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"
43 #define ERROR(errp, fmt, ...) \
45 if (errp && (*(errp) == NULL)) { \
46 error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \
50 #define RDMA_RESOLVE_TIMEOUT_MS 10000
52 /* Do not merge data if larger than this. */
53 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
54 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
56 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
59 * This is only for non-live state being migrated.
60 * Instead of RDMA_WRITE messages, we use RDMA_SEND
61 * messages for that state, which requires a different
62 * delivery design than main memory.
64 #define RDMA_SEND_INCREMENT 32768
67 * Maximum size infiniband SEND message
69 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
70 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
72 #define RDMA_CONTROL_VERSION_CURRENT 1
74 * Capabilities for negotiation.
76 #define RDMA_CAPABILITY_PIN_ALL 0x01
79 * Add the other flags above to this list of known capabilities
80 * as they are introduced.
82 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
85 * A work request ID is 64-bits and we split up these bits
88 * bits 0-15 : type of control message, 2^16
89 * bits 16-29: ram block index, 2^14
90 * bits 30-63: ram block chunk number, 2^34
92 * The last two bit ranges are only used for RDMA writes,
93 * in order to track their completion and potentially
94 * also track unregistration status of the message.
96 #define RDMA_WRID_TYPE_SHIFT 0UL
97 #define RDMA_WRID_BLOCK_SHIFT 16UL
98 #define RDMA_WRID_CHUNK_SHIFT 30UL
100 #define RDMA_WRID_TYPE_MASK \
101 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
103 #define RDMA_WRID_BLOCK_MASK \
104 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
106 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
109 * RDMA migration protocol:
110 * 1. RDMA Writes (data messages, i.e. RAM)
111 * 2. IB Send/Recv (control channel messages)
115 RDMA_WRID_RDMA_WRITE
= 1,
116 RDMA_WRID_SEND_CONTROL
= 2000,
117 RDMA_WRID_RECV_CONTROL
= 4000,
121 * Work request IDs for IB SEND messages only (not RDMA writes).
122 * This is used by the migration protocol to transmit
123 * control messages (such as device state and registration commands)
125 * We could use more WRs, but we have enough for now.
135 * SEND/RECV IB Control Messages.
138 RDMA_CONTROL_NONE
= 0,
140 RDMA_CONTROL_READY
, /* ready to receive */
141 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
142 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
143 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
144 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
145 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
146 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
147 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
148 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
149 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
154 * Memory and MR structures used to represent an IB Send/Recv work request.
155 * This is *not* used for RDMA writes, only IB Send/Recv.
158 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
159 struct ibv_mr
*control_mr
; /* registration metadata */
160 size_t control_len
; /* length of the message */
161 uint8_t *control_curr
; /* start of unconsumed bytes */
162 } RDMAWorkRequestData
;
165 * Negotiate RDMA capabilities during connection-setup time.
172 static void caps_to_network(RDMACapabilities
*cap
)
174 cap
->version
= htonl(cap
->version
);
175 cap
->flags
= htonl(cap
->flags
);
178 static void network_to_caps(RDMACapabilities
*cap
)
180 cap
->version
= ntohl(cap
->version
);
181 cap
->flags
= ntohl(cap
->flags
);
185 * Representation of a RAMBlock from an RDMA perspective.
186 * This is not transmitted, only local.
187 * This and subsequent structures cannot be linked lists
188 * because we're using a single IB message to transmit
189 * the information. It's small anyway, so a list is overkill.
191 typedef struct RDMALocalBlock
{
193 uint8_t *local_host_addr
; /* local virtual address */
194 uint64_t remote_host_addr
; /* remote virtual address */
197 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
198 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
199 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
200 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
201 int index
; /* which block are we */
202 unsigned int src_index
; /* (Only used on dest) */
205 unsigned long *transit_bitmap
;
206 unsigned long *unregister_bitmap
;
210 * Also represents a RAMblock, but only on the dest.
211 * This gets transmitted by the dest during connection-time
212 * to the source VM and then is used to populate the
213 * corresponding RDMALocalBlock with
214 * the information needed to perform the actual RDMA.
216 typedef struct QEMU_PACKED RDMADestBlock
{
217 uint64_t remote_host_addr
;
220 uint32_t remote_rkey
;
224 static const char *control_desc(unsigned int rdma_control
)
226 static const char *strs
[] = {
227 [RDMA_CONTROL_NONE
] = "NONE",
228 [RDMA_CONTROL_ERROR
] = "ERROR",
229 [RDMA_CONTROL_READY
] = "READY",
230 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
231 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
232 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
233 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
234 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
235 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
236 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
237 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
238 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
241 if (rdma_control
> RDMA_CONTROL_UNREGISTER_FINISHED
) {
242 return "??BAD CONTROL VALUE??";
245 return strs
[rdma_control
];
248 static uint64_t htonll(uint64_t v
)
250 union { uint32_t lv
[2]; uint64_t llv
; } u
;
251 u
.lv
[0] = htonl(v
>> 32);
252 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
256 static uint64_t ntohll(uint64_t v
)
258 union { uint32_t lv
[2]; uint64_t llv
; } u
;
260 return ((uint64_t)ntohl(u
.lv
[0]) << 32) | (uint64_t) ntohl(u
.lv
[1]);
263 static void dest_block_to_network(RDMADestBlock
*db
)
265 db
->remote_host_addr
= htonll(db
->remote_host_addr
);
266 db
->offset
= htonll(db
->offset
);
267 db
->length
= htonll(db
->length
);
268 db
->remote_rkey
= htonl(db
->remote_rkey
);
271 static void network_to_dest_block(RDMADestBlock
*db
)
273 db
->remote_host_addr
= ntohll(db
->remote_host_addr
);
274 db
->offset
= ntohll(db
->offset
);
275 db
->length
= ntohll(db
->length
);
276 db
->remote_rkey
= ntohl(db
->remote_rkey
);
280 * Virtual address of the above structures used for transmitting
281 * the RAMBlock descriptions at connection-time.
282 * This structure is *not* transmitted.
284 typedef struct RDMALocalBlocks
{
286 bool init
; /* main memory init complete */
287 RDMALocalBlock
*block
;
291 * Main data structure for RDMA state.
292 * While there is only one copy of this structure being allocated right now,
293 * this is the place where one would start if you wanted to consider
294 * having more than one RDMA connection open at the same time.
296 typedef struct RDMAContext
{
301 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
304 * This is used by *_exchange_send() to figure out whether or not
305 * the initial "READY" message has already been received or not.
306 * This is because other functions may potentially poll() and detect
307 * the READY message before send() does, in which case we need to
308 * know if it completed.
310 int control_ready_expected
;
312 /* number of outstanding writes */
315 /* store info about current buffer so that we can
316 merge it with future sends */
317 uint64_t current_addr
;
318 uint64_t current_length
;
319 /* index of ram block the current buffer belongs to */
321 /* index of the chunk in the current ram block */
327 * infiniband-specific variables for opening the device
328 * and maintaining connection state and so forth.
330 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
331 * cm_id->verbs, cm_id->channel, and cm_id->qp.
333 struct rdma_cm_id
*cm_id
; /* connection manager ID */
334 struct rdma_cm_id
*listen_id
;
337 struct ibv_context
*verbs
;
338 struct rdma_event_channel
*channel
;
339 struct ibv_qp
*qp
; /* queue pair */
340 struct ibv_comp_channel
*recv_comp_channel
; /* recv completion channel */
341 struct ibv_comp_channel
*send_comp_channel
; /* send completion channel */
342 struct ibv_pd
*pd
; /* protection domain */
343 struct ibv_cq
*recv_cq
; /* recvieve completion queue */
344 struct ibv_cq
*send_cq
; /* send completion queue */
347 * If a previous write failed (perhaps because of a failed
348 * memory registration, then do not attempt any future work
349 * and remember the error state.
356 * Description of ram blocks used throughout the code.
358 RDMALocalBlocks local_ram_blocks
;
359 RDMADestBlock
*dest_blocks
;
361 /* Index of the next RAMBlock received during block registration */
362 unsigned int next_src_index
;
365 * Migration on *destination* started.
366 * Then use coroutine yield function.
367 * Source runs in a thread, so we don't care.
369 int migration_started_on_destination
;
371 int total_registrations
;
374 int unregister_current
, unregister_next
;
375 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
377 GHashTable
*blockmap
;
379 /* the RDMAContext for return path */
380 struct RDMAContext
*return_path
;
384 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
385 OBJECT_DECLARE_SIMPLE_TYPE(QIOChannelRDMA
, QIO_CHANNEL_RDMA
)
389 struct QIOChannelRDMA
{
392 RDMAContext
*rdmaout
;
394 bool blocking
; /* XXX we don't actually honour this yet */
398 * Main structure for IB Send/Recv control messages.
399 * This gets prepended at the beginning of every Send/Recv.
401 typedef struct QEMU_PACKED
{
402 uint32_t len
; /* Total length of data portion */
403 uint32_t type
; /* which control command to perform */
404 uint32_t repeat
; /* number of commands in data portion of same type */
408 static void control_to_network(RDMAControlHeader
*control
)
410 control
->type
= htonl(control
->type
);
411 control
->len
= htonl(control
->len
);
412 control
->repeat
= htonl(control
->repeat
);
415 static void network_to_control(RDMAControlHeader
*control
)
417 control
->type
= ntohl(control
->type
);
418 control
->len
= ntohl(control
->len
);
419 control
->repeat
= ntohl(control
->repeat
);
423 * Register a single Chunk.
424 * Information sent by the source VM to inform the dest
425 * to register an single chunk of memory before we can perform
426 * the actual RDMA operation.
428 typedef struct QEMU_PACKED
{
430 uint64_t current_addr
; /* offset into the ram_addr_t space */
431 uint64_t chunk
; /* chunk to lookup if unregistering */
433 uint32_t current_index
; /* which ramblock the chunk belongs to */
435 uint64_t chunks
; /* how many sequential chunks to register */
438 static bool rdma_errored(RDMAContext
*rdma
)
440 if (rdma
->errored
&& !rdma
->error_reported
) {
441 error_report("RDMA is in an error state waiting migration"
443 rdma
->error_reported
= true;
445 return rdma
->errored
;
448 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
450 RDMALocalBlock
*local_block
;
451 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
453 if (local_block
->is_ram_block
) {
455 * current_addr as passed in is an address in the local ram_addr_t
456 * space, we need to translate this for the destination
458 reg
->key
.current_addr
-= local_block
->offset
;
459 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
461 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
462 reg
->current_index
= htonl(reg
->current_index
);
463 reg
->chunks
= htonll(reg
->chunks
);
466 static void network_to_register(RDMARegister
*reg
)
468 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
469 reg
->current_index
= ntohl(reg
->current_index
);
470 reg
->chunks
= ntohll(reg
->chunks
);
473 typedef struct QEMU_PACKED
{
474 uint32_t value
; /* if zero, we will madvise() */
475 uint32_t block_idx
; /* which ram block index */
476 uint64_t offset
; /* Address in remote ram_addr_t space */
477 uint64_t length
; /* length of the chunk */
480 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
482 comp
->value
= htonl(comp
->value
);
484 * comp->offset as passed in is an address in the local ram_addr_t
485 * space, we need to translate this for the destination
487 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
488 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
489 comp
->block_idx
= htonl(comp
->block_idx
);
490 comp
->offset
= htonll(comp
->offset
);
491 comp
->length
= htonll(comp
->length
);
494 static void network_to_compress(RDMACompress
*comp
)
496 comp
->value
= ntohl(comp
->value
);
497 comp
->block_idx
= ntohl(comp
->block_idx
);
498 comp
->offset
= ntohll(comp
->offset
);
499 comp
->length
= ntohll(comp
->length
);
503 * The result of the dest's memory registration produces an "rkey"
504 * which the source VM must reference in order to perform
505 * the RDMA operation.
507 typedef struct QEMU_PACKED
{
511 } RDMARegisterResult
;
513 static void result_to_network(RDMARegisterResult
*result
)
515 result
->rkey
= htonl(result
->rkey
);
516 result
->host_addr
= htonll(result
->host_addr
);
519 static void network_to_result(RDMARegisterResult
*result
)
521 result
->rkey
= ntohl(result
->rkey
);
522 result
->host_addr
= ntohll(result
->host_addr
);
525 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
526 uint8_t *data
, RDMAControlHeader
*resp
,
528 int (*callback
)(RDMAContext
*rdma
));
530 static inline uint64_t ram_chunk_index(const uint8_t *start
,
533 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
536 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
539 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
540 (i
<< RDMA_REG_CHUNK_SHIFT
));
543 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
546 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
547 (1UL << RDMA_REG_CHUNK_SHIFT
);
549 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
550 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
556 static void rdma_add_block(RDMAContext
*rdma
, const char *block_name
,
558 ram_addr_t block_offset
, uint64_t length
)
560 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
561 RDMALocalBlock
*block
;
562 RDMALocalBlock
*old
= local
->block
;
564 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
+ 1);
566 if (local
->nb_blocks
) {
569 if (rdma
->blockmap
) {
570 for (x
= 0; x
< local
->nb_blocks
; x
++) {
571 g_hash_table_remove(rdma
->blockmap
,
572 (void *)(uintptr_t)old
[x
].offset
);
573 g_hash_table_insert(rdma
->blockmap
,
574 (void *)(uintptr_t)old
[x
].offset
,
578 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
582 block
= &local
->block
[local
->nb_blocks
];
584 block
->block_name
= g_strdup(block_name
);
585 block
->local_host_addr
= host_addr
;
586 block
->offset
= block_offset
;
587 block
->length
= length
;
588 block
->index
= local
->nb_blocks
;
589 block
->src_index
= ~0U; /* Filled in by the receipt of the block list */
590 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
591 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
592 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
593 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
594 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
595 block
->remote_keys
= g_new0(uint32_t, block
->nb_chunks
);
597 block
->is_ram_block
= local
->init
? false : true;
599 if (rdma
->blockmap
) {
600 g_hash_table_insert(rdma
->blockmap
, (void *)(uintptr_t)block_offset
, block
);
603 trace_rdma_add_block(block_name
, local
->nb_blocks
,
604 (uintptr_t) block
->local_host_addr
,
605 block
->offset
, block
->length
,
606 (uintptr_t) (block
->local_host_addr
+ block
->length
),
607 BITS_TO_LONGS(block
->nb_chunks
) *
608 sizeof(unsigned long) * 8,
615 * Memory regions need to be registered with the device and queue pairs setup
616 * in advanced before the migration starts. This tells us where the RAM blocks
617 * are so that we can register them individually.
619 static int qemu_rdma_init_one_block(RAMBlock
*rb
, void *opaque
)
621 const char *block_name
= qemu_ram_get_idstr(rb
);
622 void *host_addr
= qemu_ram_get_host_addr(rb
);
623 ram_addr_t block_offset
= qemu_ram_get_offset(rb
);
624 ram_addr_t length
= qemu_ram_get_used_length(rb
);
625 rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
630 * Identify the RAMBlocks and their quantity. They will be references to
631 * identify chunk boundaries inside each RAMBlock and also be referenced
632 * during dynamic page registration.
634 static void qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
636 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
639 assert(rdma
->blockmap
== NULL
);
640 memset(local
, 0, sizeof *local
);
641 ret
= foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
643 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
644 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
645 rdma
->local_ram_blocks
.nb_blocks
);
650 * Note: If used outside of cleanup, the caller must ensure that the destination
651 * block structures are also updated
653 static void rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
655 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
656 RDMALocalBlock
*old
= local
->block
;
659 if (rdma
->blockmap
) {
660 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
665 for (j
= 0; j
< block
->nb_chunks
; j
++) {
666 if (!block
->pmr
[j
]) {
669 ibv_dereg_mr(block
->pmr
[j
]);
670 rdma
->total_registrations
--;
677 ibv_dereg_mr(block
->mr
);
678 rdma
->total_registrations
--;
682 g_free(block
->transit_bitmap
);
683 block
->transit_bitmap
= NULL
;
685 g_free(block
->unregister_bitmap
);
686 block
->unregister_bitmap
= NULL
;
688 g_free(block
->remote_keys
);
689 block
->remote_keys
= NULL
;
691 g_free(block
->block_name
);
692 block
->block_name
= NULL
;
694 if (rdma
->blockmap
) {
695 for (x
= 0; x
< local
->nb_blocks
; x
++) {
696 g_hash_table_remove(rdma
->blockmap
,
697 (void *)(uintptr_t)old
[x
].offset
);
701 if (local
->nb_blocks
> 1) {
703 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
706 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
709 if (block
->index
< (local
->nb_blocks
- 1)) {
710 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
711 sizeof(RDMALocalBlock
) *
712 (local
->nb_blocks
- (block
->index
+ 1)));
713 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
714 local
->block
[x
].index
--;
718 assert(block
== local
->block
);
722 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
723 block
->offset
, block
->length
,
724 (uintptr_t)(block
->local_host_addr
+ block
->length
),
725 BITS_TO_LONGS(block
->nb_chunks
) *
726 sizeof(unsigned long) * 8, block
->nb_chunks
);
732 if (local
->nb_blocks
&& rdma
->blockmap
) {
733 for (x
= 0; x
< local
->nb_blocks
; x
++) {
734 g_hash_table_insert(rdma
->blockmap
,
735 (void *)(uintptr_t)local
->block
[x
].offset
,
742 * Put in the log file which RDMA device was opened and the details
743 * associated with that device.
745 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
747 struct ibv_port_attr port
;
749 if (ibv_query_port(verbs
, 1, &port
)) {
750 error_report("Failed to query port information");
754 printf("%s RDMA Device opened: kernel name %s "
755 "uverbs device name %s, "
756 "infiniband_verbs class device path %s, "
757 "infiniband class device path %s, "
758 "transport: (%d) %s\n",
761 verbs
->device
->dev_name
,
762 verbs
->device
->dev_path
,
763 verbs
->device
->ibdev_path
,
765 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
766 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
767 ? "Ethernet" : "Unknown"));
771 * Put in the log file the RDMA gid addressing information,
772 * useful for folks who have trouble understanding the
773 * RDMA device hierarchy in the kernel.
775 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
779 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
780 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
781 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
785 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
786 * We will try the next addrinfo struct, and fail if there are
787 * no other valid addresses to bind against.
789 * If user is listening on '[::]', then we will not have a opened a device
790 * yet and have no way of verifying if the device is RoCE or not.
792 * In this case, the source VM will throw an error for ALL types of
793 * connections (both IPv4 and IPv6) if the destination machine does not have
794 * a regular infiniband network available for use.
796 * The only way to guarantee that an error is thrown for broken kernels is
797 * for the management software to choose a *specific* interface at bind time
798 * and validate what time of hardware it is.
800 * Unfortunately, this puts the user in a fix:
802 * If the source VM connects with an IPv4 address without knowing that the
803 * destination has bound to '[::]' the migration will unconditionally fail
804 * unless the management software is explicitly listening on the IPv4
805 * address while using a RoCE-based device.
807 * If the source VM connects with an IPv6 address, then we're OK because we can
808 * throw an error on the source (and similarly on the destination).
810 * But in mixed environments, this will be broken for a while until it is fixed
813 * We do provide a *tiny* bit of help in this function: We can list all of the
814 * devices in the system and check to see if all the devices are RoCE or
817 * If we detect that we have a *pure* RoCE environment, then we can safely
818 * thrown an error even if the management software has specified '[::]' as the
821 * However, if there is are multiple hetergeneous devices, then we cannot make
822 * this assumption and the user just has to be sure they know what they are
825 * Patches are being reviewed on linux-rdma.
827 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
829 /* This bug only exists in linux, to our knowledge. */
831 struct ibv_port_attr port_attr
;
834 * Verbs are only NULL if management has bound to '[::]'.
836 * Let's iterate through all the devices and see if there any pure IB
837 * devices (non-ethernet).
839 * If not, then we can safely proceed with the migration.
840 * Otherwise, there are no guarantees until the bug is fixed in linux.
844 struct ibv_device
**dev_list
= ibv_get_device_list(&num_devices
);
845 bool roce_found
= false;
846 bool ib_found
= false;
848 for (x
= 0; x
< num_devices
; x
++) {
849 verbs
= ibv_open_device(dev_list
[x
]);
851 * ibv_open_device() is not documented to set errno. If
852 * it does, it's somebody else's doc bug. If it doesn't,
853 * the use of errno below is wrong.
854 * TODO Find out whether ibv_open_device() sets errno.
857 if (errno
== EPERM
) {
860 error_setg_errno(errp
, errno
,
861 "could not open RDMA device context");
866 if (ibv_query_port(verbs
, 1, &port_attr
)) {
867 ibv_close_device(verbs
);
868 ERROR(errp
, "Could not query initial IB port");
872 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
874 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
878 ibv_close_device(verbs
);
884 fprintf(stderr
, "WARN: migrations may fail:"
885 " IPv6 over RoCE / iWARP in linux"
886 " is broken. But since you appear to have a"
887 " mixed RoCE / IB environment, be sure to only"
888 " migrate over the IB fabric until the kernel "
889 " fixes the bug.\n");
891 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
892 " and your management software has specified '[::]'"
893 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
902 * If we have a verbs context, that means that some other than '[::]' was
903 * used by the management software for binding. In which case we can
904 * actually warn the user about a potentially broken kernel.
907 /* IB ports start with 1, not 0 */
908 if (ibv_query_port(verbs
, 1, &port_attr
)) {
909 ERROR(errp
, "Could not query initial IB port");
913 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
914 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
915 "(but patches on linux-rdma in progress)");
925 * Figure out which RDMA device corresponds to the requested IP hostname
926 * Also create the initial connection manager identifiers for opening
929 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
932 struct rdma_addrinfo
*res
;
934 struct rdma_cm_event
*cm_event
;
935 char ip
[40] = "unknown";
936 struct rdma_addrinfo
*e
;
938 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
939 ERROR(errp
, "RDMA hostname has not been set");
943 /* create CM channel */
944 rdma
->channel
= rdma_create_event_channel();
945 if (!rdma
->channel
) {
946 ERROR(errp
, "could not create CM channel");
951 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
953 ERROR(errp
, "could not create channel id");
954 goto err_resolve_create_id
;
957 snprintf(port_str
, 16, "%d", rdma
->port
);
960 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
962 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
963 goto err_resolve_get_addr
;
966 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
967 inet_ntop(e
->ai_family
,
968 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
969 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
971 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
972 RDMA_RESOLVE_TIMEOUT_MS
);
974 if (e
->ai_family
== AF_INET6
) {
975 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
984 rdma_freeaddrinfo(res
);
985 ERROR(errp
, "could not resolve address %s", rdma
->host
);
986 goto err_resolve_get_addr
;
989 rdma_freeaddrinfo(res
);
990 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
992 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
994 ERROR(errp
, "could not perform event_addr_resolved");
995 goto err_resolve_get_addr
;
998 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
999 ERROR(errp
, "result not equal to event_addr_resolved %s",
1000 rdma_event_str(cm_event
->event
));
1001 error_report("rdma_resolve_addr");
1002 rdma_ack_cm_event(cm_event
);
1003 goto err_resolve_get_addr
;
1005 rdma_ack_cm_event(cm_event
);
1008 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1010 ERROR(errp
, "could not resolve rdma route");
1011 goto err_resolve_get_addr
;
1014 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1016 ERROR(errp
, "could not perform event_route_resolved");
1017 goto err_resolve_get_addr
;
1019 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1020 ERROR(errp
, "result not equal to event_route_resolved: %s",
1021 rdma_event_str(cm_event
->event
));
1022 rdma_ack_cm_event(cm_event
);
1023 goto err_resolve_get_addr
;
1025 rdma_ack_cm_event(cm_event
);
1026 rdma
->verbs
= rdma
->cm_id
->verbs
;
1027 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1028 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1031 err_resolve_get_addr
:
1032 rdma_destroy_id(rdma
->cm_id
);
1034 err_resolve_create_id
:
1035 rdma_destroy_event_channel(rdma
->channel
);
1036 rdma
->channel
= NULL
;
1041 * Create protection domain and completion queues
1043 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1046 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1048 error_report("failed to allocate protection domain");
1052 /* create receive completion channel */
1053 rdma
->recv_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1054 if (!rdma
->recv_comp_channel
) {
1055 error_report("failed to allocate receive completion channel");
1056 goto err_alloc_pd_cq
;
1060 * Completion queue can be filled by read work requests.
1062 rdma
->recv_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1063 NULL
, rdma
->recv_comp_channel
, 0);
1064 if (!rdma
->recv_cq
) {
1065 error_report("failed to allocate receive completion queue");
1066 goto err_alloc_pd_cq
;
1069 /* create send completion channel */
1070 rdma
->send_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1071 if (!rdma
->send_comp_channel
) {
1072 error_report("failed to allocate send completion channel");
1073 goto err_alloc_pd_cq
;
1076 rdma
->send_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1077 NULL
, rdma
->send_comp_channel
, 0);
1078 if (!rdma
->send_cq
) {
1079 error_report("failed to allocate send completion queue");
1080 goto err_alloc_pd_cq
;
1087 ibv_dealloc_pd(rdma
->pd
);
1089 if (rdma
->recv_comp_channel
) {
1090 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
1092 if (rdma
->send_comp_channel
) {
1093 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
1095 if (rdma
->recv_cq
) {
1096 ibv_destroy_cq(rdma
->recv_cq
);
1097 rdma
->recv_cq
= NULL
;
1100 rdma
->recv_comp_channel
= NULL
;
1101 rdma
->send_comp_channel
= NULL
;
1107 * Create queue pairs.
1109 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1111 struct ibv_qp_init_attr attr
= { 0 };
1114 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1115 attr
.cap
.max_recv_wr
= 3;
1116 attr
.cap
.max_send_sge
= 1;
1117 attr
.cap
.max_recv_sge
= 1;
1118 attr
.send_cq
= rdma
->send_cq
;
1119 attr
.recv_cq
= rdma
->recv_cq
;
1120 attr
.qp_type
= IBV_QPT_RC
;
1122 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1127 rdma
->qp
= rdma
->cm_id
->qp
;
1131 /* Check whether On-Demand Paging is supported by RDAM device */
1132 static bool rdma_support_odp(struct ibv_context
*dev
)
1134 struct ibv_device_attr_ex attr
= {0};
1135 int ret
= ibv_query_device_ex(dev
, NULL
, &attr
);
1140 if (attr
.odp_caps
.general_caps
& IBV_ODP_SUPPORT
) {
1148 * ibv_advise_mr to avoid RNR NAK error as far as possible.
1149 * The responder mr registering with ODP will sent RNR NAK back to
1150 * the requester in the face of the page fault.
1152 static void qemu_rdma_advise_prefetch_mr(struct ibv_pd
*pd
, uint64_t addr
,
1153 uint32_t len
, uint32_t lkey
,
1154 const char *name
, bool wr
)
1156 #ifdef HAVE_IBV_ADVISE_MR
1158 int advice
= wr
? IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE
:
1159 IBV_ADVISE_MR_ADVICE_PREFETCH
;
1160 struct ibv_sge sg_list
= {.lkey
= lkey
, .addr
= addr
, .length
= len
};
1162 ret
= ibv_advise_mr(pd
, advice
,
1163 IBV_ADVISE_MR_FLAG_FLUSH
, &sg_list
, 1);
1164 /* ignore the error */
1165 trace_qemu_rdma_advise_mr(name
, len
, addr
, strerror(ret
));
1169 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1172 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1174 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1175 int access
= IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
;
1177 local
->block
[i
].mr
=
1178 ibv_reg_mr(rdma
->pd
,
1179 local
->block
[i
].local_host_addr
,
1180 local
->block
[i
].length
, access
1183 * ibv_reg_mr() is not documented to set errno. If it does,
1184 * it's somebody else's doc bug. If it doesn't, the use of
1185 * errno below is wrong.
1186 * TODO Find out whether ibv_reg_mr() sets errno.
1188 if (!local
->block
[i
].mr
&&
1189 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1190 access
|= IBV_ACCESS_ON_DEMAND
;
1191 /* register ODP mr */
1192 local
->block
[i
].mr
=
1193 ibv_reg_mr(rdma
->pd
,
1194 local
->block
[i
].local_host_addr
,
1195 local
->block
[i
].length
, access
);
1196 trace_qemu_rdma_register_odp_mr(local
->block
[i
].block_name
);
1198 if (local
->block
[i
].mr
) {
1199 qemu_rdma_advise_prefetch_mr(rdma
->pd
,
1200 (uintptr_t)local
->block
[i
].local_host_addr
,
1201 local
->block
[i
].length
,
1202 local
->block
[i
].mr
->lkey
,
1203 local
->block
[i
].block_name
,
1208 if (!local
->block
[i
].mr
) {
1209 perror("Failed to register local dest ram block!");
1212 rdma
->total_registrations
++;
1215 if (i
>= local
->nb_blocks
) {
1219 for (i
--; i
>= 0; i
--) {
1220 ibv_dereg_mr(local
->block
[i
].mr
);
1221 local
->block
[i
].mr
= NULL
;
1222 rdma
->total_registrations
--;
1230 * Find the ram block that corresponds to the page requested to be
1231 * transmitted by QEMU.
1233 * Once the block is found, also identify which 'chunk' within that
1234 * block that the page belongs to.
1236 static void qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1237 uintptr_t block_offset
,
1240 uint64_t *block_index
,
1241 uint64_t *chunk_index
)
1243 uint64_t current_addr
= block_offset
+ offset
;
1244 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1245 (void *) block_offset
);
1247 assert(current_addr
>= block
->offset
);
1248 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1250 *block_index
= block
->index
;
1251 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1252 block
->local_host_addr
+ (current_addr
- block
->offset
));
1256 * Register a chunk with IB. If the chunk was already registered
1257 * previously, then skip.
1259 * Also return the keys associated with the registration needed
1260 * to perform the actual RDMA operation.
1262 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1263 RDMALocalBlock
*block
, uintptr_t host_addr
,
1264 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1265 uint8_t *chunk_start
, uint8_t *chunk_end
)
1269 *lkey
= block
->mr
->lkey
;
1272 *rkey
= block
->mr
->rkey
;
1277 /* allocate memory to store chunk MRs */
1279 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1283 * If 'rkey', then we're the destination, so grant access to the source.
1285 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1287 if (!block
->pmr
[chunk
]) {
1288 uint64_t len
= chunk_end
- chunk_start
;
1289 int access
= rkey
? IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
:
1292 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1294 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1296 * ibv_reg_mr() is not documented to set errno. If it does,
1297 * it's somebody else's doc bug. If it doesn't, the use of
1298 * errno below is wrong.
1299 * TODO Find out whether ibv_reg_mr() sets errno.
1301 if (!block
->pmr
[chunk
] &&
1302 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1303 access
|= IBV_ACCESS_ON_DEMAND
;
1304 /* register ODP mr */
1305 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1306 trace_qemu_rdma_register_odp_mr(block
->block_name
);
1308 if (block
->pmr
[chunk
]) {
1309 qemu_rdma_advise_prefetch_mr(rdma
->pd
, (uintptr_t)chunk_start
,
1310 len
, block
->pmr
[chunk
]->lkey
,
1311 block
->block_name
, rkey
);
1316 if (!block
->pmr
[chunk
]) {
1317 perror("Failed to register chunk!");
1318 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1319 " start %" PRIuPTR
" end %" PRIuPTR
1321 " local %" PRIuPTR
" registrations: %d\n",
1322 block
->index
, chunk
, (uintptr_t)chunk_start
,
1323 (uintptr_t)chunk_end
, host_addr
,
1324 (uintptr_t)block
->local_host_addr
,
1325 rdma
->total_registrations
);
1328 rdma
->total_registrations
++;
1331 *lkey
= block
->pmr
[chunk
]->lkey
;
1334 *rkey
= block
->pmr
[chunk
]->rkey
;
1340 * Register (at connection time) the memory used for control
1343 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1345 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1346 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1347 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1348 if (rdma
->wr_data
[idx
].control_mr
) {
1349 rdma
->total_registrations
++;
1352 error_report("qemu_rdma_reg_control failed");
1357 * Perform a non-optimized memory unregistration after every transfer
1358 * for demonstration purposes, only if pin-all is not requested.
1360 * Potential optimizations:
1361 * 1. Start a new thread to run this function continuously
1363 - and for receipt of unregister messages
1365 * 3. Use workload hints.
1367 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1369 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1371 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1373 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1375 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1376 RDMALocalBlock
*block
=
1377 &(rdma
->local_ram_blocks
.block
[index
]);
1378 RDMARegister reg
= { .current_index
= index
};
1379 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1381 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1382 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1386 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1387 rdma
->unregister_current
);
1389 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1390 rdma
->unregister_current
++;
1392 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1393 rdma
->unregister_current
= 0;
1398 * Unregistration is speculative (because migration is single-threaded
1399 * and we cannot break the protocol's inifinband message ordering).
1400 * Thus, if the memory is currently being used for transmission,
1401 * then abort the attempt to unregister and try again
1402 * later the next time a completion is received for this memory.
1404 clear_bit(chunk
, block
->unregister_bitmap
);
1406 if (test_bit(chunk
, block
->transit_bitmap
)) {
1407 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1411 trace_qemu_rdma_unregister_waiting_send(chunk
);
1413 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1414 block
->pmr
[chunk
] = NULL
;
1415 block
->remote_keys
[chunk
] = 0;
1419 * FIXME perror() is problematic, bcause ibv_dereg_mr() is
1420 * not documented to set errno. Will go away later in
1423 perror("unregistration chunk failed");
1426 rdma
->total_registrations
--;
1428 reg
.key
.chunk
= chunk
;
1429 register_to_network(rdma
, ®
);
1430 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1436 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1442 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1445 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1447 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1448 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1454 * Consult the connection manager to see a work request
1455 * (of any kind) has completed.
1456 * Return the work request ID that completed.
1458 static int qemu_rdma_poll(RDMAContext
*rdma
, struct ibv_cq
*cq
,
1459 uint64_t *wr_id_out
, uint32_t *byte_len
)
1465 ret
= ibv_poll_cq(cq
, 1, &wc
);
1468 *wr_id_out
= RDMA_WRID_NONE
;
1473 error_report("ibv_poll_cq failed");
1477 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1479 if (wc
.status
!= IBV_WC_SUCCESS
) {
1480 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1481 wc
.status
, ibv_wc_status_str(wc
.status
));
1482 fprintf(stderr
, "ibv_poll_cq wrid=%" PRIu64
"!\n", wr_id
);
1487 if (rdma
->control_ready_expected
&&
1488 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1489 trace_qemu_rdma_poll_recv(wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
,
1491 rdma
->control_ready_expected
= 0;
1494 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1496 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1498 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1499 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1501 trace_qemu_rdma_poll_write(wr_id
, rdma
->nb_sent
,
1502 index
, chunk
, block
->local_host_addr
,
1503 (void *)(uintptr_t)block
->remote_host_addr
);
1505 clear_bit(chunk
, block
->transit_bitmap
);
1507 if (rdma
->nb_sent
> 0) {
1511 trace_qemu_rdma_poll_other(wr_id
, rdma
->nb_sent
);
1514 *wr_id_out
= wc
.wr_id
;
1516 *byte_len
= wc
.byte_len
;
1522 /* Wait for activity on the completion channel.
1523 * Returns 0 on success, none-0 on error.
1525 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
,
1526 struct ibv_comp_channel
*comp_channel
)
1528 struct rdma_cm_event
*cm_event
;
1532 * Coroutine doesn't start until migration_fd_process_incoming()
1533 * so don't yield unless we know we're running inside of a coroutine.
1535 if (rdma
->migration_started_on_destination
&&
1536 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1537 yield_until_fd_readable(comp_channel
->fd
);
1539 /* This is the source side, we're in a separate thread
1540 * or destination prior to migration_fd_process_incoming()
1541 * after postcopy, the destination also in a separate thread.
1542 * we can't yield; so we have to poll the fd.
1543 * But we need to be able to handle 'cancel' or an error
1544 * without hanging forever.
1546 while (!rdma
->errored
&& !rdma
->received_error
) {
1548 pfds
[0].fd
= comp_channel
->fd
;
1549 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1550 pfds
[0].revents
= 0;
1552 pfds
[1].fd
= rdma
->channel
->fd
;
1553 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1554 pfds
[1].revents
= 0;
1556 /* 0.1s timeout, should be fine for a 'cancel' */
1557 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1559 case 1: /* fd active */
1560 if (pfds
[0].revents
) {
1564 if (pfds
[1].revents
) {
1565 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1567 error_report("failed to get cm event while wait "
1568 "completion channel");
1572 error_report("receive cm event while wait comp channel,"
1573 "cm event is %d", cm_event
->event
);
1574 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1575 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1576 rdma_ack_cm_event(cm_event
);
1579 rdma_ack_cm_event(cm_event
);
1583 case 0: /* Timeout, go around again */
1586 default: /* Error of some type -
1587 * I don't trust errno from qemu_poll_ns
1589 error_report("%s: poll failed", __func__
);
1593 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1594 /* Bail out and let the cancellation happen */
1600 if (rdma
->received_error
) {
1603 return -rdma
->errored
;
1606 static struct ibv_comp_channel
*to_channel(RDMAContext
*rdma
, uint64_t wrid
)
1608 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_comp_channel
:
1609 rdma
->recv_comp_channel
;
1612 static struct ibv_cq
*to_cq(RDMAContext
*rdma
, uint64_t wrid
)
1614 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_cq
: rdma
->recv_cq
;
1618 * Block until the next work request has completed.
1620 * First poll to see if a work request has already completed,
1623 * If we encounter completed work requests for IDs other than
1624 * the one we're interested in, then that's generally an error.
1626 * The only exception is actual RDMA Write completions. These
1627 * completions only need to be recorded, but do not actually
1628 * need further processing.
1630 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
,
1631 uint64_t wrid_requested
,
1634 int num_cq_events
= 0, ret
;
1637 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1638 struct ibv_comp_channel
*ch
= to_channel(rdma
, wrid_requested
);
1639 struct ibv_cq
*poll_cq
= to_cq(rdma
, wrid_requested
);
1641 if (ibv_req_notify_cq(poll_cq
, 0)) {
1645 while (wr_id
!= wrid_requested
) {
1646 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1651 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1653 if (wr_id
== RDMA_WRID_NONE
) {
1656 if (wr_id
!= wrid_requested
) {
1657 trace_qemu_rdma_block_for_wrid_miss(wrid_requested
, wr_id
);
1661 if (wr_id
== wrid_requested
) {
1666 ret
= qemu_rdma_wait_comp_channel(rdma
, ch
);
1668 goto err_block_for_wrid
;
1671 ret
= ibv_get_cq_event(ch
, &cq
, &cq_ctx
);
1674 * FIXME perror() is problematic, because ibv_reg_mr() is
1675 * not documented to set errno. Will go away later in
1678 perror("ibv_get_cq_event");
1679 goto err_block_for_wrid
;
1684 if (ibv_req_notify_cq(cq
, 0)) {
1685 goto err_block_for_wrid
;
1688 while (wr_id
!= wrid_requested
) {
1689 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1691 goto err_block_for_wrid
;
1694 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1696 if (wr_id
== RDMA_WRID_NONE
) {
1699 if (wr_id
!= wrid_requested
) {
1700 trace_qemu_rdma_block_for_wrid_miss(wrid_requested
, wr_id
);
1704 if (wr_id
== wrid_requested
) {
1705 goto success_block_for_wrid
;
1709 success_block_for_wrid
:
1710 if (num_cq_events
) {
1711 ibv_ack_cq_events(cq
, num_cq_events
);
1716 if (num_cq_events
) {
1717 ibv_ack_cq_events(cq
, num_cq_events
);
1720 rdma
->errored
= true;
1725 * Post a SEND message work request for the control channel
1726 * containing some data and block until the post completes.
1728 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1729 RDMAControlHeader
*head
)
1732 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1733 struct ibv_send_wr
*bad_wr
;
1734 struct ibv_sge sge
= {
1735 .addr
= (uintptr_t)(wr
->control
),
1736 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1737 .lkey
= wr
->control_mr
->lkey
,
1739 struct ibv_send_wr send_wr
= {
1740 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1741 .opcode
= IBV_WR_SEND
,
1742 .send_flags
= IBV_SEND_SIGNALED
,
1747 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1750 * We don't actually need to do a memcpy() in here if we used
1751 * the "sge" properly, but since we're only sending control messages
1752 * (not RAM in a performance-critical path), then its OK for now.
1754 * The copy makes the RDMAControlHeader simpler to manipulate
1755 * for the time being.
1757 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1758 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1759 control_to_network((void *) wr
->control
);
1762 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1766 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1769 error_report("Failed to use post IB SEND for control");
1773 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1775 error_report("rdma migration: send polling control error");
1783 * Post a RECV work request in anticipation of some future receipt
1784 * of data on the control channel.
1786 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1788 struct ibv_recv_wr
*bad_wr
;
1789 struct ibv_sge sge
= {
1790 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1791 .length
= RDMA_CONTROL_MAX_BUFFER
,
1792 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1795 struct ibv_recv_wr recv_wr
= {
1796 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1802 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1810 * Block and wait for a RECV control channel message to arrive.
1812 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1813 RDMAControlHeader
*head
, uint32_t expecting
, int idx
)
1816 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1820 error_report("rdma migration: recv polling control error!");
1824 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1825 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1827 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1829 if (expecting
== RDMA_CONTROL_NONE
) {
1830 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1832 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1833 error_report("Was expecting a %s (%d) control message"
1834 ", but got: %s (%d), length: %d",
1835 control_desc(expecting
), expecting
,
1836 control_desc(head
->type
), head
->type
, head
->len
);
1837 if (head
->type
== RDMA_CONTROL_ERROR
) {
1838 rdma
->received_error
= true;
1842 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1843 error_report("too long length: %d", head
->len
);
1846 if (sizeof(*head
) + head
->len
!= byte_len
) {
1847 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1855 * When a RECV work request has completed, the work request's
1856 * buffer is pointed at the header.
1858 * This will advance the pointer to the data portion
1859 * of the control message of the work request's buffer that
1860 * was populated after the work request finished.
1862 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1863 RDMAControlHeader
*head
)
1865 rdma
->wr_data
[idx
].control_len
= head
->len
;
1866 rdma
->wr_data
[idx
].control_curr
=
1867 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1871 * This is an 'atomic' high-level operation to deliver a single, unified
1872 * control-channel message.
1874 * Additionally, if the user is expecting some kind of reply to this message,
1875 * they can request a 'resp' response message be filled in by posting an
1876 * additional work request on behalf of the user and waiting for an additional
1879 * The extra (optional) response is used during registration to us from having
1880 * to perform an *additional* exchange of message just to provide a response by
1881 * instead piggy-backing on the acknowledgement.
1883 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1884 uint8_t *data
, RDMAControlHeader
*resp
,
1886 int (*callback
)(RDMAContext
*rdma
))
1891 * Wait until the dest is ready before attempting to deliver the message
1892 * by waiting for a READY message.
1894 if (rdma
->control_ready_expected
) {
1895 RDMAControlHeader resp_ignored
;
1897 ret
= qemu_rdma_exchange_get_response(rdma
, &resp_ignored
,
1906 * If the user is expecting a response, post a WR in anticipation of it.
1909 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1911 error_report("rdma migration: error posting"
1912 " extra control recv for anticipated result!");
1918 * Post a WR to replace the one we just consumed for the READY message.
1920 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1922 error_report("rdma migration: error posting first control recv!");
1927 * Deliver the control message that was requested.
1929 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1932 error_report("Failed to send control buffer!");
1937 * If we're expecting a response, block and wait for it.
1941 trace_qemu_rdma_exchange_send_issue_callback();
1942 ret
= callback(rdma
);
1948 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1949 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1950 resp
->type
, RDMA_WRID_DATA
);
1956 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1958 *resp_idx
= RDMA_WRID_DATA
;
1960 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1963 rdma
->control_ready_expected
= 1;
1969 * This is an 'atomic' high-level operation to receive a single, unified
1970 * control-channel message.
1972 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1975 RDMAControlHeader ready
= {
1977 .type
= RDMA_CONTROL_READY
,
1983 * Inform the source that we're ready to receive a message.
1985 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1988 error_report("Failed to send control buffer!");
1993 * Block and wait for the message.
1995 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1996 expecting
, RDMA_WRID_READY
);
2002 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
2005 * Post a new RECV work request to replace the one we just consumed.
2007 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2009 error_report("rdma migration: error posting second control recv!");
2017 * Write an actual chunk of memory using RDMA.
2019 * If we're using dynamic registration on the dest-side, we have to
2020 * send a registration command first.
2022 static int qemu_rdma_write_one(RDMAContext
*rdma
,
2023 int current_index
, uint64_t current_addr
,
2027 struct ibv_send_wr send_wr
= { 0 };
2028 struct ibv_send_wr
*bad_wr
;
2029 int reg_result_idx
, ret
, count
= 0;
2030 uint64_t chunk
, chunks
;
2031 uint8_t *chunk_start
, *chunk_end
;
2032 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
2034 RDMARegisterResult
*reg_result
;
2035 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
2036 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
2037 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
2042 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
2043 (current_addr
- block
->offset
));
2044 sge
.length
= length
;
2046 chunk
= ram_chunk_index(block
->local_host_addr
,
2047 (uint8_t *)(uintptr_t)sge
.addr
);
2048 chunk_start
= ram_chunk_start(block
, chunk
);
2050 if (block
->is_ram_block
) {
2051 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2053 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2057 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2059 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2064 trace_qemu_rdma_write_one_top(chunks
+ 1,
2066 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2068 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2071 while (test_bit(chunk
, block
->transit_bitmap
)) {
2073 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2074 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2076 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2079 error_report("Failed to Wait for previous write to complete "
2080 "block %d chunk %" PRIu64
2081 " current %" PRIu64
" len %" PRIu64
" %d",
2082 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2087 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2088 if (!block
->remote_keys
[chunk
]) {
2090 * This chunk has not yet been registered, so first check to see
2091 * if the entire chunk is zero. If so, tell the other size to
2092 * memset() + madvise() the entire chunk without RDMA.
2095 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2096 RDMACompress comp
= {
2097 .offset
= current_addr
,
2099 .block_idx
= current_index
,
2103 head
.len
= sizeof(comp
);
2104 head
.type
= RDMA_CONTROL_COMPRESS
;
2106 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2107 current_index
, current_addr
);
2109 compress_to_network(rdma
, &comp
);
2110 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2111 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2118 * TODO: Here we are sending something, but we are not
2119 * accounting for anything transferred. The following is wrong:
2121 * stat64_add(&mig_stats.rdma_bytes, sge.length);
2123 * because we are using some kind of compression. I
2124 * would think that head.len would be the more similar
2125 * thing to a correct value.
2127 stat64_add(&mig_stats
.zero_pages
,
2128 sge
.length
/ qemu_target_page_size());
2133 * Otherwise, tell other side to register.
2135 reg
.current_index
= current_index
;
2136 if (block
->is_ram_block
) {
2137 reg
.key
.current_addr
= current_addr
;
2139 reg
.key
.chunk
= chunk
;
2141 reg
.chunks
= chunks
;
2143 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2146 register_to_network(rdma
, ®
);
2147 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2148 &resp
, ®_result_idx
, NULL
);
2153 /* try to overlap this single registration with the one we sent. */
2154 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2155 &sge
.lkey
, NULL
, chunk
,
2156 chunk_start
, chunk_end
)) {
2157 error_report("cannot get lkey");
2161 reg_result
= (RDMARegisterResult
*)
2162 rdma
->wr_data
[reg_result_idx
].control_curr
;
2164 network_to_result(reg_result
);
2166 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2167 reg_result
->rkey
, chunk
);
2169 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2170 block
->remote_host_addr
= reg_result
->host_addr
;
2172 /* already registered before */
2173 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2174 &sge
.lkey
, NULL
, chunk
,
2175 chunk_start
, chunk_end
)) {
2176 error_report("cannot get lkey!");
2181 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2183 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2185 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2186 &sge
.lkey
, NULL
, chunk
,
2187 chunk_start
, chunk_end
)) {
2188 error_report("cannot get lkey!");
2194 * Encode the ram block index and chunk within this wrid.
2195 * We will use this information at the time of completion
2196 * to figure out which bitmap to check against and then which
2197 * chunk in the bitmap to look for.
2199 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2200 current_index
, chunk
);
2202 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2203 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2204 send_wr
.sg_list
= &sge
;
2205 send_wr
.num_sge
= 1;
2206 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2207 (current_addr
- block
->offset
);
2209 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2213 * ibv_post_send() does not return negative error numbers,
2214 * per the specification they are positive - no idea why.
2216 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2218 if (ret
== ENOMEM
) {
2219 trace_qemu_rdma_write_one_queue_full();
2220 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2222 error_report("rdma migration: failed to make "
2223 "room in full send queue!");
2229 } else if (ret
> 0) {
2231 * FIXME perror() is problematic, because whether
2232 * ibv_post_send() sets errno is unclear. Will go away later
2235 perror("rdma migration: post rdma write failed");
2239 set_bit(chunk
, block
->transit_bitmap
);
2240 stat64_add(&mig_stats
.normal_pages
, sge
.length
/ qemu_target_page_size());
2242 * We are adding to transferred the amount of data written, but no
2243 * overhead at all. I will asume that RDMA is magicaly and don't
2244 * need to transfer (at least) the addresses where it wants to
2245 * write the pages. Here it looks like it should be something
2247 * sizeof(send_wr) + sge.length
2248 * but this being RDMA, who knows.
2250 stat64_add(&mig_stats
.rdma_bytes
, sge
.length
);
2251 ram_transferred_add(sge
.length
);
2252 rdma
->total_writes
++;
2258 * Push out any unwritten RDMA operations.
2260 * We support sending out multiple chunks at the same time.
2261 * Not all of them need to get signaled in the completion queue.
2263 static int qemu_rdma_write_flush(RDMAContext
*rdma
)
2267 if (!rdma
->current_length
) {
2271 ret
= qemu_rdma_write_one(rdma
,
2272 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2280 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2283 rdma
->current_length
= 0;
2284 rdma
->current_addr
= 0;
2289 static inline bool qemu_rdma_buffer_mergeable(RDMAContext
*rdma
,
2290 uint64_t offset
, uint64_t len
)
2292 RDMALocalBlock
*block
;
2296 if (rdma
->current_index
< 0) {
2300 if (rdma
->current_chunk
< 0) {
2304 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2305 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2306 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2308 if (rdma
->current_length
== 0) {
2313 * Only merge into chunk sequentially.
2315 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2319 if (offset
< block
->offset
) {
2323 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2327 if ((host_addr
+ len
) > chunk_end
) {
2335 * We're not actually writing here, but doing three things:
2337 * 1. Identify the chunk the buffer belongs to.
2338 * 2. If the chunk is full or the buffer doesn't belong to the current
2339 * chunk, then start a new chunk and flush() the old chunk.
2340 * 3. To keep the hardware busy, we also group chunks into batches
2341 * and only require that a batch gets acknowledged in the completion
2342 * queue instead of each individual chunk.
2344 static int qemu_rdma_write(RDMAContext
*rdma
,
2345 uint64_t block_offset
, uint64_t offset
,
2348 uint64_t current_addr
= block_offset
+ offset
;
2349 uint64_t index
= rdma
->current_index
;
2350 uint64_t chunk
= rdma
->current_chunk
;
2353 /* If we cannot merge it, we flush the current buffer first. */
2354 if (!qemu_rdma_buffer_mergeable(rdma
, current_addr
, len
)) {
2355 ret
= qemu_rdma_write_flush(rdma
);
2359 rdma
->current_length
= 0;
2360 rdma
->current_addr
= current_addr
;
2362 qemu_rdma_search_ram_block(rdma
, block_offset
,
2363 offset
, len
, &index
, &chunk
);
2364 rdma
->current_index
= index
;
2365 rdma
->current_chunk
= chunk
;
2369 rdma
->current_length
+= len
;
2371 /* flush it if buffer is too large */
2372 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2373 return qemu_rdma_write_flush(rdma
);
2379 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2383 if (rdma
->cm_id
&& rdma
->connected
) {
2384 if ((rdma
->errored
||
2385 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2386 !rdma
->received_error
) {
2387 RDMAControlHeader head
= { .len
= 0,
2388 .type
= RDMA_CONTROL_ERROR
,
2391 error_report("Early error. Sending error.");
2392 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2395 rdma_disconnect(rdma
->cm_id
);
2396 trace_qemu_rdma_cleanup_disconnect();
2397 rdma
->connected
= false;
2400 if (rdma
->channel
) {
2401 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2403 g_free(rdma
->dest_blocks
);
2404 rdma
->dest_blocks
= NULL
;
2406 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2407 if (rdma
->wr_data
[idx
].control_mr
) {
2408 rdma
->total_registrations
--;
2409 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2411 rdma
->wr_data
[idx
].control_mr
= NULL
;
2414 if (rdma
->local_ram_blocks
.block
) {
2415 while (rdma
->local_ram_blocks
.nb_blocks
) {
2416 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2421 rdma_destroy_qp(rdma
->cm_id
);
2424 if (rdma
->recv_cq
) {
2425 ibv_destroy_cq(rdma
->recv_cq
);
2426 rdma
->recv_cq
= NULL
;
2428 if (rdma
->send_cq
) {
2429 ibv_destroy_cq(rdma
->send_cq
);
2430 rdma
->send_cq
= NULL
;
2432 if (rdma
->recv_comp_channel
) {
2433 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
2434 rdma
->recv_comp_channel
= NULL
;
2436 if (rdma
->send_comp_channel
) {
2437 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
2438 rdma
->send_comp_channel
= NULL
;
2441 ibv_dealloc_pd(rdma
->pd
);
2445 rdma_destroy_id(rdma
->cm_id
);
2449 /* the destination side, listen_id and channel is shared */
2450 if (rdma
->listen_id
) {
2451 if (!rdma
->is_return_path
) {
2452 rdma_destroy_id(rdma
->listen_id
);
2454 rdma
->listen_id
= NULL
;
2456 if (rdma
->channel
) {
2457 if (!rdma
->is_return_path
) {
2458 rdma_destroy_event_channel(rdma
->channel
);
2460 rdma
->channel
= NULL
;
2464 if (rdma
->channel
) {
2465 rdma_destroy_event_channel(rdma
->channel
);
2466 rdma
->channel
= NULL
;
2469 g_free(rdma
->host_port
);
2471 rdma
->host_port
= NULL
;
2475 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2480 * Will be validated against destination's actual capabilities
2481 * after the connect() completes.
2483 rdma
->pin_all
= pin_all
;
2485 ret
= qemu_rdma_resolve_host(rdma
, errp
);
2487 goto err_rdma_source_init
;
2490 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2492 ERROR(errp
, "rdma migration: error allocating pd and cq! Your mlock()"
2493 " limits may be too low. Please check $ ulimit -a # and "
2494 "search for 'ulimit -l' in the output");
2495 goto err_rdma_source_init
;
2498 ret
= qemu_rdma_alloc_qp(rdma
);
2500 ERROR(errp
, "rdma migration: error allocating qp!");
2501 goto err_rdma_source_init
;
2504 qemu_rdma_init_ram_blocks(rdma
);
2506 /* Build the hash that maps from offset to RAMBlock */
2507 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2508 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2509 g_hash_table_insert(rdma
->blockmap
,
2510 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2511 &rdma
->local_ram_blocks
.block
[idx
]);
2514 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2515 ret
= qemu_rdma_reg_control(rdma
, idx
);
2517 ERROR(errp
, "rdma migration: error registering %d control!",
2519 goto err_rdma_source_init
;
2525 err_rdma_source_init
:
2526 qemu_rdma_cleanup(rdma
);
2530 static int qemu_get_cm_event_timeout(RDMAContext
*rdma
,
2531 struct rdma_cm_event
**cm_event
,
2532 long msec
, Error
**errp
)
2535 struct pollfd poll_fd
= {
2536 .fd
= rdma
->channel
->fd
,
2542 ret
= poll(&poll_fd
, 1, msec
);
2543 } while (ret
< 0 && errno
== EINTR
);
2546 ERROR(errp
, "poll cm event timeout");
2548 } else if (ret
< 0) {
2549 ERROR(errp
, "failed to poll cm event, errno=%i", errno
);
2551 } else if (poll_fd
.revents
& POLLIN
) {
2552 if (rdma_get_cm_event(rdma
->channel
, cm_event
) < 0) {
2553 ERROR(errp
, "failed to get cm event");
2558 ERROR(errp
, "no POLLIN event, revent=%x", poll_fd
.revents
);
2563 static int qemu_rdma_connect(RDMAContext
*rdma
, bool return_path
,
2566 RDMACapabilities cap
= {
2567 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2570 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2572 .private_data
= &cap
,
2573 .private_data_len
= sizeof(cap
),
2575 struct rdma_cm_event
*cm_event
;
2579 * Only negotiate the capability with destination if the user
2580 * on the source first requested the capability.
2582 if (rdma
->pin_all
) {
2583 trace_qemu_rdma_connect_pin_all_requested();
2584 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2587 caps_to_network(&cap
);
2589 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2591 ERROR(errp
, "posting second control recv");
2592 goto err_rdma_source_connect
;
2595 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2597 perror("rdma_connect");
2598 ERROR(errp
, "connecting to destination!");
2599 goto err_rdma_source_connect
;
2603 ret
= qemu_get_cm_event_timeout(rdma
, &cm_event
, 5000, errp
);
2605 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2607 ERROR(errp
, "failed to get cm event");
2612 * FIXME perror() is wrong, because
2613 * qemu_get_cm_event_timeout() can fail without setting errno.
2614 * Will go away later in this series.
2616 perror("rdma_get_cm_event after rdma_connect");
2617 goto err_rdma_source_connect
;
2620 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2621 error_report("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2622 ERROR(errp
, "connecting to destination!");
2623 rdma_ack_cm_event(cm_event
);
2624 goto err_rdma_source_connect
;
2626 rdma
->connected
= true;
2628 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2629 network_to_caps(&cap
);
2632 * Verify that the *requested* capabilities are supported by the destination
2633 * and disable them otherwise.
2635 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2636 warn_report("RDMA: Server cannot support pinning all memory. "
2637 "Will register memory dynamically.");
2638 rdma
->pin_all
= false;
2641 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2643 rdma_ack_cm_event(cm_event
);
2645 rdma
->control_ready_expected
= 1;
2649 err_rdma_source_connect
:
2650 qemu_rdma_cleanup(rdma
);
2654 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2657 struct rdma_cm_id
*listen_id
;
2658 char ip
[40] = "unknown";
2659 struct rdma_addrinfo
*res
, *e
;
2663 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2664 rdma
->wr_data
[idx
].control_len
= 0;
2665 rdma
->wr_data
[idx
].control_curr
= NULL
;
2668 if (!rdma
->host
|| !rdma
->host
[0]) {
2669 ERROR(errp
, "RDMA host is not set!");
2670 rdma
->errored
= true;
2673 /* create CM channel */
2674 rdma
->channel
= rdma_create_event_channel();
2675 if (!rdma
->channel
) {
2676 ERROR(errp
, "could not create rdma event channel");
2677 rdma
->errored
= true;
2682 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2684 ERROR(errp
, "could not create cm_id!");
2685 goto err_dest_init_create_listen_id
;
2688 snprintf(port_str
, 16, "%d", rdma
->port
);
2689 port_str
[15] = '\0';
2691 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2693 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2694 goto err_dest_init_bind_addr
;
2697 ret
= rdma_set_option(listen_id
, RDMA_OPTION_ID
, RDMA_OPTION_ID_REUSEADDR
,
2698 &reuse
, sizeof reuse
);
2700 ERROR(errp
, "Error: could not set REUSEADDR option");
2701 goto err_dest_init_bind_addr
;
2703 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2704 inet_ntop(e
->ai_family
,
2705 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2706 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2707 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2711 if (e
->ai_family
== AF_INET6
) {
2712 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2720 rdma_freeaddrinfo(res
);
2722 ERROR(errp
, "Error: could not rdma_bind_addr!");
2723 goto err_dest_init_bind_addr
;
2726 rdma
->listen_id
= listen_id
;
2727 qemu_rdma_dump_gid("dest_init", listen_id
);
2730 err_dest_init_bind_addr
:
2731 rdma_destroy_id(listen_id
);
2732 err_dest_init_create_listen_id
:
2733 rdma_destroy_event_channel(rdma
->channel
);
2734 rdma
->channel
= NULL
;
2735 rdma
->errored
= true;
2740 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2745 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2746 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2747 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2750 /*the CM channel and CM id is shared*/
2751 rdma_return_path
->channel
= rdma
->channel
;
2752 rdma_return_path
->listen_id
= rdma
->listen_id
;
2754 rdma
->return_path
= rdma_return_path
;
2755 rdma_return_path
->return_path
= rdma
;
2756 rdma_return_path
->is_return_path
= true;
2759 static RDMAContext
*qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2761 RDMAContext
*rdma
= NULL
;
2762 InetSocketAddress
*addr
;
2764 rdma
= g_new0(RDMAContext
, 1);
2765 rdma
->current_index
= -1;
2766 rdma
->current_chunk
= -1;
2768 addr
= g_new(InetSocketAddress
, 1);
2769 if (!inet_parse(addr
, host_port
, NULL
)) {
2770 rdma
->port
= atoi(addr
->port
);
2771 rdma
->host
= g_strdup(addr
->host
);
2772 rdma
->host_port
= g_strdup(host_port
);
2774 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2779 qapi_free_InetSocketAddress(addr
);
2784 * QEMUFile interface to the control channel.
2785 * SEND messages for control only.
2786 * VM's ram is handled with regular RDMA messages.
2788 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2789 const struct iovec
*iov
,
2796 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2802 RCU_READ_LOCK_GUARD();
2803 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
2806 error_setg(errp
, "RDMA control channel output is not set");
2810 if (rdma
->errored
) {
2812 "RDMA is in an error state waiting migration to abort!");
2817 * Push out any writes that
2818 * we're queued up for VM's ram.
2820 ret
= qemu_rdma_write_flush(rdma
);
2822 rdma
->errored
= true;
2823 error_setg(errp
, "qemu_rdma_write_flush failed");
2827 for (i
= 0; i
< niov
; i
++) {
2828 size_t remaining
= iov
[i
].iov_len
;
2829 uint8_t * data
= (void *)iov
[i
].iov_base
;
2831 RDMAControlHeader head
= {};
2833 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2837 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2839 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2842 rdma
->errored
= true;
2843 error_setg(errp
, "qemu_rdma_exchange_send failed");
2855 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2856 size_t size
, int idx
)
2860 if (rdma
->wr_data
[idx
].control_len
) {
2861 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2863 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2864 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2865 rdma
->wr_data
[idx
].control_curr
+= len
;
2866 rdma
->wr_data
[idx
].control_len
-= len
;
2873 * QEMUFile interface to the control channel.
2874 * RDMA links don't use bytestreams, so we have to
2875 * return bytes to QEMUFile opportunistically.
2877 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2878 const struct iovec
*iov
,
2885 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2887 RDMAControlHeader head
;
2892 RCU_READ_LOCK_GUARD();
2893 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
2896 error_setg(errp
, "RDMA control channel input is not set");
2900 if (rdma
->errored
) {
2902 "RDMA is in an error state waiting migration to abort!");
2906 for (i
= 0; i
< niov
; i
++) {
2907 size_t want
= iov
[i
].iov_len
;
2908 uint8_t *data
= (void *)iov
[i
].iov_base
;
2911 * First, we hold on to the last SEND message we
2912 * were given and dish out the bytes until we run
2915 len
= qemu_rdma_fill(rdma
, data
, want
, 0);
2918 /* Got what we needed, so go to next iovec */
2923 /* If we got any data so far, then don't wait
2924 * for more, just return what we have */
2930 /* We've got nothing at all, so lets wait for
2933 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2936 rdma
->errored
= true;
2937 error_setg(errp
, "qemu_rdma_exchange_recv failed");
2942 * SEND was received with new bytes, now try again.
2944 len
= qemu_rdma_fill(rdma
, data
, want
, 0);
2948 /* Still didn't get enough, so lets just return */
2951 return QIO_CHANNEL_ERR_BLOCK
;
2961 * Block until all the outstanding chunks have been delivered by the hardware.
2963 static int qemu_rdma_drain_cq(RDMAContext
*rdma
)
2967 if (qemu_rdma_write_flush(rdma
) < 0) {
2971 while (rdma
->nb_sent
) {
2972 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2974 error_report("rdma migration: complete polling error!");
2979 qemu_rdma_unregister_waiting(rdma
);
2985 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2989 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2990 /* XXX we should make readv/writev actually honour this :-) */
2991 rioc
->blocking
= blocking
;
2996 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2997 struct QIOChannelRDMASource
{
2999 QIOChannelRDMA
*rioc
;
3000 GIOCondition condition
;
3004 qio_channel_rdma_source_prepare(GSource
*source
,
3007 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3009 GIOCondition cond
= 0;
3012 RCU_READ_LOCK_GUARD();
3013 if (rsource
->condition
== G_IO_IN
) {
3014 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3016 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3020 error_report("RDMAContext is NULL when prepare Gsource");
3024 if (rdma
->wr_data
[0].control_len
) {
3029 return cond
& rsource
->condition
;
3033 qio_channel_rdma_source_check(GSource
*source
)
3035 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3037 GIOCondition cond
= 0;
3039 RCU_READ_LOCK_GUARD();
3040 if (rsource
->condition
== G_IO_IN
) {
3041 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3043 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3047 error_report("RDMAContext is NULL when check Gsource");
3051 if (rdma
->wr_data
[0].control_len
) {
3056 return cond
& rsource
->condition
;
3060 qio_channel_rdma_source_dispatch(GSource
*source
,
3061 GSourceFunc callback
,
3064 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
3065 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3067 GIOCondition cond
= 0;
3069 RCU_READ_LOCK_GUARD();
3070 if (rsource
->condition
== G_IO_IN
) {
3071 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3073 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3077 error_report("RDMAContext is NULL when dispatch Gsource");
3081 if (rdma
->wr_data
[0].control_len
) {
3086 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
3087 (cond
& rsource
->condition
),
3092 qio_channel_rdma_source_finalize(GSource
*source
)
3094 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
3096 object_unref(OBJECT(ssource
->rioc
));
3099 static GSourceFuncs qio_channel_rdma_source_funcs
= {
3100 qio_channel_rdma_source_prepare
,
3101 qio_channel_rdma_source_check
,
3102 qio_channel_rdma_source_dispatch
,
3103 qio_channel_rdma_source_finalize
3106 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
3107 GIOCondition condition
)
3109 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3110 QIOChannelRDMASource
*ssource
;
3113 source
= g_source_new(&qio_channel_rdma_source_funcs
,
3114 sizeof(QIOChannelRDMASource
));
3115 ssource
= (QIOChannelRDMASource
*)source
;
3117 ssource
->rioc
= rioc
;
3118 object_ref(OBJECT(rioc
));
3120 ssource
->condition
= condition
;
3125 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
3126 AioContext
*read_ctx
,
3128 AioContext
*write_ctx
,
3129 IOHandler
*io_write
,
3132 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3134 aio_set_fd_handler(read_ctx
, rioc
->rdmain
->recv_comp_channel
->fd
,
3135 io_read
, io_write
, NULL
, NULL
, opaque
);
3136 aio_set_fd_handler(read_ctx
, rioc
->rdmain
->send_comp_channel
->fd
,
3137 io_read
, io_write
, NULL
, NULL
, opaque
);
3139 aio_set_fd_handler(write_ctx
, rioc
->rdmaout
->recv_comp_channel
->fd
,
3140 io_read
, io_write
, NULL
, NULL
, opaque
);
3141 aio_set_fd_handler(write_ctx
, rioc
->rdmaout
->send_comp_channel
->fd
,
3142 io_read
, io_write
, NULL
, NULL
, opaque
);
3146 struct rdma_close_rcu
{
3147 struct rcu_head rcu
;
3148 RDMAContext
*rdmain
;
3149 RDMAContext
*rdmaout
;
3152 /* callback from qio_channel_rdma_close via call_rcu */
3153 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu
*rcu
)
3156 qemu_rdma_cleanup(rcu
->rdmain
);
3160 qemu_rdma_cleanup(rcu
->rdmaout
);
3163 g_free(rcu
->rdmain
);
3164 g_free(rcu
->rdmaout
);
3168 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3171 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3172 RDMAContext
*rdmain
, *rdmaout
;
3173 struct rdma_close_rcu
*rcu
= g_new(struct rdma_close_rcu
, 1);
3175 trace_qemu_rdma_close();
3177 rdmain
= rioc
->rdmain
;
3179 qatomic_rcu_set(&rioc
->rdmain
, NULL
);
3182 rdmaout
= rioc
->rdmaout
;
3184 qatomic_rcu_set(&rioc
->rdmaout
, NULL
);
3187 rcu
->rdmain
= rdmain
;
3188 rcu
->rdmaout
= rdmaout
;
3189 call_rcu(rcu
, qio_channel_rdma_close_rcu
, rcu
);
3195 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3196 QIOChannelShutdown how
,
3199 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3200 RDMAContext
*rdmain
, *rdmaout
;
3202 RCU_READ_LOCK_GUARD();
3204 rdmain
= qatomic_rcu_read(&rioc
->rdmain
);
3205 rdmaout
= qatomic_rcu_read(&rioc
->rdmain
);
3208 case QIO_CHANNEL_SHUTDOWN_READ
:
3210 rdmain
->errored
= true;
3213 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3215 rdmaout
->errored
= true;
3218 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3221 rdmain
->errored
= true;
3224 rdmaout
->errored
= true;
3235 * This means that 'block_offset' is a full virtual address that does not
3236 * belong to a RAMBlock of the virtual machine and instead
3237 * represents a private malloc'd memory area that the caller wishes to
3241 * Offset is an offset to be added to block_offset and used
3242 * to also lookup the corresponding RAMBlock.
3244 * @size : Number of bytes to transfer
3246 * @pages_sent : User-specificed pointer to indicate how many pages were
3247 * sent. Usually, this will not be more than a few bytes of
3248 * the protocol because most transfers are sent asynchronously.
3250 static int qemu_rdma_save_page(QEMUFile
*f
, ram_addr_t block_offset
,
3251 ram_addr_t offset
, size_t size
)
3253 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3257 if (migration_in_postcopy()) {
3258 return RAM_SAVE_CONTROL_NOT_SUPP
;
3261 RCU_READ_LOCK_GUARD();
3262 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3268 if (rdma_errored(rdma
)) {
3275 * Add this page to the current 'chunk'. If the chunk
3276 * is full, or the page doesn't belong to the current chunk,
3277 * an actual RDMA write will occur and a new chunk will be formed.
3279 ret
= qemu_rdma_write(rdma
, block_offset
, offset
, size
);
3281 error_report("rdma migration: write error");
3286 * Drain the Completion Queue if possible, but do not block,
3289 * If nothing to poll, the end of the iteration will do this
3290 * again to make sure we don't overflow the request queue.
3293 uint64_t wr_id
, wr_id_in
;
3294 ret
= qemu_rdma_poll(rdma
, rdma
->recv_cq
, &wr_id_in
, NULL
);
3297 error_report("rdma migration: polling error");
3301 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3303 if (wr_id
== RDMA_WRID_NONE
) {
3309 uint64_t wr_id
, wr_id_in
;
3310 ret
= qemu_rdma_poll(rdma
, rdma
->send_cq
, &wr_id_in
, NULL
);
3313 error_report("rdma migration: polling error");
3317 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3319 if (wr_id
== RDMA_WRID_NONE
) {
3324 return RAM_SAVE_CONTROL_DELAYED
;
3327 rdma
->errored
= true;
3331 static void rdma_accept_incoming_migration(void *opaque
);
3333 static void rdma_cm_poll_handler(void *opaque
)
3335 RDMAContext
*rdma
= opaque
;
3337 struct rdma_cm_event
*cm_event
;
3338 MigrationIncomingState
*mis
= migration_incoming_get_current();
3340 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3342 error_report("get_cm_event failed %d", errno
);
3346 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3347 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3348 if (!rdma
->errored
&&
3349 migration_incoming_get_current()->state
!=
3350 MIGRATION_STATUS_COMPLETED
) {
3351 error_report("receive cm event, cm event is %d", cm_event
->event
);
3352 rdma
->errored
= true;
3353 if (rdma
->return_path
) {
3354 rdma
->return_path
->errored
= true;
3357 rdma_ack_cm_event(cm_event
);
3358 if (mis
->loadvm_co
) {
3359 qemu_coroutine_enter(mis
->loadvm_co
);
3363 rdma_ack_cm_event(cm_event
);
3366 static int qemu_rdma_accept(RDMAContext
*rdma
)
3368 RDMACapabilities cap
;
3369 struct rdma_conn_param conn_param
= {
3370 .responder_resources
= 2,
3371 .private_data
= &cap
,
3372 .private_data_len
= sizeof(cap
),
3374 RDMAContext
*rdma_return_path
= NULL
;
3375 struct rdma_cm_event
*cm_event
;
3376 struct ibv_context
*verbs
;
3380 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3382 goto err_rdma_dest_wait
;
3385 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3386 rdma_ack_cm_event(cm_event
);
3387 goto err_rdma_dest_wait
;
3391 * initialize the RDMAContext for return path for postcopy after first
3392 * connection request reached.
3394 if ((migrate_postcopy() || migrate_return_path())
3395 && !rdma
->is_return_path
) {
3396 rdma_return_path
= qemu_rdma_data_init(rdma
->host_port
, NULL
);
3397 if (rdma_return_path
== NULL
) {
3398 rdma_ack_cm_event(cm_event
);
3399 goto err_rdma_dest_wait
;
3402 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
3405 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3407 network_to_caps(&cap
);
3409 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3410 error_report("Unknown source RDMA version: %d, bailing...",
3412 rdma_ack_cm_event(cm_event
);
3413 goto err_rdma_dest_wait
;
3417 * Respond with only the capabilities this version of QEMU knows about.
3419 cap
.flags
&= known_capabilities
;
3422 * Enable the ones that we do know about.
3423 * Add other checks here as new ones are introduced.
3425 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3426 rdma
->pin_all
= true;
3429 rdma
->cm_id
= cm_event
->id
;
3430 verbs
= cm_event
->id
->verbs
;
3432 rdma_ack_cm_event(cm_event
);
3434 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3436 caps_to_network(&cap
);
3438 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3441 rdma
->verbs
= verbs
;
3442 } else if (rdma
->verbs
!= verbs
) {
3443 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3445 goto err_rdma_dest_wait
;
3448 qemu_rdma_dump_id("dest_init", verbs
);
3450 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3452 error_report("rdma migration: error allocating pd and cq!");
3453 goto err_rdma_dest_wait
;
3456 ret
= qemu_rdma_alloc_qp(rdma
);
3458 error_report("rdma migration: error allocating qp!");
3459 goto err_rdma_dest_wait
;
3462 qemu_rdma_init_ram_blocks(rdma
);
3464 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3465 ret
= qemu_rdma_reg_control(rdma
, idx
);
3467 error_report("rdma: error registering %d control", idx
);
3468 goto err_rdma_dest_wait
;
3472 /* Accept the second connection request for return path */
3473 if ((migrate_postcopy() || migrate_return_path())
3474 && !rdma
->is_return_path
) {
3475 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3477 (void *)(intptr_t)rdma
->return_path
);
3479 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3483 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3485 error_report("rdma_accept failed");
3486 goto err_rdma_dest_wait
;
3489 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3491 error_report("rdma_accept get_cm_event failed");
3492 goto err_rdma_dest_wait
;
3495 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3496 error_report("rdma_accept not event established");
3497 rdma_ack_cm_event(cm_event
);
3498 goto err_rdma_dest_wait
;
3501 rdma_ack_cm_event(cm_event
);
3502 rdma
->connected
= true;
3504 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3506 error_report("rdma migration: error posting second control recv");
3507 goto err_rdma_dest_wait
;
3510 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3515 rdma
->errored
= true;
3516 qemu_rdma_cleanup(rdma
);
3517 g_free(rdma_return_path
);
3521 static int dest_ram_sort_func(const void *a
, const void *b
)
3523 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3524 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3526 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3530 * During each iteration of the migration, we listen for instructions
3531 * by the source VM to perform dynamic page registrations before they
3532 * can perform RDMA operations.
3534 * We respond with the 'rkey'.
3536 * Keep doing this until the source tells us to stop.
3538 static int qemu_rdma_registration_handle(QEMUFile
*f
)
3540 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3541 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3544 RDMAControlHeader unreg_resp
= { .len
= 0,
3545 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3548 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3550 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3552 RDMALocalBlocks
*local
;
3553 RDMAControlHeader head
;
3554 RDMARegister
*reg
, *registers
;
3556 RDMARegisterResult
*reg_result
;
3557 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3558 RDMALocalBlock
*block
;
3565 RCU_READ_LOCK_GUARD();
3566 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3572 if (rdma_errored(rdma
)) {
3576 local
= &rdma
->local_ram_blocks
;
3578 trace_qemu_rdma_registration_handle_wait();
3580 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3586 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3587 error_report("rdma: Too many requests in this message (%d)."
3588 "Bailing.", head
.repeat
);
3592 switch (head
.type
) {
3593 case RDMA_CONTROL_COMPRESS
:
3594 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3595 network_to_compress(comp
);
3597 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3600 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3601 error_report("rdma: 'compress' bad block index %u (vs %d)",
3602 (unsigned int)comp
->block_idx
,
3603 rdma
->local_ram_blocks
.nb_blocks
);
3606 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3608 host_addr
= block
->local_host_addr
+
3609 (comp
->offset
- block
->offset
);
3611 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3614 case RDMA_CONTROL_REGISTER_FINISHED
:
3615 trace_qemu_rdma_registration_handle_finished();
3618 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3619 trace_qemu_rdma_registration_handle_ram_blocks();
3621 /* Sort our local RAM Block list so it's the same as the source,
3622 * we can do this since we've filled in a src_index in the list
3623 * as we received the RAMBlock list earlier.
3625 qsort(rdma
->local_ram_blocks
.block
,
3626 rdma
->local_ram_blocks
.nb_blocks
,
3627 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3628 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3629 local
->block
[i
].index
= i
;
3632 if (rdma
->pin_all
) {
3633 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3635 error_report("rdma migration: error dest "
3636 "registering ram blocks");
3642 * Dest uses this to prepare to transmit the RAMBlock descriptions
3643 * to the source VM after connection setup.
3644 * Both sides use the "remote" structure to communicate and update
3645 * their "local" descriptions with what was sent.
3647 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3648 rdma
->dest_blocks
[i
].remote_host_addr
=
3649 (uintptr_t)(local
->block
[i
].local_host_addr
);
3651 if (rdma
->pin_all
) {
3652 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3655 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3656 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3658 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3659 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3660 local
->block
[i
].block_name
,
3661 local
->block
[i
].offset
,
3662 local
->block
[i
].length
,
3663 local
->block
[i
].local_host_addr
,
3664 local
->block
[i
].src_index
);
3667 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3668 * sizeof(RDMADestBlock
);
3671 ret
= qemu_rdma_post_send_control(rdma
,
3672 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3675 error_report("rdma migration: error sending remote info");
3680 case RDMA_CONTROL_REGISTER_REQUEST
:
3681 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3683 reg_resp
.repeat
= head
.repeat
;
3684 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3686 for (count
= 0; count
< head
.repeat
; count
++) {
3688 uint8_t *chunk_start
, *chunk_end
;
3690 reg
= ®isters
[count
];
3691 network_to_register(reg
);
3693 reg_result
= &results
[count
];
3695 trace_qemu_rdma_registration_handle_register_loop(count
,
3696 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3698 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3699 error_report("rdma: 'register' bad block index %u (vs %d)",
3700 (unsigned int)reg
->current_index
,
3701 rdma
->local_ram_blocks
.nb_blocks
);
3704 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3705 if (block
->is_ram_block
) {
3706 if (block
->offset
> reg
->key
.current_addr
) {
3707 error_report("rdma: bad register address for block %s"
3708 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3709 block
->block_name
, block
->offset
,
3710 reg
->key
.current_addr
);
3713 host_addr
= (block
->local_host_addr
+
3714 (reg
->key
.current_addr
- block
->offset
));
3715 chunk
= ram_chunk_index(block
->local_host_addr
,
3716 (uint8_t *) host_addr
);
3718 chunk
= reg
->key
.chunk
;
3719 host_addr
= block
->local_host_addr
+
3720 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3721 /* Check for particularly bad chunk value */
3722 if (host_addr
< (void *)block
->local_host_addr
) {
3723 error_report("rdma: bad chunk for block %s"
3725 block
->block_name
, reg
->key
.chunk
);
3729 chunk_start
= ram_chunk_start(block
, chunk
);
3730 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3731 /* avoid "-Waddress-of-packed-member" warning */
3732 uint32_t tmp_rkey
= 0;
3733 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3734 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3735 chunk
, chunk_start
, chunk_end
)) {
3736 error_report("cannot get rkey");
3739 reg_result
->rkey
= tmp_rkey
;
3741 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3743 trace_qemu_rdma_registration_handle_register_rkey(
3746 result_to_network(reg_result
);
3749 ret
= qemu_rdma_post_send_control(rdma
,
3750 (uint8_t *) results
, ®_resp
);
3753 error_report("Failed to send control buffer");
3757 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3758 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3759 unreg_resp
.repeat
= head
.repeat
;
3760 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3762 for (count
= 0; count
< head
.repeat
; count
++) {
3763 reg
= ®isters
[count
];
3764 network_to_register(reg
);
3766 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3767 reg
->current_index
, reg
->key
.chunk
);
3769 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3771 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3772 block
->pmr
[reg
->key
.chunk
] = NULL
;
3775 perror("rdma unregistration chunk failed");
3779 rdma
->total_registrations
--;
3781 trace_qemu_rdma_registration_handle_unregister_success(
3785 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3788 error_report("Failed to send control buffer");
3792 case RDMA_CONTROL_REGISTER_RESULT
:
3793 error_report("Invalid RESULT message at dest.");
3796 error_report("Unknown control message %s", control_desc(head
.type
));
3802 rdma
->errored
= true;
3807 * Called via a ram_control_load_hook during the initial RAM load section which
3808 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3810 * We've already built our local RAMBlock list, but not yet sent the list to
3814 rdma_block_notification_handle(QEMUFile
*f
, const char *name
)
3817 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3821 RCU_READ_LOCK_GUARD();
3822 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3828 /* Find the matching RAMBlock in our local list */
3829 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3830 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3837 error_report("RAMBlock '%s' not found on destination", name
);
3841 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3842 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3843 rdma
->next_src_index
++;
3848 static int rdma_load_hook(QEMUFile
*f
, uint64_t flags
, void *data
)
3851 case RAM_CONTROL_BLOCK_REG
:
3852 return rdma_block_notification_handle(f
, data
);
3854 case RAM_CONTROL_HOOK
:
3855 return qemu_rdma_registration_handle(f
);
3858 /* Shouldn't be called with any other values */
3863 static int qemu_rdma_registration_start(QEMUFile
*f
,
3864 uint64_t flags
, void *data
)
3866 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3869 if (migration_in_postcopy()) {
3873 RCU_READ_LOCK_GUARD();
3874 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3879 if (rdma_errored(rdma
)) {
3883 trace_qemu_rdma_registration_start(flags
);
3884 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3891 * Inform dest that dynamic registrations are done for now.
3892 * First, flush writes, if any.
3894 static int qemu_rdma_registration_stop(QEMUFile
*f
,
3895 uint64_t flags
, void *data
)
3897 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3899 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3902 if (migration_in_postcopy()) {
3906 RCU_READ_LOCK_GUARD();
3907 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3912 if (rdma_errored(rdma
)) {
3917 ret
= qemu_rdma_drain_cq(rdma
);
3923 if (flags
== RAM_CONTROL_SETUP
) {
3924 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3925 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3926 int reg_result_idx
, i
, nb_dest_blocks
;
3928 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3929 trace_qemu_rdma_registration_stop_ram();
3932 * Make sure that we parallelize the pinning on both sides.
3933 * For very large guests, doing this serially takes a really
3934 * long time, so we have to 'interleave' the pinning locally
3935 * with the control messages by performing the pinning on this
3936 * side before we receive the control response from the other
3937 * side that the pinning has completed.
3939 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3940 ®_result_idx
, rdma
->pin_all
?
3941 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3943 fprintf(stderr
, "receiving remote info!");
3947 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3950 * The protocol uses two different sets of rkeys (mutually exclusive):
3951 * 1. One key to represent the virtual address of the entire ram block.
3952 * (dynamic chunk registration disabled - pin everything with one rkey.)
3953 * 2. One to represent individual chunks within a ram block.
3954 * (dynamic chunk registration enabled - pin individual chunks.)
3956 * Once the capability is successfully negotiated, the destination transmits
3957 * the keys to use (or sends them later) including the virtual addresses
3958 * and then propagates the remote ram block descriptions to his local copy.
3961 if (local
->nb_blocks
!= nb_dest_blocks
) {
3962 fprintf(stderr
, "ram blocks mismatch (Number of blocks %d vs %d) "
3963 "Your QEMU command line parameters are probably "
3964 "not identical on both the source and destination.",
3965 local
->nb_blocks
, nb_dest_blocks
);
3966 rdma
->errored
= true;
3970 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3971 memcpy(rdma
->dest_blocks
,
3972 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3973 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3974 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3976 /* We require that the blocks are in the same order */
3977 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3978 fprintf(stderr
, "Block %s/%d has a different length %" PRIu64
3979 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3980 local
->block
[i
].length
,
3981 rdma
->dest_blocks
[i
].length
);
3982 rdma
->errored
= true;
3985 local
->block
[i
].remote_host_addr
=
3986 rdma
->dest_blocks
[i
].remote_host_addr
;
3987 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3991 trace_qemu_rdma_registration_stop(flags
);
3993 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3994 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
4002 rdma
->errored
= true;
4006 static const QEMUFileHooks rdma_read_hooks
= {
4007 .hook_ram_load
= rdma_load_hook
,
4010 static const QEMUFileHooks rdma_write_hooks
= {
4011 .before_ram_iterate
= qemu_rdma_registration_start
,
4012 .after_ram_iterate
= qemu_rdma_registration_stop
,
4013 .save_page
= qemu_rdma_save_page
,
4017 static void qio_channel_rdma_finalize(Object
*obj
)
4019 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
4021 qemu_rdma_cleanup(rioc
->rdmain
);
4022 g_free(rioc
->rdmain
);
4023 rioc
->rdmain
= NULL
;
4025 if (rioc
->rdmaout
) {
4026 qemu_rdma_cleanup(rioc
->rdmaout
);
4027 g_free(rioc
->rdmaout
);
4028 rioc
->rdmaout
= NULL
;
4032 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
4033 void *class_data G_GNUC_UNUSED
)
4035 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
4037 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
4038 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
4039 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
4040 ioc_klass
->io_close
= qio_channel_rdma_close
;
4041 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
4042 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
4043 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
4046 static const TypeInfo qio_channel_rdma_info
= {
4047 .parent
= TYPE_QIO_CHANNEL
,
4048 .name
= TYPE_QIO_CHANNEL_RDMA
,
4049 .instance_size
= sizeof(QIOChannelRDMA
),
4050 .instance_finalize
= qio_channel_rdma_finalize
,
4051 .class_init
= qio_channel_rdma_class_init
,
4054 static void qio_channel_rdma_register_types(void)
4056 type_register_static(&qio_channel_rdma_info
);
4059 type_init(qio_channel_rdma_register_types
);
4061 static QEMUFile
*rdma_new_input(RDMAContext
*rdma
)
4063 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4065 rioc
->file
= qemu_file_new_input(QIO_CHANNEL(rioc
));
4066 rioc
->rdmain
= rdma
;
4067 rioc
->rdmaout
= rdma
->return_path
;
4068 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
4073 static QEMUFile
*rdma_new_output(RDMAContext
*rdma
)
4075 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4077 rioc
->file
= qemu_file_new_output(QIO_CHANNEL(rioc
));
4078 rioc
->rdmaout
= rdma
;
4079 rioc
->rdmain
= rdma
->return_path
;
4080 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
4085 static void rdma_accept_incoming_migration(void *opaque
)
4087 RDMAContext
*rdma
= opaque
;
4090 Error
*local_err
= NULL
;
4092 trace_qemu_rdma_accept_incoming_migration();
4093 ret
= qemu_rdma_accept(rdma
);
4096 fprintf(stderr
, "RDMA ERROR: Migration initialization failed\n");
4100 trace_qemu_rdma_accept_incoming_migration_accepted();
4102 if (rdma
->is_return_path
) {
4106 f
= rdma_new_input(rdma
);
4108 fprintf(stderr
, "RDMA ERROR: could not open RDMA for input\n");
4109 qemu_rdma_cleanup(rdma
);
4113 rdma
->migration_started_on_destination
= 1;
4114 migration_fd_process_incoming(f
, &local_err
);
4116 error_reportf_err(local_err
, "RDMA ERROR:");
4120 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4125 trace_rdma_start_incoming_migration();
4127 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4128 if (ram_block_discard_is_required()) {
4129 error_setg(errp
, "RDMA: cannot disable RAM discard");
4133 rdma
= qemu_rdma_data_init(host_port
, errp
);
4138 ret
= qemu_rdma_dest_init(rdma
, errp
);
4143 trace_rdma_start_incoming_migration_after_dest_init();
4145 ret
= rdma_listen(rdma
->listen_id
, 5);
4148 ERROR(errp
, "listening on socket!");
4152 trace_rdma_start_incoming_migration_after_rdma_listen();
4154 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4155 NULL
, (void *)(intptr_t)rdma
);
4159 qemu_rdma_cleanup(rdma
);
4163 g_free(rdma
->host_port
);
4168 void rdma_start_outgoing_migration(void *opaque
,
4169 const char *host_port
, Error
**errp
)
4171 MigrationState
*s
= opaque
;
4172 RDMAContext
*rdma_return_path
= NULL
;
4176 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4177 if (ram_block_discard_is_required()) {
4178 error_setg(errp
, "RDMA: cannot disable RAM discard");
4182 rdma
= qemu_rdma_data_init(host_port
, errp
);
4187 ret
= qemu_rdma_source_init(rdma
, migrate_rdma_pin_all(), errp
);
4193 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4194 ret
= qemu_rdma_connect(rdma
, false, errp
);
4200 /* RDMA postcopy need a separate queue pair for return path */
4201 if (migrate_postcopy() || migrate_return_path()) {
4202 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4204 if (rdma_return_path
== NULL
) {
4205 goto return_path_err
;
4208 ret
= qemu_rdma_source_init(rdma_return_path
,
4209 migrate_rdma_pin_all(), errp
);
4212 goto return_path_err
;
4215 ret
= qemu_rdma_connect(rdma_return_path
, true, errp
);
4218 goto return_path_err
;
4221 rdma
->return_path
= rdma_return_path
;
4222 rdma_return_path
->return_path
= rdma
;
4223 rdma_return_path
->is_return_path
= true;
4226 trace_rdma_start_outgoing_migration_after_rdma_connect();
4228 s
->to_dst_file
= rdma_new_output(rdma
);
4229 migrate_fd_connect(s
, NULL
);
4232 qemu_rdma_cleanup(rdma
);
4235 g_free(rdma_return_path
);