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(RAMBlock
*rb
, void *opaque
)
629 const char *block_name
= qemu_ram_get_idstr(rb
);
630 void *host_addr
= qemu_ram_get_host_addr(rb
);
631 ram_addr_t block_offset
= qemu_ram_get_offset(rb
);
632 ram_addr_t length
= qemu_ram_get_used_length(rb
);
633 return rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
637 * Identify the RAMBlocks and their quantity. They will be references to
638 * identify chunk boundaries inside each RAMBlock and also be referenced
639 * during dynamic page registration.
641 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
643 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
645 assert(rdma
->blockmap
== NULL
);
646 memset(local
, 0, sizeof *local
);
647 foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
648 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
649 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
650 rdma
->local_ram_blocks
.nb_blocks
);
656 * Note: If used outside of cleanup, the caller must ensure that the destination
657 * block structures are also updated
659 static int rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
661 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
662 RDMALocalBlock
*old
= local
->block
;
665 if (rdma
->blockmap
) {
666 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
671 for (j
= 0; j
< block
->nb_chunks
; j
++) {
672 if (!block
->pmr
[j
]) {
675 ibv_dereg_mr(block
->pmr
[j
]);
676 rdma
->total_registrations
--;
683 ibv_dereg_mr(block
->mr
);
684 rdma
->total_registrations
--;
688 g_free(block
->transit_bitmap
);
689 block
->transit_bitmap
= NULL
;
691 g_free(block
->unregister_bitmap
);
692 block
->unregister_bitmap
= NULL
;
694 g_free(block
->remote_keys
);
695 block
->remote_keys
= NULL
;
697 g_free(block
->block_name
);
698 block
->block_name
= NULL
;
700 if (rdma
->blockmap
) {
701 for (x
= 0; x
< local
->nb_blocks
; x
++) {
702 g_hash_table_remove(rdma
->blockmap
,
703 (void *)(uintptr_t)old
[x
].offset
);
707 if (local
->nb_blocks
> 1) {
709 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
712 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
715 if (block
->index
< (local
->nb_blocks
- 1)) {
716 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
717 sizeof(RDMALocalBlock
) *
718 (local
->nb_blocks
- (block
->index
+ 1)));
719 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
720 local
->block
[x
].index
--;
724 assert(block
== local
->block
);
728 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
729 block
->offset
, block
->length
,
730 (uintptr_t)(block
->local_host_addr
+ block
->length
),
731 BITS_TO_LONGS(block
->nb_chunks
) *
732 sizeof(unsigned long) * 8, block
->nb_chunks
);
738 if (local
->nb_blocks
&& rdma
->blockmap
) {
739 for (x
= 0; x
< local
->nb_blocks
; x
++) {
740 g_hash_table_insert(rdma
->blockmap
,
741 (void *)(uintptr_t)local
->block
[x
].offset
,
750 * Put in the log file which RDMA device was opened and the details
751 * associated with that device.
753 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
755 struct ibv_port_attr port
;
757 if (ibv_query_port(verbs
, 1, &port
)) {
758 error_report("Failed to query port information");
762 printf("%s RDMA Device opened: kernel name %s "
763 "uverbs device name %s, "
764 "infiniband_verbs class device path %s, "
765 "infiniband class device path %s, "
766 "transport: (%d) %s\n",
769 verbs
->device
->dev_name
,
770 verbs
->device
->dev_path
,
771 verbs
->device
->ibdev_path
,
773 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
774 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
775 ? "Ethernet" : "Unknown"));
779 * Put in the log file the RDMA gid addressing information,
780 * useful for folks who have trouble understanding the
781 * RDMA device hierarchy in the kernel.
783 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
787 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
788 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
789 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
793 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
794 * We will try the next addrinfo struct, and fail if there are
795 * no other valid addresses to bind against.
797 * If user is listening on '[::]', then we will not have a opened a device
798 * yet and have no way of verifying if the device is RoCE or not.
800 * In this case, the source VM will throw an error for ALL types of
801 * connections (both IPv4 and IPv6) if the destination machine does not have
802 * a regular infiniband network available for use.
804 * The only way to guarantee that an error is thrown for broken kernels is
805 * for the management software to choose a *specific* interface at bind time
806 * and validate what time of hardware it is.
808 * Unfortunately, this puts the user in a fix:
810 * If the source VM connects with an IPv4 address without knowing that the
811 * destination has bound to '[::]' the migration will unconditionally fail
812 * unless the management software is explicitly listening on the IPv4
813 * address while using a RoCE-based device.
815 * If the source VM connects with an IPv6 address, then we're OK because we can
816 * throw an error on the source (and similarly on the destination).
818 * But in mixed environments, this will be broken for a while until it is fixed
821 * We do provide a *tiny* bit of help in this function: We can list all of the
822 * devices in the system and check to see if all the devices are RoCE or
825 * If we detect that we have a *pure* RoCE environment, then we can safely
826 * thrown an error even if the management software has specified '[::]' as the
829 * However, if there is are multiple hetergeneous devices, then we cannot make
830 * this assumption and the user just has to be sure they know what they are
833 * Patches are being reviewed on linux-rdma.
835 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
837 struct ibv_port_attr port_attr
;
839 /* This bug only exists in linux, to our knowledge. */
843 * Verbs are only NULL if management has bound to '[::]'.
845 * Let's iterate through all the devices and see if there any pure IB
846 * devices (non-ethernet).
848 * If not, then we can safely proceed with the migration.
849 * Otherwise, there are no guarantees until the bug is fixed in linux.
853 struct ibv_device
** dev_list
= ibv_get_device_list(&num_devices
);
854 bool roce_found
= false;
855 bool ib_found
= false;
857 for (x
= 0; x
< num_devices
; x
++) {
858 verbs
= ibv_open_device(dev_list
[x
]);
860 if (errno
== EPERM
) {
867 if (ibv_query_port(verbs
, 1, &port_attr
)) {
868 ibv_close_device(verbs
);
869 ERROR(errp
, "Could not query initial IB port");
873 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
875 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
879 ibv_close_device(verbs
);
885 fprintf(stderr
, "WARN: migrations may fail:"
886 " IPv6 over RoCE / iWARP in linux"
887 " is broken. But since you appear to have a"
888 " mixed RoCE / IB environment, be sure to only"
889 " migrate over the IB fabric until the kernel "
890 " fixes the bug.\n");
892 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
893 " and your management software has specified '[::]'"
894 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
903 * If we have a verbs context, that means that some other than '[::]' was
904 * used by the management software for binding. In which case we can
905 * actually warn the user about a potentially broken kernel.
908 /* IB ports start with 1, not 0 */
909 if (ibv_query_port(verbs
, 1, &port_attr
)) {
910 ERROR(errp
, "Could not query initial IB port");
914 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
915 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
916 "(but patches on linux-rdma in progress)");
926 * Figure out which RDMA device corresponds to the requested IP hostname
927 * Also create the initial connection manager identifiers for opening
930 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
933 struct rdma_addrinfo
*res
;
935 struct rdma_cm_event
*cm_event
;
936 char ip
[40] = "unknown";
937 struct rdma_addrinfo
*e
;
939 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
940 ERROR(errp
, "RDMA hostname has not been set");
944 /* create CM channel */
945 rdma
->channel
= rdma_create_event_channel();
946 if (!rdma
->channel
) {
947 ERROR(errp
, "could not create CM channel");
952 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
954 ERROR(errp
, "could not create channel id");
955 goto err_resolve_create_id
;
958 snprintf(port_str
, 16, "%d", rdma
->port
);
961 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
963 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
964 goto err_resolve_get_addr
;
967 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
968 inet_ntop(e
->ai_family
,
969 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
970 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
972 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
973 RDMA_RESOLVE_TIMEOUT_MS
);
975 if (e
->ai_family
== AF_INET6
) {
976 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
985 ERROR(errp
, "could not resolve address %s", rdma
->host
);
986 goto err_resolve_get_addr
;
989 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
991 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
993 ERROR(errp
, "could not perform event_addr_resolved");
994 goto err_resolve_get_addr
;
997 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
998 ERROR(errp
, "result not equal to event_addr_resolved %s",
999 rdma_event_str(cm_event
->event
));
1000 perror("rdma_resolve_addr");
1001 rdma_ack_cm_event(cm_event
);
1003 goto err_resolve_get_addr
;
1005 rdma_ack_cm_event(cm_event
);
1008 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1010 ERROR(errp
, "could not resolve rdma route");
1011 goto err_resolve_get_addr
;
1014 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1016 ERROR(errp
, "could not perform event_route_resolved");
1017 goto err_resolve_get_addr
;
1019 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1020 ERROR(errp
, "result not equal to event_route_resolved: %s",
1021 rdma_event_str(cm_event
->event
));
1022 rdma_ack_cm_event(cm_event
);
1024 goto err_resolve_get_addr
;
1026 rdma_ack_cm_event(cm_event
);
1027 rdma
->verbs
= rdma
->cm_id
->verbs
;
1028 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1029 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1032 err_resolve_get_addr
:
1033 rdma_destroy_id(rdma
->cm_id
);
1035 err_resolve_create_id
:
1036 rdma_destroy_event_channel(rdma
->channel
);
1037 rdma
->channel
= NULL
;
1042 * Create protection domain and completion queues
1044 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1047 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1049 error_report("failed to allocate protection domain");
1053 /* create completion channel */
1054 rdma
->comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1055 if (!rdma
->comp_channel
) {
1056 error_report("failed to allocate completion channel");
1057 goto err_alloc_pd_cq
;
1061 * Completion queue can be filled by both read and write work requests,
1062 * so must reflect the sum of both possible queue sizes.
1064 rdma
->cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1065 NULL
, rdma
->comp_channel
, 0);
1067 error_report("failed to allocate completion queue");
1068 goto err_alloc_pd_cq
;
1075 ibv_dealloc_pd(rdma
->pd
);
1077 if (rdma
->comp_channel
) {
1078 ibv_destroy_comp_channel(rdma
->comp_channel
);
1081 rdma
->comp_channel
= NULL
;
1087 * Create queue pairs.
1089 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1091 struct ibv_qp_init_attr attr
= { 0 };
1094 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1095 attr
.cap
.max_recv_wr
= 3;
1096 attr
.cap
.max_send_sge
= 1;
1097 attr
.cap
.max_recv_sge
= 1;
1098 attr
.send_cq
= rdma
->cq
;
1099 attr
.recv_cq
= rdma
->cq
;
1100 attr
.qp_type
= IBV_QPT_RC
;
1102 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1107 rdma
->qp
= rdma
->cm_id
->qp
;
1111 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1114 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1116 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1117 local
->block
[i
].mr
=
1118 ibv_reg_mr(rdma
->pd
,
1119 local
->block
[i
].local_host_addr
,
1120 local
->block
[i
].length
,
1121 IBV_ACCESS_LOCAL_WRITE
|
1122 IBV_ACCESS_REMOTE_WRITE
1124 if (!local
->block
[i
].mr
) {
1125 perror("Failed to register local dest ram block!\n");
1128 rdma
->total_registrations
++;
1131 if (i
>= local
->nb_blocks
) {
1135 for (i
--; i
>= 0; i
--) {
1136 ibv_dereg_mr(local
->block
[i
].mr
);
1137 rdma
->total_registrations
--;
1145 * Find the ram block that corresponds to the page requested to be
1146 * transmitted by QEMU.
1148 * Once the block is found, also identify which 'chunk' within that
1149 * block that the page belongs to.
1151 * This search cannot fail or the migration will fail.
1153 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1154 uintptr_t block_offset
,
1157 uint64_t *block_index
,
1158 uint64_t *chunk_index
)
1160 uint64_t current_addr
= block_offset
+ offset
;
1161 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1162 (void *) block_offset
);
1164 assert(current_addr
>= block
->offset
);
1165 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1167 *block_index
= block
->index
;
1168 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1169 block
->local_host_addr
+ (current_addr
- block
->offset
));
1175 * Register a chunk with IB. If the chunk was already registered
1176 * previously, then skip.
1178 * Also return the keys associated with the registration needed
1179 * to perform the actual RDMA operation.
1181 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1182 RDMALocalBlock
*block
, uintptr_t host_addr
,
1183 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1184 uint8_t *chunk_start
, uint8_t *chunk_end
)
1188 *lkey
= block
->mr
->lkey
;
1191 *rkey
= block
->mr
->rkey
;
1196 /* allocate memory to store chunk MRs */
1198 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1202 * If 'rkey', then we're the destination, so grant access to the source.
1204 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1206 if (!block
->pmr
[chunk
]) {
1207 uint64_t len
= chunk_end
- chunk_start
;
1209 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1211 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
,
1213 (rkey
? (IBV_ACCESS_LOCAL_WRITE
|
1214 IBV_ACCESS_REMOTE_WRITE
) : 0));
1216 if (!block
->pmr
[chunk
]) {
1217 perror("Failed to register chunk!");
1218 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1219 " start %" PRIuPTR
" end %" PRIuPTR
1221 " local %" PRIuPTR
" registrations: %d\n",
1222 block
->index
, chunk
, (uintptr_t)chunk_start
,
1223 (uintptr_t)chunk_end
, host_addr
,
1224 (uintptr_t)block
->local_host_addr
,
1225 rdma
->total_registrations
);
1228 rdma
->total_registrations
++;
1232 *lkey
= block
->pmr
[chunk
]->lkey
;
1235 *rkey
= block
->pmr
[chunk
]->rkey
;
1241 * Register (at connection time) the memory used for control
1244 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1246 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1247 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1248 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1249 if (rdma
->wr_data
[idx
].control_mr
) {
1250 rdma
->total_registrations
++;
1253 error_report("qemu_rdma_reg_control failed");
1257 const char *print_wrid(int wrid
)
1259 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1260 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1262 return wrid_desc
[wrid
];
1266 * RDMA requires memory registration (mlock/pinning), but this is not good for
1269 * In preparation for the future where LRU information or workload-specific
1270 * writable writable working set memory access behavior is available to QEMU
1271 * it would be nice to have in place the ability to UN-register/UN-pin
1272 * particular memory regions from the RDMA hardware when it is determine that
1273 * those regions of memory will likely not be accessed again in the near future.
1275 * While we do not yet have such information right now, the following
1276 * compile-time option allows us to perform a non-optimized version of this
1279 * By uncommenting this option, you will cause *all* RDMA transfers to be
1280 * unregistered immediately after the transfer completes on both sides of the
1281 * connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
1283 * This will have a terrible impact on migration performance, so until future
1284 * workload information or LRU information is available, do not attempt to use
1285 * this feature except for basic testing.
1287 //#define RDMA_UNREGISTRATION_EXAMPLE
1290 * Perform a non-optimized memory unregistration after every transfer
1291 * for demonstration purposes, only if pin-all is not requested.
1293 * Potential optimizations:
1294 * 1. Start a new thread to run this function continuously
1296 - and for receipt of unregister messages
1298 * 3. Use workload hints.
1300 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1302 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1304 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1306 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1308 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1309 RDMALocalBlock
*block
=
1310 &(rdma
->local_ram_blocks
.block
[index
]);
1311 RDMARegister reg
= { .current_index
= index
};
1312 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1314 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1315 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1319 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1320 rdma
->unregister_current
);
1322 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1323 rdma
->unregister_current
++;
1325 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1326 rdma
->unregister_current
= 0;
1331 * Unregistration is speculative (because migration is single-threaded
1332 * and we cannot break the protocol's inifinband message ordering).
1333 * Thus, if the memory is currently being used for transmission,
1334 * then abort the attempt to unregister and try again
1335 * later the next time a completion is received for this memory.
1337 clear_bit(chunk
, block
->unregister_bitmap
);
1339 if (test_bit(chunk
, block
->transit_bitmap
)) {
1340 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1344 trace_qemu_rdma_unregister_waiting_send(chunk
);
1346 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1347 block
->pmr
[chunk
] = NULL
;
1348 block
->remote_keys
[chunk
] = 0;
1351 perror("unregistration chunk failed");
1354 rdma
->total_registrations
--;
1356 reg
.key
.chunk
= chunk
;
1357 register_to_network(rdma
, ®
);
1358 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1364 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1370 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1373 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1375 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1376 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1382 * Set bit for unregistration in the next iteration.
1383 * We cannot transmit right here, but will unpin later.
1385 static void qemu_rdma_signal_unregister(RDMAContext
*rdma
, uint64_t index
,
1386 uint64_t chunk
, uint64_t wr_id
)
1388 if (rdma
->unregistrations
[rdma
->unregister_next
] != 0) {
1389 error_report("rdma migration: queue is full");
1391 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1393 if (!test_and_set_bit(chunk
, block
->unregister_bitmap
)) {
1394 trace_qemu_rdma_signal_unregister_append(chunk
,
1395 rdma
->unregister_next
);
1397 rdma
->unregistrations
[rdma
->unregister_next
++] =
1398 qemu_rdma_make_wrid(wr_id
, index
, chunk
);
1400 if (rdma
->unregister_next
== RDMA_SIGNALED_SEND_MAX
) {
1401 rdma
->unregister_next
= 0;
1404 trace_qemu_rdma_signal_unregister_already(chunk
);
1410 * Consult the connection manager to see a work request
1411 * (of any kind) has completed.
1412 * Return the work request ID that completed.
1414 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, uint64_t *wr_id_out
,
1421 ret
= ibv_poll_cq(rdma
->cq
, 1, &wc
);
1424 *wr_id_out
= RDMA_WRID_NONE
;
1429 error_report("ibv_poll_cq return %d", ret
);
1433 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1435 if (wc
.status
!= IBV_WC_SUCCESS
) {
1436 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1437 wc
.status
, ibv_wc_status_str(wc
.status
));
1438 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1443 if (rdma
->control_ready_expected
&&
1444 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1445 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1446 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1447 rdma
->control_ready_expected
= 0;
1450 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1452 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1454 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1455 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1457 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1458 index
, chunk
, block
->local_host_addr
,
1459 (void *)(uintptr_t)block
->remote_host_addr
);
1461 clear_bit(chunk
, block
->transit_bitmap
);
1463 if (rdma
->nb_sent
> 0) {
1467 if (!rdma
->pin_all
) {
1469 * FYI: If one wanted to signal a specific chunk to be unregistered
1470 * using LRU or workload-specific information, this is the function
1471 * you would call to do so. That chunk would then get asynchronously
1472 * unregistered later.
1474 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1475 qemu_rdma_signal_unregister(rdma
, index
, chunk
, wc
.wr_id
);
1479 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1482 *wr_id_out
= wc
.wr_id
;
1484 *byte_len
= wc
.byte_len
;
1490 /* Wait for activity on the completion channel.
1491 * Returns 0 on success, none-0 on error.
1493 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
)
1495 struct rdma_cm_event
*cm_event
;
1499 * Coroutine doesn't start until migration_fd_process_incoming()
1500 * so don't yield unless we know we're running inside of a coroutine.
1502 if (rdma
->migration_started_on_destination
&&
1503 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1504 yield_until_fd_readable(rdma
->comp_channel
->fd
);
1506 /* This is the source side, we're in a separate thread
1507 * or destination prior to migration_fd_process_incoming()
1508 * after postcopy, the destination also in a seprate thread.
1509 * we can't yield; so we have to poll the fd.
1510 * But we need to be able to handle 'cancel' or an error
1511 * without hanging forever.
1513 while (!rdma
->error_state
&& !rdma
->received_error
) {
1515 pfds
[0].fd
= rdma
->comp_channel
->fd
;
1516 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1517 pfds
[0].revents
= 0;
1519 pfds
[1].fd
= rdma
->channel
->fd
;
1520 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1521 pfds
[1].revents
= 0;
1523 /* 0.1s timeout, should be fine for a 'cancel' */
1524 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1526 case 1: /* fd active */
1527 if (pfds
[0].revents
) {
1531 if (pfds
[1].revents
) {
1532 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1534 rdma_ack_cm_event(cm_event
);
1537 error_report("receive cm event while wait comp channel,"
1538 "cm event is %d", cm_event
->event
);
1539 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1540 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1546 case 0: /* Timeout, go around again */
1549 default: /* Error of some type -
1550 * I don't trust errno from qemu_poll_ns
1552 error_report("%s: poll failed", __func__
);
1556 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1557 /* Bail out and let the cancellation happen */
1563 if (rdma
->received_error
) {
1566 return rdma
->error_state
;
1570 * Block until the next work request has completed.
1572 * First poll to see if a work request has already completed,
1575 * If we encounter completed work requests for IDs other than
1576 * the one we're interested in, then that's generally an error.
1578 * The only exception is actual RDMA Write completions. These
1579 * completions only need to be recorded, but do not actually
1580 * need further processing.
1582 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1585 int num_cq_events
= 0, ret
= 0;
1588 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1590 if (ibv_req_notify_cq(rdma
->cq
, 0)) {
1594 while (wr_id
!= wrid_requested
) {
1595 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1600 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1602 if (wr_id
== RDMA_WRID_NONE
) {
1605 if (wr_id
!= wrid_requested
) {
1606 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1607 wrid_requested
, print_wrid(wr_id
), wr_id
);
1611 if (wr_id
== wrid_requested
) {
1616 ret
= qemu_rdma_wait_comp_channel(rdma
);
1618 goto err_block_for_wrid
;
1621 ret
= ibv_get_cq_event(rdma
->comp_channel
, &cq
, &cq_ctx
);
1623 perror("ibv_get_cq_event");
1624 goto err_block_for_wrid
;
1629 ret
= -ibv_req_notify_cq(cq
, 0);
1631 goto err_block_for_wrid
;
1634 while (wr_id
!= wrid_requested
) {
1635 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1637 goto err_block_for_wrid
;
1640 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1642 if (wr_id
== RDMA_WRID_NONE
) {
1645 if (wr_id
!= wrid_requested
) {
1646 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1647 wrid_requested
, print_wrid(wr_id
), wr_id
);
1651 if (wr_id
== wrid_requested
) {
1652 goto success_block_for_wrid
;
1656 success_block_for_wrid
:
1657 if (num_cq_events
) {
1658 ibv_ack_cq_events(cq
, num_cq_events
);
1663 if (num_cq_events
) {
1664 ibv_ack_cq_events(cq
, num_cq_events
);
1667 rdma
->error_state
= ret
;
1672 * Post a SEND message work request for the control channel
1673 * containing some data and block until the post completes.
1675 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1676 RDMAControlHeader
*head
)
1679 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1680 struct ibv_send_wr
*bad_wr
;
1681 struct ibv_sge sge
= {
1682 .addr
= (uintptr_t)(wr
->control
),
1683 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1684 .lkey
= wr
->control_mr
->lkey
,
1686 struct ibv_send_wr send_wr
= {
1687 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1688 .opcode
= IBV_WR_SEND
,
1689 .send_flags
= IBV_SEND_SIGNALED
,
1694 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1697 * We don't actually need to do a memcpy() in here if we used
1698 * the "sge" properly, but since we're only sending control messages
1699 * (not RAM in a performance-critical path), then its OK for now.
1701 * The copy makes the RDMAControlHeader simpler to manipulate
1702 * for the time being.
1704 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1705 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1706 control_to_network((void *) wr
->control
);
1709 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1713 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1716 error_report("Failed to use post IB SEND for control");
1720 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1722 error_report("rdma migration: send polling control error");
1729 * Post a RECV work request in anticipation of some future receipt
1730 * of data on the control channel.
1732 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1734 struct ibv_recv_wr
*bad_wr
;
1735 struct ibv_sge sge
= {
1736 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1737 .length
= RDMA_CONTROL_MAX_BUFFER
,
1738 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1741 struct ibv_recv_wr recv_wr
= {
1742 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1748 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1756 * Block and wait for a RECV control channel message to arrive.
1758 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1759 RDMAControlHeader
*head
, int expecting
, int idx
)
1762 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1766 error_report("rdma migration: recv polling control error!");
1770 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1771 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1773 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1775 if (expecting
== RDMA_CONTROL_NONE
) {
1776 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1778 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1779 error_report("Was expecting a %s (%d) control message"
1780 ", but got: %s (%d), length: %d",
1781 control_desc(expecting
), expecting
,
1782 control_desc(head
->type
), head
->type
, head
->len
);
1783 if (head
->type
== RDMA_CONTROL_ERROR
) {
1784 rdma
->received_error
= true;
1788 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1789 error_report("too long length: %d", head
->len
);
1792 if (sizeof(*head
) + head
->len
!= byte_len
) {
1793 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1801 * When a RECV work request has completed, the work request's
1802 * buffer is pointed at the header.
1804 * This will advance the pointer to the data portion
1805 * of the control message of the work request's buffer that
1806 * was populated after the work request finished.
1808 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1809 RDMAControlHeader
*head
)
1811 rdma
->wr_data
[idx
].control_len
= head
->len
;
1812 rdma
->wr_data
[idx
].control_curr
=
1813 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1817 * This is an 'atomic' high-level operation to deliver a single, unified
1818 * control-channel message.
1820 * Additionally, if the user is expecting some kind of reply to this message,
1821 * they can request a 'resp' response message be filled in by posting an
1822 * additional work request on behalf of the user and waiting for an additional
1825 * The extra (optional) response is used during registration to us from having
1826 * to perform an *additional* exchange of message just to provide a response by
1827 * instead piggy-backing on the acknowledgement.
1829 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1830 uint8_t *data
, RDMAControlHeader
*resp
,
1832 int (*callback
)(RDMAContext
*rdma
))
1837 * Wait until the dest is ready before attempting to deliver the message
1838 * by waiting for a READY message.
1840 if (rdma
->control_ready_expected
) {
1841 RDMAControlHeader resp
;
1842 ret
= qemu_rdma_exchange_get_response(rdma
,
1843 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1850 * If the user is expecting a response, post a WR in anticipation of it.
1853 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1855 error_report("rdma migration: error posting"
1856 " extra control recv for anticipated result!");
1862 * Post a WR to replace the one we just consumed for the READY message.
1864 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1866 error_report("rdma migration: error posting first control recv!");
1871 * Deliver the control message that was requested.
1873 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1876 error_report("Failed to send control buffer!");
1881 * If we're expecting a response, block and wait for it.
1885 trace_qemu_rdma_exchange_send_issue_callback();
1886 ret
= callback(rdma
);
1892 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1893 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1894 resp
->type
, RDMA_WRID_DATA
);
1900 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1902 *resp_idx
= RDMA_WRID_DATA
;
1904 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1907 rdma
->control_ready_expected
= 1;
1913 * This is an 'atomic' high-level operation to receive a single, unified
1914 * control-channel message.
1916 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1919 RDMAControlHeader ready
= {
1921 .type
= RDMA_CONTROL_READY
,
1927 * Inform the source that we're ready to receive a message.
1929 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1932 error_report("Failed to send control buffer!");
1937 * Block and wait for the message.
1939 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1940 expecting
, RDMA_WRID_READY
);
1946 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1949 * Post a new RECV work request to replace the one we just consumed.
1951 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1953 error_report("rdma migration: error posting second control recv!");
1961 * Write an actual chunk of memory using RDMA.
1963 * If we're using dynamic registration on the dest-side, we have to
1964 * send a registration command first.
1966 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
1967 int current_index
, uint64_t current_addr
,
1971 struct ibv_send_wr send_wr
= { 0 };
1972 struct ibv_send_wr
*bad_wr
;
1973 int reg_result_idx
, ret
, count
= 0;
1974 uint64_t chunk
, chunks
;
1975 uint8_t *chunk_start
, *chunk_end
;
1976 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
1978 RDMARegisterResult
*reg_result
;
1979 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
1980 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1981 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
1986 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
1987 (current_addr
- block
->offset
));
1988 sge
.length
= length
;
1990 chunk
= ram_chunk_index(block
->local_host_addr
,
1991 (uint8_t *)(uintptr_t)sge
.addr
);
1992 chunk_start
= ram_chunk_start(block
, chunk
);
1994 if (block
->is_ram_block
) {
1995 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
1997 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2001 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2003 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2008 trace_qemu_rdma_write_one_top(chunks
+ 1,
2010 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2012 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2014 if (!rdma
->pin_all
) {
2015 #ifdef RDMA_UNREGISTRATION_EXAMPLE
2016 qemu_rdma_unregister_waiting(rdma
);
2020 while (test_bit(chunk
, block
->transit_bitmap
)) {
2022 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2023 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2025 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2028 error_report("Failed to Wait for previous write to complete "
2029 "block %d chunk %" PRIu64
2030 " current %" PRIu64
" len %" PRIu64
" %d",
2031 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2036 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2037 if (!block
->remote_keys
[chunk
]) {
2039 * This chunk has not yet been registered, so first check to see
2040 * if the entire chunk is zero. If so, tell the other size to
2041 * memset() + madvise() the entire chunk without RDMA.
2044 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2045 RDMACompress comp
= {
2046 .offset
= current_addr
,
2048 .block_idx
= current_index
,
2052 head
.len
= sizeof(comp
);
2053 head
.type
= RDMA_CONTROL_COMPRESS
;
2055 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2056 current_index
, current_addr
);
2058 compress_to_network(rdma
, &comp
);
2059 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2060 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2066 acct_update_position(f
, sge
.length
, true);
2072 * Otherwise, tell other side to register.
2074 reg
.current_index
= current_index
;
2075 if (block
->is_ram_block
) {
2076 reg
.key
.current_addr
= current_addr
;
2078 reg
.key
.chunk
= chunk
;
2080 reg
.chunks
= chunks
;
2082 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2085 register_to_network(rdma
, ®
);
2086 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2087 &resp
, ®_result_idx
, NULL
);
2092 /* try to overlap this single registration with the one we sent. */
2093 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2094 &sge
.lkey
, NULL
, chunk
,
2095 chunk_start
, chunk_end
)) {
2096 error_report("cannot get lkey");
2100 reg_result
= (RDMARegisterResult
*)
2101 rdma
->wr_data
[reg_result_idx
].control_curr
;
2103 network_to_result(reg_result
);
2105 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2106 reg_result
->rkey
, chunk
);
2108 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2109 block
->remote_host_addr
= reg_result
->host_addr
;
2111 /* already registered before */
2112 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2113 &sge
.lkey
, NULL
, chunk
,
2114 chunk_start
, chunk_end
)) {
2115 error_report("cannot get lkey!");
2120 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2122 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2124 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2125 &sge
.lkey
, NULL
, chunk
,
2126 chunk_start
, chunk_end
)) {
2127 error_report("cannot get lkey!");
2133 * Encode the ram block index and chunk within this wrid.
2134 * We will use this information at the time of completion
2135 * to figure out which bitmap to check against and then which
2136 * chunk in the bitmap to look for.
2138 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2139 current_index
, chunk
);
2141 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2142 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2143 send_wr
.sg_list
= &sge
;
2144 send_wr
.num_sge
= 1;
2145 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2146 (current_addr
- block
->offset
);
2148 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2152 * ibv_post_send() does not return negative error numbers,
2153 * per the specification they are positive - no idea why.
2155 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2157 if (ret
== ENOMEM
) {
2158 trace_qemu_rdma_write_one_queue_full();
2159 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2161 error_report("rdma migration: failed to make "
2162 "room in full send queue! %d", ret
);
2168 } else if (ret
> 0) {
2169 perror("rdma migration: post rdma write failed");
2173 set_bit(chunk
, block
->transit_bitmap
);
2174 acct_update_position(f
, sge
.length
, false);
2175 rdma
->total_writes
++;
2181 * Push out any unwritten RDMA operations.
2183 * We support sending out multiple chunks at the same time.
2184 * Not all of them need to get signaled in the completion queue.
2186 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2190 if (!rdma
->current_length
) {
2194 ret
= qemu_rdma_write_one(f
, rdma
,
2195 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2203 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2206 rdma
->current_length
= 0;
2207 rdma
->current_addr
= 0;
2212 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2213 uint64_t offset
, uint64_t len
)
2215 RDMALocalBlock
*block
;
2219 if (rdma
->current_index
< 0) {
2223 if (rdma
->current_chunk
< 0) {
2227 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2228 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2229 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2231 if (rdma
->current_length
== 0) {
2236 * Only merge into chunk sequentially.
2238 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2242 if (offset
< block
->offset
) {
2246 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2250 if ((host_addr
+ len
) > chunk_end
) {
2258 * We're not actually writing here, but doing three things:
2260 * 1. Identify the chunk the buffer belongs to.
2261 * 2. If the chunk is full or the buffer doesn't belong to the current
2262 * chunk, then start a new chunk and flush() the old chunk.
2263 * 3. To keep the hardware busy, we also group chunks into batches
2264 * and only require that a batch gets acknowledged in the completion
2265 * qeueue instead of each individual chunk.
2267 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2268 uint64_t block_offset
, uint64_t offset
,
2271 uint64_t current_addr
= block_offset
+ offset
;
2272 uint64_t index
= rdma
->current_index
;
2273 uint64_t chunk
= rdma
->current_chunk
;
2276 /* If we cannot merge it, we flush the current buffer first. */
2277 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2278 ret
= qemu_rdma_write_flush(f
, rdma
);
2282 rdma
->current_length
= 0;
2283 rdma
->current_addr
= current_addr
;
2285 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2286 offset
, len
, &index
, &chunk
);
2288 error_report("ram block search failed");
2291 rdma
->current_index
= index
;
2292 rdma
->current_chunk
= chunk
;
2296 rdma
->current_length
+= len
;
2298 /* flush it if buffer is too large */
2299 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2300 return qemu_rdma_write_flush(f
, rdma
);
2306 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2310 if (rdma
->cm_id
&& rdma
->connected
) {
2311 if ((rdma
->error_state
||
2312 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2313 !rdma
->received_error
) {
2314 RDMAControlHeader head
= { .len
= 0,
2315 .type
= RDMA_CONTROL_ERROR
,
2318 error_report("Early error. Sending error.");
2319 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2322 rdma_disconnect(rdma
->cm_id
);
2323 trace_qemu_rdma_cleanup_disconnect();
2324 rdma
->connected
= false;
2327 if (rdma
->channel
) {
2328 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2330 g_free(rdma
->dest_blocks
);
2331 rdma
->dest_blocks
= NULL
;
2333 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2334 if (rdma
->wr_data
[idx
].control_mr
) {
2335 rdma
->total_registrations
--;
2336 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2338 rdma
->wr_data
[idx
].control_mr
= NULL
;
2341 if (rdma
->local_ram_blocks
.block
) {
2342 while (rdma
->local_ram_blocks
.nb_blocks
) {
2343 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2348 rdma_destroy_qp(rdma
->cm_id
);
2352 ibv_destroy_cq(rdma
->cq
);
2355 if (rdma
->comp_channel
) {
2356 ibv_destroy_comp_channel(rdma
->comp_channel
);
2357 rdma
->comp_channel
= NULL
;
2360 ibv_dealloc_pd(rdma
->pd
);
2364 rdma_destroy_id(rdma
->cm_id
);
2368 /* the destination side, listen_id and channel is shared */
2369 if (rdma
->listen_id
) {
2370 if (!rdma
->is_return_path
) {
2371 rdma_destroy_id(rdma
->listen_id
);
2373 rdma
->listen_id
= NULL
;
2375 if (rdma
->channel
) {
2376 if (!rdma
->is_return_path
) {
2377 rdma_destroy_event_channel(rdma
->channel
);
2379 rdma
->channel
= NULL
;
2383 if (rdma
->channel
) {
2384 rdma_destroy_event_channel(rdma
->channel
);
2385 rdma
->channel
= NULL
;
2392 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2395 Error
*local_err
= NULL
, **temp
= &local_err
;
2398 * Will be validated against destination's actual capabilities
2399 * after the connect() completes.
2401 rdma
->pin_all
= pin_all
;
2403 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2405 goto err_rdma_source_init
;
2408 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2410 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2411 " limits may be too low. Please check $ ulimit -a # and "
2412 "search for 'ulimit -l' in the output");
2413 goto err_rdma_source_init
;
2416 ret
= qemu_rdma_alloc_qp(rdma
);
2418 ERROR(temp
, "rdma migration: error allocating qp!");
2419 goto err_rdma_source_init
;
2422 ret
= qemu_rdma_init_ram_blocks(rdma
);
2424 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2425 goto err_rdma_source_init
;
2428 /* Build the hash that maps from offset to RAMBlock */
2429 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2430 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2431 g_hash_table_insert(rdma
->blockmap
,
2432 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2433 &rdma
->local_ram_blocks
.block
[idx
]);
2436 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2437 ret
= qemu_rdma_reg_control(rdma
, idx
);
2439 ERROR(temp
, "rdma migration: error registering %d control!",
2441 goto err_rdma_source_init
;
2447 err_rdma_source_init
:
2448 error_propagate(errp
, local_err
);
2449 qemu_rdma_cleanup(rdma
);
2453 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
)
2455 RDMACapabilities cap
= {
2456 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2459 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2461 .private_data
= &cap
,
2462 .private_data_len
= sizeof(cap
),
2464 struct rdma_cm_event
*cm_event
;
2468 * Only negotiate the capability with destination if the user
2469 * on the source first requested the capability.
2471 if (rdma
->pin_all
) {
2472 trace_qemu_rdma_connect_pin_all_requested();
2473 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2476 caps_to_network(&cap
);
2478 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2480 ERROR(errp
, "posting second control recv");
2481 goto err_rdma_source_connect
;
2484 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2486 perror("rdma_connect");
2487 ERROR(errp
, "connecting to destination!");
2488 goto err_rdma_source_connect
;
2491 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2493 perror("rdma_get_cm_event after rdma_connect");
2494 ERROR(errp
, "connecting to destination!");
2495 rdma_ack_cm_event(cm_event
);
2496 goto err_rdma_source_connect
;
2499 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2500 perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2501 ERROR(errp
, "connecting to destination!");
2502 rdma_ack_cm_event(cm_event
);
2503 goto err_rdma_source_connect
;
2505 rdma
->connected
= true;
2507 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2508 network_to_caps(&cap
);
2511 * Verify that the *requested* capabilities are supported by the destination
2512 * and disable them otherwise.
2514 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2515 ERROR(errp
, "Server cannot support pinning all memory. "
2516 "Will register memory dynamically.");
2517 rdma
->pin_all
= false;
2520 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2522 rdma_ack_cm_event(cm_event
);
2524 rdma
->control_ready_expected
= 1;
2528 err_rdma_source_connect
:
2529 qemu_rdma_cleanup(rdma
);
2533 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2536 struct rdma_cm_id
*listen_id
;
2537 char ip
[40] = "unknown";
2538 struct rdma_addrinfo
*res
, *e
;
2541 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2542 rdma
->wr_data
[idx
].control_len
= 0;
2543 rdma
->wr_data
[idx
].control_curr
= NULL
;
2546 if (!rdma
->host
|| !rdma
->host
[0]) {
2547 ERROR(errp
, "RDMA host is not set!");
2548 rdma
->error_state
= -EINVAL
;
2551 /* create CM channel */
2552 rdma
->channel
= rdma_create_event_channel();
2553 if (!rdma
->channel
) {
2554 ERROR(errp
, "could not create rdma event channel");
2555 rdma
->error_state
= -EINVAL
;
2560 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2562 ERROR(errp
, "could not create cm_id!");
2563 goto err_dest_init_create_listen_id
;
2566 snprintf(port_str
, 16, "%d", rdma
->port
);
2567 port_str
[15] = '\0';
2569 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2571 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2572 goto err_dest_init_bind_addr
;
2575 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2576 inet_ntop(e
->ai_family
,
2577 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2578 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2579 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2583 if (e
->ai_family
== AF_INET6
) {
2584 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2593 ERROR(errp
, "Error: could not rdma_bind_addr!");
2594 goto err_dest_init_bind_addr
;
2597 rdma
->listen_id
= listen_id
;
2598 qemu_rdma_dump_gid("dest_init", listen_id
);
2601 err_dest_init_bind_addr
:
2602 rdma_destroy_id(listen_id
);
2603 err_dest_init_create_listen_id
:
2604 rdma_destroy_event_channel(rdma
->channel
);
2605 rdma
->channel
= NULL
;
2606 rdma
->error_state
= ret
;
2611 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2616 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2617 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2618 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2621 /*the CM channel and CM id is shared*/
2622 rdma_return_path
->channel
= rdma
->channel
;
2623 rdma_return_path
->listen_id
= rdma
->listen_id
;
2625 rdma
->return_path
= rdma_return_path
;
2626 rdma_return_path
->return_path
= rdma
;
2627 rdma_return_path
->is_return_path
= true;
2630 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2632 RDMAContext
*rdma
= NULL
;
2633 InetSocketAddress
*addr
;
2636 rdma
= g_new0(RDMAContext
, 1);
2637 rdma
->current_index
= -1;
2638 rdma
->current_chunk
= -1;
2640 addr
= g_new(InetSocketAddress
, 1);
2641 if (!inet_parse(addr
, host_port
, NULL
)) {
2642 rdma
->port
= atoi(addr
->port
);
2643 rdma
->host
= g_strdup(addr
->host
);
2645 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2650 qapi_free_InetSocketAddress(addr
);
2657 * QEMUFile interface to the control channel.
2658 * SEND messages for control only.
2659 * VM's ram is handled with regular RDMA messages.
2661 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2662 const struct iovec
*iov
,
2668 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2669 QEMUFile
*f
= rioc
->file
;
2677 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
2684 CHECK_ERROR_STATE();
2687 * Push out any writes that
2688 * we're queued up for VM's ram.
2690 ret
= qemu_rdma_write_flush(f
, rdma
);
2692 rdma
->error_state
= ret
;
2697 for (i
= 0; i
< niov
; i
++) {
2698 size_t remaining
= iov
[i
].iov_len
;
2699 uint8_t * data
= (void *)iov
[i
].iov_base
;
2701 RDMAControlHeader head
;
2703 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2707 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2709 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2712 rdma
->error_state
= ret
;
2726 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2727 size_t size
, int idx
)
2731 if (rdma
->wr_data
[idx
].control_len
) {
2732 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2734 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2735 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2736 rdma
->wr_data
[idx
].control_curr
+= len
;
2737 rdma
->wr_data
[idx
].control_len
-= len
;
2744 * QEMUFile interface to the control channel.
2745 * RDMA links don't use bytestreams, so we have to
2746 * return bytes to QEMUFile opportunistically.
2748 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2749 const struct iovec
*iov
,
2755 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2757 RDMAControlHeader head
;
2763 rdma
= atomic_rcu_read(&rioc
->rdmain
);
2770 CHECK_ERROR_STATE();
2772 for (i
= 0; i
< niov
; i
++) {
2773 size_t want
= iov
[i
].iov_len
;
2774 uint8_t *data
= (void *)iov
[i
].iov_base
;
2777 * First, we hold on to the last SEND message we
2778 * were given and dish out the bytes until we run
2781 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2784 /* Got what we needed, so go to next iovec */
2789 /* If we got any data so far, then don't wait
2790 * for more, just return what we have */
2796 /* We've got nothing at all, so lets wait for
2799 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2802 rdma
->error_state
= ret
;
2808 * SEND was received with new bytes, now try again.
2810 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2814 /* Still didn't get enough, so lets just return */
2818 return QIO_CHANNEL_ERR_BLOCK
;
2829 * Block until all the outstanding chunks have been delivered by the hardware.
2831 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2835 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2839 while (rdma
->nb_sent
) {
2840 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2842 error_report("rdma migration: complete polling error!");
2847 qemu_rdma_unregister_waiting(rdma
);
2853 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2857 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2858 /* XXX we should make readv/writev actually honour this :-) */
2859 rioc
->blocking
= blocking
;
2864 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2865 struct QIOChannelRDMASource
{
2867 QIOChannelRDMA
*rioc
;
2868 GIOCondition condition
;
2872 qio_channel_rdma_source_prepare(GSource
*source
,
2875 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2877 GIOCondition cond
= 0;
2881 if (rsource
->condition
== G_IO_IN
) {
2882 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2884 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2888 error_report("RDMAContext is NULL when prepare Gsource");
2893 if (rdma
->wr_data
[0].control_len
) {
2899 return cond
& rsource
->condition
;
2903 qio_channel_rdma_source_check(GSource
*source
)
2905 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2907 GIOCondition cond
= 0;
2910 if (rsource
->condition
== G_IO_IN
) {
2911 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2913 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2917 error_report("RDMAContext is NULL when check Gsource");
2922 if (rdma
->wr_data
[0].control_len
) {
2928 return cond
& rsource
->condition
;
2932 qio_channel_rdma_source_dispatch(GSource
*source
,
2933 GSourceFunc callback
,
2936 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
2937 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2939 GIOCondition cond
= 0;
2942 if (rsource
->condition
== G_IO_IN
) {
2943 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2945 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2949 error_report("RDMAContext is NULL when dispatch Gsource");
2954 if (rdma
->wr_data
[0].control_len
) {
2960 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
2961 (cond
& rsource
->condition
),
2966 qio_channel_rdma_source_finalize(GSource
*source
)
2968 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
2970 object_unref(OBJECT(ssource
->rioc
));
2973 GSourceFuncs qio_channel_rdma_source_funcs
= {
2974 qio_channel_rdma_source_prepare
,
2975 qio_channel_rdma_source_check
,
2976 qio_channel_rdma_source_dispatch
,
2977 qio_channel_rdma_source_finalize
2980 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
2981 GIOCondition condition
)
2983 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2984 QIOChannelRDMASource
*ssource
;
2987 source
= g_source_new(&qio_channel_rdma_source_funcs
,
2988 sizeof(QIOChannelRDMASource
));
2989 ssource
= (QIOChannelRDMASource
*)source
;
2991 ssource
->rioc
= rioc
;
2992 object_ref(OBJECT(rioc
));
2994 ssource
->condition
= condition
;
2999 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
3002 IOHandler
*io_write
,
3005 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3007 aio_set_fd_handler(ctx
, rioc
->rdmain
->comp_channel
->fd
,
3008 false, io_read
, io_write
, NULL
, opaque
);
3010 aio_set_fd_handler(ctx
, rioc
->rdmaout
->comp_channel
->fd
,
3011 false, io_read
, io_write
, NULL
, opaque
);
3015 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3018 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3019 RDMAContext
*rdmain
, *rdmaout
;
3020 trace_qemu_rdma_close();
3022 rdmain
= rioc
->rdmain
;
3024 atomic_rcu_set(&rioc
->rdmain
, NULL
);
3027 rdmaout
= rioc
->rdmaout
;
3029 atomic_rcu_set(&rioc
->rdmaout
, NULL
);
3035 qemu_rdma_cleanup(rdmain
);
3039 qemu_rdma_cleanup(rdmaout
);
3049 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3050 QIOChannelShutdown how
,
3053 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3054 RDMAContext
*rdmain
, *rdmaout
;
3058 rdmain
= atomic_rcu_read(&rioc
->rdmain
);
3059 rdmaout
= atomic_rcu_read(&rioc
->rdmain
);
3062 case QIO_CHANNEL_SHUTDOWN_READ
:
3064 rdmain
->error_state
= -1;
3067 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3069 rdmaout
->error_state
= -1;
3072 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3075 rdmain
->error_state
= -1;
3078 rdmaout
->error_state
= -1;
3090 * This means that 'block_offset' is a full virtual address that does not
3091 * belong to a RAMBlock of the virtual machine and instead
3092 * represents a private malloc'd memory area that the caller wishes to
3096 * Offset is an offset to be added to block_offset and used
3097 * to also lookup the corresponding RAMBlock.
3100 * Initiate an transfer this size.
3103 * A 'hint' or 'advice' that means that we wish to speculatively
3104 * and asynchronously unregister this memory. In this case, there is no
3105 * guarantee that the unregister will actually happen, for example,
3106 * if the memory is being actively transmitted. Additionally, the memory
3107 * may be re-registered at any future time if a write within the same
3108 * chunk was requested again, even if you attempted to unregister it
3111 * @size < 0 : TODO, not yet supported
3112 * Unregister the memory NOW. This means that the caller does not
3113 * expect there to be any future RDMA transfers and we just want to clean
3114 * things up. This is used in case the upper layer owns the memory and
3115 * cannot wait for qemu_fclose() to occur.
3117 * @bytes_sent : User-specificed pointer to indicate how many bytes were
3118 * sent. Usually, this will not be more than a few bytes of
3119 * the protocol because most transfers are sent asynchronously.
3121 static size_t qemu_rdma_save_page(QEMUFile
*f
, void *opaque
,
3122 ram_addr_t block_offset
, ram_addr_t offset
,
3123 size_t size
, uint64_t *bytes_sent
)
3125 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3130 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3137 CHECK_ERROR_STATE();
3139 if (migrate_get_current()->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
3141 return RAM_SAVE_CONTROL_NOT_SUPP
;
3148 * Add this page to the current 'chunk'. If the chunk
3149 * is full, or the page doen't belong to the current chunk,
3150 * an actual RDMA write will occur and a new chunk will be formed.
3152 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
3154 error_report("rdma migration: write error! %d", ret
);
3159 * We always return 1 bytes because the RDMA
3160 * protocol is completely asynchronous. We do not yet know
3161 * whether an identified chunk is zero or not because we're
3162 * waiting for other pages to potentially be merged with
3163 * the current chunk. So, we have to call qemu_update_position()
3164 * later on when the actual write occurs.
3170 uint64_t index
, chunk
;
3172 /* TODO: Change QEMUFileOps prototype to be signed: size_t => long
3174 ret = qemu_rdma_drain_cq(f, rdma);
3176 fprintf(stderr, "rdma: failed to synchronously drain"
3177 " completion queue before unregistration.\n");
3183 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
3184 offset
, size
, &index
, &chunk
);
3187 error_report("ram block search failed");
3191 qemu_rdma_signal_unregister(rdma
, index
, chunk
, 0);
3194 * TODO: Synchronous, guaranteed unregistration (should not occur during
3195 * fast-path). Otherwise, unregisters will process on the next call to
3196 * qemu_rdma_drain_cq()
3198 qemu_rdma_unregister_waiting(rdma);
3204 * Drain the Completion Queue if possible, but do not block,
3207 * If nothing to poll, the end of the iteration will do this
3208 * again to make sure we don't overflow the request queue.
3211 uint64_t wr_id
, wr_id_in
;
3212 int ret
= qemu_rdma_poll(rdma
, &wr_id_in
, NULL
);
3214 error_report("rdma migration: polling error! %d", ret
);
3218 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3220 if (wr_id
== RDMA_WRID_NONE
) {
3226 return RAM_SAVE_CONTROL_DELAYED
;
3228 rdma
->error_state
= ret
;
3233 static void rdma_accept_incoming_migration(void *opaque
);
3235 static void rdma_cm_poll_handler(void *opaque
)
3237 RDMAContext
*rdma
= opaque
;
3239 struct rdma_cm_event
*cm_event
;
3240 MigrationIncomingState
*mis
= migration_incoming_get_current();
3242 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3244 error_report("get_cm_event failed %d", errno
);
3247 rdma_ack_cm_event(cm_event
);
3249 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3250 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3251 error_report("receive cm event, cm event is %d", cm_event
->event
);
3252 rdma
->error_state
= -EPIPE
;
3253 if (rdma
->return_path
) {
3254 rdma
->return_path
->error_state
= -EPIPE
;
3257 if (mis
->migration_incoming_co
) {
3258 qemu_coroutine_enter(mis
->migration_incoming_co
);
3264 static int qemu_rdma_accept(RDMAContext
*rdma
)
3266 RDMACapabilities cap
;
3267 struct rdma_conn_param conn_param
= {
3268 .responder_resources
= 2,
3269 .private_data
= &cap
,
3270 .private_data_len
= sizeof(cap
),
3272 struct rdma_cm_event
*cm_event
;
3273 struct ibv_context
*verbs
;
3277 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3279 goto err_rdma_dest_wait
;
3282 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3283 rdma_ack_cm_event(cm_event
);
3284 goto err_rdma_dest_wait
;
3287 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3289 network_to_caps(&cap
);
3291 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3292 error_report("Unknown source RDMA version: %d, bailing...",
3294 rdma_ack_cm_event(cm_event
);
3295 goto err_rdma_dest_wait
;
3299 * Respond with only the capabilities this version of QEMU knows about.
3301 cap
.flags
&= known_capabilities
;
3304 * Enable the ones that we do know about.
3305 * Add other checks here as new ones are introduced.
3307 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3308 rdma
->pin_all
= true;
3311 rdma
->cm_id
= cm_event
->id
;
3312 verbs
= cm_event
->id
->verbs
;
3314 rdma_ack_cm_event(cm_event
);
3316 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3318 caps_to_network(&cap
);
3320 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3323 rdma
->verbs
= verbs
;
3324 } else if (rdma
->verbs
!= verbs
) {
3325 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3327 goto err_rdma_dest_wait
;
3330 qemu_rdma_dump_id("dest_init", verbs
);
3332 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3334 error_report("rdma migration: error allocating pd and cq!");
3335 goto err_rdma_dest_wait
;
3338 ret
= qemu_rdma_alloc_qp(rdma
);
3340 error_report("rdma migration: error allocating qp!");
3341 goto err_rdma_dest_wait
;
3344 ret
= qemu_rdma_init_ram_blocks(rdma
);
3346 error_report("rdma migration: error initializing ram blocks!");
3347 goto err_rdma_dest_wait
;
3350 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3351 ret
= qemu_rdma_reg_control(rdma
, idx
);
3353 error_report("rdma: error registering %d control", idx
);
3354 goto err_rdma_dest_wait
;
3358 /* Accept the second connection request for return path */
3359 if (migrate_postcopy() && !rdma
->is_return_path
) {
3360 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3362 (void *)(intptr_t)rdma
->return_path
);
3364 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3368 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3370 error_report("rdma_accept returns %d", ret
);
3371 goto err_rdma_dest_wait
;
3374 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3376 error_report("rdma_accept get_cm_event failed %d", ret
);
3377 goto err_rdma_dest_wait
;
3380 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3381 error_report("rdma_accept not event established");
3382 rdma_ack_cm_event(cm_event
);
3383 goto err_rdma_dest_wait
;
3386 rdma_ack_cm_event(cm_event
);
3387 rdma
->connected
= true;
3389 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3391 error_report("rdma migration: error posting second control recv");
3392 goto err_rdma_dest_wait
;
3395 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3400 rdma
->error_state
= ret
;
3401 qemu_rdma_cleanup(rdma
);
3405 static int dest_ram_sort_func(const void *a
, const void *b
)
3407 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3408 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3410 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3414 * During each iteration of the migration, we listen for instructions
3415 * by the source VM to perform dynamic page registrations before they
3416 * can perform RDMA operations.
3418 * We respond with the 'rkey'.
3420 * Keep doing this until the source tells us to stop.
3422 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
)
3424 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3425 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3428 RDMAControlHeader unreg_resp
= { .len
= 0,
3429 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3432 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3434 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3436 RDMALocalBlocks
*local
;
3437 RDMAControlHeader head
;
3438 RDMARegister
*reg
, *registers
;
3440 RDMARegisterResult
*reg_result
;
3441 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3442 RDMALocalBlock
*block
;
3450 rdma
= atomic_rcu_read(&rioc
->rdmain
);
3457 CHECK_ERROR_STATE();
3459 local
= &rdma
->local_ram_blocks
;
3461 trace_qemu_rdma_registration_handle_wait();
3463 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3469 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3470 error_report("rdma: Too many requests in this message (%d)."
3471 "Bailing.", head
.repeat
);
3476 switch (head
.type
) {
3477 case RDMA_CONTROL_COMPRESS
:
3478 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3479 network_to_compress(comp
);
3481 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3484 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3485 error_report("rdma: 'compress' bad block index %u (vs %d)",
3486 (unsigned int)comp
->block_idx
,
3487 rdma
->local_ram_blocks
.nb_blocks
);
3491 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3493 host_addr
= block
->local_host_addr
+
3494 (comp
->offset
- block
->offset
);
3496 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3499 case RDMA_CONTROL_REGISTER_FINISHED
:
3500 trace_qemu_rdma_registration_handle_finished();
3503 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3504 trace_qemu_rdma_registration_handle_ram_blocks();
3506 /* Sort our local RAM Block list so it's the same as the source,
3507 * we can do this since we've filled in a src_index in the list
3508 * as we received the RAMBlock list earlier.
3510 qsort(rdma
->local_ram_blocks
.block
,
3511 rdma
->local_ram_blocks
.nb_blocks
,
3512 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3513 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3514 local
->block
[i
].index
= i
;
3517 if (rdma
->pin_all
) {
3518 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3520 error_report("rdma migration: error dest "
3521 "registering ram blocks");
3527 * Dest uses this to prepare to transmit the RAMBlock descriptions
3528 * to the source VM after connection setup.
3529 * Both sides use the "remote" structure to communicate and update
3530 * their "local" descriptions with what was sent.
3532 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3533 rdma
->dest_blocks
[i
].remote_host_addr
=
3534 (uintptr_t)(local
->block
[i
].local_host_addr
);
3536 if (rdma
->pin_all
) {
3537 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3540 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3541 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3543 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3544 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3545 local
->block
[i
].block_name
,
3546 local
->block
[i
].offset
,
3547 local
->block
[i
].length
,
3548 local
->block
[i
].local_host_addr
,
3549 local
->block
[i
].src_index
);
3552 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3553 * sizeof(RDMADestBlock
);
3556 ret
= qemu_rdma_post_send_control(rdma
,
3557 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3560 error_report("rdma migration: error sending remote info");
3565 case RDMA_CONTROL_REGISTER_REQUEST
:
3566 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3568 reg_resp
.repeat
= head
.repeat
;
3569 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3571 for (count
= 0; count
< head
.repeat
; count
++) {
3573 uint8_t *chunk_start
, *chunk_end
;
3575 reg
= ®isters
[count
];
3576 network_to_register(reg
);
3578 reg_result
= &results
[count
];
3580 trace_qemu_rdma_registration_handle_register_loop(count
,
3581 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3583 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3584 error_report("rdma: 'register' bad block index %u (vs %d)",
3585 (unsigned int)reg
->current_index
,
3586 rdma
->local_ram_blocks
.nb_blocks
);
3590 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3591 if (block
->is_ram_block
) {
3592 if (block
->offset
> reg
->key
.current_addr
) {
3593 error_report("rdma: bad register address for block %s"
3594 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3595 block
->block_name
, block
->offset
,
3596 reg
->key
.current_addr
);
3600 host_addr
= (block
->local_host_addr
+
3601 (reg
->key
.current_addr
- block
->offset
));
3602 chunk
= ram_chunk_index(block
->local_host_addr
,
3603 (uint8_t *) host_addr
);
3605 chunk
= reg
->key
.chunk
;
3606 host_addr
= block
->local_host_addr
+
3607 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3608 /* Check for particularly bad chunk value */
3609 if (host_addr
< (void *)block
->local_host_addr
) {
3610 error_report("rdma: bad chunk for block %s"
3612 block
->block_name
, reg
->key
.chunk
);
3617 chunk_start
= ram_chunk_start(block
, chunk
);
3618 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3619 /* avoid "-Waddress-of-packed-member" warning */
3620 uint32_t tmp_rkey
= 0;
3621 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3622 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3623 chunk
, chunk_start
, chunk_end
)) {
3624 error_report("cannot get rkey");
3628 reg_result
->rkey
= tmp_rkey
;
3630 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3632 trace_qemu_rdma_registration_handle_register_rkey(
3635 result_to_network(reg_result
);
3638 ret
= qemu_rdma_post_send_control(rdma
,
3639 (uint8_t *) results
, ®_resp
);
3642 error_report("Failed to send control buffer");
3646 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3647 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3648 unreg_resp
.repeat
= head
.repeat
;
3649 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3651 for (count
= 0; count
< head
.repeat
; count
++) {
3652 reg
= ®isters
[count
];
3653 network_to_register(reg
);
3655 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3656 reg
->current_index
, reg
->key
.chunk
);
3658 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3660 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3661 block
->pmr
[reg
->key
.chunk
] = NULL
;
3664 perror("rdma unregistration chunk failed");
3669 rdma
->total_registrations
--;
3671 trace_qemu_rdma_registration_handle_unregister_success(
3675 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3678 error_report("Failed to send control buffer");
3682 case RDMA_CONTROL_REGISTER_RESULT
:
3683 error_report("Invalid RESULT message at dest.");
3687 error_report("Unknown control message %s", control_desc(head
.type
));
3694 rdma
->error_state
= ret
;
3701 * Called via a ram_control_load_hook during the initial RAM load section which
3702 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3704 * We've already built our local RAMBlock list, but not yet sent the list to
3708 rdma_block_notification_handle(QIOChannelRDMA
*rioc
, const char *name
)
3715 rdma
= atomic_rcu_read(&rioc
->rdmain
);
3722 /* Find the matching RAMBlock in our local list */
3723 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3724 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3731 error_report("RAMBlock '%s' not found on destination", name
);
3736 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3737 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3738 rdma
->next_src_index
++;
3744 static int rdma_load_hook(QEMUFile
*f
, void *opaque
, uint64_t flags
, void *data
)
3747 case RAM_CONTROL_BLOCK_REG
:
3748 return rdma_block_notification_handle(opaque
, data
);
3750 case RAM_CONTROL_HOOK
:
3751 return qemu_rdma_registration_handle(f
, opaque
);
3754 /* Shouldn't be called with any other values */
3759 static int qemu_rdma_registration_start(QEMUFile
*f
, void *opaque
,
3760 uint64_t flags
, void *data
)
3762 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3766 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3772 CHECK_ERROR_STATE();
3774 if (migrate_get_current()->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
3779 trace_qemu_rdma_registration_start(flags
);
3780 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3788 * Inform dest that dynamic registrations are done for now.
3789 * First, flush writes, if any.
3791 static int qemu_rdma_registration_stop(QEMUFile
*f
, void *opaque
,
3792 uint64_t flags
, void *data
)
3794 Error
*local_err
= NULL
, **errp
= &local_err
;
3795 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3797 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3801 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3807 CHECK_ERROR_STATE();
3809 if (migrate_get_current()->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
3815 ret
= qemu_rdma_drain_cq(f
, rdma
);
3821 if (flags
== RAM_CONTROL_SETUP
) {
3822 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3823 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3824 int reg_result_idx
, i
, nb_dest_blocks
;
3826 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3827 trace_qemu_rdma_registration_stop_ram();
3830 * Make sure that we parallelize the pinning on both sides.
3831 * For very large guests, doing this serially takes a really
3832 * long time, so we have to 'interleave' the pinning locally
3833 * with the control messages by performing the pinning on this
3834 * side before we receive the control response from the other
3835 * side that the pinning has completed.
3837 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3838 ®_result_idx
, rdma
->pin_all
?
3839 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3841 ERROR(errp
, "receiving remote info!");
3846 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3849 * The protocol uses two different sets of rkeys (mutually exclusive):
3850 * 1. One key to represent the virtual address of the entire ram block.
3851 * (dynamic chunk registration disabled - pin everything with one rkey.)
3852 * 2. One to represent individual chunks within a ram block.
3853 * (dynamic chunk registration enabled - pin individual chunks.)
3855 * Once the capability is successfully negotiated, the destination transmits
3856 * the keys to use (or sends them later) including the virtual addresses
3857 * and then propagates the remote ram block descriptions to his local copy.
3860 if (local
->nb_blocks
!= nb_dest_blocks
) {
3861 ERROR(errp
, "ram blocks mismatch (Number of blocks %d vs %d) "
3862 "Your QEMU command line parameters are probably "
3863 "not identical on both the source and destination.",
3864 local
->nb_blocks
, nb_dest_blocks
);
3865 rdma
->error_state
= -EINVAL
;
3870 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3871 memcpy(rdma
->dest_blocks
,
3872 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3873 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3874 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3876 /* We require that the blocks are in the same order */
3877 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3878 ERROR(errp
, "Block %s/%d has a different length %" PRIu64
3879 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3880 local
->block
[i
].length
,
3881 rdma
->dest_blocks
[i
].length
);
3882 rdma
->error_state
= -EINVAL
;
3886 local
->block
[i
].remote_host_addr
=
3887 rdma
->dest_blocks
[i
].remote_host_addr
;
3888 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3892 trace_qemu_rdma_registration_stop(flags
);
3894 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3895 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3904 rdma
->error_state
= ret
;
3909 static const QEMUFileHooks rdma_read_hooks
= {
3910 .hook_ram_load
= rdma_load_hook
,
3913 static const QEMUFileHooks rdma_write_hooks
= {
3914 .before_ram_iterate
= qemu_rdma_registration_start
,
3915 .after_ram_iterate
= qemu_rdma_registration_stop
,
3916 .save_page
= qemu_rdma_save_page
,
3920 static void qio_channel_rdma_finalize(Object
*obj
)
3922 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
3924 qemu_rdma_cleanup(rioc
->rdmain
);
3925 g_free(rioc
->rdmain
);
3926 rioc
->rdmain
= NULL
;
3928 if (rioc
->rdmaout
) {
3929 qemu_rdma_cleanup(rioc
->rdmaout
);
3930 g_free(rioc
->rdmaout
);
3931 rioc
->rdmaout
= NULL
;
3935 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
3936 void *class_data G_GNUC_UNUSED
)
3938 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
3940 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
3941 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
3942 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
3943 ioc_klass
->io_close
= qio_channel_rdma_close
;
3944 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
3945 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
3946 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
3949 static const TypeInfo qio_channel_rdma_info
= {
3950 .parent
= TYPE_QIO_CHANNEL
,
3951 .name
= TYPE_QIO_CHANNEL_RDMA
,
3952 .instance_size
= sizeof(QIOChannelRDMA
),
3953 .instance_finalize
= qio_channel_rdma_finalize
,
3954 .class_init
= qio_channel_rdma_class_init
,
3957 static void qio_channel_rdma_register_types(void)
3959 type_register_static(&qio_channel_rdma_info
);
3962 type_init(qio_channel_rdma_register_types
);
3964 static QEMUFile
*qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
3966 QIOChannelRDMA
*rioc
;
3968 if (qemu_file_mode_is_not_valid(mode
)) {
3972 rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
3974 if (mode
[0] == 'w') {
3975 rioc
->file
= qemu_fopen_channel_output(QIO_CHANNEL(rioc
));
3976 rioc
->rdmaout
= rdma
;
3977 rioc
->rdmain
= rdma
->return_path
;
3978 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
3980 rioc
->file
= qemu_fopen_channel_input(QIO_CHANNEL(rioc
));
3981 rioc
->rdmain
= rdma
;
3982 rioc
->rdmaout
= rdma
->return_path
;
3983 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
3989 static void rdma_accept_incoming_migration(void *opaque
)
3991 RDMAContext
*rdma
= opaque
;
3994 Error
*local_err
= NULL
, **errp
= &local_err
;
3996 trace_qemu_rdma_accept_incoming_migration();
3997 ret
= qemu_rdma_accept(rdma
);
4000 ERROR(errp
, "RDMA Migration initialization failed!");
4004 trace_qemu_rdma_accept_incoming_migration_accepted();
4006 if (rdma
->is_return_path
) {
4010 f
= qemu_fopen_rdma(rdma
, "rb");
4012 ERROR(errp
, "could not qemu_fopen_rdma!");
4013 qemu_rdma_cleanup(rdma
);
4017 rdma
->migration_started_on_destination
= 1;
4018 migration_fd_process_incoming(f
);
4021 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4024 RDMAContext
*rdma
, *rdma_return_path
= NULL
;
4025 Error
*local_err
= NULL
;
4027 trace_rdma_start_incoming_migration();
4028 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
4034 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
4040 trace_rdma_start_incoming_migration_after_dest_init();
4042 ret
= rdma_listen(rdma
->listen_id
, 5);
4045 ERROR(errp
, "listening on socket!");
4049 trace_rdma_start_incoming_migration_after_rdma_listen();
4051 /* initialize the RDMAContext for return path */
4052 if (migrate_postcopy()) {
4053 rdma_return_path
= qemu_rdma_data_init(host_port
, &local_err
);
4055 if (rdma_return_path
== NULL
) {
4059 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
4062 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4063 NULL
, (void *)(intptr_t)rdma
);
4066 error_propagate(errp
, local_err
);
4068 g_free(rdma_return_path
);
4071 void rdma_start_outgoing_migration(void *opaque
,
4072 const char *host_port
, Error
**errp
)
4074 MigrationState
*s
= opaque
;
4075 RDMAContext
*rdma
= qemu_rdma_data_init(host_port
, errp
);
4076 RDMAContext
*rdma_return_path
= NULL
;
4083 ret
= qemu_rdma_source_init(rdma
,
4084 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4090 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4091 ret
= qemu_rdma_connect(rdma
, errp
);
4097 /* RDMA postcopy need a seprate queue pair for return path */
4098 if (migrate_postcopy()) {
4099 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4101 if (rdma_return_path
== NULL
) {
4105 ret
= qemu_rdma_source_init(rdma_return_path
,
4106 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4112 ret
= qemu_rdma_connect(rdma_return_path
, errp
);
4118 rdma
->return_path
= rdma_return_path
;
4119 rdma_return_path
->return_path
= rdma
;
4120 rdma_return_path
->is_return_path
= true;
4123 trace_rdma_start_outgoing_migration_after_rdma_connect();
4125 s
->to_dst_file
= qemu_fopen_rdma(rdma
, "wb");
4126 migrate_fd_connect(s
, NULL
);
4130 g_free(rdma_return_path
);