2 * RDMA protocol and interfaces
4 * Copyright IBM, Corp. 2010-2013
5 * Copyright Red Hat, Inc. 2015-2016
8 * Michael R. Hines <mrhines@us.ibm.com>
9 * Jiuxing Liu <jl@us.ibm.com>
10 * Daniel P. Berrange <berrange@redhat.com>
12 * This work is licensed under the terms of the GNU GPL, version 2 or
13 * later. See the COPYING file in the top-level directory.
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "qemu-common.h"
19 #include "qemu/cutils.h"
21 #include "migration.h"
22 #include "qemu-file.h"
24 #include "qemu-file-channel.h"
25 #include "qemu/error-report.h"
26 #include "qemu/main-loop.h"
27 #include "qemu/sockets.h"
28 #include "qemu/bitmap.h"
29 #include "qemu/coroutine.h"
30 #include <sys/socket.h>
32 #include <arpa/inet.h>
33 #include <rdma/rdma_cma.h>
37 * Print and error on both the Monitor and the Log file.
39 #define ERROR(errp, fmt, ...) \
41 fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \
42 if (errp && (*(errp) == NULL)) { \
43 error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \
47 #define RDMA_RESOLVE_TIMEOUT_MS 10000
49 /* Do not merge data if larger than this. */
50 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
51 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
53 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
56 * This is only for non-live state being migrated.
57 * Instead of RDMA_WRITE messages, we use RDMA_SEND
58 * messages for that state, which requires a different
59 * delivery design than main memory.
61 #define RDMA_SEND_INCREMENT 32768
64 * Maximum size infiniband SEND message
66 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
67 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
69 #define RDMA_CONTROL_VERSION_CURRENT 1
71 * Capabilities for negotiation.
73 #define RDMA_CAPABILITY_PIN_ALL 0x01
76 * Add the other flags above to this list of known capabilities
77 * as they are introduced.
79 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
81 #define CHECK_ERROR_STATE() \
83 if (rdma->error_state) { \
84 if (!rdma->error_reported) { \
85 error_report("RDMA is in an error state waiting migration" \
87 rdma->error_reported = 1; \
90 return rdma->error_state; \
95 * A work request ID is 64-bits and we split up these bits
98 * bits 0-15 : type of control message, 2^16
99 * bits 16-29: ram block index, 2^14
100 * bits 30-63: ram block chunk number, 2^34
102 * The last two bit ranges are only used for RDMA writes,
103 * in order to track their completion and potentially
104 * also track unregistration status of the message.
106 #define RDMA_WRID_TYPE_SHIFT 0UL
107 #define RDMA_WRID_BLOCK_SHIFT 16UL
108 #define RDMA_WRID_CHUNK_SHIFT 30UL
110 #define RDMA_WRID_TYPE_MASK \
111 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
113 #define RDMA_WRID_BLOCK_MASK \
114 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
116 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
119 * RDMA migration protocol:
120 * 1. RDMA Writes (data messages, i.e. RAM)
121 * 2. IB Send/Recv (control channel messages)
125 RDMA_WRID_RDMA_WRITE
= 1,
126 RDMA_WRID_SEND_CONTROL
= 2000,
127 RDMA_WRID_RECV_CONTROL
= 4000,
130 static const char *wrid_desc
[] = {
131 [RDMA_WRID_NONE
] = "NONE",
132 [RDMA_WRID_RDMA_WRITE
] = "WRITE RDMA",
133 [RDMA_WRID_SEND_CONTROL
] = "CONTROL SEND",
134 [RDMA_WRID_RECV_CONTROL
] = "CONTROL RECV",
138 * Work request IDs for IB SEND messages only (not RDMA writes).
139 * This is used by the migration protocol to transmit
140 * control messages (such as device state and registration commands)
142 * We could use more WRs, but we have enough for now.
152 * SEND/RECV IB Control Messages.
155 RDMA_CONTROL_NONE
= 0,
157 RDMA_CONTROL_READY
, /* ready to receive */
158 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
159 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
160 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
161 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
162 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
163 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
164 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
165 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
166 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
171 * Memory and MR structures used to represent an IB Send/Recv work request.
172 * This is *not* used for RDMA writes, only IB Send/Recv.
175 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
176 struct ibv_mr
*control_mr
; /* registration metadata */
177 size_t control_len
; /* length of the message */
178 uint8_t *control_curr
; /* start of unconsumed bytes */
179 } RDMAWorkRequestData
;
182 * Negotiate RDMA capabilities during connection-setup time.
189 static void caps_to_network(RDMACapabilities
*cap
)
191 cap
->version
= htonl(cap
->version
);
192 cap
->flags
= htonl(cap
->flags
);
195 static void network_to_caps(RDMACapabilities
*cap
)
197 cap
->version
= ntohl(cap
->version
);
198 cap
->flags
= ntohl(cap
->flags
);
202 * Representation of a RAMBlock from an RDMA perspective.
203 * This is not transmitted, only local.
204 * This and subsequent structures cannot be linked lists
205 * because we're using a single IB message to transmit
206 * the information. It's small anyway, so a list is overkill.
208 typedef struct RDMALocalBlock
{
210 uint8_t *local_host_addr
; /* local virtual address */
211 uint64_t remote_host_addr
; /* remote virtual address */
214 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
215 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
216 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
217 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
218 int index
; /* which block are we */
219 unsigned int src_index
; /* (Only used on dest) */
222 unsigned long *transit_bitmap
;
223 unsigned long *unregister_bitmap
;
227 * Also represents a RAMblock, but only on the dest.
228 * This gets transmitted by the dest during connection-time
229 * to the source VM and then is used to populate the
230 * corresponding RDMALocalBlock with
231 * the information needed to perform the actual RDMA.
233 typedef struct QEMU_PACKED RDMADestBlock
{
234 uint64_t remote_host_addr
;
237 uint32_t remote_rkey
;
241 static const char *control_desc(unsigned int rdma_control
)
243 static const char *strs
[] = {
244 [RDMA_CONTROL_NONE
] = "NONE",
245 [RDMA_CONTROL_ERROR
] = "ERROR",
246 [RDMA_CONTROL_READY
] = "READY",
247 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
248 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
249 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
250 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
251 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
252 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
253 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
254 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
255 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
258 if (rdma_control
> RDMA_CONTROL_UNREGISTER_FINISHED
) {
259 return "??BAD CONTROL VALUE??";
262 return strs
[rdma_control
];
265 static uint64_t htonll(uint64_t v
)
267 union { uint32_t lv
[2]; uint64_t llv
; } u
;
268 u
.lv
[0] = htonl(v
>> 32);
269 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
273 static uint64_t ntohll(uint64_t v
) {
274 union { uint32_t lv
[2]; uint64_t llv
; } u
;
276 return ((uint64_t)ntohl(u
.lv
[0]) << 32) | (uint64_t) ntohl(u
.lv
[1]);
279 static void dest_block_to_network(RDMADestBlock
*db
)
281 db
->remote_host_addr
= htonll(db
->remote_host_addr
);
282 db
->offset
= htonll(db
->offset
);
283 db
->length
= htonll(db
->length
);
284 db
->remote_rkey
= htonl(db
->remote_rkey
);
287 static void network_to_dest_block(RDMADestBlock
*db
)
289 db
->remote_host_addr
= ntohll(db
->remote_host_addr
);
290 db
->offset
= ntohll(db
->offset
);
291 db
->length
= ntohll(db
->length
);
292 db
->remote_rkey
= ntohl(db
->remote_rkey
);
296 * Virtual address of the above structures used for transmitting
297 * the RAMBlock descriptions at connection-time.
298 * This structure is *not* transmitted.
300 typedef struct RDMALocalBlocks
{
302 bool init
; /* main memory init complete */
303 RDMALocalBlock
*block
;
307 * Main data structure for RDMA state.
308 * While there is only one copy of this structure being allocated right now,
309 * this is the place where one would start if you wanted to consider
310 * having more than one RDMA connection open at the same time.
312 typedef struct RDMAContext
{
316 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
319 * This is used by *_exchange_send() to figure out whether or not
320 * the initial "READY" message has already been received or not.
321 * This is because other functions may potentially poll() and detect
322 * the READY message before send() does, in which case we need to
323 * know if it completed.
325 int control_ready_expected
;
327 /* number of outstanding writes */
330 /* store info about current buffer so that we can
331 merge it with future sends */
332 uint64_t current_addr
;
333 uint64_t current_length
;
334 /* index of ram block the current buffer belongs to */
336 /* index of the chunk in the current ram block */
342 * infiniband-specific variables for opening the device
343 * and maintaining connection state and so forth.
345 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
346 * cm_id->verbs, cm_id->channel, and cm_id->qp.
348 struct rdma_cm_id
*cm_id
; /* connection manager ID */
349 struct rdma_cm_id
*listen_id
;
352 struct ibv_context
*verbs
;
353 struct rdma_event_channel
*channel
;
354 struct ibv_qp
*qp
; /* queue pair */
355 struct ibv_comp_channel
*comp_channel
; /* completion channel */
356 struct ibv_pd
*pd
; /* protection domain */
357 struct ibv_cq
*cq
; /* completion queue */
360 * If a previous write failed (perhaps because of a failed
361 * memory registration, then do not attempt any future work
362 * and remember the error state.
369 * Description of ram blocks used throughout the code.
371 RDMALocalBlocks local_ram_blocks
;
372 RDMADestBlock
*dest_blocks
;
374 /* Index of the next RAMBlock received during block registration */
375 unsigned int next_src_index
;
378 * Migration on *destination* started.
379 * Then use coroutine yield function.
380 * Source runs in a thread, so we don't care.
382 int migration_started_on_destination
;
384 int total_registrations
;
387 int unregister_current
, unregister_next
;
388 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
390 GHashTable
*blockmap
;
392 /* the RDMAContext for return path */
393 struct RDMAContext
*return_path
;
397 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
398 #define QIO_CHANNEL_RDMA(obj) \
399 OBJECT_CHECK(QIOChannelRDMA, (obj), TYPE_QIO_CHANNEL_RDMA)
401 typedef struct QIOChannelRDMA QIOChannelRDMA
;
404 struct QIOChannelRDMA
{
407 RDMAContext
*rdmaout
;
409 bool blocking
; /* XXX we don't actually honour this yet */
413 * Main structure for IB Send/Recv control messages.
414 * This gets prepended at the beginning of every Send/Recv.
416 typedef struct QEMU_PACKED
{
417 uint32_t len
; /* Total length of data portion */
418 uint32_t type
; /* which control command to perform */
419 uint32_t repeat
; /* number of commands in data portion of same type */
423 static void control_to_network(RDMAControlHeader
*control
)
425 control
->type
= htonl(control
->type
);
426 control
->len
= htonl(control
->len
);
427 control
->repeat
= htonl(control
->repeat
);
430 static void network_to_control(RDMAControlHeader
*control
)
432 control
->type
= ntohl(control
->type
);
433 control
->len
= ntohl(control
->len
);
434 control
->repeat
= ntohl(control
->repeat
);
438 * Register a single Chunk.
439 * Information sent by the source VM to inform the dest
440 * to register an single chunk of memory before we can perform
441 * the actual RDMA operation.
443 typedef struct QEMU_PACKED
{
445 uint64_t current_addr
; /* offset into the ram_addr_t space */
446 uint64_t chunk
; /* chunk to lookup if unregistering */
448 uint32_t current_index
; /* which ramblock the chunk belongs to */
450 uint64_t chunks
; /* how many sequential chunks to register */
453 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
455 RDMALocalBlock
*local_block
;
456 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
458 if (local_block
->is_ram_block
) {
460 * current_addr as passed in is an address in the local ram_addr_t
461 * space, we need to translate this for the destination
463 reg
->key
.current_addr
-= local_block
->offset
;
464 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
466 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
467 reg
->current_index
= htonl(reg
->current_index
);
468 reg
->chunks
= htonll(reg
->chunks
);
471 static void network_to_register(RDMARegister
*reg
)
473 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
474 reg
->current_index
= ntohl(reg
->current_index
);
475 reg
->chunks
= ntohll(reg
->chunks
);
478 typedef struct QEMU_PACKED
{
479 uint32_t value
; /* if zero, we will madvise() */
480 uint32_t block_idx
; /* which ram block index */
481 uint64_t offset
; /* Address in remote ram_addr_t space */
482 uint64_t length
; /* length of the chunk */
485 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
487 comp
->value
= htonl(comp
->value
);
489 * comp->offset as passed in is an address in the local ram_addr_t
490 * space, we need to translate this for the destination
492 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
493 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
494 comp
->block_idx
= htonl(comp
->block_idx
);
495 comp
->offset
= htonll(comp
->offset
);
496 comp
->length
= htonll(comp
->length
);
499 static void network_to_compress(RDMACompress
*comp
)
501 comp
->value
= ntohl(comp
->value
);
502 comp
->block_idx
= ntohl(comp
->block_idx
);
503 comp
->offset
= ntohll(comp
->offset
);
504 comp
->length
= ntohll(comp
->length
);
508 * The result of the dest's memory registration produces an "rkey"
509 * which the source VM must reference in order to perform
510 * the RDMA operation.
512 typedef struct QEMU_PACKED
{
516 } RDMARegisterResult
;
518 static void result_to_network(RDMARegisterResult
*result
)
520 result
->rkey
= htonl(result
->rkey
);
521 result
->host_addr
= htonll(result
->host_addr
);
524 static void network_to_result(RDMARegisterResult
*result
)
526 result
->rkey
= ntohl(result
->rkey
);
527 result
->host_addr
= ntohll(result
->host_addr
);
530 const char *print_wrid(int wrid
);
531 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
532 uint8_t *data
, RDMAControlHeader
*resp
,
534 int (*callback
)(RDMAContext
*rdma
));
536 static inline uint64_t ram_chunk_index(const uint8_t *start
,
539 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
542 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
545 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
546 (i
<< RDMA_REG_CHUNK_SHIFT
));
549 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
552 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
553 (1UL << RDMA_REG_CHUNK_SHIFT
);
555 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
556 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
562 static int rdma_add_block(RDMAContext
*rdma
, const char *block_name
,
564 ram_addr_t block_offset
, uint64_t length
)
566 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
567 RDMALocalBlock
*block
;
568 RDMALocalBlock
*old
= local
->block
;
570 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
+ 1);
572 if (local
->nb_blocks
) {
575 if (rdma
->blockmap
) {
576 for (x
= 0; x
< local
->nb_blocks
; x
++) {
577 g_hash_table_remove(rdma
->blockmap
,
578 (void *)(uintptr_t)old
[x
].offset
);
579 g_hash_table_insert(rdma
->blockmap
,
580 (void *)(uintptr_t)old
[x
].offset
,
584 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
588 block
= &local
->block
[local
->nb_blocks
];
590 block
->block_name
= g_strdup(block_name
);
591 block
->local_host_addr
= host_addr
;
592 block
->offset
= block_offset
;
593 block
->length
= length
;
594 block
->index
= local
->nb_blocks
;
595 block
->src_index
= ~0U; /* Filled in by the receipt of the block list */
596 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
597 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
598 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
599 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
600 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
601 block
->remote_keys
= g_new0(uint32_t, block
->nb_chunks
);
603 block
->is_ram_block
= local
->init
? false : true;
605 if (rdma
->blockmap
) {
606 g_hash_table_insert(rdma
->blockmap
, (void *)(uintptr_t)block_offset
, block
);
609 trace_rdma_add_block(block_name
, local
->nb_blocks
,
610 (uintptr_t) block
->local_host_addr
,
611 block
->offset
, block
->length
,
612 (uintptr_t) (block
->local_host_addr
+ block
->length
),
613 BITS_TO_LONGS(block
->nb_chunks
) *
614 sizeof(unsigned long) * 8,
623 * Memory regions need to be registered with the device and queue pairs setup
624 * in advanced before the migration starts. This tells us where the RAM blocks
625 * are so that we can register them individually.
627 static int qemu_rdma_init_one_block(const char *block_name
, void *host_addr
,
628 ram_addr_t block_offset
, ram_addr_t length
, void *opaque
)
630 return rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
634 * Identify the RAMBlocks and their quantity. They will be references to
635 * identify chunk boundaries inside each RAMBlock and also be referenced
636 * during dynamic page registration.
638 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
640 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
642 assert(rdma
->blockmap
== NULL
);
643 memset(local
, 0, sizeof *local
);
644 qemu_ram_foreach_migratable_block(qemu_rdma_init_one_block
, rdma
);
645 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
646 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
647 rdma
->local_ram_blocks
.nb_blocks
);
653 * Note: If used outside of cleanup, the caller must ensure that the destination
654 * block structures are also updated
656 static int rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
658 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
659 RDMALocalBlock
*old
= local
->block
;
662 if (rdma
->blockmap
) {
663 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
668 for (j
= 0; j
< block
->nb_chunks
; j
++) {
669 if (!block
->pmr
[j
]) {
672 ibv_dereg_mr(block
->pmr
[j
]);
673 rdma
->total_registrations
--;
680 ibv_dereg_mr(block
->mr
);
681 rdma
->total_registrations
--;
685 g_free(block
->transit_bitmap
);
686 block
->transit_bitmap
= NULL
;
688 g_free(block
->unregister_bitmap
);
689 block
->unregister_bitmap
= NULL
;
691 g_free(block
->remote_keys
);
692 block
->remote_keys
= NULL
;
694 g_free(block
->block_name
);
695 block
->block_name
= NULL
;
697 if (rdma
->blockmap
) {
698 for (x
= 0; x
< local
->nb_blocks
; x
++) {
699 g_hash_table_remove(rdma
->blockmap
,
700 (void *)(uintptr_t)old
[x
].offset
);
704 if (local
->nb_blocks
> 1) {
706 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
709 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
712 if (block
->index
< (local
->nb_blocks
- 1)) {
713 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
714 sizeof(RDMALocalBlock
) *
715 (local
->nb_blocks
- (block
->index
+ 1)));
716 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
717 local
->block
[x
].index
--;
721 assert(block
== local
->block
);
725 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
726 block
->offset
, block
->length
,
727 (uintptr_t)(block
->local_host_addr
+ block
->length
),
728 BITS_TO_LONGS(block
->nb_chunks
) *
729 sizeof(unsigned long) * 8, block
->nb_chunks
);
735 if (local
->nb_blocks
&& rdma
->blockmap
) {
736 for (x
= 0; x
< local
->nb_blocks
; x
++) {
737 g_hash_table_insert(rdma
->blockmap
,
738 (void *)(uintptr_t)local
->block
[x
].offset
,
747 * Put in the log file which RDMA device was opened and the details
748 * associated with that device.
750 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
752 struct ibv_port_attr port
;
754 if (ibv_query_port(verbs
, 1, &port
)) {
755 error_report("Failed to query port information");
759 printf("%s RDMA Device opened: kernel name %s "
760 "uverbs device name %s, "
761 "infiniband_verbs class device path %s, "
762 "infiniband class device path %s, "
763 "transport: (%d) %s\n",
766 verbs
->device
->dev_name
,
767 verbs
->device
->dev_path
,
768 verbs
->device
->ibdev_path
,
770 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
771 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
772 ? "Ethernet" : "Unknown"));
776 * Put in the log file the RDMA gid addressing information,
777 * useful for folks who have trouble understanding the
778 * RDMA device hierarchy in the kernel.
780 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
784 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
785 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
786 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
790 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
791 * We will try the next addrinfo struct, and fail if there are
792 * no other valid addresses to bind against.
794 * If user is listening on '[::]', then we will not have a opened a device
795 * yet and have no way of verifying if the device is RoCE or not.
797 * In this case, the source VM will throw an error for ALL types of
798 * connections (both IPv4 and IPv6) if the destination machine does not have
799 * a regular infiniband network available for use.
801 * The only way to guarantee that an error is thrown for broken kernels is
802 * for the management software to choose a *specific* interface at bind time
803 * and validate what time of hardware it is.
805 * Unfortunately, this puts the user in a fix:
807 * If the source VM connects with an IPv4 address without knowing that the
808 * destination has bound to '[::]' the migration will unconditionally fail
809 * unless the management software is explicitly listening on the IPv4
810 * address while using a RoCE-based device.
812 * If the source VM connects with an IPv6 address, then we're OK because we can
813 * throw an error on the source (and similarly on the destination).
815 * But in mixed environments, this will be broken for a while until it is fixed
818 * We do provide a *tiny* bit of help in this function: We can list all of the
819 * devices in the system and check to see if all the devices are RoCE or
822 * If we detect that we have a *pure* RoCE environment, then we can safely
823 * thrown an error even if the management software has specified '[::]' as the
826 * However, if there is are multiple hetergeneous devices, then we cannot make
827 * this assumption and the user just has to be sure they know what they are
830 * Patches are being reviewed on linux-rdma.
832 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
834 struct ibv_port_attr port_attr
;
836 /* This bug only exists in linux, to our knowledge. */
840 * Verbs are only NULL if management has bound to '[::]'.
842 * Let's iterate through all the devices and see if there any pure IB
843 * devices (non-ethernet).
845 * If not, then we can safely proceed with the migration.
846 * Otherwise, there are no guarantees until the bug is fixed in linux.
850 struct ibv_device
** dev_list
= ibv_get_device_list(&num_devices
);
851 bool roce_found
= false;
852 bool ib_found
= false;
854 for (x
= 0; x
< num_devices
; x
++) {
855 verbs
= ibv_open_device(dev_list
[x
]);
857 if (errno
== EPERM
) {
864 if (ibv_query_port(verbs
, 1, &port_attr
)) {
865 ibv_close_device(verbs
);
866 ERROR(errp
, "Could not query initial IB port");
870 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
872 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
876 ibv_close_device(verbs
);
882 fprintf(stderr
, "WARN: migrations may fail:"
883 " IPv6 over RoCE / iWARP in linux"
884 " is broken. But since you appear to have a"
885 " mixed RoCE / IB environment, be sure to only"
886 " migrate over the IB fabric until the kernel "
887 " fixes the bug.\n");
889 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
890 " and your management software has specified '[::]'"
891 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
900 * If we have a verbs context, that means that some other than '[::]' was
901 * used by the management software for binding. In which case we can
902 * actually warn the user about a potentially broken kernel.
905 /* IB ports start with 1, not 0 */
906 if (ibv_query_port(verbs
, 1, &port_attr
)) {
907 ERROR(errp
, "Could not query initial IB port");
911 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
912 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
913 "(but patches on linux-rdma in progress)");
923 * Figure out which RDMA device corresponds to the requested IP hostname
924 * Also create the initial connection manager identifiers for opening
927 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
930 struct rdma_addrinfo
*res
;
932 struct rdma_cm_event
*cm_event
;
933 char ip
[40] = "unknown";
934 struct rdma_addrinfo
*e
;
936 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
937 ERROR(errp
, "RDMA hostname has not been set");
941 /* create CM channel */
942 rdma
->channel
= rdma_create_event_channel();
943 if (!rdma
->channel
) {
944 ERROR(errp
, "could not create CM channel");
949 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
951 ERROR(errp
, "could not create channel id");
952 goto err_resolve_create_id
;
955 snprintf(port_str
, 16, "%d", rdma
->port
);
958 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
960 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
961 goto err_resolve_get_addr
;
964 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
965 inet_ntop(e
->ai_family
,
966 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
967 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
969 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
970 RDMA_RESOLVE_TIMEOUT_MS
);
972 if (e
->ai_family
== AF_INET6
) {
973 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
982 ERROR(errp
, "could not resolve address %s", rdma
->host
);
983 goto err_resolve_get_addr
;
986 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
988 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
990 ERROR(errp
, "could not perform event_addr_resolved");
991 goto err_resolve_get_addr
;
994 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
995 ERROR(errp
, "result not equal to event_addr_resolved %s",
996 rdma_event_str(cm_event
->event
));
997 perror("rdma_resolve_addr");
998 rdma_ack_cm_event(cm_event
);
1000 goto err_resolve_get_addr
;
1002 rdma_ack_cm_event(cm_event
);
1005 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1007 ERROR(errp
, "could not resolve rdma route");
1008 goto err_resolve_get_addr
;
1011 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1013 ERROR(errp
, "could not perform event_route_resolved");
1014 goto err_resolve_get_addr
;
1016 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1017 ERROR(errp
, "result not equal to event_route_resolved: %s",
1018 rdma_event_str(cm_event
->event
));
1019 rdma_ack_cm_event(cm_event
);
1021 goto err_resolve_get_addr
;
1023 rdma_ack_cm_event(cm_event
);
1024 rdma
->verbs
= rdma
->cm_id
->verbs
;
1025 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1026 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1029 err_resolve_get_addr
:
1030 rdma_destroy_id(rdma
->cm_id
);
1032 err_resolve_create_id
:
1033 rdma_destroy_event_channel(rdma
->channel
);
1034 rdma
->channel
= NULL
;
1039 * Create protection domain and completion queues
1041 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1044 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1046 error_report("failed to allocate protection domain");
1050 /* create completion channel */
1051 rdma
->comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1052 if (!rdma
->comp_channel
) {
1053 error_report("failed to allocate completion channel");
1054 goto err_alloc_pd_cq
;
1058 * Completion queue can be filled by both read and write work requests,
1059 * so must reflect the sum of both possible queue sizes.
1061 rdma
->cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1062 NULL
, rdma
->comp_channel
, 0);
1064 error_report("failed to allocate completion queue");
1065 goto err_alloc_pd_cq
;
1072 ibv_dealloc_pd(rdma
->pd
);
1074 if (rdma
->comp_channel
) {
1075 ibv_destroy_comp_channel(rdma
->comp_channel
);
1078 rdma
->comp_channel
= NULL
;
1084 * Create queue pairs.
1086 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1088 struct ibv_qp_init_attr attr
= { 0 };
1091 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1092 attr
.cap
.max_recv_wr
= 3;
1093 attr
.cap
.max_send_sge
= 1;
1094 attr
.cap
.max_recv_sge
= 1;
1095 attr
.send_cq
= rdma
->cq
;
1096 attr
.recv_cq
= rdma
->cq
;
1097 attr
.qp_type
= IBV_QPT_RC
;
1099 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1104 rdma
->qp
= rdma
->cm_id
->qp
;
1108 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1111 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1113 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1114 local
->block
[i
].mr
=
1115 ibv_reg_mr(rdma
->pd
,
1116 local
->block
[i
].local_host_addr
,
1117 local
->block
[i
].length
,
1118 IBV_ACCESS_LOCAL_WRITE
|
1119 IBV_ACCESS_REMOTE_WRITE
1121 if (!local
->block
[i
].mr
) {
1122 perror("Failed to register local dest ram block!\n");
1125 rdma
->total_registrations
++;
1128 if (i
>= local
->nb_blocks
) {
1132 for (i
--; i
>= 0; i
--) {
1133 ibv_dereg_mr(local
->block
[i
].mr
);
1134 rdma
->total_registrations
--;
1142 * Find the ram block that corresponds to the page requested to be
1143 * transmitted by QEMU.
1145 * Once the block is found, also identify which 'chunk' within that
1146 * block that the page belongs to.
1148 * This search cannot fail or the migration will fail.
1150 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1151 uintptr_t block_offset
,
1154 uint64_t *block_index
,
1155 uint64_t *chunk_index
)
1157 uint64_t current_addr
= block_offset
+ offset
;
1158 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1159 (void *) block_offset
);
1161 assert(current_addr
>= block
->offset
);
1162 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1164 *block_index
= block
->index
;
1165 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1166 block
->local_host_addr
+ (current_addr
- block
->offset
));
1172 * Register a chunk with IB. If the chunk was already registered
1173 * previously, then skip.
1175 * Also return the keys associated with the registration needed
1176 * to perform the actual RDMA operation.
1178 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1179 RDMALocalBlock
*block
, uintptr_t host_addr
,
1180 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1181 uint8_t *chunk_start
, uint8_t *chunk_end
)
1185 *lkey
= block
->mr
->lkey
;
1188 *rkey
= block
->mr
->rkey
;
1193 /* allocate memory to store chunk MRs */
1195 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1199 * If 'rkey', then we're the destination, so grant access to the source.
1201 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1203 if (!block
->pmr
[chunk
]) {
1204 uint64_t len
= chunk_end
- chunk_start
;
1206 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1208 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
,
1210 (rkey
? (IBV_ACCESS_LOCAL_WRITE
|
1211 IBV_ACCESS_REMOTE_WRITE
) : 0));
1213 if (!block
->pmr
[chunk
]) {
1214 perror("Failed to register chunk!");
1215 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1216 " start %" PRIuPTR
" end %" PRIuPTR
1218 " local %" PRIuPTR
" registrations: %d\n",
1219 block
->index
, chunk
, (uintptr_t)chunk_start
,
1220 (uintptr_t)chunk_end
, host_addr
,
1221 (uintptr_t)block
->local_host_addr
,
1222 rdma
->total_registrations
);
1225 rdma
->total_registrations
++;
1229 *lkey
= block
->pmr
[chunk
]->lkey
;
1232 *rkey
= block
->pmr
[chunk
]->rkey
;
1238 * Register (at connection time) the memory used for control
1241 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1243 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1244 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1245 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1246 if (rdma
->wr_data
[idx
].control_mr
) {
1247 rdma
->total_registrations
++;
1250 error_report("qemu_rdma_reg_control failed");
1254 const char *print_wrid(int wrid
)
1256 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1257 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1259 return wrid_desc
[wrid
];
1263 * RDMA requires memory registration (mlock/pinning), but this is not good for
1266 * In preparation for the future where LRU information or workload-specific
1267 * writable writable working set memory access behavior is available to QEMU
1268 * it would be nice to have in place the ability to UN-register/UN-pin
1269 * particular memory regions from the RDMA hardware when it is determine that
1270 * those regions of memory will likely not be accessed again in the near future.
1272 * While we do not yet have such information right now, the following
1273 * compile-time option allows us to perform a non-optimized version of this
1276 * By uncommenting this option, you will cause *all* RDMA transfers to be
1277 * unregistered immediately after the transfer completes on both sides of the
1278 * connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
1280 * This will have a terrible impact on migration performance, so until future
1281 * workload information or LRU information is available, do not attempt to use
1282 * this feature except for basic testing.
1284 //#define RDMA_UNREGISTRATION_EXAMPLE
1287 * Perform a non-optimized memory unregistration after every transfer
1288 * for demonstration purposes, only if pin-all is not requested.
1290 * Potential optimizations:
1291 * 1. Start a new thread to run this function continuously
1293 - and for receipt of unregister messages
1295 * 3. Use workload hints.
1297 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1299 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1301 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1303 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1305 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1306 RDMALocalBlock
*block
=
1307 &(rdma
->local_ram_blocks
.block
[index
]);
1308 RDMARegister reg
= { .current_index
= index
};
1309 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1311 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1312 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1316 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1317 rdma
->unregister_current
);
1319 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1320 rdma
->unregister_current
++;
1322 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1323 rdma
->unregister_current
= 0;
1328 * Unregistration is speculative (because migration is single-threaded
1329 * and we cannot break the protocol's inifinband message ordering).
1330 * Thus, if the memory is currently being used for transmission,
1331 * then abort the attempt to unregister and try again
1332 * later the next time a completion is received for this memory.
1334 clear_bit(chunk
, block
->unregister_bitmap
);
1336 if (test_bit(chunk
, block
->transit_bitmap
)) {
1337 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1341 trace_qemu_rdma_unregister_waiting_send(chunk
);
1343 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1344 block
->pmr
[chunk
] = NULL
;
1345 block
->remote_keys
[chunk
] = 0;
1348 perror("unregistration chunk failed");
1351 rdma
->total_registrations
--;
1353 reg
.key
.chunk
= chunk
;
1354 register_to_network(rdma
, ®
);
1355 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1361 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1367 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1370 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1372 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1373 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1379 * Set bit for unregistration in the next iteration.
1380 * We cannot transmit right here, but will unpin later.
1382 static void qemu_rdma_signal_unregister(RDMAContext
*rdma
, uint64_t index
,
1383 uint64_t chunk
, uint64_t wr_id
)
1385 if (rdma
->unregistrations
[rdma
->unregister_next
] != 0) {
1386 error_report("rdma migration: queue is full");
1388 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1390 if (!test_and_set_bit(chunk
, block
->unregister_bitmap
)) {
1391 trace_qemu_rdma_signal_unregister_append(chunk
,
1392 rdma
->unregister_next
);
1394 rdma
->unregistrations
[rdma
->unregister_next
++] =
1395 qemu_rdma_make_wrid(wr_id
, index
, chunk
);
1397 if (rdma
->unregister_next
== RDMA_SIGNALED_SEND_MAX
) {
1398 rdma
->unregister_next
= 0;
1401 trace_qemu_rdma_signal_unregister_already(chunk
);
1407 * Consult the connection manager to see a work request
1408 * (of any kind) has completed.
1409 * Return the work request ID that completed.
1411 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, uint64_t *wr_id_out
,
1418 ret
= ibv_poll_cq(rdma
->cq
, 1, &wc
);
1421 *wr_id_out
= RDMA_WRID_NONE
;
1426 error_report("ibv_poll_cq return %d", ret
);
1430 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1432 if (wc
.status
!= IBV_WC_SUCCESS
) {
1433 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1434 wc
.status
, ibv_wc_status_str(wc
.status
));
1435 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1440 if (rdma
->control_ready_expected
&&
1441 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1442 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1443 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1444 rdma
->control_ready_expected
= 0;
1447 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1449 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1451 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1452 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1454 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1455 index
, chunk
, block
->local_host_addr
,
1456 (void *)(uintptr_t)block
->remote_host_addr
);
1458 clear_bit(chunk
, block
->transit_bitmap
);
1460 if (rdma
->nb_sent
> 0) {
1464 if (!rdma
->pin_all
) {
1466 * FYI: If one wanted to signal a specific chunk to be unregistered
1467 * using LRU or workload-specific information, this is the function
1468 * you would call to do so. That chunk would then get asynchronously
1469 * unregistered later.
1471 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1472 qemu_rdma_signal_unregister(rdma
, index
, chunk
, wc
.wr_id
);
1476 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1479 *wr_id_out
= wc
.wr_id
;
1481 *byte_len
= wc
.byte_len
;
1487 /* Wait for activity on the completion channel.
1488 * Returns 0 on success, none-0 on error.
1490 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
)
1492 struct rdma_cm_event
*cm_event
;
1496 * Coroutine doesn't start until migration_fd_process_incoming()
1497 * so don't yield unless we know we're running inside of a coroutine.
1499 if (rdma
->migration_started_on_destination
&&
1500 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1501 yield_until_fd_readable(rdma
->comp_channel
->fd
);
1503 /* This is the source side, we're in a separate thread
1504 * or destination prior to migration_fd_process_incoming()
1505 * after postcopy, the destination also in a seprate thread.
1506 * we can't yield; so we have to poll the fd.
1507 * But we need to be able to handle 'cancel' or an error
1508 * without hanging forever.
1510 while (!rdma
->error_state
&& !rdma
->received_error
) {
1512 pfds
[0].fd
= rdma
->comp_channel
->fd
;
1513 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1514 pfds
[0].revents
= 0;
1516 pfds
[1].fd
= rdma
->channel
->fd
;
1517 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1518 pfds
[1].revents
= 0;
1520 /* 0.1s timeout, should be fine for a 'cancel' */
1521 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1523 case 1: /* fd active */
1524 if (pfds
[0].revents
) {
1528 if (pfds
[1].revents
) {
1529 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1531 rdma_ack_cm_event(cm_event
);
1534 error_report("receive cm event while wait comp channel,"
1535 "cm event is %d", cm_event
->event
);
1536 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1537 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1543 case 0: /* Timeout, go around again */
1546 default: /* Error of some type -
1547 * I don't trust errno from qemu_poll_ns
1549 error_report("%s: poll failed", __func__
);
1553 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1554 /* Bail out and let the cancellation happen */
1560 if (rdma
->received_error
) {
1563 return rdma
->error_state
;
1567 * Block until the next work request has completed.
1569 * First poll to see if a work request has already completed,
1572 * If we encounter completed work requests for IDs other than
1573 * the one we're interested in, then that's generally an error.
1575 * The only exception is actual RDMA Write completions. These
1576 * completions only need to be recorded, but do not actually
1577 * need further processing.
1579 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1582 int num_cq_events
= 0, ret
= 0;
1585 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1587 if (ibv_req_notify_cq(rdma
->cq
, 0)) {
1591 while (wr_id
!= wrid_requested
) {
1592 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1597 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1599 if (wr_id
== RDMA_WRID_NONE
) {
1602 if (wr_id
!= wrid_requested
) {
1603 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1604 wrid_requested
, print_wrid(wr_id
), wr_id
);
1608 if (wr_id
== wrid_requested
) {
1613 ret
= qemu_rdma_wait_comp_channel(rdma
);
1615 goto err_block_for_wrid
;
1618 ret
= ibv_get_cq_event(rdma
->comp_channel
, &cq
, &cq_ctx
);
1620 perror("ibv_get_cq_event");
1621 goto err_block_for_wrid
;
1626 ret
= -ibv_req_notify_cq(cq
, 0);
1628 goto err_block_for_wrid
;
1631 while (wr_id
!= wrid_requested
) {
1632 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1634 goto err_block_for_wrid
;
1637 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1639 if (wr_id
== RDMA_WRID_NONE
) {
1642 if (wr_id
!= wrid_requested
) {
1643 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1644 wrid_requested
, print_wrid(wr_id
), wr_id
);
1648 if (wr_id
== wrid_requested
) {
1649 goto success_block_for_wrid
;
1653 success_block_for_wrid
:
1654 if (num_cq_events
) {
1655 ibv_ack_cq_events(cq
, num_cq_events
);
1660 if (num_cq_events
) {
1661 ibv_ack_cq_events(cq
, num_cq_events
);
1664 rdma
->error_state
= ret
;
1669 * Post a SEND message work request for the control channel
1670 * containing some data and block until the post completes.
1672 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1673 RDMAControlHeader
*head
)
1676 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1677 struct ibv_send_wr
*bad_wr
;
1678 struct ibv_sge sge
= {
1679 .addr
= (uintptr_t)(wr
->control
),
1680 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1681 .lkey
= wr
->control_mr
->lkey
,
1683 struct ibv_send_wr send_wr
= {
1684 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1685 .opcode
= IBV_WR_SEND
,
1686 .send_flags
= IBV_SEND_SIGNALED
,
1691 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1694 * We don't actually need to do a memcpy() in here if we used
1695 * the "sge" properly, but since we're only sending control messages
1696 * (not RAM in a performance-critical path), then its OK for now.
1698 * The copy makes the RDMAControlHeader simpler to manipulate
1699 * for the time being.
1701 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1702 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1703 control_to_network((void *) wr
->control
);
1706 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1710 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1713 error_report("Failed to use post IB SEND for control");
1717 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1719 error_report("rdma migration: send polling control error");
1726 * Post a RECV work request in anticipation of some future receipt
1727 * of data on the control channel.
1729 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1731 struct ibv_recv_wr
*bad_wr
;
1732 struct ibv_sge sge
= {
1733 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1734 .length
= RDMA_CONTROL_MAX_BUFFER
,
1735 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1738 struct ibv_recv_wr recv_wr
= {
1739 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1745 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1753 * Block and wait for a RECV control channel message to arrive.
1755 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1756 RDMAControlHeader
*head
, int expecting
, int idx
)
1759 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1763 error_report("rdma migration: recv polling control error!");
1767 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1768 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1770 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1772 if (expecting
== RDMA_CONTROL_NONE
) {
1773 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1775 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1776 error_report("Was expecting a %s (%d) control message"
1777 ", but got: %s (%d), length: %d",
1778 control_desc(expecting
), expecting
,
1779 control_desc(head
->type
), head
->type
, head
->len
);
1780 if (head
->type
== RDMA_CONTROL_ERROR
) {
1781 rdma
->received_error
= true;
1785 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1786 error_report("too long length: %d", head
->len
);
1789 if (sizeof(*head
) + head
->len
!= byte_len
) {
1790 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1798 * When a RECV work request has completed, the work request's
1799 * buffer is pointed at the header.
1801 * This will advance the pointer to the data portion
1802 * of the control message of the work request's buffer that
1803 * was populated after the work request finished.
1805 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1806 RDMAControlHeader
*head
)
1808 rdma
->wr_data
[idx
].control_len
= head
->len
;
1809 rdma
->wr_data
[idx
].control_curr
=
1810 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1814 * This is an 'atomic' high-level operation to deliver a single, unified
1815 * control-channel message.
1817 * Additionally, if the user is expecting some kind of reply to this message,
1818 * they can request a 'resp' response message be filled in by posting an
1819 * additional work request on behalf of the user and waiting for an additional
1822 * The extra (optional) response is used during registration to us from having
1823 * to perform an *additional* exchange of message just to provide a response by
1824 * instead piggy-backing on the acknowledgement.
1826 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1827 uint8_t *data
, RDMAControlHeader
*resp
,
1829 int (*callback
)(RDMAContext
*rdma
))
1834 * Wait until the dest is ready before attempting to deliver the message
1835 * by waiting for a READY message.
1837 if (rdma
->control_ready_expected
) {
1838 RDMAControlHeader resp
;
1839 ret
= qemu_rdma_exchange_get_response(rdma
,
1840 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1847 * If the user is expecting a response, post a WR in anticipation of it.
1850 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1852 error_report("rdma migration: error posting"
1853 " extra control recv for anticipated result!");
1859 * Post a WR to replace the one we just consumed for the READY message.
1861 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1863 error_report("rdma migration: error posting first control recv!");
1868 * Deliver the control message that was requested.
1870 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1873 error_report("Failed to send control buffer!");
1878 * If we're expecting a response, block and wait for it.
1882 trace_qemu_rdma_exchange_send_issue_callback();
1883 ret
= callback(rdma
);
1889 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1890 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1891 resp
->type
, RDMA_WRID_DATA
);
1897 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1899 *resp_idx
= RDMA_WRID_DATA
;
1901 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1904 rdma
->control_ready_expected
= 1;
1910 * This is an 'atomic' high-level operation to receive a single, unified
1911 * control-channel message.
1913 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1916 RDMAControlHeader ready
= {
1918 .type
= RDMA_CONTROL_READY
,
1924 * Inform the source that we're ready to receive a message.
1926 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1929 error_report("Failed to send control buffer!");
1934 * Block and wait for the message.
1936 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1937 expecting
, RDMA_WRID_READY
);
1943 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1946 * Post a new RECV work request to replace the one we just consumed.
1948 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1950 error_report("rdma migration: error posting second control recv!");
1958 * Write an actual chunk of memory using RDMA.
1960 * If we're using dynamic registration on the dest-side, we have to
1961 * send a registration command first.
1963 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
1964 int current_index
, uint64_t current_addr
,
1968 struct ibv_send_wr send_wr
= { 0 };
1969 struct ibv_send_wr
*bad_wr
;
1970 int reg_result_idx
, ret
, count
= 0;
1971 uint64_t chunk
, chunks
;
1972 uint8_t *chunk_start
, *chunk_end
;
1973 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
1975 RDMARegisterResult
*reg_result
;
1976 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
1977 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1978 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
1983 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
1984 (current_addr
- block
->offset
));
1985 sge
.length
= length
;
1987 chunk
= ram_chunk_index(block
->local_host_addr
,
1988 (uint8_t *)(uintptr_t)sge
.addr
);
1989 chunk_start
= ram_chunk_start(block
, chunk
);
1991 if (block
->is_ram_block
) {
1992 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
1994 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
1998 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2000 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2005 trace_qemu_rdma_write_one_top(chunks
+ 1,
2007 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2009 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2011 if (!rdma
->pin_all
) {
2012 #ifdef RDMA_UNREGISTRATION_EXAMPLE
2013 qemu_rdma_unregister_waiting(rdma
);
2017 while (test_bit(chunk
, block
->transit_bitmap
)) {
2019 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2020 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2022 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2025 error_report("Failed to Wait for previous write to complete "
2026 "block %d chunk %" PRIu64
2027 " current %" PRIu64
" len %" PRIu64
" %d",
2028 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2033 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2034 if (!block
->remote_keys
[chunk
]) {
2036 * This chunk has not yet been registered, so first check to see
2037 * if the entire chunk is zero. If so, tell the other size to
2038 * memset() + madvise() the entire chunk without RDMA.
2041 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2042 RDMACompress comp
= {
2043 .offset
= current_addr
,
2045 .block_idx
= current_index
,
2049 head
.len
= sizeof(comp
);
2050 head
.type
= RDMA_CONTROL_COMPRESS
;
2052 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2053 current_index
, current_addr
);
2055 compress_to_network(rdma
, &comp
);
2056 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2057 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2063 acct_update_position(f
, sge
.length
, true);
2069 * Otherwise, tell other side to register.
2071 reg
.current_index
= current_index
;
2072 if (block
->is_ram_block
) {
2073 reg
.key
.current_addr
= current_addr
;
2075 reg
.key
.chunk
= chunk
;
2077 reg
.chunks
= chunks
;
2079 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2082 register_to_network(rdma
, ®
);
2083 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2084 &resp
, ®_result_idx
, NULL
);
2089 /* try to overlap this single registration with the one we sent. */
2090 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2091 &sge
.lkey
, NULL
, chunk
,
2092 chunk_start
, chunk_end
)) {
2093 error_report("cannot get lkey");
2097 reg_result
= (RDMARegisterResult
*)
2098 rdma
->wr_data
[reg_result_idx
].control_curr
;
2100 network_to_result(reg_result
);
2102 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2103 reg_result
->rkey
, chunk
);
2105 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2106 block
->remote_host_addr
= reg_result
->host_addr
;
2108 /* already registered before */
2109 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2110 &sge
.lkey
, NULL
, chunk
,
2111 chunk_start
, chunk_end
)) {
2112 error_report("cannot get lkey!");
2117 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2119 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2121 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2122 &sge
.lkey
, NULL
, chunk
,
2123 chunk_start
, chunk_end
)) {
2124 error_report("cannot get lkey!");
2130 * Encode the ram block index and chunk within this wrid.
2131 * We will use this information at the time of completion
2132 * to figure out which bitmap to check against and then which
2133 * chunk in the bitmap to look for.
2135 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2136 current_index
, chunk
);
2138 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2139 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2140 send_wr
.sg_list
= &sge
;
2141 send_wr
.num_sge
= 1;
2142 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2143 (current_addr
- block
->offset
);
2145 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2149 * ibv_post_send() does not return negative error numbers,
2150 * per the specification they are positive - no idea why.
2152 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2154 if (ret
== ENOMEM
) {
2155 trace_qemu_rdma_write_one_queue_full();
2156 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2158 error_report("rdma migration: failed to make "
2159 "room in full send queue! %d", ret
);
2165 } else if (ret
> 0) {
2166 perror("rdma migration: post rdma write failed");
2170 set_bit(chunk
, block
->transit_bitmap
);
2171 acct_update_position(f
, sge
.length
, false);
2172 rdma
->total_writes
++;
2178 * Push out any unwritten RDMA operations.
2180 * We support sending out multiple chunks at the same time.
2181 * Not all of them need to get signaled in the completion queue.
2183 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2187 if (!rdma
->current_length
) {
2191 ret
= qemu_rdma_write_one(f
, rdma
,
2192 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2200 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2203 rdma
->current_length
= 0;
2204 rdma
->current_addr
= 0;
2209 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2210 uint64_t offset
, uint64_t len
)
2212 RDMALocalBlock
*block
;
2216 if (rdma
->current_index
< 0) {
2220 if (rdma
->current_chunk
< 0) {
2224 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2225 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2226 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2228 if (rdma
->current_length
== 0) {
2233 * Only merge into chunk sequentially.
2235 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2239 if (offset
< block
->offset
) {
2243 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2247 if ((host_addr
+ len
) > chunk_end
) {
2255 * We're not actually writing here, but doing three things:
2257 * 1. Identify the chunk the buffer belongs to.
2258 * 2. If the chunk is full or the buffer doesn't belong to the current
2259 * chunk, then start a new chunk and flush() the old chunk.
2260 * 3. To keep the hardware busy, we also group chunks into batches
2261 * and only require that a batch gets acknowledged in the completion
2262 * qeueue instead of each individual chunk.
2264 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2265 uint64_t block_offset
, uint64_t offset
,
2268 uint64_t current_addr
= block_offset
+ offset
;
2269 uint64_t index
= rdma
->current_index
;
2270 uint64_t chunk
= rdma
->current_chunk
;
2273 /* If we cannot merge it, we flush the current buffer first. */
2274 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2275 ret
= qemu_rdma_write_flush(f
, rdma
);
2279 rdma
->current_length
= 0;
2280 rdma
->current_addr
= current_addr
;
2282 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2283 offset
, len
, &index
, &chunk
);
2285 error_report("ram block search failed");
2288 rdma
->current_index
= index
;
2289 rdma
->current_chunk
= chunk
;
2293 rdma
->current_length
+= len
;
2295 /* flush it if buffer is too large */
2296 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2297 return qemu_rdma_write_flush(f
, rdma
);
2303 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2307 if (rdma
->cm_id
&& rdma
->connected
) {
2308 if ((rdma
->error_state
||
2309 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2310 !rdma
->received_error
) {
2311 RDMAControlHeader head
= { .len
= 0,
2312 .type
= RDMA_CONTROL_ERROR
,
2315 error_report("Early error. Sending error.");
2316 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2319 rdma_disconnect(rdma
->cm_id
);
2320 trace_qemu_rdma_cleanup_disconnect();
2321 rdma
->connected
= false;
2324 if (rdma
->channel
) {
2325 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2327 g_free(rdma
->dest_blocks
);
2328 rdma
->dest_blocks
= NULL
;
2330 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2331 if (rdma
->wr_data
[idx
].control_mr
) {
2332 rdma
->total_registrations
--;
2333 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2335 rdma
->wr_data
[idx
].control_mr
= NULL
;
2338 if (rdma
->local_ram_blocks
.block
) {
2339 while (rdma
->local_ram_blocks
.nb_blocks
) {
2340 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2345 rdma_destroy_qp(rdma
->cm_id
);
2349 ibv_destroy_cq(rdma
->cq
);
2352 if (rdma
->comp_channel
) {
2353 ibv_destroy_comp_channel(rdma
->comp_channel
);
2354 rdma
->comp_channel
= NULL
;
2357 ibv_dealloc_pd(rdma
->pd
);
2361 rdma_destroy_id(rdma
->cm_id
);
2365 /* the destination side, listen_id and channel is shared */
2366 if (rdma
->listen_id
) {
2367 if (!rdma
->is_return_path
) {
2368 rdma_destroy_id(rdma
->listen_id
);
2370 rdma
->listen_id
= NULL
;
2372 if (rdma
->channel
) {
2373 if (!rdma
->is_return_path
) {
2374 rdma_destroy_event_channel(rdma
->channel
);
2376 rdma
->channel
= NULL
;
2380 if (rdma
->channel
) {
2381 rdma_destroy_event_channel(rdma
->channel
);
2382 rdma
->channel
= NULL
;
2389 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2392 Error
*local_err
= NULL
, **temp
= &local_err
;
2395 * Will be validated against destination's actual capabilities
2396 * after the connect() completes.
2398 rdma
->pin_all
= pin_all
;
2400 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2402 goto err_rdma_source_init
;
2405 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2407 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2408 " limits may be too low. Please check $ ulimit -a # and "
2409 "search for 'ulimit -l' in the output");
2410 goto err_rdma_source_init
;
2413 ret
= qemu_rdma_alloc_qp(rdma
);
2415 ERROR(temp
, "rdma migration: error allocating qp!");
2416 goto err_rdma_source_init
;
2419 ret
= qemu_rdma_init_ram_blocks(rdma
);
2421 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2422 goto err_rdma_source_init
;
2425 /* Build the hash that maps from offset to RAMBlock */
2426 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2427 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2428 g_hash_table_insert(rdma
->blockmap
,
2429 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2430 &rdma
->local_ram_blocks
.block
[idx
]);
2433 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2434 ret
= qemu_rdma_reg_control(rdma
, idx
);
2436 ERROR(temp
, "rdma migration: error registering %d control!",
2438 goto err_rdma_source_init
;
2444 err_rdma_source_init
:
2445 error_propagate(errp
, local_err
);
2446 qemu_rdma_cleanup(rdma
);
2450 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
)
2452 RDMACapabilities cap
= {
2453 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2456 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2458 .private_data
= &cap
,
2459 .private_data_len
= sizeof(cap
),
2461 struct rdma_cm_event
*cm_event
;
2465 * Only negotiate the capability with destination if the user
2466 * on the source first requested the capability.
2468 if (rdma
->pin_all
) {
2469 trace_qemu_rdma_connect_pin_all_requested();
2470 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2473 caps_to_network(&cap
);
2475 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2477 ERROR(errp
, "posting second control recv");
2478 goto err_rdma_source_connect
;
2481 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2483 perror("rdma_connect");
2484 ERROR(errp
, "connecting to destination!");
2485 goto err_rdma_source_connect
;
2488 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2490 perror("rdma_get_cm_event after rdma_connect");
2491 ERROR(errp
, "connecting to destination!");
2492 rdma_ack_cm_event(cm_event
);
2493 goto err_rdma_source_connect
;
2496 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2497 perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2498 ERROR(errp
, "connecting to destination!");
2499 rdma_ack_cm_event(cm_event
);
2500 goto err_rdma_source_connect
;
2502 rdma
->connected
= true;
2504 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2505 network_to_caps(&cap
);
2508 * Verify that the *requested* capabilities are supported by the destination
2509 * and disable them otherwise.
2511 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2512 ERROR(errp
, "Server cannot support pinning all memory. "
2513 "Will register memory dynamically.");
2514 rdma
->pin_all
= false;
2517 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2519 rdma_ack_cm_event(cm_event
);
2521 rdma
->control_ready_expected
= 1;
2525 err_rdma_source_connect
:
2526 qemu_rdma_cleanup(rdma
);
2530 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2533 struct rdma_cm_id
*listen_id
;
2534 char ip
[40] = "unknown";
2535 struct rdma_addrinfo
*res
, *e
;
2538 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2539 rdma
->wr_data
[idx
].control_len
= 0;
2540 rdma
->wr_data
[idx
].control_curr
= NULL
;
2543 if (!rdma
->host
|| !rdma
->host
[0]) {
2544 ERROR(errp
, "RDMA host is not set!");
2545 rdma
->error_state
= -EINVAL
;
2548 /* create CM channel */
2549 rdma
->channel
= rdma_create_event_channel();
2550 if (!rdma
->channel
) {
2551 ERROR(errp
, "could not create rdma event channel");
2552 rdma
->error_state
= -EINVAL
;
2557 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2559 ERROR(errp
, "could not create cm_id!");
2560 goto err_dest_init_create_listen_id
;
2563 snprintf(port_str
, 16, "%d", rdma
->port
);
2564 port_str
[15] = '\0';
2566 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2568 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2569 goto err_dest_init_bind_addr
;
2572 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2573 inet_ntop(e
->ai_family
,
2574 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2575 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2576 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2580 if (e
->ai_family
== AF_INET6
) {
2581 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2590 ERROR(errp
, "Error: could not rdma_bind_addr!");
2591 goto err_dest_init_bind_addr
;
2594 rdma
->listen_id
= listen_id
;
2595 qemu_rdma_dump_gid("dest_init", listen_id
);
2598 err_dest_init_bind_addr
:
2599 rdma_destroy_id(listen_id
);
2600 err_dest_init_create_listen_id
:
2601 rdma_destroy_event_channel(rdma
->channel
);
2602 rdma
->channel
= NULL
;
2603 rdma
->error_state
= ret
;
2608 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2613 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2614 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2615 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2618 /*the CM channel and CM id is shared*/
2619 rdma_return_path
->channel
= rdma
->channel
;
2620 rdma_return_path
->listen_id
= rdma
->listen_id
;
2622 rdma
->return_path
= rdma_return_path
;
2623 rdma_return_path
->return_path
= rdma
;
2624 rdma_return_path
->is_return_path
= true;
2627 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2629 RDMAContext
*rdma
= NULL
;
2630 InetSocketAddress
*addr
;
2633 rdma
= g_new0(RDMAContext
, 1);
2634 rdma
->current_index
= -1;
2635 rdma
->current_chunk
= -1;
2637 addr
= g_new(InetSocketAddress
, 1);
2638 if (!inet_parse(addr
, host_port
, NULL
)) {
2639 rdma
->port
= atoi(addr
->port
);
2640 rdma
->host
= g_strdup(addr
->host
);
2642 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2647 qapi_free_InetSocketAddress(addr
);
2654 * QEMUFile interface to the control channel.
2655 * SEND messages for control only.
2656 * VM's ram is handled with regular RDMA messages.
2658 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2659 const struct iovec
*iov
,
2665 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2666 QEMUFile
*f
= rioc
->file
;
2674 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
2681 CHECK_ERROR_STATE();
2684 * Push out any writes that
2685 * we're queued up for VM's ram.
2687 ret
= qemu_rdma_write_flush(f
, rdma
);
2689 rdma
->error_state
= ret
;
2694 for (i
= 0; i
< niov
; i
++) {
2695 size_t remaining
= iov
[i
].iov_len
;
2696 uint8_t * data
= (void *)iov
[i
].iov_base
;
2698 RDMAControlHeader head
;
2700 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2704 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2706 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2709 rdma
->error_state
= ret
;
2723 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2724 size_t size
, int idx
)
2728 if (rdma
->wr_data
[idx
].control_len
) {
2729 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2731 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2732 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2733 rdma
->wr_data
[idx
].control_curr
+= len
;
2734 rdma
->wr_data
[idx
].control_len
-= len
;
2741 * QEMUFile interface to the control channel.
2742 * RDMA links don't use bytestreams, so we have to
2743 * return bytes to QEMUFile opportunistically.
2745 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2746 const struct iovec
*iov
,
2752 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2754 RDMAControlHeader head
;
2760 rdma
= atomic_rcu_read(&rioc
->rdmain
);
2767 CHECK_ERROR_STATE();
2769 for (i
= 0; i
< niov
; i
++) {
2770 size_t want
= iov
[i
].iov_len
;
2771 uint8_t *data
= (void *)iov
[i
].iov_base
;
2774 * First, we hold on to the last SEND message we
2775 * were given and dish out the bytes until we run
2778 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2781 /* Got what we needed, so go to next iovec */
2786 /* If we got any data so far, then don't wait
2787 * for more, just return what we have */
2793 /* We've got nothing at all, so lets wait for
2796 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2799 rdma
->error_state
= ret
;
2805 * SEND was received with new bytes, now try again.
2807 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2811 /* Still didn't get enough, so lets just return */
2815 return QIO_CHANNEL_ERR_BLOCK
;
2826 * Block until all the outstanding chunks have been delivered by the hardware.
2828 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2832 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2836 while (rdma
->nb_sent
) {
2837 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2839 error_report("rdma migration: complete polling error!");
2844 qemu_rdma_unregister_waiting(rdma
);
2850 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2854 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2855 /* XXX we should make readv/writev actually honour this :-) */
2856 rioc
->blocking
= blocking
;
2861 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2862 struct QIOChannelRDMASource
{
2864 QIOChannelRDMA
*rioc
;
2865 GIOCondition condition
;
2869 qio_channel_rdma_source_prepare(GSource
*source
,
2872 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2874 GIOCondition cond
= 0;
2878 if (rsource
->condition
== G_IO_IN
) {
2879 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2881 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2885 error_report("RDMAContext is NULL when prepare Gsource");
2890 if (rdma
->wr_data
[0].control_len
) {
2896 return cond
& rsource
->condition
;
2900 qio_channel_rdma_source_check(GSource
*source
)
2902 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2904 GIOCondition cond
= 0;
2907 if (rsource
->condition
== G_IO_IN
) {
2908 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2910 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2914 error_report("RDMAContext is NULL when check Gsource");
2919 if (rdma
->wr_data
[0].control_len
) {
2925 return cond
& rsource
->condition
;
2929 qio_channel_rdma_source_dispatch(GSource
*source
,
2930 GSourceFunc callback
,
2933 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
2934 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2936 GIOCondition cond
= 0;
2939 if (rsource
->condition
== G_IO_IN
) {
2940 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2942 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2946 error_report("RDMAContext is NULL when dispatch Gsource");
2951 if (rdma
->wr_data
[0].control_len
) {
2957 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
2958 (cond
& rsource
->condition
),
2963 qio_channel_rdma_source_finalize(GSource
*source
)
2965 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
2967 object_unref(OBJECT(ssource
->rioc
));
2970 GSourceFuncs qio_channel_rdma_source_funcs
= {
2971 qio_channel_rdma_source_prepare
,
2972 qio_channel_rdma_source_check
,
2973 qio_channel_rdma_source_dispatch
,
2974 qio_channel_rdma_source_finalize
2977 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
2978 GIOCondition condition
)
2980 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2981 QIOChannelRDMASource
*ssource
;
2984 source
= g_source_new(&qio_channel_rdma_source_funcs
,
2985 sizeof(QIOChannelRDMASource
));
2986 ssource
= (QIOChannelRDMASource
*)source
;
2988 ssource
->rioc
= rioc
;
2989 object_ref(OBJECT(rioc
));
2991 ssource
->condition
= condition
;
2996 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
2999 IOHandler
*io_write
,
3002 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3004 aio_set_fd_handler(ctx
, rioc
->rdmain
->comp_channel
->fd
,
3005 false, io_read
, io_write
, NULL
, opaque
);
3007 aio_set_fd_handler(ctx
, rioc
->rdmaout
->comp_channel
->fd
,
3008 false, io_read
, io_write
, NULL
, opaque
);
3012 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3015 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3016 RDMAContext
*rdmain
, *rdmaout
;
3017 trace_qemu_rdma_close();
3019 rdmain
= rioc
->rdmain
;
3021 atomic_rcu_set(&rioc
->rdmain
, NULL
);
3024 rdmaout
= rioc
->rdmaout
;
3026 atomic_rcu_set(&rioc
->rdmaout
, NULL
);
3032 qemu_rdma_cleanup(rdmain
);
3036 qemu_rdma_cleanup(rdmaout
);
3046 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3047 QIOChannelShutdown how
,
3050 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3051 RDMAContext
*rdmain
, *rdmaout
;
3055 rdmain
= atomic_rcu_read(&rioc
->rdmain
);
3056 rdmaout
= atomic_rcu_read(&rioc
->rdmain
);
3059 case QIO_CHANNEL_SHUTDOWN_READ
:
3061 rdmain
->error_state
= -1;
3064 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3066 rdmaout
->error_state
= -1;
3069 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3072 rdmain
->error_state
= -1;
3075 rdmaout
->error_state
= -1;
3087 * This means that 'block_offset' is a full virtual address that does not
3088 * belong to a RAMBlock of the virtual machine and instead
3089 * represents a private malloc'd memory area that the caller wishes to
3093 * Offset is an offset to be added to block_offset and used
3094 * to also lookup the corresponding RAMBlock.
3097 * Initiate an transfer this size.
3100 * A 'hint' or 'advice' that means that we wish to speculatively
3101 * and asynchronously unregister this memory. In this case, there is no
3102 * guarantee that the unregister will actually happen, for example,
3103 * if the memory is being actively transmitted. Additionally, the memory
3104 * may be re-registered at any future time if a write within the same
3105 * chunk was requested again, even if you attempted to unregister it
3108 * @size < 0 : TODO, not yet supported
3109 * Unregister the memory NOW. This means that the caller does not
3110 * expect there to be any future RDMA transfers and we just want to clean
3111 * things up. This is used in case the upper layer owns the memory and
3112 * cannot wait for qemu_fclose() to occur.
3114 * @bytes_sent : User-specificed pointer to indicate how many bytes were
3115 * sent. Usually, this will not be more than a few bytes of
3116 * the protocol because most transfers are sent asynchronously.
3118 static size_t qemu_rdma_save_page(QEMUFile
*f
, void *opaque
,
3119 ram_addr_t block_offset
, ram_addr_t offset
,
3120 size_t size
, uint64_t *bytes_sent
)
3122 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3127 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3134 CHECK_ERROR_STATE();
3136 if (migrate_get_current()->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
3138 return RAM_SAVE_CONTROL_NOT_SUPP
;
3145 * Add this page to the current 'chunk'. If the chunk
3146 * is full, or the page doen't belong to the current chunk,
3147 * an actual RDMA write will occur and a new chunk will be formed.
3149 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
3151 error_report("rdma migration: write error! %d", ret
);
3156 * We always return 1 bytes because the RDMA
3157 * protocol is completely asynchronous. We do not yet know
3158 * whether an identified chunk is zero or not because we're
3159 * waiting for other pages to potentially be merged with
3160 * the current chunk. So, we have to call qemu_update_position()
3161 * later on when the actual write occurs.
3167 uint64_t index
, chunk
;
3169 /* TODO: Change QEMUFileOps prototype to be signed: size_t => long
3171 ret = qemu_rdma_drain_cq(f, rdma);
3173 fprintf(stderr, "rdma: failed to synchronously drain"
3174 " completion queue before unregistration.\n");
3180 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
3181 offset
, size
, &index
, &chunk
);
3184 error_report("ram block search failed");
3188 qemu_rdma_signal_unregister(rdma
, index
, chunk
, 0);
3191 * TODO: Synchronous, guaranteed unregistration (should not occur during
3192 * fast-path). Otherwise, unregisters will process on the next call to
3193 * qemu_rdma_drain_cq()
3195 qemu_rdma_unregister_waiting(rdma);
3201 * Drain the Completion Queue if possible, but do not block,
3204 * If nothing to poll, the end of the iteration will do this
3205 * again to make sure we don't overflow the request queue.
3208 uint64_t wr_id
, wr_id_in
;
3209 int ret
= qemu_rdma_poll(rdma
, &wr_id_in
, NULL
);
3211 error_report("rdma migration: polling error! %d", ret
);
3215 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3217 if (wr_id
== RDMA_WRID_NONE
) {
3223 return RAM_SAVE_CONTROL_DELAYED
;
3225 rdma
->error_state
= ret
;
3230 static void rdma_accept_incoming_migration(void *opaque
);
3232 static void rdma_cm_poll_handler(void *opaque
)
3234 RDMAContext
*rdma
= opaque
;
3236 struct rdma_cm_event
*cm_event
;
3237 MigrationIncomingState
*mis
= migration_incoming_get_current();
3239 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3241 error_report("get_cm_event failed %d", errno
);
3244 rdma_ack_cm_event(cm_event
);
3246 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3247 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3248 error_report("receive cm event, cm event is %d", cm_event
->event
);
3249 rdma
->error_state
= -EPIPE
;
3250 if (rdma
->return_path
) {
3251 rdma
->return_path
->error_state
= -EPIPE
;
3254 if (mis
->migration_incoming_co
) {
3255 qemu_coroutine_enter(mis
->migration_incoming_co
);
3261 static int qemu_rdma_accept(RDMAContext
*rdma
)
3263 RDMACapabilities cap
;
3264 struct rdma_conn_param conn_param
= {
3265 .responder_resources
= 2,
3266 .private_data
= &cap
,
3267 .private_data_len
= sizeof(cap
),
3269 struct rdma_cm_event
*cm_event
;
3270 struct ibv_context
*verbs
;
3274 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3276 goto err_rdma_dest_wait
;
3279 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3280 rdma_ack_cm_event(cm_event
);
3281 goto err_rdma_dest_wait
;
3284 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3286 network_to_caps(&cap
);
3288 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3289 error_report("Unknown source RDMA version: %d, bailing...",
3291 rdma_ack_cm_event(cm_event
);
3292 goto err_rdma_dest_wait
;
3296 * Respond with only the capabilities this version of QEMU knows about.
3298 cap
.flags
&= known_capabilities
;
3301 * Enable the ones that we do know about.
3302 * Add other checks here as new ones are introduced.
3304 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3305 rdma
->pin_all
= true;
3308 rdma
->cm_id
= cm_event
->id
;
3309 verbs
= cm_event
->id
->verbs
;
3311 rdma_ack_cm_event(cm_event
);
3313 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3315 caps_to_network(&cap
);
3317 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3320 rdma
->verbs
= verbs
;
3321 } else if (rdma
->verbs
!= verbs
) {
3322 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3324 goto err_rdma_dest_wait
;
3327 qemu_rdma_dump_id("dest_init", verbs
);
3329 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3331 error_report("rdma migration: error allocating pd and cq!");
3332 goto err_rdma_dest_wait
;
3335 ret
= qemu_rdma_alloc_qp(rdma
);
3337 error_report("rdma migration: error allocating qp!");
3338 goto err_rdma_dest_wait
;
3341 ret
= qemu_rdma_init_ram_blocks(rdma
);
3343 error_report("rdma migration: error initializing ram blocks!");
3344 goto err_rdma_dest_wait
;
3347 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3348 ret
= qemu_rdma_reg_control(rdma
, idx
);
3350 error_report("rdma: error registering %d control", idx
);
3351 goto err_rdma_dest_wait
;
3355 /* Accept the second connection request for return path */
3356 if (migrate_postcopy() && !rdma
->is_return_path
) {
3357 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3359 (void *)(intptr_t)rdma
->return_path
);
3361 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3365 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3367 error_report("rdma_accept returns %d", ret
);
3368 goto err_rdma_dest_wait
;
3371 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3373 error_report("rdma_accept get_cm_event failed %d", ret
);
3374 goto err_rdma_dest_wait
;
3377 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3378 error_report("rdma_accept not event established");
3379 rdma_ack_cm_event(cm_event
);
3380 goto err_rdma_dest_wait
;
3383 rdma_ack_cm_event(cm_event
);
3384 rdma
->connected
= true;
3386 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3388 error_report("rdma migration: error posting second control recv");
3389 goto err_rdma_dest_wait
;
3392 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3397 rdma
->error_state
= ret
;
3398 qemu_rdma_cleanup(rdma
);
3402 static int dest_ram_sort_func(const void *a
, const void *b
)
3404 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3405 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3407 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3411 * During each iteration of the migration, we listen for instructions
3412 * by the source VM to perform dynamic page registrations before they
3413 * can perform RDMA operations.
3415 * We respond with the 'rkey'.
3417 * Keep doing this until the source tells us to stop.
3419 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
)
3421 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3422 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3425 RDMAControlHeader unreg_resp
= { .len
= 0,
3426 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3429 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3431 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3433 RDMALocalBlocks
*local
;
3434 RDMAControlHeader head
;
3435 RDMARegister
*reg
, *registers
;
3437 RDMARegisterResult
*reg_result
;
3438 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3439 RDMALocalBlock
*block
;
3447 rdma
= atomic_rcu_read(&rioc
->rdmain
);
3454 CHECK_ERROR_STATE();
3456 local
= &rdma
->local_ram_blocks
;
3458 trace_qemu_rdma_registration_handle_wait();
3460 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3466 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3467 error_report("rdma: Too many requests in this message (%d)."
3468 "Bailing.", head
.repeat
);
3473 switch (head
.type
) {
3474 case RDMA_CONTROL_COMPRESS
:
3475 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3476 network_to_compress(comp
);
3478 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3481 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3482 error_report("rdma: 'compress' bad block index %u (vs %d)",
3483 (unsigned int)comp
->block_idx
,
3484 rdma
->local_ram_blocks
.nb_blocks
);
3488 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3490 host_addr
= block
->local_host_addr
+
3491 (comp
->offset
- block
->offset
);
3493 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3496 case RDMA_CONTROL_REGISTER_FINISHED
:
3497 trace_qemu_rdma_registration_handle_finished();
3500 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3501 trace_qemu_rdma_registration_handle_ram_blocks();
3503 /* Sort our local RAM Block list so it's the same as the source,
3504 * we can do this since we've filled in a src_index in the list
3505 * as we received the RAMBlock list earlier.
3507 qsort(rdma
->local_ram_blocks
.block
,
3508 rdma
->local_ram_blocks
.nb_blocks
,
3509 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3510 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3511 local
->block
[i
].index
= i
;
3514 if (rdma
->pin_all
) {
3515 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3517 error_report("rdma migration: error dest "
3518 "registering ram blocks");
3524 * Dest uses this to prepare to transmit the RAMBlock descriptions
3525 * to the source VM after connection setup.
3526 * Both sides use the "remote" structure to communicate and update
3527 * their "local" descriptions with what was sent.
3529 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3530 rdma
->dest_blocks
[i
].remote_host_addr
=
3531 (uintptr_t)(local
->block
[i
].local_host_addr
);
3533 if (rdma
->pin_all
) {
3534 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3537 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3538 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3540 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3541 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3542 local
->block
[i
].block_name
,
3543 local
->block
[i
].offset
,
3544 local
->block
[i
].length
,
3545 local
->block
[i
].local_host_addr
,
3546 local
->block
[i
].src_index
);
3549 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3550 * sizeof(RDMADestBlock
);
3553 ret
= qemu_rdma_post_send_control(rdma
,
3554 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3557 error_report("rdma migration: error sending remote info");
3562 case RDMA_CONTROL_REGISTER_REQUEST
:
3563 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3565 reg_resp
.repeat
= head
.repeat
;
3566 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3568 for (count
= 0; count
< head
.repeat
; count
++) {
3570 uint8_t *chunk_start
, *chunk_end
;
3572 reg
= ®isters
[count
];
3573 network_to_register(reg
);
3575 reg_result
= &results
[count
];
3577 trace_qemu_rdma_registration_handle_register_loop(count
,
3578 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3580 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3581 error_report("rdma: 'register' bad block index %u (vs %d)",
3582 (unsigned int)reg
->current_index
,
3583 rdma
->local_ram_blocks
.nb_blocks
);
3587 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3588 if (block
->is_ram_block
) {
3589 if (block
->offset
> reg
->key
.current_addr
) {
3590 error_report("rdma: bad register address for block %s"
3591 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3592 block
->block_name
, block
->offset
,
3593 reg
->key
.current_addr
);
3597 host_addr
= (block
->local_host_addr
+
3598 (reg
->key
.current_addr
- block
->offset
));
3599 chunk
= ram_chunk_index(block
->local_host_addr
,
3600 (uint8_t *) host_addr
);
3602 chunk
= reg
->key
.chunk
;
3603 host_addr
= block
->local_host_addr
+
3604 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3605 /* Check for particularly bad chunk value */
3606 if (host_addr
< (void *)block
->local_host_addr
) {
3607 error_report("rdma: bad chunk for block %s"
3609 block
->block_name
, reg
->key
.chunk
);
3614 chunk_start
= ram_chunk_start(block
, chunk
);
3615 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3616 /* avoid "-Waddress-of-packed-member" warning */
3617 uint32_t tmp_rkey
= 0;
3618 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3619 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3620 chunk
, chunk_start
, chunk_end
)) {
3621 error_report("cannot get rkey");
3625 reg_result
->rkey
= tmp_rkey
;
3627 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3629 trace_qemu_rdma_registration_handle_register_rkey(
3632 result_to_network(reg_result
);
3635 ret
= qemu_rdma_post_send_control(rdma
,
3636 (uint8_t *) results
, ®_resp
);
3639 error_report("Failed to send control buffer");
3643 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3644 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3645 unreg_resp
.repeat
= head
.repeat
;
3646 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3648 for (count
= 0; count
< head
.repeat
; count
++) {
3649 reg
= ®isters
[count
];
3650 network_to_register(reg
);
3652 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3653 reg
->current_index
, reg
->key
.chunk
);
3655 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3657 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3658 block
->pmr
[reg
->key
.chunk
] = NULL
;
3661 perror("rdma unregistration chunk failed");
3666 rdma
->total_registrations
--;
3668 trace_qemu_rdma_registration_handle_unregister_success(
3672 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3675 error_report("Failed to send control buffer");
3679 case RDMA_CONTROL_REGISTER_RESULT
:
3680 error_report("Invalid RESULT message at dest.");
3684 error_report("Unknown control message %s", control_desc(head
.type
));
3691 rdma
->error_state
= ret
;
3698 * Called via a ram_control_load_hook during the initial RAM load section which
3699 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3701 * We've already built our local RAMBlock list, but not yet sent the list to
3705 rdma_block_notification_handle(QIOChannelRDMA
*rioc
, const char *name
)
3712 rdma
= atomic_rcu_read(&rioc
->rdmain
);
3719 /* Find the matching RAMBlock in our local list */
3720 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3721 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3728 error_report("RAMBlock '%s' not found on destination", name
);
3733 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3734 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3735 rdma
->next_src_index
++;
3741 static int rdma_load_hook(QEMUFile
*f
, void *opaque
, uint64_t flags
, void *data
)
3744 case RAM_CONTROL_BLOCK_REG
:
3745 return rdma_block_notification_handle(opaque
, data
);
3747 case RAM_CONTROL_HOOK
:
3748 return qemu_rdma_registration_handle(f
, opaque
);
3751 /* Shouldn't be called with any other values */
3756 static int qemu_rdma_registration_start(QEMUFile
*f
, void *opaque
,
3757 uint64_t flags
, void *data
)
3759 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3763 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3769 CHECK_ERROR_STATE();
3771 if (migrate_get_current()->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
3776 trace_qemu_rdma_registration_start(flags
);
3777 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3785 * Inform dest that dynamic registrations are done for now.
3786 * First, flush writes, if any.
3788 static int qemu_rdma_registration_stop(QEMUFile
*f
, void *opaque
,
3789 uint64_t flags
, void *data
)
3791 Error
*local_err
= NULL
, **errp
= &local_err
;
3792 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3794 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3798 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3804 CHECK_ERROR_STATE();
3806 if (migrate_get_current()->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
3812 ret
= qemu_rdma_drain_cq(f
, rdma
);
3818 if (flags
== RAM_CONTROL_SETUP
) {
3819 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3820 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3821 int reg_result_idx
, i
, nb_dest_blocks
;
3823 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3824 trace_qemu_rdma_registration_stop_ram();
3827 * Make sure that we parallelize the pinning on both sides.
3828 * For very large guests, doing this serially takes a really
3829 * long time, so we have to 'interleave' the pinning locally
3830 * with the control messages by performing the pinning on this
3831 * side before we receive the control response from the other
3832 * side that the pinning has completed.
3834 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3835 ®_result_idx
, rdma
->pin_all
?
3836 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3838 ERROR(errp
, "receiving remote info!");
3843 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3846 * The protocol uses two different sets of rkeys (mutually exclusive):
3847 * 1. One key to represent the virtual address of the entire ram block.
3848 * (dynamic chunk registration disabled - pin everything with one rkey.)
3849 * 2. One to represent individual chunks within a ram block.
3850 * (dynamic chunk registration enabled - pin individual chunks.)
3852 * Once the capability is successfully negotiated, the destination transmits
3853 * the keys to use (or sends them later) including the virtual addresses
3854 * and then propagates the remote ram block descriptions to his local copy.
3857 if (local
->nb_blocks
!= nb_dest_blocks
) {
3858 ERROR(errp
, "ram blocks mismatch (Number of blocks %d vs %d) "
3859 "Your QEMU command line parameters are probably "
3860 "not identical on both the source and destination.",
3861 local
->nb_blocks
, nb_dest_blocks
);
3862 rdma
->error_state
= -EINVAL
;
3867 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3868 memcpy(rdma
->dest_blocks
,
3869 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3870 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3871 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3873 /* We require that the blocks are in the same order */
3874 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3875 ERROR(errp
, "Block %s/%d has a different length %" PRIu64
3876 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3877 local
->block
[i
].length
,
3878 rdma
->dest_blocks
[i
].length
);
3879 rdma
->error_state
= -EINVAL
;
3883 local
->block
[i
].remote_host_addr
=
3884 rdma
->dest_blocks
[i
].remote_host_addr
;
3885 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3889 trace_qemu_rdma_registration_stop(flags
);
3891 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3892 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3901 rdma
->error_state
= ret
;
3906 static const QEMUFileHooks rdma_read_hooks
= {
3907 .hook_ram_load
= rdma_load_hook
,
3910 static const QEMUFileHooks rdma_write_hooks
= {
3911 .before_ram_iterate
= qemu_rdma_registration_start
,
3912 .after_ram_iterate
= qemu_rdma_registration_stop
,
3913 .save_page
= qemu_rdma_save_page
,
3917 static void qio_channel_rdma_finalize(Object
*obj
)
3919 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
3921 qemu_rdma_cleanup(rioc
->rdmain
);
3922 g_free(rioc
->rdmain
);
3923 rioc
->rdmain
= NULL
;
3925 if (rioc
->rdmaout
) {
3926 qemu_rdma_cleanup(rioc
->rdmaout
);
3927 g_free(rioc
->rdmaout
);
3928 rioc
->rdmaout
= NULL
;
3932 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
3933 void *class_data G_GNUC_UNUSED
)
3935 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
3937 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
3938 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
3939 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
3940 ioc_klass
->io_close
= qio_channel_rdma_close
;
3941 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
3942 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
3943 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
3946 static const TypeInfo qio_channel_rdma_info
= {
3947 .parent
= TYPE_QIO_CHANNEL
,
3948 .name
= TYPE_QIO_CHANNEL_RDMA
,
3949 .instance_size
= sizeof(QIOChannelRDMA
),
3950 .instance_finalize
= qio_channel_rdma_finalize
,
3951 .class_init
= qio_channel_rdma_class_init
,
3954 static void qio_channel_rdma_register_types(void)
3956 type_register_static(&qio_channel_rdma_info
);
3959 type_init(qio_channel_rdma_register_types
);
3961 static QEMUFile
*qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
3963 QIOChannelRDMA
*rioc
;
3965 if (qemu_file_mode_is_not_valid(mode
)) {
3969 rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
3971 if (mode
[0] == 'w') {
3972 rioc
->file
= qemu_fopen_channel_output(QIO_CHANNEL(rioc
));
3973 rioc
->rdmaout
= rdma
;
3974 rioc
->rdmain
= rdma
->return_path
;
3975 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
3977 rioc
->file
= qemu_fopen_channel_input(QIO_CHANNEL(rioc
));
3978 rioc
->rdmain
= rdma
;
3979 rioc
->rdmaout
= rdma
->return_path
;
3980 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
3986 static void rdma_accept_incoming_migration(void *opaque
)
3988 RDMAContext
*rdma
= opaque
;
3991 Error
*local_err
= NULL
, **errp
= &local_err
;
3993 trace_qemu_rdma_accept_incoming_migration();
3994 ret
= qemu_rdma_accept(rdma
);
3997 ERROR(errp
, "RDMA Migration initialization failed!");
4001 trace_qemu_rdma_accept_incoming_migration_accepted();
4003 if (rdma
->is_return_path
) {
4007 f
= qemu_fopen_rdma(rdma
, "rb");
4009 ERROR(errp
, "could not qemu_fopen_rdma!");
4010 qemu_rdma_cleanup(rdma
);
4014 rdma
->migration_started_on_destination
= 1;
4015 migration_fd_process_incoming(f
);
4018 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4021 RDMAContext
*rdma
, *rdma_return_path
= NULL
;
4022 Error
*local_err
= NULL
;
4024 trace_rdma_start_incoming_migration();
4025 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
4031 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
4037 trace_rdma_start_incoming_migration_after_dest_init();
4039 ret
= rdma_listen(rdma
->listen_id
, 5);
4042 ERROR(errp
, "listening on socket!");
4046 trace_rdma_start_incoming_migration_after_rdma_listen();
4048 /* initialize the RDMAContext for return path */
4049 if (migrate_postcopy()) {
4050 rdma_return_path
= qemu_rdma_data_init(host_port
, &local_err
);
4052 if (rdma_return_path
== NULL
) {
4056 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
4059 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4060 NULL
, (void *)(intptr_t)rdma
);
4063 error_propagate(errp
, local_err
);
4065 g_free(rdma_return_path
);
4068 void rdma_start_outgoing_migration(void *opaque
,
4069 const char *host_port
, Error
**errp
)
4071 MigrationState
*s
= opaque
;
4072 RDMAContext
*rdma
= qemu_rdma_data_init(host_port
, errp
);
4073 RDMAContext
*rdma_return_path
= NULL
;
4080 ret
= qemu_rdma_source_init(rdma
,
4081 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4087 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4088 ret
= qemu_rdma_connect(rdma
, errp
);
4094 /* RDMA postcopy need a seprate queue pair for return path */
4095 if (migrate_postcopy()) {
4096 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4098 if (rdma_return_path
== NULL
) {
4102 ret
= qemu_rdma_source_init(rdma_return_path
,
4103 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4109 ret
= qemu_rdma_connect(rdma_return_path
, errp
);
4115 rdma
->return_path
= rdma_return_path
;
4116 rdma_return_path
->return_path
= rdma
;
4117 rdma_return_path
->is_return_path
= true;
4120 trace_rdma_start_outgoing_migration_after_rdma_connect();
4122 s
->to_dst_file
= qemu_fopen_rdma(rdma
, "wb");
4123 migrate_fd_connect(s
, NULL
);
4127 g_free(rdma_return_path
);