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.
17 #include "qemu/osdep.h"
18 #include "qapi/error.h"
19 #include "qemu/cutils.h"
20 #include "exec/target_page.h"
22 #include "migration.h"
23 #include "migration-stats.h"
24 #include "qemu-file.h"
26 #include "qemu/error-report.h"
27 #include "qemu/main-loop.h"
28 #include "qemu/module.h"
30 #include "qemu/sockets.h"
31 #include "qemu/bitmap.h"
32 #include "qemu/coroutine.h"
33 #include "exec/memory.h"
34 #include <sys/socket.h>
36 #include <arpa/inet.h>
37 #include <rdma/rdma_cma.h>
39 #include "qom/object.h"
44 * Print and error on both the Monitor and the Log file.
46 #define ERROR(errp, fmt, ...) \
48 fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \
49 if (errp && (*(errp) == NULL)) { \
50 error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \
54 #define RDMA_RESOLVE_TIMEOUT_MS 10000
56 /* Do not merge data if larger than this. */
57 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
58 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
60 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
63 * This is only for non-live state being migrated.
64 * Instead of RDMA_WRITE messages, we use RDMA_SEND
65 * messages for that state, which requires a different
66 * delivery design than main memory.
68 #define RDMA_SEND_INCREMENT 32768
71 * Maximum size infiniband SEND message
73 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
74 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
76 #define RDMA_CONTROL_VERSION_CURRENT 1
78 * Capabilities for negotiation.
80 #define RDMA_CAPABILITY_PIN_ALL 0x01
83 * Add the other flags above to this list of known capabilities
84 * as they are introduced.
86 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
88 #define CHECK_ERROR_STATE() \
90 if (rdma->error_state) { \
91 if (!rdma->error_reported) { \
92 error_report("RDMA is in an error state waiting migration" \
94 rdma->error_reported = 1; \
96 return rdma->error_state; \
101 * A work request ID is 64-bits and we split up these bits
104 * bits 0-15 : type of control message, 2^16
105 * bits 16-29: ram block index, 2^14
106 * bits 30-63: ram block chunk number, 2^34
108 * The last two bit ranges are only used for RDMA writes,
109 * in order to track their completion and potentially
110 * also track unregistration status of the message.
112 #define RDMA_WRID_TYPE_SHIFT 0UL
113 #define RDMA_WRID_BLOCK_SHIFT 16UL
114 #define RDMA_WRID_CHUNK_SHIFT 30UL
116 #define RDMA_WRID_TYPE_MASK \
117 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
119 #define RDMA_WRID_BLOCK_MASK \
120 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
122 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
125 * RDMA migration protocol:
126 * 1. RDMA Writes (data messages, i.e. RAM)
127 * 2. IB Send/Recv (control channel messages)
131 RDMA_WRID_RDMA_WRITE
= 1,
132 RDMA_WRID_SEND_CONTROL
= 2000,
133 RDMA_WRID_RECV_CONTROL
= 4000,
137 * Work request IDs for IB SEND messages only (not RDMA writes).
138 * This is used by the migration protocol to transmit
139 * control messages (such as device state and registration commands)
141 * We could use more WRs, but we have enough for now.
151 * SEND/RECV IB Control Messages.
154 RDMA_CONTROL_NONE
= 0,
156 RDMA_CONTROL_READY
, /* ready to receive */
157 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
158 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
159 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
160 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
161 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
162 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
163 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
164 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
165 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
170 * Memory and MR structures used to represent an IB Send/Recv work request.
171 * This is *not* used for RDMA writes, only IB Send/Recv.
174 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
175 struct ibv_mr
*control_mr
; /* registration metadata */
176 size_t control_len
; /* length of the message */
177 uint8_t *control_curr
; /* start of unconsumed bytes */
178 } RDMAWorkRequestData
;
181 * Negotiate RDMA capabilities during connection-setup time.
188 static void caps_to_network(RDMACapabilities
*cap
)
190 cap
->version
= htonl(cap
->version
);
191 cap
->flags
= htonl(cap
->flags
);
194 static void network_to_caps(RDMACapabilities
*cap
)
196 cap
->version
= ntohl(cap
->version
);
197 cap
->flags
= ntohl(cap
->flags
);
201 * Representation of a RAMBlock from an RDMA perspective.
202 * This is not transmitted, only local.
203 * This and subsequent structures cannot be linked lists
204 * because we're using a single IB message to transmit
205 * the information. It's small anyway, so a list is overkill.
207 typedef struct RDMALocalBlock
{
209 uint8_t *local_host_addr
; /* local virtual address */
210 uint64_t remote_host_addr
; /* remote virtual address */
213 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
214 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
215 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
216 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
217 int index
; /* which block are we */
218 unsigned int src_index
; /* (Only used on dest) */
221 unsigned long *transit_bitmap
;
222 unsigned long *unregister_bitmap
;
226 * Also represents a RAMblock, but only on the dest.
227 * This gets transmitted by the dest during connection-time
228 * to the source VM and then is used to populate the
229 * corresponding RDMALocalBlock with
230 * the information needed to perform the actual RDMA.
232 typedef struct QEMU_PACKED RDMADestBlock
{
233 uint64_t remote_host_addr
;
236 uint32_t remote_rkey
;
240 static const char *control_desc(unsigned int rdma_control
)
242 static const char *strs
[] = {
243 [RDMA_CONTROL_NONE
] = "NONE",
244 [RDMA_CONTROL_ERROR
] = "ERROR",
245 [RDMA_CONTROL_READY
] = "READY",
246 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
247 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
248 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
249 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
250 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
251 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
252 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
253 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
254 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
257 if (rdma_control
> RDMA_CONTROL_UNREGISTER_FINISHED
) {
258 return "??BAD CONTROL VALUE??";
261 return strs
[rdma_control
];
264 static uint64_t htonll(uint64_t v
)
266 union { uint32_t lv
[2]; uint64_t llv
; } u
;
267 u
.lv
[0] = htonl(v
>> 32);
268 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
272 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
{
317 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
320 * This is used by *_exchange_send() to figure out whether or not
321 * the initial "READY" message has already been received or not.
322 * This is because other functions may potentially poll() and detect
323 * the READY message before send() does, in which case we need to
324 * know if it completed.
326 int control_ready_expected
;
328 /* number of outstanding writes */
331 /* store info about current buffer so that we can
332 merge it with future sends */
333 uint64_t current_addr
;
334 uint64_t current_length
;
335 /* index of ram block the current buffer belongs to */
337 /* index of the chunk in the current ram block */
343 * infiniband-specific variables for opening the device
344 * and maintaining connection state and so forth.
346 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
347 * cm_id->verbs, cm_id->channel, and cm_id->qp.
349 struct rdma_cm_id
*cm_id
; /* connection manager ID */
350 struct rdma_cm_id
*listen_id
;
353 struct ibv_context
*verbs
;
354 struct rdma_event_channel
*channel
;
355 struct ibv_qp
*qp
; /* queue pair */
356 struct ibv_comp_channel
*recv_comp_channel
; /* recv completion channel */
357 struct ibv_comp_channel
*send_comp_channel
; /* send completion channel */
358 struct ibv_pd
*pd
; /* protection domain */
359 struct ibv_cq
*recv_cq
; /* recvieve completion queue */
360 struct ibv_cq
*send_cq
; /* send completion queue */
363 * If a previous write failed (perhaps because of a failed
364 * memory registration, then do not attempt any future work
365 * and remember the error state.
372 * Description of ram blocks used throughout the code.
374 RDMALocalBlocks local_ram_blocks
;
375 RDMADestBlock
*dest_blocks
;
377 /* Index of the next RAMBlock received during block registration */
378 unsigned int next_src_index
;
381 * Migration on *destination* started.
382 * Then use coroutine yield function.
383 * Source runs in a thread, so we don't care.
385 int migration_started_on_destination
;
387 int total_registrations
;
390 int unregister_current
, unregister_next
;
391 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
393 GHashTable
*blockmap
;
395 /* the RDMAContext for return path */
396 struct RDMAContext
*return_path
;
400 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
401 OBJECT_DECLARE_SIMPLE_TYPE(QIOChannelRDMA
, QIO_CHANNEL_RDMA
)
405 struct QIOChannelRDMA
{
408 RDMAContext
*rdmaout
;
410 bool blocking
; /* XXX we don't actually honour this yet */
414 * Main structure for IB Send/Recv control messages.
415 * This gets prepended at the beginning of every Send/Recv.
417 typedef struct QEMU_PACKED
{
418 uint32_t len
; /* Total length of data portion */
419 uint32_t type
; /* which control command to perform */
420 uint32_t repeat
; /* number of commands in data portion of same type */
424 static void control_to_network(RDMAControlHeader
*control
)
426 control
->type
= htonl(control
->type
);
427 control
->len
= htonl(control
->len
);
428 control
->repeat
= htonl(control
->repeat
);
431 static void network_to_control(RDMAControlHeader
*control
)
433 control
->type
= ntohl(control
->type
);
434 control
->len
= ntohl(control
->len
);
435 control
->repeat
= ntohl(control
->repeat
);
439 * Register a single Chunk.
440 * Information sent by the source VM to inform the dest
441 * to register an single chunk of memory before we can perform
442 * the actual RDMA operation.
444 typedef struct QEMU_PACKED
{
446 uint64_t current_addr
; /* offset into the ram_addr_t space */
447 uint64_t chunk
; /* chunk to lookup if unregistering */
449 uint32_t current_index
; /* which ramblock the chunk belongs to */
451 uint64_t chunks
; /* how many sequential chunks to register */
454 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
456 RDMALocalBlock
*local_block
;
457 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
459 if (local_block
->is_ram_block
) {
461 * current_addr as passed in is an address in the local ram_addr_t
462 * space, we need to translate this for the destination
464 reg
->key
.current_addr
-= local_block
->offset
;
465 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
467 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
468 reg
->current_index
= htonl(reg
->current_index
);
469 reg
->chunks
= htonll(reg
->chunks
);
472 static void network_to_register(RDMARegister
*reg
)
474 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
475 reg
->current_index
= ntohl(reg
->current_index
);
476 reg
->chunks
= ntohll(reg
->chunks
);
479 typedef struct QEMU_PACKED
{
480 uint32_t value
; /* if zero, we will madvise() */
481 uint32_t block_idx
; /* which ram block index */
482 uint64_t offset
; /* Address in remote ram_addr_t space */
483 uint64_t length
; /* length of the chunk */
486 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
488 comp
->value
= htonl(comp
->value
);
490 * comp->offset as passed in is an address in the local ram_addr_t
491 * space, we need to translate this for the destination
493 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
494 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
495 comp
->block_idx
= htonl(comp
->block_idx
);
496 comp
->offset
= htonll(comp
->offset
);
497 comp
->length
= htonll(comp
->length
);
500 static void network_to_compress(RDMACompress
*comp
)
502 comp
->value
= ntohl(comp
->value
);
503 comp
->block_idx
= ntohl(comp
->block_idx
);
504 comp
->offset
= ntohll(comp
->offset
);
505 comp
->length
= ntohll(comp
->length
);
509 * The result of the dest's memory registration produces an "rkey"
510 * which the source VM must reference in order to perform
511 * the RDMA operation.
513 typedef struct QEMU_PACKED
{
517 } RDMARegisterResult
;
519 static void result_to_network(RDMARegisterResult
*result
)
521 result
->rkey
= htonl(result
->rkey
);
522 result
->host_addr
= htonll(result
->host_addr
);
525 static void network_to_result(RDMARegisterResult
*result
)
527 result
->rkey
= ntohl(result
->rkey
);
528 result
->host_addr
= ntohll(result
->host_addr
);
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
;
646 assert(rdma
->blockmap
== NULL
);
647 memset(local
, 0, sizeof *local
);
648 ret
= foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
652 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
653 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
654 rdma
->local_ram_blocks
.nb_blocks
);
660 * Note: If used outside of cleanup, the caller must ensure that the destination
661 * block structures are also updated
663 static void rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
665 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
666 RDMALocalBlock
*old
= local
->block
;
669 if (rdma
->blockmap
) {
670 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
675 for (j
= 0; j
< block
->nb_chunks
; j
++) {
676 if (!block
->pmr
[j
]) {
679 ibv_dereg_mr(block
->pmr
[j
]);
680 rdma
->total_registrations
--;
687 ibv_dereg_mr(block
->mr
);
688 rdma
->total_registrations
--;
692 g_free(block
->transit_bitmap
);
693 block
->transit_bitmap
= NULL
;
695 g_free(block
->unregister_bitmap
);
696 block
->unregister_bitmap
= NULL
;
698 g_free(block
->remote_keys
);
699 block
->remote_keys
= NULL
;
701 g_free(block
->block_name
);
702 block
->block_name
= NULL
;
704 if (rdma
->blockmap
) {
705 for (x
= 0; x
< local
->nb_blocks
; x
++) {
706 g_hash_table_remove(rdma
->blockmap
,
707 (void *)(uintptr_t)old
[x
].offset
);
711 if (local
->nb_blocks
> 1) {
713 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
716 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
719 if (block
->index
< (local
->nb_blocks
- 1)) {
720 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
721 sizeof(RDMALocalBlock
) *
722 (local
->nb_blocks
- (block
->index
+ 1)));
723 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
724 local
->block
[x
].index
--;
728 assert(block
== local
->block
);
732 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
733 block
->offset
, block
->length
,
734 (uintptr_t)(block
->local_host_addr
+ block
->length
),
735 BITS_TO_LONGS(block
->nb_chunks
) *
736 sizeof(unsigned long) * 8, block
->nb_chunks
);
742 if (local
->nb_blocks
&& rdma
->blockmap
) {
743 for (x
= 0; x
< local
->nb_blocks
; x
++) {
744 g_hash_table_insert(rdma
->blockmap
,
745 (void *)(uintptr_t)local
->block
[x
].offset
,
752 * Put in the log file which RDMA device was opened and the details
753 * associated with that device.
755 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
757 struct ibv_port_attr port
;
759 if (ibv_query_port(verbs
, 1, &port
)) {
760 error_report("Failed to query port information");
764 printf("%s RDMA Device opened: kernel name %s "
765 "uverbs device name %s, "
766 "infiniband_verbs class device path %s, "
767 "infiniband class device path %s, "
768 "transport: (%d) %s\n",
771 verbs
->device
->dev_name
,
772 verbs
->device
->dev_path
,
773 verbs
->device
->ibdev_path
,
775 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
776 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
777 ? "Ethernet" : "Unknown"));
781 * Put in the log file the RDMA gid addressing information,
782 * useful for folks who have trouble understanding the
783 * RDMA device hierarchy in the kernel.
785 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
789 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
790 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
791 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
795 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
796 * We will try the next addrinfo struct, and fail if there are
797 * no other valid addresses to bind against.
799 * If user is listening on '[::]', then we will not have a opened a device
800 * yet and have no way of verifying if the device is RoCE or not.
802 * In this case, the source VM will throw an error for ALL types of
803 * connections (both IPv4 and IPv6) if the destination machine does not have
804 * a regular infiniband network available for use.
806 * The only way to guarantee that an error is thrown for broken kernels is
807 * for the management software to choose a *specific* interface at bind time
808 * and validate what time of hardware it is.
810 * Unfortunately, this puts the user in a fix:
812 * If the source VM connects with an IPv4 address without knowing that the
813 * destination has bound to '[::]' the migration will unconditionally fail
814 * unless the management software is explicitly listening on the IPv4
815 * address while using a RoCE-based device.
817 * If the source VM connects with an IPv6 address, then we're OK because we can
818 * throw an error on the source (and similarly on the destination).
820 * But in mixed environments, this will be broken for a while until it is fixed
823 * We do provide a *tiny* bit of help in this function: We can list all of the
824 * devices in the system and check to see if all the devices are RoCE or
827 * If we detect that we have a *pure* RoCE environment, then we can safely
828 * thrown an error even if the management software has specified '[::]' as the
831 * However, if there is are multiple hetergeneous devices, then we cannot make
832 * this assumption and the user just has to be sure they know what they are
835 * Patches are being reviewed on linux-rdma.
837 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
839 /* This bug only exists in linux, to our knowledge. */
841 struct ibv_port_attr port_attr
;
844 * Verbs are only NULL if management has bound to '[::]'.
846 * Let's iterate through all the devices and see if there any pure IB
847 * devices (non-ethernet).
849 * If not, then we can safely proceed with the migration.
850 * Otherwise, there are no guarantees until the bug is fixed in linux.
854 struct ibv_device
**dev_list
= ibv_get_device_list(&num_devices
);
855 bool roce_found
= false;
856 bool ib_found
= false;
858 for (x
= 0; x
< num_devices
; x
++) {
859 verbs
= ibv_open_device(dev_list
[x
]);
861 if (errno
== EPERM
) {
868 if (ibv_query_port(verbs
, 1, &port_attr
)) {
869 ibv_close_device(verbs
);
870 ERROR(errp
, "Could not query initial IB port");
874 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
876 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
880 ibv_close_device(verbs
);
886 fprintf(stderr
, "WARN: migrations may fail:"
887 " IPv6 over RoCE / iWARP in linux"
888 " is broken. But since you appear to have a"
889 " mixed RoCE / IB environment, be sure to only"
890 " migrate over the IB fabric until the kernel "
891 " fixes the bug.\n");
893 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
894 " and your management software has specified '[::]'"
895 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
904 * If we have a verbs context, that means that some other than '[::]' was
905 * used by the management software for binding. In which case we can
906 * actually warn the user about a potentially broken kernel.
909 /* IB ports start with 1, not 0 */
910 if (ibv_query_port(verbs
, 1, &port_attr
)) {
911 ERROR(errp
, "Could not query initial IB port");
915 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
916 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
917 "(but patches on linux-rdma in progress)");
927 * Figure out which RDMA device corresponds to the requested IP hostname
928 * Also create the initial connection manager identifiers for opening
931 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
934 struct rdma_addrinfo
*res
;
936 struct rdma_cm_event
*cm_event
;
937 char ip
[40] = "unknown";
938 struct rdma_addrinfo
*e
;
940 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
941 ERROR(errp
, "RDMA hostname has not been set");
945 /* create CM channel */
946 rdma
->channel
= rdma_create_event_channel();
947 if (!rdma
->channel
) {
948 ERROR(errp
, "could not create CM channel");
953 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
955 ERROR(errp
, "could not create channel id");
956 goto err_resolve_create_id
;
959 snprintf(port_str
, 16, "%d", rdma
->port
);
962 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
964 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
965 goto err_resolve_get_addr
;
968 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
969 inet_ntop(e
->ai_family
,
970 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
971 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
973 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
974 RDMA_RESOLVE_TIMEOUT_MS
);
976 if (e
->ai_family
== AF_INET6
) {
977 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
986 rdma_freeaddrinfo(res
);
987 ERROR(errp
, "could not resolve address %s", rdma
->host
);
988 goto err_resolve_get_addr
;
991 rdma_freeaddrinfo(res
);
992 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
994 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
996 ERROR(errp
, "could not perform event_addr_resolved");
997 goto err_resolve_get_addr
;
1000 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
1001 ERROR(errp
, "result not equal to event_addr_resolved %s",
1002 rdma_event_str(cm_event
->event
));
1003 error_report("rdma_resolve_addr");
1004 rdma_ack_cm_event(cm_event
);
1006 goto err_resolve_get_addr
;
1008 rdma_ack_cm_event(cm_event
);
1011 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1013 ERROR(errp
, "could not resolve rdma route");
1014 goto err_resolve_get_addr
;
1017 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1019 ERROR(errp
, "could not perform event_route_resolved");
1020 goto err_resolve_get_addr
;
1022 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1023 ERROR(errp
, "result not equal to event_route_resolved: %s",
1024 rdma_event_str(cm_event
->event
));
1025 rdma_ack_cm_event(cm_event
);
1027 goto err_resolve_get_addr
;
1029 rdma_ack_cm_event(cm_event
);
1030 rdma
->verbs
= rdma
->cm_id
->verbs
;
1031 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1032 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1035 err_resolve_get_addr
:
1036 rdma_destroy_id(rdma
->cm_id
);
1038 err_resolve_create_id
:
1039 rdma_destroy_event_channel(rdma
->channel
);
1040 rdma
->channel
= NULL
;
1045 * Create protection domain and completion queues
1047 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1050 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1052 error_report("failed to allocate protection domain");
1056 /* create receive completion channel */
1057 rdma
->recv_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1058 if (!rdma
->recv_comp_channel
) {
1059 error_report("failed to allocate receive completion channel");
1060 goto err_alloc_pd_cq
;
1064 * Completion queue can be filled by read work requests.
1066 rdma
->recv_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1067 NULL
, rdma
->recv_comp_channel
, 0);
1068 if (!rdma
->recv_cq
) {
1069 error_report("failed to allocate receive completion queue");
1070 goto err_alloc_pd_cq
;
1073 /* create send completion channel */
1074 rdma
->send_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1075 if (!rdma
->send_comp_channel
) {
1076 error_report("failed to allocate send completion channel");
1077 goto err_alloc_pd_cq
;
1080 rdma
->send_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1081 NULL
, rdma
->send_comp_channel
, 0);
1082 if (!rdma
->send_cq
) {
1083 error_report("failed to allocate send completion queue");
1084 goto err_alloc_pd_cq
;
1091 ibv_dealloc_pd(rdma
->pd
);
1093 if (rdma
->recv_comp_channel
) {
1094 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
1096 if (rdma
->send_comp_channel
) {
1097 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
1099 if (rdma
->recv_cq
) {
1100 ibv_destroy_cq(rdma
->recv_cq
);
1101 rdma
->recv_cq
= NULL
;
1104 rdma
->recv_comp_channel
= NULL
;
1105 rdma
->send_comp_channel
= NULL
;
1111 * Create queue pairs.
1113 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1115 struct ibv_qp_init_attr attr
= { 0 };
1118 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1119 attr
.cap
.max_recv_wr
= 3;
1120 attr
.cap
.max_send_sge
= 1;
1121 attr
.cap
.max_recv_sge
= 1;
1122 attr
.send_cq
= rdma
->send_cq
;
1123 attr
.recv_cq
= rdma
->recv_cq
;
1124 attr
.qp_type
= IBV_QPT_RC
;
1126 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1131 rdma
->qp
= rdma
->cm_id
->qp
;
1135 /* Check whether On-Demand Paging is supported by RDAM device */
1136 static bool rdma_support_odp(struct ibv_context
*dev
)
1138 struct ibv_device_attr_ex attr
= {0};
1139 int ret
= ibv_query_device_ex(dev
, NULL
, &attr
);
1144 if (attr
.odp_caps
.general_caps
& IBV_ODP_SUPPORT
) {
1152 * ibv_advise_mr to avoid RNR NAK error as far as possible.
1153 * The responder mr registering with ODP will sent RNR NAK back to
1154 * the requester in the face of the page fault.
1156 static void qemu_rdma_advise_prefetch_mr(struct ibv_pd
*pd
, uint64_t addr
,
1157 uint32_t len
, uint32_t lkey
,
1158 const char *name
, bool wr
)
1160 #ifdef HAVE_IBV_ADVISE_MR
1162 int advice
= wr
? IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE
:
1163 IBV_ADVISE_MR_ADVICE_PREFETCH
;
1164 struct ibv_sge sg_list
= {.lkey
= lkey
, .addr
= addr
, .length
= len
};
1166 ret
= ibv_advise_mr(pd
, advice
,
1167 IBV_ADVISE_MR_FLAG_FLUSH
, &sg_list
, 1);
1168 /* ignore the error */
1170 trace_qemu_rdma_advise_mr(name
, len
, addr
, strerror(errno
));
1172 trace_qemu_rdma_advise_mr(name
, len
, addr
, "successed");
1177 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1180 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1182 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1183 int access
= IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
;
1185 local
->block
[i
].mr
=
1186 ibv_reg_mr(rdma
->pd
,
1187 local
->block
[i
].local_host_addr
,
1188 local
->block
[i
].length
, access
1191 if (!local
->block
[i
].mr
&&
1192 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1193 access
|= IBV_ACCESS_ON_DEMAND
;
1194 /* register ODP mr */
1195 local
->block
[i
].mr
=
1196 ibv_reg_mr(rdma
->pd
,
1197 local
->block
[i
].local_host_addr
,
1198 local
->block
[i
].length
, access
);
1199 trace_qemu_rdma_register_odp_mr(local
->block
[i
].block_name
);
1201 if (local
->block
[i
].mr
) {
1202 qemu_rdma_advise_prefetch_mr(rdma
->pd
,
1203 (uintptr_t)local
->block
[i
].local_host_addr
,
1204 local
->block
[i
].length
,
1205 local
->block
[i
].mr
->lkey
,
1206 local
->block
[i
].block_name
,
1211 if (!local
->block
[i
].mr
) {
1212 perror("Failed to register local dest ram block!");
1215 rdma
->total_registrations
++;
1218 if (i
>= local
->nb_blocks
) {
1222 for (i
--; i
>= 0; i
--) {
1223 ibv_dereg_mr(local
->block
[i
].mr
);
1224 local
->block
[i
].mr
= NULL
;
1225 rdma
->total_registrations
--;
1233 * Find the ram block that corresponds to the page requested to be
1234 * transmitted by QEMU.
1236 * Once the block is found, also identify which 'chunk' within that
1237 * block that the page belongs to.
1239 * This search cannot fail or the migration will fail.
1241 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1242 uintptr_t block_offset
,
1245 uint64_t *block_index
,
1246 uint64_t *chunk_index
)
1248 uint64_t current_addr
= block_offset
+ offset
;
1249 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1250 (void *) block_offset
);
1252 assert(current_addr
>= block
->offset
);
1253 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1255 *block_index
= block
->index
;
1256 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1257 block
->local_host_addr
+ (current_addr
- block
->offset
));
1263 * Register a chunk with IB. If the chunk was already registered
1264 * previously, then skip.
1266 * Also return the keys associated with the registration needed
1267 * to perform the actual RDMA operation.
1269 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1270 RDMALocalBlock
*block
, uintptr_t host_addr
,
1271 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1272 uint8_t *chunk_start
, uint8_t *chunk_end
)
1276 *lkey
= block
->mr
->lkey
;
1279 *rkey
= block
->mr
->rkey
;
1284 /* allocate memory to store chunk MRs */
1286 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1290 * If 'rkey', then we're the destination, so grant access to the source.
1292 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1294 if (!block
->pmr
[chunk
]) {
1295 uint64_t len
= chunk_end
- chunk_start
;
1296 int access
= rkey
? IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
:
1299 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1301 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1302 if (!block
->pmr
[chunk
] &&
1303 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1304 access
|= IBV_ACCESS_ON_DEMAND
;
1305 /* register ODP mr */
1306 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1307 trace_qemu_rdma_register_odp_mr(block
->block_name
);
1309 if (block
->pmr
[chunk
]) {
1310 qemu_rdma_advise_prefetch_mr(rdma
->pd
, (uintptr_t)chunk_start
,
1311 len
, block
->pmr
[chunk
]->lkey
,
1312 block
->block_name
, rkey
);
1317 if (!block
->pmr
[chunk
]) {
1318 perror("Failed to register chunk!");
1319 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1320 " start %" PRIuPTR
" end %" PRIuPTR
1322 " local %" PRIuPTR
" registrations: %d\n",
1323 block
->index
, chunk
, (uintptr_t)chunk_start
,
1324 (uintptr_t)chunk_end
, host_addr
,
1325 (uintptr_t)block
->local_host_addr
,
1326 rdma
->total_registrations
);
1329 rdma
->total_registrations
++;
1332 *lkey
= block
->pmr
[chunk
]->lkey
;
1335 *rkey
= block
->pmr
[chunk
]->rkey
;
1341 * Register (at connection time) the memory used for control
1344 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1346 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1347 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1348 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1349 if (rdma
->wr_data
[idx
].control_mr
) {
1350 rdma
->total_registrations
++;
1353 error_report("qemu_rdma_reg_control failed");
1358 * Perform a non-optimized memory unregistration after every transfer
1359 * for demonstration purposes, only if pin-all is not requested.
1361 * Potential optimizations:
1362 * 1. Start a new thread to run this function continuously
1364 - and for receipt of unregister messages
1366 * 3. Use workload hints.
1368 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1370 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1372 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1374 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1376 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1377 RDMALocalBlock
*block
=
1378 &(rdma
->local_ram_blocks
.block
[index
]);
1379 RDMARegister reg
= { .current_index
= index
};
1380 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1382 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1383 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1387 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1388 rdma
->unregister_current
);
1390 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1391 rdma
->unregister_current
++;
1393 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1394 rdma
->unregister_current
= 0;
1399 * Unregistration is speculative (because migration is single-threaded
1400 * and we cannot break the protocol's inifinband message ordering).
1401 * Thus, if the memory is currently being used for transmission,
1402 * then abort the attempt to unregister and try again
1403 * later the next time a completion is received for this memory.
1405 clear_bit(chunk
, block
->unregister_bitmap
);
1407 if (test_bit(chunk
, block
->transit_bitmap
)) {
1408 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1412 trace_qemu_rdma_unregister_waiting_send(chunk
);
1414 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1415 block
->pmr
[chunk
] = NULL
;
1416 block
->remote_keys
[chunk
] = 0;
1419 perror("unregistration chunk failed");
1422 rdma
->total_registrations
--;
1424 reg
.key
.chunk
= chunk
;
1425 register_to_network(rdma
, ®
);
1426 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1432 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1438 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1441 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1443 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1444 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1450 * Consult the connection manager to see a work request
1451 * (of any kind) has completed.
1452 * Return the work request ID that completed.
1454 static int qemu_rdma_poll(RDMAContext
*rdma
, struct ibv_cq
*cq
,
1455 uint64_t *wr_id_out
, uint32_t *byte_len
)
1461 ret
= ibv_poll_cq(cq
, 1, &wc
);
1464 *wr_id_out
= RDMA_WRID_NONE
;
1469 error_report("ibv_poll_cq return %d", ret
);
1473 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1475 if (wc
.status
!= IBV_WC_SUCCESS
) {
1476 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1477 wc
.status
, ibv_wc_status_str(wc
.status
));
1478 fprintf(stderr
, "ibv_poll_cq wrid=%" PRIu64
"!\n", wr_id
);
1483 if (rdma
->control_ready_expected
&&
1484 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1485 trace_qemu_rdma_poll_recv(wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
,
1487 rdma
->control_ready_expected
= 0;
1490 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1492 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1494 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1495 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1497 trace_qemu_rdma_poll_write(wr_id
, rdma
->nb_sent
,
1498 index
, chunk
, block
->local_host_addr
,
1499 (void *)(uintptr_t)block
->remote_host_addr
);
1501 clear_bit(chunk
, block
->transit_bitmap
);
1503 if (rdma
->nb_sent
> 0) {
1507 trace_qemu_rdma_poll_other(wr_id
, rdma
->nb_sent
);
1510 *wr_id_out
= wc
.wr_id
;
1512 *byte_len
= wc
.byte_len
;
1518 /* Wait for activity on the completion channel.
1519 * Returns 0 on success, none-0 on error.
1521 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
,
1522 struct ibv_comp_channel
*comp_channel
)
1524 struct rdma_cm_event
*cm_event
;
1528 * Coroutine doesn't start until migration_fd_process_incoming()
1529 * so don't yield unless we know we're running inside of a coroutine.
1531 if (rdma
->migration_started_on_destination
&&
1532 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1533 yield_until_fd_readable(comp_channel
->fd
);
1535 /* This is the source side, we're in a separate thread
1536 * or destination prior to migration_fd_process_incoming()
1537 * after postcopy, the destination also in a separate thread.
1538 * we can't yield; so we have to poll the fd.
1539 * But we need to be able to handle 'cancel' or an error
1540 * without hanging forever.
1542 while (!rdma
->error_state
&& !rdma
->received_error
) {
1544 pfds
[0].fd
= comp_channel
->fd
;
1545 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1546 pfds
[0].revents
= 0;
1548 pfds
[1].fd
= rdma
->channel
->fd
;
1549 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1550 pfds
[1].revents
= 0;
1552 /* 0.1s timeout, should be fine for a 'cancel' */
1553 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1555 case 1: /* fd active */
1556 if (pfds
[0].revents
) {
1560 if (pfds
[1].revents
) {
1561 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1563 error_report("failed to get cm event while wait "
1564 "completion channel");
1568 error_report("receive cm event while wait comp channel,"
1569 "cm event is %d", cm_event
->event
);
1570 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1571 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1572 rdma_ack_cm_event(cm_event
);
1575 rdma_ack_cm_event(cm_event
);
1579 case 0: /* Timeout, go around again */
1582 default: /* Error of some type -
1583 * I don't trust errno from qemu_poll_ns
1585 error_report("%s: poll failed", __func__
);
1589 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1590 /* Bail out and let the cancellation happen */
1596 if (rdma
->received_error
) {
1599 return rdma
->error_state
;
1602 static struct ibv_comp_channel
*to_channel(RDMAContext
*rdma
, uint64_t wrid
)
1604 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_comp_channel
:
1605 rdma
->recv_comp_channel
;
1608 static struct ibv_cq
*to_cq(RDMAContext
*rdma
, uint64_t wrid
)
1610 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_cq
: rdma
->recv_cq
;
1614 * Block until the next work request has completed.
1616 * First poll to see if a work request has already completed,
1619 * If we encounter completed work requests for IDs other than
1620 * the one we're interested in, then that's generally an error.
1622 * The only exception is actual RDMA Write completions. These
1623 * completions only need to be recorded, but do not actually
1624 * need further processing.
1626 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
,
1627 uint64_t wrid_requested
,
1630 int num_cq_events
= 0, ret
= 0;
1633 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1634 struct ibv_comp_channel
*ch
= to_channel(rdma
, wrid_requested
);
1635 struct ibv_cq
*poll_cq
= to_cq(rdma
, wrid_requested
);
1637 if (ibv_req_notify_cq(poll_cq
, 0)) {
1641 while (wr_id
!= wrid_requested
) {
1642 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1647 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1649 if (wr_id
== RDMA_WRID_NONE
) {
1652 if (wr_id
!= wrid_requested
) {
1653 trace_qemu_rdma_block_for_wrid_miss(wrid_requested
, wr_id
);
1657 if (wr_id
== wrid_requested
) {
1662 ret
= qemu_rdma_wait_comp_channel(rdma
, ch
);
1664 goto err_block_for_wrid
;
1667 ret
= ibv_get_cq_event(ch
, &cq
, &cq_ctx
);
1669 perror("ibv_get_cq_event");
1670 goto err_block_for_wrid
;
1675 ret
= -ibv_req_notify_cq(cq
, 0);
1677 goto err_block_for_wrid
;
1680 while (wr_id
!= wrid_requested
) {
1681 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1683 goto err_block_for_wrid
;
1686 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1688 if (wr_id
== RDMA_WRID_NONE
) {
1691 if (wr_id
!= wrid_requested
) {
1692 trace_qemu_rdma_block_for_wrid_miss(wrid_requested
, wr_id
);
1696 if (wr_id
== wrid_requested
) {
1697 goto success_block_for_wrid
;
1701 success_block_for_wrid
:
1702 if (num_cq_events
) {
1703 ibv_ack_cq_events(cq
, num_cq_events
);
1708 if (num_cq_events
) {
1709 ibv_ack_cq_events(cq
, num_cq_events
);
1712 rdma
->error_state
= ret
;
1717 * Post a SEND message work request for the control channel
1718 * containing some data and block until the post completes.
1720 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1721 RDMAControlHeader
*head
)
1724 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1725 struct ibv_send_wr
*bad_wr
;
1726 struct ibv_sge sge
= {
1727 .addr
= (uintptr_t)(wr
->control
),
1728 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1729 .lkey
= wr
->control_mr
->lkey
,
1731 struct ibv_send_wr send_wr
= {
1732 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1733 .opcode
= IBV_WR_SEND
,
1734 .send_flags
= IBV_SEND_SIGNALED
,
1739 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1742 * We don't actually need to do a memcpy() in here if we used
1743 * the "sge" properly, but since we're only sending control messages
1744 * (not RAM in a performance-critical path), then its OK for now.
1746 * The copy makes the RDMAControlHeader simpler to manipulate
1747 * for the time being.
1749 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1750 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1751 control_to_network((void *) wr
->control
);
1754 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1758 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1761 error_report("Failed to use post IB SEND for control");
1765 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1767 error_report("rdma migration: send polling control error");
1774 * Post a RECV work request in anticipation of some future receipt
1775 * of data on the control channel.
1777 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1779 struct ibv_recv_wr
*bad_wr
;
1780 struct ibv_sge sge
= {
1781 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1782 .length
= RDMA_CONTROL_MAX_BUFFER
,
1783 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1786 struct ibv_recv_wr recv_wr
= {
1787 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1793 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1801 * Block and wait for a RECV control channel message to arrive.
1803 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1804 RDMAControlHeader
*head
, uint32_t expecting
, int idx
)
1807 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1811 error_report("rdma migration: recv polling control error!");
1815 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1816 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1818 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1820 if (expecting
== RDMA_CONTROL_NONE
) {
1821 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1823 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1824 error_report("Was expecting a %s (%d) control message"
1825 ", but got: %s (%d), length: %d",
1826 control_desc(expecting
), expecting
,
1827 control_desc(head
->type
), head
->type
, head
->len
);
1828 if (head
->type
== RDMA_CONTROL_ERROR
) {
1829 rdma
->received_error
= true;
1833 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1834 error_report("too long length: %d", head
->len
);
1837 if (sizeof(*head
) + head
->len
!= byte_len
) {
1838 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1846 * When a RECV work request has completed, the work request's
1847 * buffer is pointed at the header.
1849 * This will advance the pointer to the data portion
1850 * of the control message of the work request's buffer that
1851 * was populated after the work request finished.
1853 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1854 RDMAControlHeader
*head
)
1856 rdma
->wr_data
[idx
].control_len
= head
->len
;
1857 rdma
->wr_data
[idx
].control_curr
=
1858 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1862 * This is an 'atomic' high-level operation to deliver a single, unified
1863 * control-channel message.
1865 * Additionally, if the user is expecting some kind of reply to this message,
1866 * they can request a 'resp' response message be filled in by posting an
1867 * additional work request on behalf of the user and waiting for an additional
1870 * The extra (optional) response is used during registration to us from having
1871 * to perform an *additional* exchange of message just to provide a response by
1872 * instead piggy-backing on the acknowledgement.
1874 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1875 uint8_t *data
, RDMAControlHeader
*resp
,
1877 int (*callback
)(RDMAContext
*rdma
))
1882 * Wait until the dest is ready before attempting to deliver the message
1883 * by waiting for a READY message.
1885 if (rdma
->control_ready_expected
) {
1886 RDMAControlHeader resp_ignored
;
1888 ret
= qemu_rdma_exchange_get_response(rdma
, &resp_ignored
,
1897 * If the user is expecting a response, post a WR in anticipation of it.
1900 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1902 error_report("rdma migration: error posting"
1903 " extra control recv for anticipated result!");
1909 * Post a WR to replace the one we just consumed for the READY message.
1911 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1913 error_report("rdma migration: error posting first control recv!");
1918 * Deliver the control message that was requested.
1920 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1923 error_report("Failed to send control buffer!");
1928 * If we're expecting a response, block and wait for it.
1932 trace_qemu_rdma_exchange_send_issue_callback();
1933 ret
= callback(rdma
);
1939 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1940 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1941 resp
->type
, RDMA_WRID_DATA
);
1947 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1949 *resp_idx
= RDMA_WRID_DATA
;
1951 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1954 rdma
->control_ready_expected
= 1;
1960 * This is an 'atomic' high-level operation to receive a single, unified
1961 * control-channel message.
1963 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1966 RDMAControlHeader ready
= {
1968 .type
= RDMA_CONTROL_READY
,
1974 * Inform the source that we're ready to receive a message.
1976 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1979 error_report("Failed to send control buffer!");
1984 * Block and wait for the message.
1986 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1987 expecting
, RDMA_WRID_READY
);
1993 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1996 * Post a new RECV work request to replace the one we just consumed.
1998 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2000 error_report("rdma migration: error posting second control recv!");
2008 * Write an actual chunk of memory using RDMA.
2010 * If we're using dynamic registration on the dest-side, we have to
2011 * send a registration command first.
2013 static int qemu_rdma_write_one(RDMAContext
*rdma
,
2014 int current_index
, uint64_t current_addr
,
2018 struct ibv_send_wr send_wr
= { 0 };
2019 struct ibv_send_wr
*bad_wr
;
2020 int reg_result_idx
, ret
, count
= 0;
2021 uint64_t chunk
, chunks
;
2022 uint8_t *chunk_start
, *chunk_end
;
2023 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
2025 RDMARegisterResult
*reg_result
;
2026 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
2027 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
2028 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
2033 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
2034 (current_addr
- block
->offset
));
2035 sge
.length
= length
;
2037 chunk
= ram_chunk_index(block
->local_host_addr
,
2038 (uint8_t *)(uintptr_t)sge
.addr
);
2039 chunk_start
= ram_chunk_start(block
, chunk
);
2041 if (block
->is_ram_block
) {
2042 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2044 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2048 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2050 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2055 trace_qemu_rdma_write_one_top(chunks
+ 1,
2057 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2059 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2062 while (test_bit(chunk
, block
->transit_bitmap
)) {
2064 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2065 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2067 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2070 error_report("Failed to Wait for previous write to complete "
2071 "block %d chunk %" PRIu64
2072 " current %" PRIu64
" len %" PRIu64
" %d",
2073 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2078 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2079 if (!block
->remote_keys
[chunk
]) {
2081 * This chunk has not yet been registered, so first check to see
2082 * if the entire chunk is zero. If so, tell the other size to
2083 * memset() + madvise() the entire chunk without RDMA.
2086 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2087 RDMACompress comp
= {
2088 .offset
= current_addr
,
2090 .block_idx
= current_index
,
2094 head
.len
= sizeof(comp
);
2095 head
.type
= RDMA_CONTROL_COMPRESS
;
2097 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2098 current_index
, current_addr
);
2100 compress_to_network(rdma
, &comp
);
2101 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2102 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2109 * TODO: Here we are sending something, but we are not
2110 * accounting for anything transferred. The following is wrong:
2112 * stat64_add(&mig_stats.rdma_bytes, sge.length);
2114 * because we are using some kind of compression. I
2115 * would think that head.len would be the more similar
2116 * thing to a correct value.
2118 stat64_add(&mig_stats
.zero_pages
,
2119 sge
.length
/ qemu_target_page_size());
2124 * Otherwise, tell other side to register.
2126 reg
.current_index
= current_index
;
2127 if (block
->is_ram_block
) {
2128 reg
.key
.current_addr
= current_addr
;
2130 reg
.key
.chunk
= chunk
;
2132 reg
.chunks
= chunks
;
2134 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2137 register_to_network(rdma
, ®
);
2138 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2139 &resp
, ®_result_idx
, NULL
);
2144 /* try to overlap this single registration with the one we sent. */
2145 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2146 &sge
.lkey
, NULL
, chunk
,
2147 chunk_start
, chunk_end
)) {
2148 error_report("cannot get lkey");
2152 reg_result
= (RDMARegisterResult
*)
2153 rdma
->wr_data
[reg_result_idx
].control_curr
;
2155 network_to_result(reg_result
);
2157 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2158 reg_result
->rkey
, chunk
);
2160 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2161 block
->remote_host_addr
= reg_result
->host_addr
;
2163 /* already registered before */
2164 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2165 &sge
.lkey
, NULL
, chunk
,
2166 chunk_start
, chunk_end
)) {
2167 error_report("cannot get lkey!");
2172 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2174 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2176 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2177 &sge
.lkey
, NULL
, chunk
,
2178 chunk_start
, chunk_end
)) {
2179 error_report("cannot get lkey!");
2185 * Encode the ram block index and chunk within this wrid.
2186 * We will use this information at the time of completion
2187 * to figure out which bitmap to check against and then which
2188 * chunk in the bitmap to look for.
2190 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2191 current_index
, chunk
);
2193 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2194 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2195 send_wr
.sg_list
= &sge
;
2196 send_wr
.num_sge
= 1;
2197 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2198 (current_addr
- block
->offset
);
2200 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2204 * ibv_post_send() does not return negative error numbers,
2205 * per the specification they are positive - no idea why.
2207 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2209 if (ret
== ENOMEM
) {
2210 trace_qemu_rdma_write_one_queue_full();
2211 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2213 error_report("rdma migration: failed to make "
2214 "room in full send queue! %d", ret
);
2220 } else if (ret
> 0) {
2221 perror("rdma migration: post rdma write failed");
2225 set_bit(chunk
, block
->transit_bitmap
);
2226 stat64_add(&mig_stats
.normal_pages
, sge
.length
/ qemu_target_page_size());
2228 * We are adding to transferred the amount of data written, but no
2229 * overhead at all. I will asume that RDMA is magicaly and don't
2230 * need to transfer (at least) the addresses where it wants to
2231 * write the pages. Here it looks like it should be something
2233 * sizeof(send_wr) + sge.length
2234 * but this being RDMA, who knows.
2236 stat64_add(&mig_stats
.rdma_bytes
, sge
.length
);
2237 ram_transferred_add(sge
.length
);
2238 rdma
->total_writes
++;
2244 * Push out any unwritten RDMA operations.
2246 * We support sending out multiple chunks at the same time.
2247 * Not all of them need to get signaled in the completion queue.
2249 static int qemu_rdma_write_flush(RDMAContext
*rdma
)
2253 if (!rdma
->current_length
) {
2257 ret
= qemu_rdma_write_one(rdma
,
2258 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2266 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2269 rdma
->current_length
= 0;
2270 rdma
->current_addr
= 0;
2275 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2276 uint64_t offset
, uint64_t len
)
2278 RDMALocalBlock
*block
;
2282 if (rdma
->current_index
< 0) {
2286 if (rdma
->current_chunk
< 0) {
2290 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2291 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2292 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2294 if (rdma
->current_length
== 0) {
2299 * Only merge into chunk sequentially.
2301 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2305 if (offset
< block
->offset
) {
2309 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2313 if ((host_addr
+ len
) > chunk_end
) {
2321 * We're not actually writing here, but doing three things:
2323 * 1. Identify the chunk the buffer belongs to.
2324 * 2. If the chunk is full or the buffer doesn't belong to the current
2325 * chunk, then start a new chunk and flush() the old chunk.
2326 * 3. To keep the hardware busy, we also group chunks into batches
2327 * and only require that a batch gets acknowledged in the completion
2328 * queue instead of each individual chunk.
2330 static int qemu_rdma_write(RDMAContext
*rdma
,
2331 uint64_t block_offset
, uint64_t offset
,
2334 uint64_t current_addr
= block_offset
+ offset
;
2335 uint64_t index
= rdma
->current_index
;
2336 uint64_t chunk
= rdma
->current_chunk
;
2339 /* If we cannot merge it, we flush the current buffer first. */
2340 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2341 ret
= qemu_rdma_write_flush(rdma
);
2345 rdma
->current_length
= 0;
2346 rdma
->current_addr
= current_addr
;
2348 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2349 offset
, len
, &index
, &chunk
);
2351 error_report("ram block search failed");
2354 rdma
->current_index
= index
;
2355 rdma
->current_chunk
= chunk
;
2359 rdma
->current_length
+= len
;
2361 /* flush it if buffer is too large */
2362 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2363 return qemu_rdma_write_flush(rdma
);
2369 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2373 if (rdma
->cm_id
&& rdma
->connected
) {
2374 if ((rdma
->error_state
||
2375 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2376 !rdma
->received_error
) {
2377 RDMAControlHeader head
= { .len
= 0,
2378 .type
= RDMA_CONTROL_ERROR
,
2381 error_report("Early error. Sending error.");
2382 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2385 rdma_disconnect(rdma
->cm_id
);
2386 trace_qemu_rdma_cleanup_disconnect();
2387 rdma
->connected
= false;
2390 if (rdma
->channel
) {
2391 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2393 g_free(rdma
->dest_blocks
);
2394 rdma
->dest_blocks
= NULL
;
2396 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2397 if (rdma
->wr_data
[idx
].control_mr
) {
2398 rdma
->total_registrations
--;
2399 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2401 rdma
->wr_data
[idx
].control_mr
= NULL
;
2404 if (rdma
->local_ram_blocks
.block
) {
2405 while (rdma
->local_ram_blocks
.nb_blocks
) {
2406 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2411 rdma_destroy_qp(rdma
->cm_id
);
2414 if (rdma
->recv_cq
) {
2415 ibv_destroy_cq(rdma
->recv_cq
);
2416 rdma
->recv_cq
= NULL
;
2418 if (rdma
->send_cq
) {
2419 ibv_destroy_cq(rdma
->send_cq
);
2420 rdma
->send_cq
= NULL
;
2422 if (rdma
->recv_comp_channel
) {
2423 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
2424 rdma
->recv_comp_channel
= NULL
;
2426 if (rdma
->send_comp_channel
) {
2427 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
2428 rdma
->send_comp_channel
= NULL
;
2431 ibv_dealloc_pd(rdma
->pd
);
2435 rdma_destroy_id(rdma
->cm_id
);
2439 /* the destination side, listen_id and channel is shared */
2440 if (rdma
->listen_id
) {
2441 if (!rdma
->is_return_path
) {
2442 rdma_destroy_id(rdma
->listen_id
);
2444 rdma
->listen_id
= NULL
;
2446 if (rdma
->channel
) {
2447 if (!rdma
->is_return_path
) {
2448 rdma_destroy_event_channel(rdma
->channel
);
2450 rdma
->channel
= NULL
;
2454 if (rdma
->channel
) {
2455 rdma_destroy_event_channel(rdma
->channel
);
2456 rdma
->channel
= NULL
;
2459 g_free(rdma
->host_port
);
2461 rdma
->host_port
= NULL
;
2465 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2468 Error
*local_err
= NULL
, **temp
= &local_err
;
2471 * Will be validated against destination's actual capabilities
2472 * after the connect() completes.
2474 rdma
->pin_all
= pin_all
;
2476 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2478 goto err_rdma_source_init
;
2481 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2483 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2484 " limits may be too low. Please check $ ulimit -a # and "
2485 "search for 'ulimit -l' in the output");
2486 goto err_rdma_source_init
;
2489 ret
= qemu_rdma_alloc_qp(rdma
);
2491 ERROR(temp
, "rdma migration: error allocating qp!");
2492 goto err_rdma_source_init
;
2495 ret
= qemu_rdma_init_ram_blocks(rdma
);
2497 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2498 goto err_rdma_source_init
;
2501 /* Build the hash that maps from offset to RAMBlock */
2502 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2503 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2504 g_hash_table_insert(rdma
->blockmap
,
2505 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2506 &rdma
->local_ram_blocks
.block
[idx
]);
2509 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2510 ret
= qemu_rdma_reg_control(rdma
, idx
);
2512 ERROR(temp
, "rdma migration: error registering %d control!",
2514 goto err_rdma_source_init
;
2520 err_rdma_source_init
:
2521 error_propagate(errp
, local_err
);
2522 qemu_rdma_cleanup(rdma
);
2526 static int qemu_get_cm_event_timeout(RDMAContext
*rdma
,
2527 struct rdma_cm_event
**cm_event
,
2528 long msec
, Error
**errp
)
2531 struct pollfd poll_fd
= {
2532 .fd
= rdma
->channel
->fd
,
2538 ret
= poll(&poll_fd
, 1, msec
);
2539 } while (ret
< 0 && errno
== EINTR
);
2542 ERROR(errp
, "poll cm event timeout");
2544 } else if (ret
< 0) {
2545 ERROR(errp
, "failed to poll cm event, errno=%i", errno
);
2547 } else if (poll_fd
.revents
& POLLIN
) {
2548 return rdma_get_cm_event(rdma
->channel
, cm_event
);
2550 ERROR(errp
, "no POLLIN event, revent=%x", poll_fd
.revents
);
2555 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
, bool return_path
)
2557 RDMACapabilities cap
= {
2558 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2561 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2563 .private_data
= &cap
,
2564 .private_data_len
= sizeof(cap
),
2566 struct rdma_cm_event
*cm_event
;
2570 * Only negotiate the capability with destination if the user
2571 * on the source first requested the capability.
2573 if (rdma
->pin_all
) {
2574 trace_qemu_rdma_connect_pin_all_requested();
2575 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2578 caps_to_network(&cap
);
2580 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2582 ERROR(errp
, "posting second control recv");
2583 goto err_rdma_source_connect
;
2586 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2588 perror("rdma_connect");
2589 ERROR(errp
, "connecting to destination!");
2590 goto err_rdma_source_connect
;
2594 ret
= qemu_get_cm_event_timeout(rdma
, &cm_event
, 5000, errp
);
2596 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2599 perror("rdma_get_cm_event after rdma_connect");
2600 ERROR(errp
, "connecting to destination!");
2601 goto err_rdma_source_connect
;
2604 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2605 error_report("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2606 ERROR(errp
, "connecting to destination!");
2607 rdma_ack_cm_event(cm_event
);
2608 goto err_rdma_source_connect
;
2610 rdma
->connected
= true;
2612 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2613 network_to_caps(&cap
);
2616 * Verify that the *requested* capabilities are supported by the destination
2617 * and disable them otherwise.
2619 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2620 ERROR(errp
, "Server cannot support pinning all memory. "
2621 "Will register memory dynamically.");
2622 rdma
->pin_all
= false;
2625 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2627 rdma_ack_cm_event(cm_event
);
2629 rdma
->control_ready_expected
= 1;
2633 err_rdma_source_connect
:
2634 qemu_rdma_cleanup(rdma
);
2638 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2641 struct rdma_cm_id
*listen_id
;
2642 char ip
[40] = "unknown";
2643 struct rdma_addrinfo
*res
, *e
;
2647 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2648 rdma
->wr_data
[idx
].control_len
= 0;
2649 rdma
->wr_data
[idx
].control_curr
= NULL
;
2652 if (!rdma
->host
|| !rdma
->host
[0]) {
2653 ERROR(errp
, "RDMA host is not set!");
2654 rdma
->error_state
= -EINVAL
;
2657 /* create CM channel */
2658 rdma
->channel
= rdma_create_event_channel();
2659 if (!rdma
->channel
) {
2660 ERROR(errp
, "could not create rdma event channel");
2661 rdma
->error_state
= -EINVAL
;
2666 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2668 ERROR(errp
, "could not create cm_id!");
2669 goto err_dest_init_create_listen_id
;
2672 snprintf(port_str
, 16, "%d", rdma
->port
);
2673 port_str
[15] = '\0';
2675 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2677 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2678 goto err_dest_init_bind_addr
;
2681 ret
= rdma_set_option(listen_id
, RDMA_OPTION_ID
, RDMA_OPTION_ID_REUSEADDR
,
2682 &reuse
, sizeof reuse
);
2684 ERROR(errp
, "Error: could not set REUSEADDR option");
2685 goto err_dest_init_bind_addr
;
2687 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2688 inet_ntop(e
->ai_family
,
2689 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2690 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2691 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2695 if (e
->ai_family
== AF_INET6
) {
2696 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2704 rdma_freeaddrinfo(res
);
2706 ERROR(errp
, "Error: could not rdma_bind_addr!");
2707 goto err_dest_init_bind_addr
;
2710 rdma
->listen_id
= listen_id
;
2711 qemu_rdma_dump_gid("dest_init", listen_id
);
2714 err_dest_init_bind_addr
:
2715 rdma_destroy_id(listen_id
);
2716 err_dest_init_create_listen_id
:
2717 rdma_destroy_event_channel(rdma
->channel
);
2718 rdma
->channel
= NULL
;
2719 rdma
->error_state
= ret
;
2724 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2729 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2730 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2731 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2734 /*the CM channel and CM id is shared*/
2735 rdma_return_path
->channel
= rdma
->channel
;
2736 rdma_return_path
->listen_id
= rdma
->listen_id
;
2738 rdma
->return_path
= rdma_return_path
;
2739 rdma_return_path
->return_path
= rdma
;
2740 rdma_return_path
->is_return_path
= true;
2743 static RDMAContext
*qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2745 RDMAContext
*rdma
= NULL
;
2746 InetSocketAddress
*addr
;
2749 rdma
= g_new0(RDMAContext
, 1);
2750 rdma
->current_index
= -1;
2751 rdma
->current_chunk
= -1;
2753 addr
= g_new(InetSocketAddress
, 1);
2754 if (!inet_parse(addr
, host_port
, NULL
)) {
2755 rdma
->port
= atoi(addr
->port
);
2756 rdma
->host
= g_strdup(addr
->host
);
2757 rdma
->host_port
= g_strdup(host_port
);
2759 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2764 qapi_free_InetSocketAddress(addr
);
2771 * QEMUFile interface to the control channel.
2772 * SEND messages for control only.
2773 * VM's ram is handled with regular RDMA messages.
2775 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2776 const struct iovec
*iov
,
2783 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2789 RCU_READ_LOCK_GUARD();
2790 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
2793 error_setg(errp
, "RDMA control channel output is not set");
2797 CHECK_ERROR_STATE();
2800 * Push out any writes that
2801 * we're queued up for VM's ram.
2803 ret
= qemu_rdma_write_flush(rdma
);
2805 rdma
->error_state
= ret
;
2806 error_setg(errp
, "qemu_rdma_write_flush returned %d", ret
);
2810 for (i
= 0; i
< niov
; i
++) {
2811 size_t remaining
= iov
[i
].iov_len
;
2812 uint8_t * data
= (void *)iov
[i
].iov_base
;
2814 RDMAControlHeader head
= {};
2816 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2820 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2822 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2825 rdma
->error_state
= ret
;
2826 error_setg(errp
, "qemu_rdma_exchange_send returned %d", ret
);
2838 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2839 size_t size
, int idx
)
2843 if (rdma
->wr_data
[idx
].control_len
) {
2844 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2846 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2847 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2848 rdma
->wr_data
[idx
].control_curr
+= len
;
2849 rdma
->wr_data
[idx
].control_len
-= len
;
2856 * QEMUFile interface to the control channel.
2857 * RDMA links don't use bytestreams, so we have to
2858 * return bytes to QEMUFile opportunistically.
2860 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2861 const struct iovec
*iov
,
2868 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2870 RDMAControlHeader head
;
2875 RCU_READ_LOCK_GUARD();
2876 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
2879 error_setg(errp
, "RDMA control channel input is not set");
2883 CHECK_ERROR_STATE();
2885 for (i
= 0; i
< niov
; i
++) {
2886 size_t want
= iov
[i
].iov_len
;
2887 uint8_t *data
= (void *)iov
[i
].iov_base
;
2890 * First, we hold on to the last SEND message we
2891 * were given and dish out the bytes until we run
2894 len
= qemu_rdma_fill(rdma
, data
, want
, 0);
2897 /* Got what we needed, so go to next iovec */
2902 /* If we got any data so far, then don't wait
2903 * for more, just return what we have */
2909 /* We've got nothing at all, so lets wait for
2912 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2915 rdma
->error_state
= ret
;
2916 error_setg(errp
, "qemu_rdma_exchange_recv returned %d", ret
);
2921 * SEND was received with new bytes, now try again.
2923 len
= qemu_rdma_fill(rdma
, data
, want
, 0);
2927 /* Still didn't get enough, so lets just return */
2930 return QIO_CHANNEL_ERR_BLOCK
;
2940 * Block until all the outstanding chunks have been delivered by the hardware.
2942 static int qemu_rdma_drain_cq(RDMAContext
*rdma
)
2946 if (qemu_rdma_write_flush(rdma
) < 0) {
2950 while (rdma
->nb_sent
) {
2951 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2953 error_report("rdma migration: complete polling error!");
2958 qemu_rdma_unregister_waiting(rdma
);
2964 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2968 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2969 /* XXX we should make readv/writev actually honour this :-) */
2970 rioc
->blocking
= blocking
;
2975 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2976 struct QIOChannelRDMASource
{
2978 QIOChannelRDMA
*rioc
;
2979 GIOCondition condition
;
2983 qio_channel_rdma_source_prepare(GSource
*source
,
2986 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2988 GIOCondition cond
= 0;
2991 RCU_READ_LOCK_GUARD();
2992 if (rsource
->condition
== G_IO_IN
) {
2993 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
2995 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
2999 error_report("RDMAContext is NULL when prepare Gsource");
3003 if (rdma
->wr_data
[0].control_len
) {
3008 return cond
& rsource
->condition
;
3012 qio_channel_rdma_source_check(GSource
*source
)
3014 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3016 GIOCondition cond
= 0;
3018 RCU_READ_LOCK_GUARD();
3019 if (rsource
->condition
== G_IO_IN
) {
3020 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3022 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3026 error_report("RDMAContext is NULL when check Gsource");
3030 if (rdma
->wr_data
[0].control_len
) {
3035 return cond
& rsource
->condition
;
3039 qio_channel_rdma_source_dispatch(GSource
*source
,
3040 GSourceFunc callback
,
3043 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
3044 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3046 GIOCondition cond
= 0;
3048 RCU_READ_LOCK_GUARD();
3049 if (rsource
->condition
== G_IO_IN
) {
3050 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3052 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3056 error_report("RDMAContext is NULL when dispatch Gsource");
3060 if (rdma
->wr_data
[0].control_len
) {
3065 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
3066 (cond
& rsource
->condition
),
3071 qio_channel_rdma_source_finalize(GSource
*source
)
3073 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
3075 object_unref(OBJECT(ssource
->rioc
));
3078 static GSourceFuncs qio_channel_rdma_source_funcs
= {
3079 qio_channel_rdma_source_prepare
,
3080 qio_channel_rdma_source_check
,
3081 qio_channel_rdma_source_dispatch
,
3082 qio_channel_rdma_source_finalize
3085 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
3086 GIOCondition condition
)
3088 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3089 QIOChannelRDMASource
*ssource
;
3092 source
= g_source_new(&qio_channel_rdma_source_funcs
,
3093 sizeof(QIOChannelRDMASource
));
3094 ssource
= (QIOChannelRDMASource
*)source
;
3096 ssource
->rioc
= rioc
;
3097 object_ref(OBJECT(rioc
));
3099 ssource
->condition
= condition
;
3104 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
3105 AioContext
*read_ctx
,
3107 AioContext
*write_ctx
,
3108 IOHandler
*io_write
,
3111 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3113 aio_set_fd_handler(read_ctx
, rioc
->rdmain
->recv_comp_channel
->fd
,
3114 io_read
, io_write
, NULL
, NULL
, opaque
);
3115 aio_set_fd_handler(read_ctx
, rioc
->rdmain
->send_comp_channel
->fd
,
3116 io_read
, io_write
, NULL
, NULL
, opaque
);
3118 aio_set_fd_handler(write_ctx
, rioc
->rdmaout
->recv_comp_channel
->fd
,
3119 io_read
, io_write
, NULL
, NULL
, opaque
);
3120 aio_set_fd_handler(write_ctx
, rioc
->rdmaout
->send_comp_channel
->fd
,
3121 io_read
, io_write
, NULL
, NULL
, opaque
);
3125 struct rdma_close_rcu
{
3126 struct rcu_head rcu
;
3127 RDMAContext
*rdmain
;
3128 RDMAContext
*rdmaout
;
3131 /* callback from qio_channel_rdma_close via call_rcu */
3132 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu
*rcu
)
3135 qemu_rdma_cleanup(rcu
->rdmain
);
3139 qemu_rdma_cleanup(rcu
->rdmaout
);
3142 g_free(rcu
->rdmain
);
3143 g_free(rcu
->rdmaout
);
3147 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3150 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3151 RDMAContext
*rdmain
, *rdmaout
;
3152 struct rdma_close_rcu
*rcu
= g_new(struct rdma_close_rcu
, 1);
3154 trace_qemu_rdma_close();
3156 rdmain
= rioc
->rdmain
;
3158 qatomic_rcu_set(&rioc
->rdmain
, NULL
);
3161 rdmaout
= rioc
->rdmaout
;
3163 qatomic_rcu_set(&rioc
->rdmaout
, NULL
);
3166 rcu
->rdmain
= rdmain
;
3167 rcu
->rdmaout
= rdmaout
;
3168 call_rcu(rcu
, qio_channel_rdma_close_rcu
, rcu
);
3174 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3175 QIOChannelShutdown how
,
3178 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3179 RDMAContext
*rdmain
, *rdmaout
;
3181 RCU_READ_LOCK_GUARD();
3183 rdmain
= qatomic_rcu_read(&rioc
->rdmain
);
3184 rdmaout
= qatomic_rcu_read(&rioc
->rdmain
);
3187 case QIO_CHANNEL_SHUTDOWN_READ
:
3189 rdmain
->error_state
= -1;
3192 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3194 rdmaout
->error_state
= -1;
3197 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3200 rdmain
->error_state
= -1;
3203 rdmaout
->error_state
= -1;
3214 * This means that 'block_offset' is a full virtual address that does not
3215 * belong to a RAMBlock of the virtual machine and instead
3216 * represents a private malloc'd memory area that the caller wishes to
3220 * Offset is an offset to be added to block_offset and used
3221 * to also lookup the corresponding RAMBlock.
3223 * @size : Number of bytes to transfer
3225 * @pages_sent : User-specificed pointer to indicate how many pages were
3226 * sent. Usually, this will not be more than a few bytes of
3227 * the protocol because most transfers are sent asynchronously.
3229 static int qemu_rdma_save_page(QEMUFile
*f
, ram_addr_t block_offset
,
3230 ram_addr_t offset
, size_t size
)
3232 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3236 if (migration_in_postcopy()) {
3237 return RAM_SAVE_CONTROL_NOT_SUPP
;
3240 RCU_READ_LOCK_GUARD();
3241 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3247 CHECK_ERROR_STATE();
3252 * Add this page to the current 'chunk'. If the chunk
3253 * is full, or the page doesn't belong to the current chunk,
3254 * an actual RDMA write will occur and a new chunk will be formed.
3256 ret
= qemu_rdma_write(rdma
, block_offset
, offset
, size
);
3258 error_report("rdma migration: write error! %d", ret
);
3263 * Drain the Completion Queue if possible, but do not block,
3266 * If nothing to poll, the end of the iteration will do this
3267 * again to make sure we don't overflow the request queue.
3270 uint64_t wr_id
, wr_id_in
;
3271 ret
= qemu_rdma_poll(rdma
, rdma
->recv_cq
, &wr_id_in
, NULL
);
3274 error_report("rdma migration: polling error! %d", ret
);
3278 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3280 if (wr_id
== RDMA_WRID_NONE
) {
3286 uint64_t wr_id
, wr_id_in
;
3287 ret
= qemu_rdma_poll(rdma
, rdma
->send_cq
, &wr_id_in
, NULL
);
3290 error_report("rdma migration: polling error! %d", ret
);
3294 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3296 if (wr_id
== RDMA_WRID_NONE
) {
3301 return RAM_SAVE_CONTROL_DELAYED
;
3303 rdma
->error_state
= ret
;
3307 static void rdma_accept_incoming_migration(void *opaque
);
3309 static void rdma_cm_poll_handler(void *opaque
)
3311 RDMAContext
*rdma
= opaque
;
3313 struct rdma_cm_event
*cm_event
;
3314 MigrationIncomingState
*mis
= migration_incoming_get_current();
3316 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3318 error_report("get_cm_event failed %d", errno
);
3322 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3323 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3324 if (!rdma
->error_state
&&
3325 migration_incoming_get_current()->state
!=
3326 MIGRATION_STATUS_COMPLETED
) {
3327 error_report("receive cm event, cm event is %d", cm_event
->event
);
3328 rdma
->error_state
= -EPIPE
;
3329 if (rdma
->return_path
) {
3330 rdma
->return_path
->error_state
= -EPIPE
;
3333 rdma_ack_cm_event(cm_event
);
3334 if (mis
->loadvm_co
) {
3335 qemu_coroutine_enter(mis
->loadvm_co
);
3339 rdma_ack_cm_event(cm_event
);
3342 static int qemu_rdma_accept(RDMAContext
*rdma
)
3344 RDMACapabilities cap
;
3345 struct rdma_conn_param conn_param
= {
3346 .responder_resources
= 2,
3347 .private_data
= &cap
,
3348 .private_data_len
= sizeof(cap
),
3350 RDMAContext
*rdma_return_path
= NULL
;
3351 struct rdma_cm_event
*cm_event
;
3352 struct ibv_context
*verbs
;
3356 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3358 goto err_rdma_dest_wait
;
3361 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3362 rdma_ack_cm_event(cm_event
);
3364 goto err_rdma_dest_wait
;
3368 * initialize the RDMAContext for return path for postcopy after first
3369 * connection request reached.
3371 if ((migrate_postcopy() || migrate_return_path())
3372 && !rdma
->is_return_path
) {
3373 rdma_return_path
= qemu_rdma_data_init(rdma
->host_port
, NULL
);
3374 if (rdma_return_path
== NULL
) {
3375 rdma_ack_cm_event(cm_event
);
3377 goto err_rdma_dest_wait
;
3380 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
3383 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3385 network_to_caps(&cap
);
3387 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3388 error_report("Unknown source RDMA version: %d, bailing...",
3390 rdma_ack_cm_event(cm_event
);
3392 goto err_rdma_dest_wait
;
3396 * Respond with only the capabilities this version of QEMU knows about.
3398 cap
.flags
&= known_capabilities
;
3401 * Enable the ones that we do know about.
3402 * Add other checks here as new ones are introduced.
3404 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3405 rdma
->pin_all
= true;
3408 rdma
->cm_id
= cm_event
->id
;
3409 verbs
= cm_event
->id
->verbs
;
3411 rdma_ack_cm_event(cm_event
);
3413 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3415 caps_to_network(&cap
);
3417 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3420 rdma
->verbs
= verbs
;
3421 } else if (rdma
->verbs
!= verbs
) {
3422 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3425 goto err_rdma_dest_wait
;
3428 qemu_rdma_dump_id("dest_init", verbs
);
3430 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3432 error_report("rdma migration: error allocating pd and cq!");
3433 goto err_rdma_dest_wait
;
3436 ret
= qemu_rdma_alloc_qp(rdma
);
3438 error_report("rdma migration: error allocating qp!");
3439 goto err_rdma_dest_wait
;
3442 ret
= qemu_rdma_init_ram_blocks(rdma
);
3444 error_report("rdma migration: error initializing ram blocks!");
3445 goto err_rdma_dest_wait
;
3448 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3449 ret
= qemu_rdma_reg_control(rdma
, idx
);
3451 error_report("rdma: error registering %d control", idx
);
3452 goto err_rdma_dest_wait
;
3456 /* Accept the second connection request for return path */
3457 if ((migrate_postcopy() || migrate_return_path())
3458 && !rdma
->is_return_path
) {
3459 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3461 (void *)(intptr_t)rdma
->return_path
);
3463 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3467 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3469 error_report("rdma_accept returns %d", ret
);
3470 goto err_rdma_dest_wait
;
3473 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3475 error_report("rdma_accept get_cm_event failed %d", ret
);
3476 goto err_rdma_dest_wait
;
3479 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3480 error_report("rdma_accept not event established");
3481 rdma_ack_cm_event(cm_event
);
3483 goto err_rdma_dest_wait
;
3486 rdma_ack_cm_event(cm_event
);
3487 rdma
->connected
= true;
3489 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3491 error_report("rdma migration: error posting second control recv");
3492 goto err_rdma_dest_wait
;
3495 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3500 rdma
->error_state
= ret
;
3501 qemu_rdma_cleanup(rdma
);
3502 g_free(rdma_return_path
);
3506 static int dest_ram_sort_func(const void *a
, const void *b
)
3508 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3509 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3511 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3515 * During each iteration of the migration, we listen for instructions
3516 * by the source VM to perform dynamic page registrations before they
3517 * can perform RDMA operations.
3519 * We respond with the 'rkey'.
3521 * Keep doing this until the source tells us to stop.
3523 static int qemu_rdma_registration_handle(QEMUFile
*f
)
3525 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3526 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3529 RDMAControlHeader unreg_resp
= { .len
= 0,
3530 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3533 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3535 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3537 RDMALocalBlocks
*local
;
3538 RDMAControlHeader head
;
3539 RDMARegister
*reg
, *registers
;
3541 RDMARegisterResult
*reg_result
;
3542 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3543 RDMALocalBlock
*block
;
3550 RCU_READ_LOCK_GUARD();
3551 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3557 CHECK_ERROR_STATE();
3559 local
= &rdma
->local_ram_blocks
;
3561 trace_qemu_rdma_registration_handle_wait();
3563 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3569 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3570 error_report("rdma: Too many requests in this message (%d)."
3571 "Bailing.", head
.repeat
);
3576 switch (head
.type
) {
3577 case RDMA_CONTROL_COMPRESS
:
3578 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3579 network_to_compress(comp
);
3581 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3584 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3585 error_report("rdma: 'compress' bad block index %u (vs %d)",
3586 (unsigned int)comp
->block_idx
,
3587 rdma
->local_ram_blocks
.nb_blocks
);
3591 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3593 host_addr
= block
->local_host_addr
+
3594 (comp
->offset
- block
->offset
);
3596 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3599 case RDMA_CONTROL_REGISTER_FINISHED
:
3600 trace_qemu_rdma_registration_handle_finished();
3603 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3604 trace_qemu_rdma_registration_handle_ram_blocks();
3606 /* Sort our local RAM Block list so it's the same as the source,
3607 * we can do this since we've filled in a src_index in the list
3608 * as we received the RAMBlock list earlier.
3610 qsort(rdma
->local_ram_blocks
.block
,
3611 rdma
->local_ram_blocks
.nb_blocks
,
3612 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3613 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3614 local
->block
[i
].index
= i
;
3617 if (rdma
->pin_all
) {
3618 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3620 error_report("rdma migration: error dest "
3621 "registering ram blocks");
3627 * Dest uses this to prepare to transmit the RAMBlock descriptions
3628 * to the source VM after connection setup.
3629 * Both sides use the "remote" structure to communicate and update
3630 * their "local" descriptions with what was sent.
3632 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3633 rdma
->dest_blocks
[i
].remote_host_addr
=
3634 (uintptr_t)(local
->block
[i
].local_host_addr
);
3636 if (rdma
->pin_all
) {
3637 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3640 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3641 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3643 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3644 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3645 local
->block
[i
].block_name
,
3646 local
->block
[i
].offset
,
3647 local
->block
[i
].length
,
3648 local
->block
[i
].local_host_addr
,
3649 local
->block
[i
].src_index
);
3652 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3653 * sizeof(RDMADestBlock
);
3656 ret
= qemu_rdma_post_send_control(rdma
,
3657 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3660 error_report("rdma migration: error sending remote info");
3665 case RDMA_CONTROL_REGISTER_REQUEST
:
3666 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3668 reg_resp
.repeat
= head
.repeat
;
3669 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3671 for (count
= 0; count
< head
.repeat
; count
++) {
3673 uint8_t *chunk_start
, *chunk_end
;
3675 reg
= ®isters
[count
];
3676 network_to_register(reg
);
3678 reg_result
= &results
[count
];
3680 trace_qemu_rdma_registration_handle_register_loop(count
,
3681 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3683 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3684 error_report("rdma: 'register' bad block index %u (vs %d)",
3685 (unsigned int)reg
->current_index
,
3686 rdma
->local_ram_blocks
.nb_blocks
);
3690 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3691 if (block
->is_ram_block
) {
3692 if (block
->offset
> reg
->key
.current_addr
) {
3693 error_report("rdma: bad register address for block %s"
3694 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3695 block
->block_name
, block
->offset
,
3696 reg
->key
.current_addr
);
3700 host_addr
= (block
->local_host_addr
+
3701 (reg
->key
.current_addr
- block
->offset
));
3702 chunk
= ram_chunk_index(block
->local_host_addr
,
3703 (uint8_t *) host_addr
);
3705 chunk
= reg
->key
.chunk
;
3706 host_addr
= block
->local_host_addr
+
3707 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3708 /* Check for particularly bad chunk value */
3709 if (host_addr
< (void *)block
->local_host_addr
) {
3710 error_report("rdma: bad chunk for block %s"
3712 block
->block_name
, reg
->key
.chunk
);
3717 chunk_start
= ram_chunk_start(block
, chunk
);
3718 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3719 /* avoid "-Waddress-of-packed-member" warning */
3720 uint32_t tmp_rkey
= 0;
3721 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3722 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3723 chunk
, chunk_start
, chunk_end
)) {
3724 error_report("cannot get rkey");
3728 reg_result
->rkey
= tmp_rkey
;
3730 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3732 trace_qemu_rdma_registration_handle_register_rkey(
3735 result_to_network(reg_result
);
3738 ret
= qemu_rdma_post_send_control(rdma
,
3739 (uint8_t *) results
, ®_resp
);
3742 error_report("Failed to send control buffer");
3746 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3747 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3748 unreg_resp
.repeat
= head
.repeat
;
3749 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3751 for (count
= 0; count
< head
.repeat
; count
++) {
3752 reg
= ®isters
[count
];
3753 network_to_register(reg
);
3755 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3756 reg
->current_index
, reg
->key
.chunk
);
3758 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3760 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3761 block
->pmr
[reg
->key
.chunk
] = NULL
;
3764 perror("rdma unregistration chunk failed");
3769 rdma
->total_registrations
--;
3771 trace_qemu_rdma_registration_handle_unregister_success(
3775 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3778 error_report("Failed to send control buffer");
3782 case RDMA_CONTROL_REGISTER_RESULT
:
3783 error_report("Invalid RESULT message at dest.");
3787 error_report("Unknown control message %s", control_desc(head
.type
));
3794 rdma
->error_state
= ret
;
3800 * Called via a ram_control_load_hook during the initial RAM load section which
3801 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3803 * We've already built our local RAMBlock list, but not yet sent the list to
3807 rdma_block_notification_handle(QEMUFile
*f
, const char *name
)
3810 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3814 RCU_READ_LOCK_GUARD();
3815 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3821 /* Find the matching RAMBlock in our local list */
3822 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3823 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3830 error_report("RAMBlock '%s' not found on destination", name
);
3834 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3835 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3836 rdma
->next_src_index
++;
3841 static int rdma_load_hook(QEMUFile
*f
, uint64_t flags
, void *data
)
3844 case RAM_CONTROL_BLOCK_REG
:
3845 return rdma_block_notification_handle(f
, data
);
3847 case RAM_CONTROL_HOOK
:
3848 return qemu_rdma_registration_handle(f
);
3851 /* Shouldn't be called with any other values */
3856 static int qemu_rdma_registration_start(QEMUFile
*f
,
3857 uint64_t flags
, void *data
)
3859 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3862 if (migration_in_postcopy()) {
3866 RCU_READ_LOCK_GUARD();
3867 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3872 CHECK_ERROR_STATE();
3874 trace_qemu_rdma_registration_start(flags
);
3875 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3882 * Inform dest that dynamic registrations are done for now.
3883 * First, flush writes, if any.
3885 static int qemu_rdma_registration_stop(QEMUFile
*f
,
3886 uint64_t flags
, void *data
)
3888 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3890 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3893 if (migration_in_postcopy()) {
3897 RCU_READ_LOCK_GUARD();
3898 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3903 CHECK_ERROR_STATE();
3906 ret
= qemu_rdma_drain_cq(rdma
);
3912 if (flags
== RAM_CONTROL_SETUP
) {
3913 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3914 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3915 int reg_result_idx
, i
, nb_dest_blocks
;
3917 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3918 trace_qemu_rdma_registration_stop_ram();
3921 * Make sure that we parallelize the pinning on both sides.
3922 * For very large guests, doing this serially takes a really
3923 * long time, so we have to 'interleave' the pinning locally
3924 * with the control messages by performing the pinning on this
3925 * side before we receive the control response from the other
3926 * side that the pinning has completed.
3928 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3929 ®_result_idx
, rdma
->pin_all
?
3930 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3932 fprintf(stderr
, "receiving remote info!");
3936 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3939 * The protocol uses two different sets of rkeys (mutually exclusive):
3940 * 1. One key to represent the virtual address of the entire ram block.
3941 * (dynamic chunk registration disabled - pin everything with one rkey.)
3942 * 2. One to represent individual chunks within a ram block.
3943 * (dynamic chunk registration enabled - pin individual chunks.)
3945 * Once the capability is successfully negotiated, the destination transmits
3946 * the keys to use (or sends them later) including the virtual addresses
3947 * and then propagates the remote ram block descriptions to his local copy.
3950 if (local
->nb_blocks
!= nb_dest_blocks
) {
3951 fprintf(stderr
, "ram blocks mismatch (Number of blocks %d vs %d) "
3952 "Your QEMU command line parameters are probably "
3953 "not identical on both the source and destination.",
3954 local
->nb_blocks
, nb_dest_blocks
);
3955 rdma
->error_state
= -EINVAL
;
3959 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3960 memcpy(rdma
->dest_blocks
,
3961 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3962 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3963 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3965 /* We require that the blocks are in the same order */
3966 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3967 fprintf(stderr
, "Block %s/%d has a different length %" PRIu64
3968 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3969 local
->block
[i
].length
,
3970 rdma
->dest_blocks
[i
].length
);
3971 rdma
->error_state
= -EINVAL
;
3974 local
->block
[i
].remote_host_addr
=
3975 rdma
->dest_blocks
[i
].remote_host_addr
;
3976 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3980 trace_qemu_rdma_registration_stop(flags
);
3982 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3983 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3991 rdma
->error_state
= ret
;
3995 static const QEMUFileHooks rdma_read_hooks
= {
3996 .hook_ram_load
= rdma_load_hook
,
3999 static const QEMUFileHooks rdma_write_hooks
= {
4000 .before_ram_iterate
= qemu_rdma_registration_start
,
4001 .after_ram_iterate
= qemu_rdma_registration_stop
,
4002 .save_page
= qemu_rdma_save_page
,
4006 static void qio_channel_rdma_finalize(Object
*obj
)
4008 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
4010 qemu_rdma_cleanup(rioc
->rdmain
);
4011 g_free(rioc
->rdmain
);
4012 rioc
->rdmain
= NULL
;
4014 if (rioc
->rdmaout
) {
4015 qemu_rdma_cleanup(rioc
->rdmaout
);
4016 g_free(rioc
->rdmaout
);
4017 rioc
->rdmaout
= NULL
;
4021 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
4022 void *class_data G_GNUC_UNUSED
)
4024 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
4026 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
4027 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
4028 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
4029 ioc_klass
->io_close
= qio_channel_rdma_close
;
4030 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
4031 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
4032 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
4035 static const TypeInfo qio_channel_rdma_info
= {
4036 .parent
= TYPE_QIO_CHANNEL
,
4037 .name
= TYPE_QIO_CHANNEL_RDMA
,
4038 .instance_size
= sizeof(QIOChannelRDMA
),
4039 .instance_finalize
= qio_channel_rdma_finalize
,
4040 .class_init
= qio_channel_rdma_class_init
,
4043 static void qio_channel_rdma_register_types(void)
4045 type_register_static(&qio_channel_rdma_info
);
4048 type_init(qio_channel_rdma_register_types
);
4050 static QEMUFile
*rdma_new_input(RDMAContext
*rdma
)
4052 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4054 rioc
->file
= qemu_file_new_input(QIO_CHANNEL(rioc
));
4055 rioc
->rdmain
= rdma
;
4056 rioc
->rdmaout
= rdma
->return_path
;
4057 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
4062 static QEMUFile
*rdma_new_output(RDMAContext
*rdma
)
4064 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4066 rioc
->file
= qemu_file_new_output(QIO_CHANNEL(rioc
));
4067 rioc
->rdmaout
= rdma
;
4068 rioc
->rdmain
= rdma
->return_path
;
4069 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
4074 static void rdma_accept_incoming_migration(void *opaque
)
4076 RDMAContext
*rdma
= opaque
;
4079 Error
*local_err
= NULL
;
4081 trace_qemu_rdma_accept_incoming_migration();
4082 ret
= qemu_rdma_accept(rdma
);
4085 fprintf(stderr
, "RDMA ERROR: Migration initialization failed\n");
4089 trace_qemu_rdma_accept_incoming_migration_accepted();
4091 if (rdma
->is_return_path
) {
4095 f
= rdma_new_input(rdma
);
4097 fprintf(stderr
, "RDMA ERROR: could not open RDMA for input\n");
4098 qemu_rdma_cleanup(rdma
);
4102 rdma
->migration_started_on_destination
= 1;
4103 migration_fd_process_incoming(f
, &local_err
);
4105 error_reportf_err(local_err
, "RDMA ERROR:");
4109 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4113 Error
*local_err
= NULL
;
4115 trace_rdma_start_incoming_migration();
4117 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4118 if (ram_block_discard_is_required()) {
4119 error_setg(errp
, "RDMA: cannot disable RAM discard");
4123 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
4128 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
4134 trace_rdma_start_incoming_migration_after_dest_init();
4136 ret
= rdma_listen(rdma
->listen_id
, 5);
4139 ERROR(errp
, "listening on socket!");
4143 trace_rdma_start_incoming_migration_after_rdma_listen();
4145 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4146 NULL
, (void *)(intptr_t)rdma
);
4150 qemu_rdma_cleanup(rdma
);
4152 error_propagate(errp
, local_err
);
4155 g_free(rdma
->host_port
);
4160 void rdma_start_outgoing_migration(void *opaque
,
4161 const char *host_port
, Error
**errp
)
4163 MigrationState
*s
= opaque
;
4164 RDMAContext
*rdma_return_path
= NULL
;
4168 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4169 if (ram_block_discard_is_required()) {
4170 error_setg(errp
, "RDMA: cannot disable RAM discard");
4174 rdma
= qemu_rdma_data_init(host_port
, errp
);
4179 ret
= qemu_rdma_source_init(rdma
, migrate_rdma_pin_all(), errp
);
4185 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4186 ret
= qemu_rdma_connect(rdma
, errp
, false);
4192 /* RDMA postcopy need a separate queue pair for return path */
4193 if (migrate_postcopy() || migrate_return_path()) {
4194 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4196 if (rdma_return_path
== NULL
) {
4197 goto return_path_err
;
4200 ret
= qemu_rdma_source_init(rdma_return_path
,
4201 migrate_rdma_pin_all(), errp
);
4204 goto return_path_err
;
4207 ret
= qemu_rdma_connect(rdma_return_path
, errp
, true);
4210 goto return_path_err
;
4213 rdma
->return_path
= rdma_return_path
;
4214 rdma_return_path
->return_path
= rdma
;
4215 rdma_return_path
->is_return_path
= true;
4218 trace_rdma_start_outgoing_migration_after_rdma_connect();
4220 s
->to_dst_file
= rdma_new_output(rdma
);
4221 migrate_fd_connect(s
, NULL
);
4224 qemu_rdma_cleanup(rdma
);
4227 g_free(rdma_return_path
);