rdma
->error_state
= ret
;
2585 * SEND was received with new bytes, now try again.
2587 return qemu_rdma_fill(r
->rdma
, buf
, size
, 0);
2591 * Block until all the outstanding chunks have been delivered by the hardware.
2593 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2597 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2601 while (rdma
->nb_sent
) {
2602 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2604 error_report("rdma migration: complete polling error!");
2609 qemu_rdma_unregister_waiting(rdma
);
2614 static int qemu_rdma_close(void *opaque
)
2616 trace_qemu_rdma_close();
2617 QEMUFileRDMA
*r
= opaque
;
2619 qemu_rdma_cleanup(r
->rdma
);
2629 * This means that 'block_offset' is a full virtual address that does not
2630 * belong to a RAMBlock of the virtual machine and instead
2631 * represents a private malloc'd memory area that the caller wishes to
2635 * Offset is an offset to be added to block_offset and used
2636 * to also lookup the corresponding RAMBlock.
2639 * Initiate an transfer this size.
2642 * A 'hint' or 'advice' that means that we wish to speculatively
2643 * and asynchronously unregister this memory. In this case, there is no
2644 * guarantee that the unregister will actually happen, for example,
2645 * if the memory is being actively transmitted. Additionally, the memory
2646 * may be re-registered at any future time if a write within the same
2647 * chunk was requested again, even if you attempted to unregister it
2650 * @size < 0 : TODO, not yet supported
2651 * Unregister the memory NOW. This means that the caller does not
2652 * expect there to be any future RDMA transfers and we just want to clean
2653 * things up. This is used in case the upper layer owns the memory and
2654 * cannot wait for qemu_fclose() to occur.
2656 * @bytes_sent : User-specificed pointer to indicate how many bytes were
2657 * sent. Usually, this will not be more than a few bytes of
2658 * the protocol because most transfers are sent asynchronously.
2660 static size_t qemu_rdma_save_page(QEMUFile
*f
, void *opaque
,
2661 ram_addr_t block_offset
, ram_addr_t offset
,
2662 size_t size
, int *bytes_sent
)
2664 QEMUFileRDMA
*rfile
= opaque
;
2665 RDMAContext
*rdma
= rfile
->rdma
;
2668 CHECK_ERROR_STATE();
2674 * Add this page to the current 'chunk'. If the chunk
2675 * is full, or the page doen't belong to the current chunk,
2676 * an actual RDMA write will occur and a new chunk will be formed.
2678 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
2680 error_report("rdma migration: write error! %d", ret
);
2685 * We always return 1 bytes because the RDMA
2686 * protocol is completely asynchronous. We do not yet know
2687 * whether an identified chunk is zero or not because we're
2688 * waiting for other pages to potentially be merged with
2689 * the current chunk. So, we have to call qemu_update_position()
2690 * later on when the actual write occurs.
2696 uint64_t index
, chunk
;
2698 /* TODO: Change QEMUFileOps prototype to be signed: size_t => long
2700 ret = qemu_rdma_drain_cq(f, rdma);
2702 fprintf(stderr, "rdma: failed to synchronously drain"
2703 " completion queue before unregistration.\n");
2709 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2710 offset
, size
, &index
, &chunk
);
2713 error_report("ram block search failed");
2717 qemu_rdma_signal_unregister(rdma
, index
, chunk
, 0);
2720 * TODO: Synchronous, guaranteed unregistration (should not occur during
2721 * fast-path). Otherwise, unregisters will process on the next call to
2722 * qemu_rdma_drain_cq()
2724 qemu_rdma_unregister_waiting(rdma);
2730 * Drain the Completion Queue if possible, but do not block,
2733 * If nothing to poll, the end of the iteration will do this
2734 * again to make sure we don't overflow the request queue.
2737 uint64_t wr_id
, wr_id_in
;
2738 int ret
= qemu_rdma_poll(rdma
, &wr_id_in
, NULL
);
2740 error_report("rdma migration: polling error! %d", ret
);
2744 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
2746 if (wr_id
== RDMA_WRID_NONE
) {
2751 return RAM_SAVE_CONTROL_DELAYED
;
2753 rdma
->error_state
= ret
;
2757 static int qemu_rdma_accept(RDMAContext
*rdma
)
2759 RDMACapabilities cap
;
2760 struct rdma_conn_param conn_param
= {
2761 .responder_resources
= 2,
2762 .private_data
= &cap
,
2763 .private_data_len
= sizeof(cap
),
2765 struct rdma_cm_event
*cm_event
;
2766 struct ibv_context
*verbs
;
2770 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2772 goto err_rdma_dest_wait
;
2775 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
2776 rdma_ack_cm_event(cm_event
);
2777 goto err_rdma_dest_wait
;
2780 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2782 network_to_caps(&cap
);
2784 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
2785 error_report("Unknown source RDMA version: %d, bailing...",
2787 rdma_ack_cm_event(cm_event
);
2788 goto err_rdma_dest_wait
;
2792 * Respond with only the capabilities this version of QEMU knows about.
2794 cap
.flags
&= known_capabilities
;
2797 * Enable the ones that we do know about.
2798 * Add other checks here as new ones are introduced.
2800 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
2801 rdma
->pin_all
= true;
2804 rdma
->cm_id
= cm_event
->id
;
2805 verbs
= cm_event
->id
->verbs
;
2807 rdma_ack_cm_event(cm_event
);
2809 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
2811 caps_to_network(&cap
);
2813 trace_qemu_rdma_accept_pin_verbsc(verbs
);
2816 rdma
->verbs
= verbs
;
2817 } else if (rdma
->verbs
!= verbs
) {
2818 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
2820 goto err_rdma_dest_wait
;
2823 qemu_rdma_dump_id("dest_init", verbs
);
2825 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2827 error_report("rdma migration: error allocating pd and cq!");
2828 goto err_rdma_dest_wait
;
2831 ret
= qemu_rdma_alloc_qp(rdma
);
2833 error_report("rdma migration: error allocating qp!");
2834 goto err_rdma_dest_wait
;
2837 ret
= qemu_rdma_init_ram_blocks(rdma
);
2839 error_report("rdma migration: error initializing ram blocks!");
2840 goto err_rdma_dest_wait
;
2843 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2844 ret
= qemu_rdma_reg_control(rdma
, idx
);
2846 error_report("rdma: error registering %d control", idx
);
2847 goto err_rdma_dest_wait
;
2851 qemu_set_fd_handler2(rdma
->channel
->fd
, NULL
, NULL
, NULL
, NULL
);
2853 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
2855 error_report("rdma_accept returns %d", ret
);
2856 goto err_rdma_dest_wait
;
2859 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2861 error_report("rdma_accept get_cm_event failed %d", ret
);
2862 goto err_rdma_dest_wait
;
2865 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2866 error_report("rdma_accept not event established");
2867 rdma_ack_cm_event(cm_event
);
2868 goto err_rdma_dest_wait
;
2871 rdma_ack_cm_event(cm_event
);
2872 rdma
->connected
= true;
2874 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2876 error_report("rdma migration: error posting second control recv");
2877 goto err_rdma_dest_wait
;
2880 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
2885 rdma
->error_state
= ret
;
2886 qemu_rdma_cleanup(rdma
);
2891 * During each iteration of the migration, we listen for instructions
2892 * by the source VM to perform dynamic page registrations before they
2893 * can perform RDMA operations.
2895 * We respond with the 'rkey'.
2897 * Keep doing this until the source tells us to stop.
2899 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
,
2902 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
2903 .type
= RDMA_CONTROL_REGISTER_RESULT
,
2906 RDMAControlHeader unreg_resp
= { .len
= 0,
2907 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
2910 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
2912 QEMUFileRDMA
*rfile
= opaque
;
2913 RDMAContext
*rdma
= rfile
->rdma
;
2914 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
2915 RDMAControlHeader head
;
2916 RDMARegister
*reg
, *registers
;
2918 RDMARegisterResult
*reg_result
;
2919 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
2920 RDMALocalBlock
*block
;
2927 CHECK_ERROR_STATE();
2930 trace_qemu_rdma_registration_handle_wait(flags
);
2932 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
2938 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
2939 error_report("rdma: Too many requests in this message (%d)."
2940 "Bailing.", head
.repeat
);
2945 switch (head
.type
) {
2946 case RDMA_CONTROL_COMPRESS
:
2947 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
2948 network_to_compress(comp
);
2950 trace_qemu_rdma_registration_handle_compress(comp
->length
,
2953 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
2955 host_addr
= block
->local_host_addr
+
2956 (comp
->offset
- block
->offset
);
2958 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
2961 case RDMA_CONTROL_REGISTER_FINISHED
:
2962 trace_qemu_rdma_registration_handle_finished();
2965 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
2966 trace_qemu_rdma_registration_handle_ram_blocks();
2968 if (rdma
->pin_all
) {
2969 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
2971 error_report("rdma migration: error dest "
2972 "registering ram blocks");
2978 * Dest uses this to prepare to transmit the RAMBlock descriptions
2979 * to the source VM after connection setup.
2980 * Both sides use the "remote" structure to communicate and update
2981 * their "local" descriptions with what was sent.
2983 for (i
= 0; i
< local
->nb_blocks
; i
++) {
2984 rdma
->block
[i
].remote_host_addr
=
2985 (uint64_t)(local
->block
[i
].local_host_addr
);
2987 if (rdma
->pin_all
) {
2988 rdma
->block
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
2991 rdma
->block
[i
].offset
= local
->block
[i
].offset
;
2992 rdma
->block
[i
].length
= local
->block
[i
].length
;
2994 remote_block_to_network(&rdma
->block
[i
]);
2997 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
2998 * sizeof(RDMARemoteBlock
);
3001 ret
= qemu_rdma_post_send_control(rdma
,
3002 (uint8_t *) rdma
->block
, &blocks
);
3005 error_report("rdma migration: error sending remote info");
3010 case RDMA_CONTROL_REGISTER_REQUEST
:
3011 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3013 reg_resp
.repeat
= head
.repeat
;
3014 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3016 for (count
= 0; count
< head
.repeat
; count
++) {
3018 uint8_t *chunk_start
, *chunk_end
;
3020 reg
= ®isters
[count
];
3021 network_to_register(reg
);
3023 reg_result
= &results
[count
];
3025 trace_qemu_rdma_registration_handle_register_loop(count
,
3026 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3028 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3029 if (block
->is_ram_block
) {
3030 host_addr
= (block
->local_host_addr
+
3031 (reg
->key
.current_addr
- block
->offset
));
3032 chunk
= ram_chunk_index(block
->local_host_addr
,
3033 (uint8_t *) host_addr
);
3035 chunk
= reg
->key
.chunk
;
3036 host_addr
= block
->local_host_addr
+
3037 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3039 chunk_start
= ram_chunk_start(block
, chunk
);
3040 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3041 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3042 (uint8_t *)host_addr
, NULL
, ®_result
->rkey
,
3043 chunk
, chunk_start
, chunk_end
)) {
3044 error_report("cannot get rkey");
3049 reg_result
->host_addr
= (uint64_t) block
->local_host_addr
;
3051 trace_qemu_rdma_registration_handle_register_rkey(
3054 result_to_network(reg_result
);
3057 ret
= qemu_rdma_post_send_control(rdma
,
3058 (uint8_t *) results
, ®_resp
);
3061 error_report("Failed to send control buffer");
3065 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3066 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3067 unreg_resp
.repeat
= head
.repeat
;
3068 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3070 for (count
= 0; count
< head
.repeat
; count
++) {
3071 reg
= ®isters
[count
];
3072 network_to_register(reg
);
3074 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3075 reg
->current_index
, reg
->key
.chunk
);
3077 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3079 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3080 block
->pmr
[reg
->key
.chunk
] = NULL
;
3083 perror("rdma unregistration chunk failed");
3088 rdma
->total_registrations
--;
3090 trace_qemu_rdma_registration_handle_unregister_success(
3094 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3097 error_report("Failed to send control buffer");
3101 case RDMA_CONTROL_REGISTER_RESULT
:
3102 error_report("Invalid RESULT message at dest.");
3106 error_report("Unknown control message %s", control_desc
[head
.type
]);
3113 rdma
->error_state
= ret
;
3118 static int qemu_rdma_registration_start(QEMUFile
*f
, void *opaque
,
3121 QEMUFileRDMA
*rfile
= opaque
;
3122 RDMAContext
*rdma
= rfile
->rdma
;
3124 CHECK_ERROR_STATE();
3126 trace_qemu_rdma_registration_start(flags
);
3127 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3134 * Inform dest that dynamic registrations are done for now.
3135 * First, flush writes, if any.
3137 static int qemu_rdma_registration_stop(QEMUFile
*f
, void *opaque
,
3140 Error
*local_err
= NULL
, **errp
= &local_err
;
3141 QEMUFileRDMA
*rfile
= opaque
;
3142 RDMAContext
*rdma
= rfile
->rdma
;
3143 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3146 CHECK_ERROR_STATE();
3149 ret
= qemu_rdma_drain_cq(f
, rdma
);
3155 if (flags
== RAM_CONTROL_SETUP
) {
3156 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3157 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3158 int reg_result_idx
, i
, j
, nb_remote_blocks
;
3160 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3161 trace_qemu_rdma_registration_stop_ram();
3164 * Make sure that we parallelize the pinning on both sides.
3165 * For very large guests, doing this serially takes a really
3166 * long time, so we have to 'interleave' the pinning locally
3167 * with the control messages by performing the pinning on this
3168 * side before we receive the control response from the other
3169 * side that the pinning has completed.
3171 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3172 ®_result_idx
, rdma
->pin_all
?
3173 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3175 ERROR(errp
, "receiving remote info!");
3179 nb_remote_blocks
= resp
.len
/ sizeof(RDMARemoteBlock
);
3182 * The protocol uses two different sets of rkeys (mutually exclusive):
3183 * 1. One key to represent the virtual address of the entire ram block.
3184 * (dynamic chunk registration disabled - pin everything with one rkey.)
3185 * 2. One to represent individual chunks within a ram block.
3186 * (dynamic chunk registration enabled - pin individual chunks.)
3188 * Once the capability is successfully negotiated, the destination transmits
3189 * the keys to use (or sends them later) including the virtual addresses
3190 * and then propagates the remote ram block descriptions to his local copy.
3193 if (local
->nb_blocks
!= nb_remote_blocks
) {
3194 ERROR(errp
, "ram blocks mismatch #1! "
3195 "Your QEMU command line parameters are probably "
3196 "not identical on both the source and destination.");
3200 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3202 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3203 for (i
= 0; i
< nb_remote_blocks
; i
++) {
3204 network_to_remote_block(&rdma
->block
[i
]);
3206 /* search local ram blocks */
3207 for (j
= 0; j
< local
->nb_blocks
; j
++) {
3208 if (rdma
->block
[i
].offset
!= local
->block
[j
].offset
) {
3212 if (rdma
->block
[i
].length
!= local
->block
[j
].length
) {
3213 ERROR(errp
, "ram blocks mismatch #2! "
3214 "Your QEMU command line parameters are probably "
3215 "not identical on both the source and destination.");
3218 local
->block
[j
].remote_host_addr
=
3219 rdma
->block
[i
].remote_host_addr
;
3220 local
->block
[j
].remote_rkey
= rdma
->block
[i
].remote_rkey
;
3224 if (j
>= local
->nb_blocks
) {
3225 ERROR(errp
, "ram blocks mismatch #3! "
3226 "Your QEMU command line parameters are probably "
3227 "not identical on both the source and destination.");
3233 trace_qemu_rdma_registration_stop(flags
);
3235 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3236 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3244 rdma
->error_state
= ret
;
3248 static int qemu_rdma_get_fd(void *opaque
)
3250 QEMUFileRDMA
*rfile
= opaque
;
3251 RDMAContext
*rdma
= rfile
->rdma
;
3253 return rdma
->comp_channel
->fd
;
3256 const QEMUFileOps rdma_read_ops
= {
3257 .get_buffer
= qemu_rdma_get_buffer
,
3258 .get_fd
= qemu_rdma_get_fd
,
3259 .close
= qemu_rdma_close
,
3260 .hook_ram_load
= qemu_rdma_registration_handle
,
3263 const QEMUFileOps rdma_write_ops
= {
3264 .put_buffer
= qemu_rdma_put_buffer
,
3265 .close
= qemu_rdma_close
,
3266 .before_ram_iterate
= qemu_rdma_registration_start
,
3267 .after_ram_iterate
= qemu_rdma_registration_stop
,
3268 .save_page
= qemu_rdma_save_page
,
3271 static void *qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
3273 QEMUFileRDMA
*r
= g_malloc0(sizeof(QEMUFileRDMA
));
3275 if (qemu_file_mode_is_not_valid(mode
)) {
3281 if (mode
[0] == 'w') {
3282 r
->file
= qemu_fopen_ops(r
, &rdma_write_ops
);
3284 r
->file
= qemu_fopen_ops(r
, &rdma_read_ops
);
3290 static void rdma_accept_incoming_migration(void *opaque
)
3292 RDMAContext
*rdma
= opaque
;
3295 Error
*local_err
= NULL
, **errp
= &local_err
;
3297 trace_qemu_dma_accept_incoming_migration();
3298 ret
= qemu_rdma_accept(rdma
);
3301 ERROR(errp
, "RDMA Migration initialization failed!");
3305 trace_qemu_dma_accept_incoming_migration_accepted();
3307 f
= qemu_fopen_rdma(rdma
, "rb");
3309 ERROR(errp
, "could not qemu_fopen_rdma!");
3310 qemu_rdma_cleanup(rdma
);
3314 rdma
->migration_started_on_destination
= 1;
3315 process_incoming_migration(f
);
3318 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
3322 Error
*local_err
= NULL
;
3324 trace_rdma_start_incoming_migration();
3325 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
3331 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
3337 trace_rdma_start_incoming_migration_after_dest_init();
3339 ret
= rdma_listen(rdma
->listen_id
, 5);
3342 ERROR(errp
, "listening on socket!");
3346 trace_rdma_start_incoming_migration_after_rdma_listen();
3348 qemu_set_fd_handler2(rdma
->channel
->fd
, NULL
,
3349 rdma_accept_incoming_migration
, NULL
,
3350 (void *)(intptr_t) rdma
);
3353 error_propagate(errp
, local_err
);
3357 void rdma_start_outgoing_migration(void *opaque
,
3358 const char *host_port
, Error
**errp
)
3360 MigrationState
*s
= opaque
;
3361 Error
*local_err
= NULL
, **temp
= &local_err
;
3362 RDMAContext
*rdma
= qemu_rdma_data_init(host_port
, &local_err
);
3366 ERROR(temp
, "Failed to initialize RDMA data structures! %d", ret
);
3370 ret
= qemu_rdma_source_init(rdma
, &local_err
,
3371 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
]);
3377 trace_rdma_start_outgoing_migration_after_rdma_source_init();
3378 ret
= qemu_rdma_connect(rdma
, &local_err
);
3384 trace_rdma_start_outgoing_migration_after_rdma_connect();
3386 s
->file
= qemu_fopen_rdma(rdma
, "wb");
3387 migrate_fd_connect(s
);
3390 error_propagate(errp
, local_err
);
3392 migrate_fd_error(s
);