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"
21 #include "migration.h"
22 #include "qemu-file.h"
24 #include "qemu-file-channel.h"
25 #include "qemu/error-report.h"
26 #include "qemu/main-loop.h"
27 #include "qemu/module.h"
29 #include "qemu/sockets.h"
30 #include "qemu/bitmap.h"
31 #include "qemu/coroutine.h"
32 #include "exec/memory.h"
33 #include <sys/socket.h>
35 #include <arpa/inet.h>
36 #include <rdma/rdma_cma.h>
38 #include "qom/object.h"
41 * Print and error on both the Monitor and the Log file.
43 #define ERROR(errp, fmt, ...) \
45 fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \
46 if (errp && (*(errp) == NULL)) { \
47 error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \
51 #define RDMA_RESOLVE_TIMEOUT_MS 10000
53 /* Do not merge data if larger than this. */
54 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
55 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
57 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
60 * This is only for non-live state being migrated.
61 * Instead of RDMA_WRITE messages, we use RDMA_SEND
62 * messages for that state, which requires a different
63 * delivery design than main memory.
65 #define RDMA_SEND_INCREMENT 32768
68 * Maximum size infiniband SEND message
70 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
71 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
73 #define RDMA_CONTROL_VERSION_CURRENT 1
75 * Capabilities for negotiation.
77 #define RDMA_CAPABILITY_PIN_ALL 0x01
80 * Add the other flags above to this list of known capabilities
81 * as they are introduced.
83 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
85 #define CHECK_ERROR_STATE() \
87 if (rdma->error_state) { \
88 if (!rdma->error_reported) { \
89 error_report("RDMA is in an error state waiting migration" \
91 rdma->error_reported = 1; \
93 return rdma->error_state; \
98 * A work request ID is 64-bits and we split up these bits
101 * bits 0-15 : type of control message, 2^16
102 * bits 16-29: ram block index, 2^14
103 * bits 30-63: ram block chunk number, 2^34
105 * The last two bit ranges are only used for RDMA writes,
106 * in order to track their completion and potentially
107 * also track unregistration status of the message.
109 #define RDMA_WRID_TYPE_SHIFT 0UL
110 #define RDMA_WRID_BLOCK_SHIFT 16UL
111 #define RDMA_WRID_CHUNK_SHIFT 30UL
113 #define RDMA_WRID_TYPE_MASK \
114 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
116 #define RDMA_WRID_BLOCK_MASK \
117 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
119 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
122 * RDMA migration protocol:
123 * 1. RDMA Writes (data messages, i.e. RAM)
124 * 2. IB Send/Recv (control channel messages)
128 RDMA_WRID_RDMA_WRITE
= 1,
129 RDMA_WRID_SEND_CONTROL
= 2000,
130 RDMA_WRID_RECV_CONTROL
= 4000,
133 static const char *wrid_desc
[] = {
134 [RDMA_WRID_NONE
] = "NONE",
135 [RDMA_WRID_RDMA_WRITE
] = "WRITE RDMA",
136 [RDMA_WRID_SEND_CONTROL
] = "CONTROL SEND",
137 [RDMA_WRID_RECV_CONTROL
] = "CONTROL RECV",
141 * Work request IDs for IB SEND messages only (not RDMA writes).
142 * This is used by the migration protocol to transmit
143 * control messages (such as device state and registration commands)
145 * We could use more WRs, but we have enough for now.
155 * SEND/RECV IB Control Messages.
158 RDMA_CONTROL_NONE
= 0,
160 RDMA_CONTROL_READY
, /* ready to receive */
161 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
162 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
163 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
164 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
165 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
166 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
167 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
168 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
169 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
174 * Memory and MR structures used to represent an IB Send/Recv work request.
175 * This is *not* used for RDMA writes, only IB Send/Recv.
178 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
179 struct ibv_mr
*control_mr
; /* registration metadata */
180 size_t control_len
; /* length of the message */
181 uint8_t *control_curr
; /* start of unconsumed bytes */
182 } RDMAWorkRequestData
;
185 * Negotiate RDMA capabilities during connection-setup time.
192 static void caps_to_network(RDMACapabilities
*cap
)
194 cap
->version
= htonl(cap
->version
);
195 cap
->flags
= htonl(cap
->flags
);
198 static void network_to_caps(RDMACapabilities
*cap
)
200 cap
->version
= ntohl(cap
->version
);
201 cap
->flags
= ntohl(cap
->flags
);
205 * Representation of a RAMBlock from an RDMA perspective.
206 * This is not transmitted, only local.
207 * This and subsequent structures cannot be linked lists
208 * because we're using a single IB message to transmit
209 * the information. It's small anyway, so a list is overkill.
211 typedef struct RDMALocalBlock
{
213 uint8_t *local_host_addr
; /* local virtual address */
214 uint64_t remote_host_addr
; /* remote virtual address */
217 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
218 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
219 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
220 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
221 int index
; /* which block are we */
222 unsigned int src_index
; /* (Only used on dest) */
225 unsigned long *transit_bitmap
;
226 unsigned long *unregister_bitmap
;
230 * Also represents a RAMblock, but only on the dest.
231 * This gets transmitted by the dest during connection-time
232 * to the source VM and then is used to populate the
233 * corresponding RDMALocalBlock with
234 * the information needed to perform the actual RDMA.
236 typedef struct QEMU_PACKED RDMADestBlock
{
237 uint64_t remote_host_addr
;
240 uint32_t remote_rkey
;
244 static const char *control_desc(unsigned int rdma_control
)
246 static const char *strs
[] = {
247 [RDMA_CONTROL_NONE
] = "NONE",
248 [RDMA_CONTROL_ERROR
] = "ERROR",
249 [RDMA_CONTROL_READY
] = "READY",
250 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
251 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
252 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
253 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
254 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
255 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
256 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
257 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
258 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
261 if (rdma_control
> RDMA_CONTROL_UNREGISTER_FINISHED
) {
262 return "??BAD CONTROL VALUE??";
265 return strs
[rdma_control
];
268 static uint64_t htonll(uint64_t v
)
270 union { uint32_t lv
[2]; uint64_t llv
; } u
;
271 u
.lv
[0] = htonl(v
>> 32);
272 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
276 static uint64_t ntohll(uint64_t v
) {
277 union { uint32_t lv
[2]; uint64_t llv
; } u
;
279 return ((uint64_t)ntohl(u
.lv
[0]) << 32) | (uint64_t) ntohl(u
.lv
[1]);
282 static void dest_block_to_network(RDMADestBlock
*db
)
284 db
->remote_host_addr
= htonll(db
->remote_host_addr
);
285 db
->offset
= htonll(db
->offset
);
286 db
->length
= htonll(db
->length
);
287 db
->remote_rkey
= htonl(db
->remote_rkey
);
290 static void network_to_dest_block(RDMADestBlock
*db
)
292 db
->remote_host_addr
= ntohll(db
->remote_host_addr
);
293 db
->offset
= ntohll(db
->offset
);
294 db
->length
= ntohll(db
->length
);
295 db
->remote_rkey
= ntohl(db
->remote_rkey
);
299 * Virtual address of the above structures used for transmitting
300 * the RAMBlock descriptions at connection-time.
301 * This structure is *not* transmitted.
303 typedef struct RDMALocalBlocks
{
305 bool init
; /* main memory init complete */
306 RDMALocalBlock
*block
;
310 * Main data structure for RDMA state.
311 * While there is only one copy of this structure being allocated right now,
312 * this is the place where one would start if you wanted to consider
313 * having more than one RDMA connection open at the same time.
315 typedef struct RDMAContext
{
319 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
322 * This is used by *_exchange_send() to figure out whether or not
323 * the initial "READY" message has already been received or not.
324 * This is because other functions may potentially poll() and detect
325 * the READY message before send() does, in which case we need to
326 * know if it completed.
328 int control_ready_expected
;
330 /* number of outstanding writes */
333 /* store info about current buffer so that we can
334 merge it with future sends */
335 uint64_t current_addr
;
336 uint64_t current_length
;
337 /* index of ram block the current buffer belongs to */
339 /* index of the chunk in the current ram block */
345 * infiniband-specific variables for opening the device
346 * and maintaining connection state and so forth.
348 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
349 * cm_id->verbs, cm_id->channel, and cm_id->qp.
351 struct rdma_cm_id
*cm_id
; /* connection manager ID */
352 struct rdma_cm_id
*listen_id
;
355 struct ibv_context
*verbs
;
356 struct rdma_event_channel
*channel
;
357 struct ibv_qp
*qp
; /* queue pair */
358 struct ibv_comp_channel
*comp_channel
; /* completion channel */
359 struct ibv_pd
*pd
; /* protection domain */
360 struct ibv_cq
*cq
; /* completion queue */
363 * If a previous write failed (perhaps because of a failed
364 * memory registration, then do not attempt any future work
365 * and remember the error state.
372 * Description of ram blocks used throughout the code.
374 RDMALocalBlocks local_ram_blocks
;
375 RDMADestBlock
*dest_blocks
;
377 /* Index of the next RAMBlock received during block registration */
378 unsigned int next_src_index
;
381 * Migration on *destination* started.
382 * Then use coroutine yield function.
383 * Source runs in a thread, so we don't care.
385 int migration_started_on_destination
;
387 int total_registrations
;
390 int unregister_current
, unregister_next
;
391 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
393 GHashTable
*blockmap
;
395 /* the RDMAContext for return path */
396 struct RDMAContext
*return_path
;
400 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
401 typedef struct QIOChannelRDMA QIOChannelRDMA
;
402 DECLARE_INSTANCE_CHECKER(QIOChannelRDMA
, QIO_CHANNEL_RDMA
,
403 TYPE_QIO_CHANNEL_RDMA
)
407 struct QIOChannelRDMA
{
410 RDMAContext
*rdmaout
;
412 bool blocking
; /* XXX we don't actually honour this yet */
416 * Main structure for IB Send/Recv control messages.
417 * This gets prepended at the beginning of every Send/Recv.
419 typedef struct QEMU_PACKED
{
420 uint32_t len
; /* Total length of data portion */
421 uint32_t type
; /* which control command to perform */
422 uint32_t repeat
; /* number of commands in data portion of same type */
426 static void control_to_network(RDMAControlHeader
*control
)
428 control
->type
= htonl(control
->type
);
429 control
->len
= htonl(control
->len
);
430 control
->repeat
= htonl(control
->repeat
);
433 static void network_to_control(RDMAControlHeader
*control
)
435 control
->type
= ntohl(control
->type
);
436 control
->len
= ntohl(control
->len
);
437 control
->repeat
= ntohl(control
->repeat
);
441 * Register a single Chunk.
442 * Information sent by the source VM to inform the dest
443 * to register an single chunk of memory before we can perform
444 * the actual RDMA operation.
446 typedef struct QEMU_PACKED
{
448 uint64_t current_addr
; /* offset into the ram_addr_t space */
449 uint64_t chunk
; /* chunk to lookup if unregistering */
451 uint32_t current_index
; /* which ramblock the chunk belongs to */
453 uint64_t chunks
; /* how many sequential chunks to register */
456 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
458 RDMALocalBlock
*local_block
;
459 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
461 if (local_block
->is_ram_block
) {
463 * current_addr as passed in is an address in the local ram_addr_t
464 * space, we need to translate this for the destination
466 reg
->key
.current_addr
-= local_block
->offset
;
467 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
469 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
470 reg
->current_index
= htonl(reg
->current_index
);
471 reg
->chunks
= htonll(reg
->chunks
);
474 static void network_to_register(RDMARegister
*reg
)
476 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
477 reg
->current_index
= ntohl(reg
->current_index
);
478 reg
->chunks
= ntohll(reg
->chunks
);
481 typedef struct QEMU_PACKED
{
482 uint32_t value
; /* if zero, we will madvise() */
483 uint32_t block_idx
; /* which ram block index */
484 uint64_t offset
; /* Address in remote ram_addr_t space */
485 uint64_t length
; /* length of the chunk */
488 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
490 comp
->value
= htonl(comp
->value
);
492 * comp->offset as passed in is an address in the local ram_addr_t
493 * space, we need to translate this for the destination
495 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
496 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
497 comp
->block_idx
= htonl(comp
->block_idx
);
498 comp
->offset
= htonll(comp
->offset
);
499 comp
->length
= htonll(comp
->length
);
502 static void network_to_compress(RDMACompress
*comp
)
504 comp
->value
= ntohl(comp
->value
);
505 comp
->block_idx
= ntohl(comp
->block_idx
);
506 comp
->offset
= ntohll(comp
->offset
);
507 comp
->length
= ntohll(comp
->length
);
511 * The result of the dest's memory registration produces an "rkey"
512 * which the source VM must reference in order to perform
513 * the RDMA operation.
515 typedef struct QEMU_PACKED
{
519 } RDMARegisterResult
;
521 static void result_to_network(RDMARegisterResult
*result
)
523 result
->rkey
= htonl(result
->rkey
);
524 result
->host_addr
= htonll(result
->host_addr
);
527 static void network_to_result(RDMARegisterResult
*result
)
529 result
->rkey
= ntohl(result
->rkey
);
530 result
->host_addr
= ntohll(result
->host_addr
);
533 const char *print_wrid(int wrid
);
534 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
535 uint8_t *data
, RDMAControlHeader
*resp
,
537 int (*callback
)(RDMAContext
*rdma
));
539 static inline uint64_t ram_chunk_index(const uint8_t *start
,
542 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
545 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
548 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
549 (i
<< RDMA_REG_CHUNK_SHIFT
));
552 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
555 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
556 (1UL << RDMA_REG_CHUNK_SHIFT
);
558 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
559 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
565 static int rdma_add_block(RDMAContext
*rdma
, const char *block_name
,
567 ram_addr_t block_offset
, uint64_t length
)
569 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
570 RDMALocalBlock
*block
;
571 RDMALocalBlock
*old
= local
->block
;
573 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
+ 1);
575 if (local
->nb_blocks
) {
578 if (rdma
->blockmap
) {
579 for (x
= 0; x
< local
->nb_blocks
; x
++) {
580 g_hash_table_remove(rdma
->blockmap
,
581 (void *)(uintptr_t)old
[x
].offset
);
582 g_hash_table_insert(rdma
->blockmap
,
583 (void *)(uintptr_t)old
[x
].offset
,
587 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
591 block
= &local
->block
[local
->nb_blocks
];
593 block
->block_name
= g_strdup(block_name
);
594 block
->local_host_addr
= host_addr
;
595 block
->offset
= block_offset
;
596 block
->length
= length
;
597 block
->index
= local
->nb_blocks
;
598 block
->src_index
= ~0U; /* Filled in by the receipt of the block list */
599 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
600 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
601 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
602 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
603 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
604 block
->remote_keys
= g_new0(uint32_t, block
->nb_chunks
);
606 block
->is_ram_block
= local
->init
? false : true;
608 if (rdma
->blockmap
) {
609 g_hash_table_insert(rdma
->blockmap
, (void *)(uintptr_t)block_offset
, block
);
612 trace_rdma_add_block(block_name
, local
->nb_blocks
,
613 (uintptr_t) block
->local_host_addr
,
614 block
->offset
, block
->length
,
615 (uintptr_t) (block
->local_host_addr
+ block
->length
),
616 BITS_TO_LONGS(block
->nb_chunks
) *
617 sizeof(unsigned long) * 8,
626 * Memory regions need to be registered with the device and queue pairs setup
627 * in advanced before the migration starts. This tells us where the RAM blocks
628 * are so that we can register them individually.
630 static int qemu_rdma_init_one_block(RAMBlock
*rb
, void *opaque
)
632 const char *block_name
= qemu_ram_get_idstr(rb
);
633 void *host_addr
= qemu_ram_get_host_addr(rb
);
634 ram_addr_t block_offset
= qemu_ram_get_offset(rb
);
635 ram_addr_t length
= qemu_ram_get_used_length(rb
);
636 return rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
640 * Identify the RAMBlocks and their quantity. They will be references to
641 * identify chunk boundaries inside each RAMBlock and also be referenced
642 * during dynamic page registration.
644 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
646 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
649 assert(rdma
->blockmap
== NULL
);
650 memset(local
, 0, sizeof *local
);
651 ret
= foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
655 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
656 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
657 rdma
->local_ram_blocks
.nb_blocks
);
663 * Note: If used outside of cleanup, the caller must ensure that the destination
664 * block structures are also updated
666 static int rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
668 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
669 RDMALocalBlock
*old
= local
->block
;
672 if (rdma
->blockmap
) {
673 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
678 for (j
= 0; j
< block
->nb_chunks
; j
++) {
679 if (!block
->pmr
[j
]) {
682 ibv_dereg_mr(block
->pmr
[j
]);
683 rdma
->total_registrations
--;
690 ibv_dereg_mr(block
->mr
);
691 rdma
->total_registrations
--;
695 g_free(block
->transit_bitmap
);
696 block
->transit_bitmap
= NULL
;
698 g_free(block
->unregister_bitmap
);
699 block
->unregister_bitmap
= NULL
;
701 g_free(block
->remote_keys
);
702 block
->remote_keys
= NULL
;
704 g_free(block
->block_name
);
705 block
->block_name
= NULL
;
707 if (rdma
->blockmap
) {
708 for (x
= 0; x
< local
->nb_blocks
; x
++) {
709 g_hash_table_remove(rdma
->blockmap
,
710 (void *)(uintptr_t)old
[x
].offset
);
714 if (local
->nb_blocks
> 1) {
716 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
719 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
722 if (block
->index
< (local
->nb_blocks
- 1)) {
723 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
724 sizeof(RDMALocalBlock
) *
725 (local
->nb_blocks
- (block
->index
+ 1)));
726 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
727 local
->block
[x
].index
--;
731 assert(block
== local
->block
);
735 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
736 block
->offset
, block
->length
,
737 (uintptr_t)(block
->local_host_addr
+ block
->length
),
738 BITS_TO_LONGS(block
->nb_chunks
) *
739 sizeof(unsigned long) * 8, block
->nb_chunks
);
745 if (local
->nb_blocks
&& rdma
->blockmap
) {
746 for (x
= 0; x
< local
->nb_blocks
; x
++) {
747 g_hash_table_insert(rdma
->blockmap
,
748 (void *)(uintptr_t)local
->block
[x
].offset
,
757 * Put in the log file which RDMA device was opened and the details
758 * associated with that device.
760 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
762 struct ibv_port_attr port
;
764 if (ibv_query_port(verbs
, 1, &port
)) {
765 error_report("Failed to query port information");
769 printf("%s RDMA Device opened: kernel name %s "
770 "uverbs device name %s, "
771 "infiniband_verbs class device path %s, "
772 "infiniband class device path %s, "
773 "transport: (%d) %s\n",
776 verbs
->device
->dev_name
,
777 verbs
->device
->dev_path
,
778 verbs
->device
->ibdev_path
,
780 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
781 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
782 ? "Ethernet" : "Unknown"));
786 * Put in the log file the RDMA gid addressing information,
787 * useful for folks who have trouble understanding the
788 * RDMA device hierarchy in the kernel.
790 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
794 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
795 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
796 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
800 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
801 * We will try the next addrinfo struct, and fail if there are
802 * no other valid addresses to bind against.
804 * If user is listening on '[::]', then we will not have a opened a device
805 * yet and have no way of verifying if the device is RoCE or not.
807 * In this case, the source VM will throw an error for ALL types of
808 * connections (both IPv4 and IPv6) if the destination machine does not have
809 * a regular infiniband network available for use.
811 * The only way to guarantee that an error is thrown for broken kernels is
812 * for the management software to choose a *specific* interface at bind time
813 * and validate what time of hardware it is.
815 * Unfortunately, this puts the user in a fix:
817 * If the source VM connects with an IPv4 address without knowing that the
818 * destination has bound to '[::]' the migration will unconditionally fail
819 * unless the management software is explicitly listening on the IPv4
820 * address while using a RoCE-based device.
822 * If the source VM connects with an IPv6 address, then we're OK because we can
823 * throw an error on the source (and similarly on the destination).
825 * But in mixed environments, this will be broken for a while until it is fixed
828 * We do provide a *tiny* bit of help in this function: We can list all of the
829 * devices in the system and check to see if all the devices are RoCE or
832 * If we detect that we have a *pure* RoCE environment, then we can safely
833 * thrown an error even if the management software has specified '[::]' as the
836 * However, if there is are multiple hetergeneous devices, then we cannot make
837 * this assumption and the user just has to be sure they know what they are
840 * Patches are being reviewed on linux-rdma.
842 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
844 /* This bug only exists in linux, to our knowledge. */
846 struct ibv_port_attr port_attr
;
849 * Verbs are only NULL if management has bound to '[::]'.
851 * Let's iterate through all the devices and see if there any pure IB
852 * devices (non-ethernet).
854 * If not, then we can safely proceed with the migration.
855 * Otherwise, there are no guarantees until the bug is fixed in linux.
859 struct ibv_device
** dev_list
= ibv_get_device_list(&num_devices
);
860 bool roce_found
= false;
861 bool ib_found
= false;
863 for (x
= 0; x
< num_devices
; x
++) {
864 verbs
= ibv_open_device(dev_list
[x
]);
866 if (errno
== EPERM
) {
873 if (ibv_query_port(verbs
, 1, &port_attr
)) {
874 ibv_close_device(verbs
);
875 ERROR(errp
, "Could not query initial IB port");
879 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
881 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
885 ibv_close_device(verbs
);
891 fprintf(stderr
, "WARN: migrations may fail:"
892 " IPv6 over RoCE / iWARP in linux"
893 " is broken. But since you appear to have a"
894 " mixed RoCE / IB environment, be sure to only"
895 " migrate over the IB fabric until the kernel "
896 " fixes the bug.\n");
898 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
899 " and your management software has specified '[::]'"
900 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
909 * If we have a verbs context, that means that some other than '[::]' was
910 * used by the management software for binding. In which case we can
911 * actually warn the user about a potentially broken kernel.
914 /* IB ports start with 1, not 0 */
915 if (ibv_query_port(verbs
, 1, &port_attr
)) {
916 ERROR(errp
, "Could not query initial IB port");
920 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
921 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
922 "(but patches on linux-rdma in progress)");
932 * Figure out which RDMA device corresponds to the requested IP hostname
933 * Also create the initial connection manager identifiers for opening
936 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
939 struct rdma_addrinfo
*res
;
941 struct rdma_cm_event
*cm_event
;
942 char ip
[40] = "unknown";
943 struct rdma_addrinfo
*e
;
945 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
946 ERROR(errp
, "RDMA hostname has not been set");
950 /* create CM channel */
951 rdma
->channel
= rdma_create_event_channel();
952 if (!rdma
->channel
) {
953 ERROR(errp
, "could not create CM channel");
958 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
960 ERROR(errp
, "could not create channel id");
961 goto err_resolve_create_id
;
964 snprintf(port_str
, 16, "%d", rdma
->port
);
967 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
969 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
970 goto err_resolve_get_addr
;
973 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
974 inet_ntop(e
->ai_family
,
975 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
976 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
978 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
979 RDMA_RESOLVE_TIMEOUT_MS
);
981 if (e
->ai_family
== AF_INET6
) {
982 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
991 ERROR(errp
, "could not resolve address %s", rdma
->host
);
992 goto err_resolve_get_addr
;
995 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
997 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
999 ERROR(errp
, "could not perform event_addr_resolved");
1000 goto err_resolve_get_addr
;
1003 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
1004 ERROR(errp
, "result not equal to event_addr_resolved %s",
1005 rdma_event_str(cm_event
->event
));
1006 perror("rdma_resolve_addr");
1007 rdma_ack_cm_event(cm_event
);
1009 goto err_resolve_get_addr
;
1011 rdma_ack_cm_event(cm_event
);
1014 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1016 ERROR(errp
, "could not resolve rdma route");
1017 goto err_resolve_get_addr
;
1020 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1022 ERROR(errp
, "could not perform event_route_resolved");
1023 goto err_resolve_get_addr
;
1025 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1026 ERROR(errp
, "result not equal to event_route_resolved: %s",
1027 rdma_event_str(cm_event
->event
));
1028 rdma_ack_cm_event(cm_event
);
1030 goto err_resolve_get_addr
;
1032 rdma_ack_cm_event(cm_event
);
1033 rdma
->verbs
= rdma
->cm_id
->verbs
;
1034 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1035 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1038 err_resolve_get_addr
:
1039 rdma_destroy_id(rdma
->cm_id
);
1041 err_resolve_create_id
:
1042 rdma_destroy_event_channel(rdma
->channel
);
1043 rdma
->channel
= NULL
;
1048 * Create protection domain and completion queues
1050 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1053 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1055 error_report("failed to allocate protection domain");
1059 /* create completion channel */
1060 rdma
->comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1061 if (!rdma
->comp_channel
) {
1062 error_report("failed to allocate completion channel");
1063 goto err_alloc_pd_cq
;
1067 * Completion queue can be filled by both read and write work requests,
1068 * so must reflect the sum of both possible queue sizes.
1070 rdma
->cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1071 NULL
, rdma
->comp_channel
, 0);
1073 error_report("failed to allocate completion queue");
1074 goto err_alloc_pd_cq
;
1081 ibv_dealloc_pd(rdma
->pd
);
1083 if (rdma
->comp_channel
) {
1084 ibv_destroy_comp_channel(rdma
->comp_channel
);
1087 rdma
->comp_channel
= NULL
;
1093 * Create queue pairs.
1095 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1097 struct ibv_qp_init_attr attr
= { 0 };
1100 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1101 attr
.cap
.max_recv_wr
= 3;
1102 attr
.cap
.max_send_sge
= 1;
1103 attr
.cap
.max_recv_sge
= 1;
1104 attr
.send_cq
= rdma
->cq
;
1105 attr
.recv_cq
= rdma
->cq
;
1106 attr
.qp_type
= IBV_QPT_RC
;
1108 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1113 rdma
->qp
= rdma
->cm_id
->qp
;
1117 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1120 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1122 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1123 local
->block
[i
].mr
=
1124 ibv_reg_mr(rdma
->pd
,
1125 local
->block
[i
].local_host_addr
,
1126 local
->block
[i
].length
,
1127 IBV_ACCESS_LOCAL_WRITE
|
1128 IBV_ACCESS_REMOTE_WRITE
1130 if (!local
->block
[i
].mr
) {
1131 perror("Failed to register local dest ram block!\n");
1134 rdma
->total_registrations
++;
1137 if (i
>= local
->nb_blocks
) {
1141 for (i
--; i
>= 0; i
--) {
1142 ibv_dereg_mr(local
->block
[i
].mr
);
1143 rdma
->total_registrations
--;
1151 * Find the ram block that corresponds to the page requested to be
1152 * transmitted by QEMU.
1154 * Once the block is found, also identify which 'chunk' within that
1155 * block that the page belongs to.
1157 * This search cannot fail or the migration will fail.
1159 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1160 uintptr_t block_offset
,
1163 uint64_t *block_index
,
1164 uint64_t *chunk_index
)
1166 uint64_t current_addr
= block_offset
+ offset
;
1167 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1168 (void *) block_offset
);
1170 assert(current_addr
>= block
->offset
);
1171 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1173 *block_index
= block
->index
;
1174 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1175 block
->local_host_addr
+ (current_addr
- block
->offset
));
1181 * Register a chunk with IB. If the chunk was already registered
1182 * previously, then skip.
1184 * Also return the keys associated with the registration needed
1185 * to perform the actual RDMA operation.
1187 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1188 RDMALocalBlock
*block
, uintptr_t host_addr
,
1189 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1190 uint8_t *chunk_start
, uint8_t *chunk_end
)
1194 *lkey
= block
->mr
->lkey
;
1197 *rkey
= block
->mr
->rkey
;
1202 /* allocate memory to store chunk MRs */
1204 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1208 * If 'rkey', then we're the destination, so grant access to the source.
1210 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1212 if (!block
->pmr
[chunk
]) {
1213 uint64_t len
= chunk_end
- chunk_start
;
1215 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1217 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
,
1219 (rkey
? (IBV_ACCESS_LOCAL_WRITE
|
1220 IBV_ACCESS_REMOTE_WRITE
) : 0));
1222 if (!block
->pmr
[chunk
]) {
1223 perror("Failed to register chunk!");
1224 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1225 " start %" PRIuPTR
" end %" PRIuPTR
1227 " local %" PRIuPTR
" registrations: %d\n",
1228 block
->index
, chunk
, (uintptr_t)chunk_start
,
1229 (uintptr_t)chunk_end
, host_addr
,
1230 (uintptr_t)block
->local_host_addr
,
1231 rdma
->total_registrations
);
1234 rdma
->total_registrations
++;
1238 *lkey
= block
->pmr
[chunk
]->lkey
;
1241 *rkey
= block
->pmr
[chunk
]->rkey
;
1247 * Register (at connection time) the memory used for control
1250 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1252 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1253 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1254 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1255 if (rdma
->wr_data
[idx
].control_mr
) {
1256 rdma
->total_registrations
++;
1259 error_report("qemu_rdma_reg_control failed");
1263 const char *print_wrid(int wrid
)
1265 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1266 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1268 return wrid_desc
[wrid
];
1272 * RDMA requires memory registration (mlock/pinning), but this is not good for
1275 * In preparation for the future where LRU information or workload-specific
1276 * writable writable working set memory access behavior is available to QEMU
1277 * it would be nice to have in place the ability to UN-register/UN-pin
1278 * particular memory regions from the RDMA hardware when it is determine that
1279 * those regions of memory will likely not be accessed again in the near future.
1281 * While we do not yet have such information right now, the following
1282 * compile-time option allows us to perform a non-optimized version of this
1285 * By uncommenting this option, you will cause *all* RDMA transfers to be
1286 * unregistered immediately after the transfer completes on both sides of the
1287 * connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
1289 * This will have a terrible impact on migration performance, so until future
1290 * workload information or LRU information is available, do not attempt to use
1291 * this feature except for basic testing.
1293 //#define RDMA_UNREGISTRATION_EXAMPLE
1296 * Perform a non-optimized memory unregistration after every transfer
1297 * for demonstration purposes, only if pin-all is not requested.
1299 * Potential optimizations:
1300 * 1. Start a new thread to run this function continuously
1302 - and for receipt of unregister messages
1304 * 3. Use workload hints.
1306 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1308 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1310 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1312 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1314 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1315 RDMALocalBlock
*block
=
1316 &(rdma
->local_ram_blocks
.block
[index
]);
1317 RDMARegister reg
= { .current_index
= index
};
1318 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1320 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1321 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1325 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1326 rdma
->unregister_current
);
1328 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1329 rdma
->unregister_current
++;
1331 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1332 rdma
->unregister_current
= 0;
1337 * Unregistration is speculative (because migration is single-threaded
1338 * and we cannot break the protocol's inifinband message ordering).
1339 * Thus, if the memory is currently being used for transmission,
1340 * then abort the attempt to unregister and try again
1341 * later the next time a completion is received for this memory.
1343 clear_bit(chunk
, block
->unregister_bitmap
);
1345 if (test_bit(chunk
, block
->transit_bitmap
)) {
1346 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1350 trace_qemu_rdma_unregister_waiting_send(chunk
);
1352 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1353 block
->pmr
[chunk
] = NULL
;
1354 block
->remote_keys
[chunk
] = 0;
1357 perror("unregistration chunk failed");
1360 rdma
->total_registrations
--;
1362 reg
.key
.chunk
= chunk
;
1363 register_to_network(rdma
, ®
);
1364 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1370 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1376 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1379 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1381 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1382 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1388 * Set bit for unregistration in the next iteration.
1389 * We cannot transmit right here, but will unpin later.
1391 static void qemu_rdma_signal_unregister(RDMAContext
*rdma
, uint64_t index
,
1392 uint64_t chunk
, uint64_t wr_id
)
1394 if (rdma
->unregistrations
[rdma
->unregister_next
] != 0) {
1395 error_report("rdma migration: queue is full");
1397 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1399 if (!test_and_set_bit(chunk
, block
->unregister_bitmap
)) {
1400 trace_qemu_rdma_signal_unregister_append(chunk
,
1401 rdma
->unregister_next
);
1403 rdma
->unregistrations
[rdma
->unregister_next
++] =
1404 qemu_rdma_make_wrid(wr_id
, index
, chunk
);
1406 if (rdma
->unregister_next
== RDMA_SIGNALED_SEND_MAX
) {
1407 rdma
->unregister_next
= 0;
1410 trace_qemu_rdma_signal_unregister_already(chunk
);
1416 * Consult the connection manager to see a work request
1417 * (of any kind) has completed.
1418 * Return the work request ID that completed.
1420 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, uint64_t *wr_id_out
,
1427 ret
= ibv_poll_cq(rdma
->cq
, 1, &wc
);
1430 *wr_id_out
= RDMA_WRID_NONE
;
1435 error_report("ibv_poll_cq return %d", ret
);
1439 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1441 if (wc
.status
!= IBV_WC_SUCCESS
) {
1442 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1443 wc
.status
, ibv_wc_status_str(wc
.status
));
1444 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1449 if (rdma
->control_ready_expected
&&
1450 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1451 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1452 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1453 rdma
->control_ready_expected
= 0;
1456 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1458 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1460 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1461 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1463 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1464 index
, chunk
, block
->local_host_addr
,
1465 (void *)(uintptr_t)block
->remote_host_addr
);
1467 clear_bit(chunk
, block
->transit_bitmap
);
1469 if (rdma
->nb_sent
> 0) {
1473 if (!rdma
->pin_all
) {
1475 * FYI: If one wanted to signal a specific chunk to be unregistered
1476 * using LRU or workload-specific information, this is the function
1477 * you would call to do so. That chunk would then get asynchronously
1478 * unregistered later.
1480 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1481 qemu_rdma_signal_unregister(rdma
, index
, chunk
, wc
.wr_id
);
1485 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1488 *wr_id_out
= wc
.wr_id
;
1490 *byte_len
= wc
.byte_len
;
1496 /* Wait for activity on the completion channel.
1497 * Returns 0 on success, none-0 on error.
1499 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
)
1501 struct rdma_cm_event
*cm_event
;
1505 * Coroutine doesn't start until migration_fd_process_incoming()
1506 * so don't yield unless we know we're running inside of a coroutine.
1508 if (rdma
->migration_started_on_destination
&&
1509 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1510 yield_until_fd_readable(rdma
->comp_channel
->fd
);
1512 /* This is the source side, we're in a separate thread
1513 * or destination prior to migration_fd_process_incoming()
1514 * after postcopy, the destination also in a seprate thread.
1515 * we can't yield; so we have to poll the fd.
1516 * But we need to be able to handle 'cancel' or an error
1517 * without hanging forever.
1519 while (!rdma
->error_state
&& !rdma
->received_error
) {
1521 pfds
[0].fd
= rdma
->comp_channel
->fd
;
1522 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1523 pfds
[0].revents
= 0;
1525 pfds
[1].fd
= rdma
->channel
->fd
;
1526 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1527 pfds
[1].revents
= 0;
1529 /* 0.1s timeout, should be fine for a 'cancel' */
1530 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1532 case 1: /* fd active */
1533 if (pfds
[0].revents
) {
1537 if (pfds
[1].revents
) {
1538 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1540 rdma_ack_cm_event(cm_event
);
1543 error_report("receive cm event while wait comp channel,"
1544 "cm event is %d", cm_event
->event
);
1545 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1546 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1552 case 0: /* Timeout, go around again */
1555 default: /* Error of some type -
1556 * I don't trust errno from qemu_poll_ns
1558 error_report("%s: poll failed", __func__
);
1562 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1563 /* Bail out and let the cancellation happen */
1569 if (rdma
->received_error
) {
1572 return rdma
->error_state
;
1576 * Block until the next work request has completed.
1578 * First poll to see if a work request has already completed,
1581 * If we encounter completed work requests for IDs other than
1582 * the one we're interested in, then that's generally an error.
1584 * The only exception is actual RDMA Write completions. These
1585 * completions only need to be recorded, but do not actually
1586 * need further processing.
1588 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1591 int num_cq_events
= 0, ret
= 0;
1594 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1596 if (ibv_req_notify_cq(rdma
->cq
, 0)) {
1600 while (wr_id
!= wrid_requested
) {
1601 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1606 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1608 if (wr_id
== RDMA_WRID_NONE
) {
1611 if (wr_id
!= wrid_requested
) {
1612 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1613 wrid_requested
, print_wrid(wr_id
), wr_id
);
1617 if (wr_id
== wrid_requested
) {
1622 ret
= qemu_rdma_wait_comp_channel(rdma
);
1624 goto err_block_for_wrid
;
1627 ret
= ibv_get_cq_event(rdma
->comp_channel
, &cq
, &cq_ctx
);
1629 perror("ibv_get_cq_event");
1630 goto err_block_for_wrid
;
1635 ret
= -ibv_req_notify_cq(cq
, 0);
1637 goto err_block_for_wrid
;
1640 while (wr_id
!= wrid_requested
) {
1641 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1643 goto err_block_for_wrid
;
1646 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1648 if (wr_id
== RDMA_WRID_NONE
) {
1651 if (wr_id
!= wrid_requested
) {
1652 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1653 wrid_requested
, print_wrid(wr_id
), wr_id
);
1657 if (wr_id
== wrid_requested
) {
1658 goto success_block_for_wrid
;
1662 success_block_for_wrid
:
1663 if (num_cq_events
) {
1664 ibv_ack_cq_events(cq
, num_cq_events
);
1669 if (num_cq_events
) {
1670 ibv_ack_cq_events(cq
, num_cq_events
);
1673 rdma
->error_state
= ret
;
1678 * Post a SEND message work request for the control channel
1679 * containing some data and block until the post completes.
1681 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1682 RDMAControlHeader
*head
)
1685 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1686 struct ibv_send_wr
*bad_wr
;
1687 struct ibv_sge sge
= {
1688 .addr
= (uintptr_t)(wr
->control
),
1689 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1690 .lkey
= wr
->control_mr
->lkey
,
1692 struct ibv_send_wr send_wr
= {
1693 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1694 .opcode
= IBV_WR_SEND
,
1695 .send_flags
= IBV_SEND_SIGNALED
,
1700 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1703 * We don't actually need to do a memcpy() in here if we used
1704 * the "sge" properly, but since we're only sending control messages
1705 * (not RAM in a performance-critical path), then its OK for now.
1707 * The copy makes the RDMAControlHeader simpler to manipulate
1708 * for the time being.
1710 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1711 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1712 control_to_network((void *) wr
->control
);
1715 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1719 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1722 error_report("Failed to use post IB SEND for control");
1726 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1728 error_report("rdma migration: send polling control error");
1735 * Post a RECV work request in anticipation of some future receipt
1736 * of data on the control channel.
1738 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1740 struct ibv_recv_wr
*bad_wr
;
1741 struct ibv_sge sge
= {
1742 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1743 .length
= RDMA_CONTROL_MAX_BUFFER
,
1744 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1747 struct ibv_recv_wr recv_wr
= {
1748 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1754 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1762 * Block and wait for a RECV control channel message to arrive.
1764 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1765 RDMAControlHeader
*head
, int expecting
, int idx
)
1768 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1772 error_report("rdma migration: recv polling control error!");
1776 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1777 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1779 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1781 if (expecting
== RDMA_CONTROL_NONE
) {
1782 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1784 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1785 error_report("Was expecting a %s (%d) control message"
1786 ", but got: %s (%d), length: %d",
1787 control_desc(expecting
), expecting
,
1788 control_desc(head
->type
), head
->type
, head
->len
);
1789 if (head
->type
== RDMA_CONTROL_ERROR
) {
1790 rdma
->received_error
= true;
1794 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1795 error_report("too long length: %d", head
->len
);
1798 if (sizeof(*head
) + head
->len
!= byte_len
) {
1799 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1807 * When a RECV work request has completed, the work request's
1808 * buffer is pointed at the header.
1810 * This will advance the pointer to the data portion
1811 * of the control message of the work request's buffer that
1812 * was populated after the work request finished.
1814 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1815 RDMAControlHeader
*head
)
1817 rdma
->wr_data
[idx
].control_len
= head
->len
;
1818 rdma
->wr_data
[idx
].control_curr
=
1819 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1823 * This is an 'atomic' high-level operation to deliver a single, unified
1824 * control-channel message.
1826 * Additionally, if the user is expecting some kind of reply to this message,
1827 * they can request a 'resp' response message be filled in by posting an
1828 * additional work request on behalf of the user and waiting for an additional
1831 * The extra (optional) response is used during registration to us from having
1832 * to perform an *additional* exchange of message just to provide a response by
1833 * instead piggy-backing on the acknowledgement.
1835 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1836 uint8_t *data
, RDMAControlHeader
*resp
,
1838 int (*callback
)(RDMAContext
*rdma
))
1843 * Wait until the dest is ready before attempting to deliver the message
1844 * by waiting for a READY message.
1846 if (rdma
->control_ready_expected
) {
1847 RDMAControlHeader resp
;
1848 ret
= qemu_rdma_exchange_get_response(rdma
,
1849 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1856 * If the user is expecting a response, post a WR in anticipation of it.
1859 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1861 error_report("rdma migration: error posting"
1862 " extra control recv for anticipated result!");
1868 * Post a WR to replace the one we just consumed for the READY message.
1870 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1872 error_report("rdma migration: error posting first control recv!");
1877 * Deliver the control message that was requested.
1879 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1882 error_report("Failed to send control buffer!");
1887 * If we're expecting a response, block and wait for it.
1891 trace_qemu_rdma_exchange_send_issue_callback();
1892 ret
= callback(rdma
);
1898 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1899 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1900 resp
->type
, RDMA_WRID_DATA
);
1906 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1908 *resp_idx
= RDMA_WRID_DATA
;
1910 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1913 rdma
->control_ready_expected
= 1;
1919 * This is an 'atomic' high-level operation to receive a single, unified
1920 * control-channel message.
1922 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1925 RDMAControlHeader ready
= {
1927 .type
= RDMA_CONTROL_READY
,
1933 * Inform the source that we're ready to receive a message.
1935 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1938 error_report("Failed to send control buffer!");
1943 * Block and wait for the message.
1945 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1946 expecting
, RDMA_WRID_READY
);
1952 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1955 * Post a new RECV work request to replace the one we just consumed.
1957 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1959 error_report("rdma migration: error posting second control recv!");
1967 * Write an actual chunk of memory using RDMA.
1969 * If we're using dynamic registration on the dest-side, we have to
1970 * send a registration command first.
1972 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
1973 int current_index
, uint64_t current_addr
,
1977 struct ibv_send_wr send_wr
= { 0 };
1978 struct ibv_send_wr
*bad_wr
;
1979 int reg_result_idx
, ret
, count
= 0;
1980 uint64_t chunk
, chunks
;
1981 uint8_t *chunk_start
, *chunk_end
;
1982 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
1984 RDMARegisterResult
*reg_result
;
1985 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
1986 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1987 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
1992 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
1993 (current_addr
- block
->offset
));
1994 sge
.length
= length
;
1996 chunk
= ram_chunk_index(block
->local_host_addr
,
1997 (uint8_t *)(uintptr_t)sge
.addr
);
1998 chunk_start
= ram_chunk_start(block
, chunk
);
2000 if (block
->is_ram_block
) {
2001 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2003 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2007 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2009 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2014 trace_qemu_rdma_write_one_top(chunks
+ 1,
2016 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2018 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2020 if (!rdma
->pin_all
) {
2021 #ifdef RDMA_UNREGISTRATION_EXAMPLE
2022 qemu_rdma_unregister_waiting(rdma
);
2026 while (test_bit(chunk
, block
->transit_bitmap
)) {
2028 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2029 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2031 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2034 error_report("Failed to Wait for previous write to complete "
2035 "block %d chunk %" PRIu64
2036 " current %" PRIu64
" len %" PRIu64
" %d",
2037 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2042 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2043 if (!block
->remote_keys
[chunk
]) {
2045 * This chunk has not yet been registered, so first check to see
2046 * if the entire chunk is zero. If so, tell the other size to
2047 * memset() + madvise() the entire chunk without RDMA.
2050 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2051 RDMACompress comp
= {
2052 .offset
= current_addr
,
2054 .block_idx
= current_index
,
2058 head
.len
= sizeof(comp
);
2059 head
.type
= RDMA_CONTROL_COMPRESS
;
2061 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2062 current_index
, current_addr
);
2064 compress_to_network(rdma
, &comp
);
2065 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2066 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2072 acct_update_position(f
, sge
.length
, true);
2078 * Otherwise, tell other side to register.
2080 reg
.current_index
= current_index
;
2081 if (block
->is_ram_block
) {
2082 reg
.key
.current_addr
= current_addr
;
2084 reg
.key
.chunk
= chunk
;
2086 reg
.chunks
= chunks
;
2088 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2091 register_to_network(rdma
, ®
);
2092 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2093 &resp
, ®_result_idx
, NULL
);
2098 /* try to overlap this single registration with the one we sent. */
2099 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2100 &sge
.lkey
, NULL
, chunk
,
2101 chunk_start
, chunk_end
)) {
2102 error_report("cannot get lkey");
2106 reg_result
= (RDMARegisterResult
*)
2107 rdma
->wr_data
[reg_result_idx
].control_curr
;
2109 network_to_result(reg_result
);
2111 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2112 reg_result
->rkey
, chunk
);
2114 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2115 block
->remote_host_addr
= reg_result
->host_addr
;
2117 /* already registered before */
2118 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2119 &sge
.lkey
, NULL
, chunk
,
2120 chunk_start
, chunk_end
)) {
2121 error_report("cannot get lkey!");
2126 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2128 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2130 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2131 &sge
.lkey
, NULL
, chunk
,
2132 chunk_start
, chunk_end
)) {
2133 error_report("cannot get lkey!");
2139 * Encode the ram block index and chunk within this wrid.
2140 * We will use this information at the time of completion
2141 * to figure out which bitmap to check against and then which
2142 * chunk in the bitmap to look for.
2144 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2145 current_index
, chunk
);
2147 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2148 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2149 send_wr
.sg_list
= &sge
;
2150 send_wr
.num_sge
= 1;
2151 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2152 (current_addr
- block
->offset
);
2154 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2158 * ibv_post_send() does not return negative error numbers,
2159 * per the specification they are positive - no idea why.
2161 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2163 if (ret
== ENOMEM
) {
2164 trace_qemu_rdma_write_one_queue_full();
2165 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2167 error_report("rdma migration: failed to make "
2168 "room in full send queue! %d", ret
);
2174 } else if (ret
> 0) {
2175 perror("rdma migration: post rdma write failed");
2179 set_bit(chunk
, block
->transit_bitmap
);
2180 acct_update_position(f
, sge
.length
, false);
2181 rdma
->total_writes
++;
2187 * Push out any unwritten RDMA operations.
2189 * We support sending out multiple chunks at the same time.
2190 * Not all of them need to get signaled in the completion queue.
2192 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2196 if (!rdma
->current_length
) {
2200 ret
= qemu_rdma_write_one(f
, rdma
,
2201 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2209 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2212 rdma
->current_length
= 0;
2213 rdma
->current_addr
= 0;
2218 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2219 uint64_t offset
, uint64_t len
)
2221 RDMALocalBlock
*block
;
2225 if (rdma
->current_index
< 0) {
2229 if (rdma
->current_chunk
< 0) {
2233 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2234 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2235 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2237 if (rdma
->current_length
== 0) {
2242 * Only merge into chunk sequentially.
2244 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2248 if (offset
< block
->offset
) {
2252 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2256 if ((host_addr
+ len
) > chunk_end
) {
2264 * We're not actually writing here, but doing three things:
2266 * 1. Identify the chunk the buffer belongs to.
2267 * 2. If the chunk is full or the buffer doesn't belong to the current
2268 * chunk, then start a new chunk and flush() the old chunk.
2269 * 3. To keep the hardware busy, we also group chunks into batches
2270 * and only require that a batch gets acknowledged in the completion
2271 * qeueue instead of each individual chunk.
2273 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2274 uint64_t block_offset
, uint64_t offset
,
2277 uint64_t current_addr
= block_offset
+ offset
;
2278 uint64_t index
= rdma
->current_index
;
2279 uint64_t chunk
= rdma
->current_chunk
;
2282 /* If we cannot merge it, we flush the current buffer first. */
2283 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2284 ret
= qemu_rdma_write_flush(f
, rdma
);
2288 rdma
->current_length
= 0;
2289 rdma
->current_addr
= current_addr
;
2291 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2292 offset
, len
, &index
, &chunk
);
2294 error_report("ram block search failed");
2297 rdma
->current_index
= index
;
2298 rdma
->current_chunk
= chunk
;
2302 rdma
->current_length
+= len
;
2304 /* flush it if buffer is too large */
2305 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2306 return qemu_rdma_write_flush(f
, rdma
);
2312 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2316 if (rdma
->cm_id
&& rdma
->connected
) {
2317 if ((rdma
->error_state
||
2318 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2319 !rdma
->received_error
) {
2320 RDMAControlHeader head
= { .len
= 0,
2321 .type
= RDMA_CONTROL_ERROR
,
2324 error_report("Early error. Sending error.");
2325 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2328 rdma_disconnect(rdma
->cm_id
);
2329 trace_qemu_rdma_cleanup_disconnect();
2330 rdma
->connected
= false;
2333 if (rdma
->channel
) {
2334 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2336 g_free(rdma
->dest_blocks
);
2337 rdma
->dest_blocks
= NULL
;
2339 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2340 if (rdma
->wr_data
[idx
].control_mr
) {
2341 rdma
->total_registrations
--;
2342 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2344 rdma
->wr_data
[idx
].control_mr
= NULL
;
2347 if (rdma
->local_ram_blocks
.block
) {
2348 while (rdma
->local_ram_blocks
.nb_blocks
) {
2349 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2354 rdma_destroy_qp(rdma
->cm_id
);
2358 ibv_destroy_cq(rdma
->cq
);
2361 if (rdma
->comp_channel
) {
2362 ibv_destroy_comp_channel(rdma
->comp_channel
);
2363 rdma
->comp_channel
= NULL
;
2366 ibv_dealloc_pd(rdma
->pd
);
2370 rdma_destroy_id(rdma
->cm_id
);
2374 /* the destination side, listen_id and channel is shared */
2375 if (rdma
->listen_id
) {
2376 if (!rdma
->is_return_path
) {
2377 rdma_destroy_id(rdma
->listen_id
);
2379 rdma
->listen_id
= NULL
;
2381 if (rdma
->channel
) {
2382 if (!rdma
->is_return_path
) {
2383 rdma_destroy_event_channel(rdma
->channel
);
2385 rdma
->channel
= NULL
;
2389 if (rdma
->channel
) {
2390 rdma_destroy_event_channel(rdma
->channel
);
2391 rdma
->channel
= NULL
;
2398 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2401 Error
*local_err
= NULL
, **temp
= &local_err
;
2404 * Will be validated against destination's actual capabilities
2405 * after the connect() completes.
2407 rdma
->pin_all
= pin_all
;
2409 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2411 goto err_rdma_source_init
;
2414 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2416 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2417 " limits may be too low. Please check $ ulimit -a # and "
2418 "search for 'ulimit -l' in the output");
2419 goto err_rdma_source_init
;
2422 ret
= qemu_rdma_alloc_qp(rdma
);
2424 ERROR(temp
, "rdma migration: error allocating qp!");
2425 goto err_rdma_source_init
;
2428 ret
= qemu_rdma_init_ram_blocks(rdma
);
2430 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2431 goto err_rdma_source_init
;
2434 /* Build the hash that maps from offset to RAMBlock */
2435 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2436 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2437 g_hash_table_insert(rdma
->blockmap
,
2438 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2439 &rdma
->local_ram_blocks
.block
[idx
]);
2442 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2443 ret
= qemu_rdma_reg_control(rdma
, idx
);
2445 ERROR(temp
, "rdma migration: error registering %d control!",
2447 goto err_rdma_source_init
;
2453 err_rdma_source_init
:
2454 error_propagate(errp
, local_err
);
2455 qemu_rdma_cleanup(rdma
);
2459 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
)
2461 RDMACapabilities cap
= {
2462 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2465 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2467 .private_data
= &cap
,
2468 .private_data_len
= sizeof(cap
),
2470 struct rdma_cm_event
*cm_event
;
2474 * Only negotiate the capability with destination if the user
2475 * on the source first requested the capability.
2477 if (rdma
->pin_all
) {
2478 trace_qemu_rdma_connect_pin_all_requested();
2479 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2482 caps_to_network(&cap
);
2484 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2486 ERROR(errp
, "posting second control recv");
2487 goto err_rdma_source_connect
;
2490 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2492 perror("rdma_connect");
2493 ERROR(errp
, "connecting to destination!");
2494 goto err_rdma_source_connect
;
2497 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2499 perror("rdma_get_cm_event after rdma_connect");
2500 ERROR(errp
, "connecting to destination!");
2501 rdma_ack_cm_event(cm_event
);
2502 goto err_rdma_source_connect
;
2505 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2506 perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2507 ERROR(errp
, "connecting to destination!");
2508 rdma_ack_cm_event(cm_event
);
2509 goto err_rdma_source_connect
;
2511 rdma
->connected
= true;
2513 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2514 network_to_caps(&cap
);
2517 * Verify that the *requested* capabilities are supported by the destination
2518 * and disable them otherwise.
2520 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2521 ERROR(errp
, "Server cannot support pinning all memory. "
2522 "Will register memory dynamically.");
2523 rdma
->pin_all
= false;
2526 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2528 rdma_ack_cm_event(cm_event
);
2530 rdma
->control_ready_expected
= 1;
2534 err_rdma_source_connect
:
2535 qemu_rdma_cleanup(rdma
);
2539 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2542 struct rdma_cm_id
*listen_id
;
2543 char ip
[40] = "unknown";
2544 struct rdma_addrinfo
*res
, *e
;
2547 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2548 rdma
->wr_data
[idx
].control_len
= 0;
2549 rdma
->wr_data
[idx
].control_curr
= NULL
;
2552 if (!rdma
->host
|| !rdma
->host
[0]) {
2553 ERROR(errp
, "RDMA host is not set!");
2554 rdma
->error_state
= -EINVAL
;
2557 /* create CM channel */
2558 rdma
->channel
= rdma_create_event_channel();
2559 if (!rdma
->channel
) {
2560 ERROR(errp
, "could not create rdma event channel");
2561 rdma
->error_state
= -EINVAL
;
2566 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2568 ERROR(errp
, "could not create cm_id!");
2569 goto err_dest_init_create_listen_id
;
2572 snprintf(port_str
, 16, "%d", rdma
->port
);
2573 port_str
[15] = '\0';
2575 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2577 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2578 goto err_dest_init_bind_addr
;
2581 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2582 inet_ntop(e
->ai_family
,
2583 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2584 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2585 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2589 if (e
->ai_family
== AF_INET6
) {
2590 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2599 ERROR(errp
, "Error: could not rdma_bind_addr!");
2600 goto err_dest_init_bind_addr
;
2603 rdma
->listen_id
= listen_id
;
2604 qemu_rdma_dump_gid("dest_init", listen_id
);
2607 err_dest_init_bind_addr
:
2608 rdma_destroy_id(listen_id
);
2609 err_dest_init_create_listen_id
:
2610 rdma_destroy_event_channel(rdma
->channel
);
2611 rdma
->channel
= NULL
;
2612 rdma
->error_state
= ret
;
2617 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2622 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2623 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2624 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2627 /*the CM channel and CM id is shared*/
2628 rdma_return_path
->channel
= rdma
->channel
;
2629 rdma_return_path
->listen_id
= rdma
->listen_id
;
2631 rdma
->return_path
= rdma_return_path
;
2632 rdma_return_path
->return_path
= rdma
;
2633 rdma_return_path
->is_return_path
= true;
2636 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2638 RDMAContext
*rdma
= NULL
;
2639 InetSocketAddress
*addr
;
2642 rdma
= g_new0(RDMAContext
, 1);
2643 rdma
->current_index
= -1;
2644 rdma
->current_chunk
= -1;
2646 addr
= g_new(InetSocketAddress
, 1);
2647 if (!inet_parse(addr
, host_port
, NULL
)) {
2648 rdma
->port
= atoi(addr
->port
);
2649 rdma
->host
= g_strdup(addr
->host
);
2651 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2656 qapi_free_InetSocketAddress(addr
);
2663 * QEMUFile interface to the control channel.
2664 * SEND messages for control only.
2665 * VM's ram is handled with regular RDMA messages.
2667 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2668 const struct iovec
*iov
,
2674 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2675 QEMUFile
*f
= rioc
->file
;
2682 RCU_READ_LOCK_GUARD();
2683 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
2689 CHECK_ERROR_STATE();
2692 * Push out any writes that
2693 * we're queued up for VM's ram.
2695 ret
= qemu_rdma_write_flush(f
, rdma
);
2697 rdma
->error_state
= ret
;
2701 for (i
= 0; i
< niov
; i
++) {
2702 size_t remaining
= iov
[i
].iov_len
;
2703 uint8_t * data
= (void *)iov
[i
].iov_base
;
2705 RDMAControlHeader head
;
2707 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2711 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2713 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2716 rdma
->error_state
= ret
;
2728 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2729 size_t size
, int idx
)
2733 if (rdma
->wr_data
[idx
].control_len
) {
2734 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2736 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2737 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2738 rdma
->wr_data
[idx
].control_curr
+= len
;
2739 rdma
->wr_data
[idx
].control_len
-= len
;
2746 * QEMUFile interface to the control channel.
2747 * RDMA links don't use bytestreams, so we have to
2748 * return bytes to QEMUFile opportunistically.
2750 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2751 const struct iovec
*iov
,
2757 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2759 RDMAControlHeader head
;
2764 RCU_READ_LOCK_GUARD();
2765 rdma
= atomic_rcu_read(&rioc
->rdmain
);
2771 CHECK_ERROR_STATE();
2773 for (i
= 0; i
< niov
; i
++) {
2774 size_t want
= iov
[i
].iov_len
;
2775 uint8_t *data
= (void *)iov
[i
].iov_base
;
2778 * First, we hold on to the last SEND message we
2779 * were given and dish out the bytes until we run
2782 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2785 /* Got what we needed, so go to next iovec */
2790 /* If we got any data so far, then don't wait
2791 * for more, just return what we have */
2797 /* We've got nothing at all, so lets wait for
2800 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2803 rdma
->error_state
= ret
;
2808 * SEND was received with new bytes, now try again.
2810 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2814 /* Still didn't get enough, so lets just return */
2817 return QIO_CHANNEL_ERR_BLOCK
;
2827 * Block until all the outstanding chunks have been delivered by the hardware.
2829 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2833 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2837 while (rdma
->nb_sent
) {
2838 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2840 error_report("rdma migration: complete polling error!");
2845 qemu_rdma_unregister_waiting(rdma
);
2851 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2855 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2856 /* XXX we should make readv/writev actually honour this :-) */
2857 rioc
->blocking
= blocking
;
2862 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2863 struct QIOChannelRDMASource
{
2865 QIOChannelRDMA
*rioc
;
2866 GIOCondition condition
;
2870 qio_channel_rdma_source_prepare(GSource
*source
,
2873 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2875 GIOCondition cond
= 0;
2878 RCU_READ_LOCK_GUARD();
2879 if (rsource
->condition
== G_IO_IN
) {
2880 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2882 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2886 error_report("RDMAContext is NULL when prepare Gsource");
2890 if (rdma
->wr_data
[0].control_len
) {
2895 return cond
& rsource
->condition
;
2899 qio_channel_rdma_source_check(GSource
*source
)
2901 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2903 GIOCondition cond
= 0;
2905 RCU_READ_LOCK_GUARD();
2906 if (rsource
->condition
== G_IO_IN
) {
2907 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2909 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2913 error_report("RDMAContext is NULL when check Gsource");
2917 if (rdma
->wr_data
[0].control_len
) {
2922 return cond
& rsource
->condition
;
2926 qio_channel_rdma_source_dispatch(GSource
*source
,
2927 GSourceFunc callback
,
2930 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
2931 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2933 GIOCondition cond
= 0;
2935 RCU_READ_LOCK_GUARD();
2936 if (rsource
->condition
== G_IO_IN
) {
2937 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmain
);
2939 rdma
= atomic_rcu_read(&rsource
->rioc
->rdmaout
);
2943 error_report("RDMAContext is NULL when dispatch Gsource");
2947 if (rdma
->wr_data
[0].control_len
) {
2952 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
2953 (cond
& rsource
->condition
),
2958 qio_channel_rdma_source_finalize(GSource
*source
)
2960 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
2962 object_unref(OBJECT(ssource
->rioc
));
2965 GSourceFuncs qio_channel_rdma_source_funcs
= {
2966 qio_channel_rdma_source_prepare
,
2967 qio_channel_rdma_source_check
,
2968 qio_channel_rdma_source_dispatch
,
2969 qio_channel_rdma_source_finalize
2972 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
2973 GIOCondition condition
)
2975 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2976 QIOChannelRDMASource
*ssource
;
2979 source
= g_source_new(&qio_channel_rdma_source_funcs
,
2980 sizeof(QIOChannelRDMASource
));
2981 ssource
= (QIOChannelRDMASource
*)source
;
2983 ssource
->rioc
= rioc
;
2984 object_ref(OBJECT(rioc
));
2986 ssource
->condition
= condition
;
2991 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
2994 IOHandler
*io_write
,
2997 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2999 aio_set_fd_handler(ctx
, rioc
->rdmain
->comp_channel
->fd
,
3000 false, io_read
, io_write
, NULL
, opaque
);
3002 aio_set_fd_handler(ctx
, rioc
->rdmaout
->comp_channel
->fd
,
3003 false, io_read
, io_write
, NULL
, opaque
);
3007 struct rdma_close_rcu
{
3008 struct rcu_head rcu
;
3009 RDMAContext
*rdmain
;
3010 RDMAContext
*rdmaout
;
3013 /* callback from qio_channel_rdma_close via call_rcu */
3014 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu
*rcu
)
3017 qemu_rdma_cleanup(rcu
->rdmain
);
3021 qemu_rdma_cleanup(rcu
->rdmaout
);
3024 g_free(rcu
->rdmain
);
3025 g_free(rcu
->rdmaout
);
3029 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3032 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3033 RDMAContext
*rdmain
, *rdmaout
;
3034 struct rdma_close_rcu
*rcu
= g_new(struct rdma_close_rcu
, 1);
3036 trace_qemu_rdma_close();
3038 rdmain
= rioc
->rdmain
;
3040 atomic_rcu_set(&rioc
->rdmain
, NULL
);
3043 rdmaout
= rioc
->rdmaout
;
3045 atomic_rcu_set(&rioc
->rdmaout
, NULL
);
3048 rcu
->rdmain
= rdmain
;
3049 rcu
->rdmaout
= rdmaout
;
3050 call_rcu(rcu
, qio_channel_rdma_close_rcu
, rcu
);
3056 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3057 QIOChannelShutdown how
,
3060 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3061 RDMAContext
*rdmain
, *rdmaout
;
3063 RCU_READ_LOCK_GUARD();
3065 rdmain
= atomic_rcu_read(&rioc
->rdmain
);
3066 rdmaout
= atomic_rcu_read(&rioc
->rdmain
);
3069 case QIO_CHANNEL_SHUTDOWN_READ
:
3071 rdmain
->error_state
= -1;
3074 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3076 rdmaout
->error_state
= -1;
3079 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3082 rdmain
->error_state
= -1;
3085 rdmaout
->error_state
= -1;
3096 * This means that 'block_offset' is a full virtual address that does not
3097 * belong to a RAMBlock of the virtual machine and instead
3098 * represents a private malloc'd memory area that the caller wishes to
3102 * Offset is an offset to be added to block_offset and used
3103 * to also lookup the corresponding RAMBlock.
3106 * Initiate an transfer this size.
3109 * A 'hint' or 'advice' that means that we wish to speculatively
3110 * and asynchronously unregister this memory. In this case, there is no
3111 * guarantee that the unregister will actually happen, for example,
3112 * if the memory is being actively transmitted. Additionally, the memory
3113 * may be re-registered at any future time if a write within the same
3114 * chunk was requested again, even if you attempted to unregister it
3117 * @size < 0 : TODO, not yet supported
3118 * Unregister the memory NOW. This means that the caller does not
3119 * expect there to be any future RDMA transfers and we just want to clean
3120 * things up. This is used in case the upper layer owns the memory and
3121 * cannot wait for qemu_fclose() to occur.
3123 * @bytes_sent : User-specificed pointer to indicate how many bytes were
3124 * sent. Usually, this will not be more than a few bytes of
3125 * the protocol because most transfers are sent asynchronously.
3127 static size_t qemu_rdma_save_page(QEMUFile
*f
, void *opaque
,
3128 ram_addr_t block_offset
, ram_addr_t offset
,
3129 size_t size
, uint64_t *bytes_sent
)
3131 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3135 RCU_READ_LOCK_GUARD();
3136 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3142 CHECK_ERROR_STATE();
3144 if (migration_in_postcopy()) {
3145 return RAM_SAVE_CONTROL_NOT_SUPP
;
3152 * Add this page to the current 'chunk'. If the chunk
3153 * is full, or the page doen't belong to the current chunk,
3154 * an actual RDMA write will occur and a new chunk will be formed.
3156 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
3158 error_report("rdma migration: write error! %d", ret
);
3163 * We always return 1 bytes because the RDMA
3164 * protocol is completely asynchronous. We do not yet know
3165 * whether an identified chunk is zero or not because we're
3166 * waiting for other pages to potentially be merged with
3167 * the current chunk. So, we have to call qemu_update_position()
3168 * later on when the actual write occurs.
3174 uint64_t index
, chunk
;
3176 /* TODO: Change QEMUFileOps prototype to be signed: size_t => long
3178 ret = qemu_rdma_drain_cq(f, rdma);
3180 fprintf(stderr, "rdma: failed to synchronously drain"
3181 " completion queue before unregistration.\n");
3187 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
3188 offset
, size
, &index
, &chunk
);
3191 error_report("ram block search failed");
3195 qemu_rdma_signal_unregister(rdma
, index
, chunk
, 0);
3198 * TODO: Synchronous, guaranteed unregistration (should not occur during
3199 * fast-path). Otherwise, unregisters will process on the next call to
3200 * qemu_rdma_drain_cq()
3202 qemu_rdma_unregister_waiting(rdma);
3208 * Drain the Completion Queue if possible, but do not block,
3211 * If nothing to poll, the end of the iteration will do this
3212 * again to make sure we don't overflow the request queue.
3215 uint64_t wr_id
, wr_id_in
;
3216 int ret
= qemu_rdma_poll(rdma
, &wr_id_in
, NULL
);
3218 error_report("rdma migration: polling error! %d", ret
);
3222 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3224 if (wr_id
== RDMA_WRID_NONE
) {
3229 return RAM_SAVE_CONTROL_DELAYED
;
3231 rdma
->error_state
= ret
;
3235 static void rdma_accept_incoming_migration(void *opaque
);
3237 static void rdma_cm_poll_handler(void *opaque
)
3239 RDMAContext
*rdma
= opaque
;
3241 struct rdma_cm_event
*cm_event
;
3242 MigrationIncomingState
*mis
= migration_incoming_get_current();
3244 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3246 error_report("get_cm_event failed %d", errno
);
3249 rdma_ack_cm_event(cm_event
);
3251 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3252 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3253 if (!rdma
->error_state
&&
3254 migration_incoming_get_current()->state
!=
3255 MIGRATION_STATUS_COMPLETED
) {
3256 error_report("receive cm event, cm event is %d", cm_event
->event
);
3257 rdma
->error_state
= -EPIPE
;
3258 if (rdma
->return_path
) {
3259 rdma
->return_path
->error_state
= -EPIPE
;
3263 if (mis
->migration_incoming_co
) {
3264 qemu_coroutine_enter(mis
->migration_incoming_co
);
3270 static int qemu_rdma_accept(RDMAContext
*rdma
)
3272 RDMACapabilities cap
;
3273 struct rdma_conn_param conn_param
= {
3274 .responder_resources
= 2,
3275 .private_data
= &cap
,
3276 .private_data_len
= sizeof(cap
),
3278 struct rdma_cm_event
*cm_event
;
3279 struct ibv_context
*verbs
;
3283 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3285 goto err_rdma_dest_wait
;
3288 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3289 rdma_ack_cm_event(cm_event
);
3290 goto err_rdma_dest_wait
;
3293 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3295 network_to_caps(&cap
);
3297 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3298 error_report("Unknown source RDMA version: %d, bailing...",
3300 rdma_ack_cm_event(cm_event
);
3301 goto err_rdma_dest_wait
;
3305 * Respond with only the capabilities this version of QEMU knows about.
3307 cap
.flags
&= known_capabilities
;
3310 * Enable the ones that we do know about.
3311 * Add other checks here as new ones are introduced.
3313 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3314 rdma
->pin_all
= true;
3317 rdma
->cm_id
= cm_event
->id
;
3318 verbs
= cm_event
->id
->verbs
;
3320 rdma_ack_cm_event(cm_event
);
3322 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3324 caps_to_network(&cap
);
3326 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3329 rdma
->verbs
= verbs
;
3330 } else if (rdma
->verbs
!= verbs
) {
3331 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3333 goto err_rdma_dest_wait
;
3336 qemu_rdma_dump_id("dest_init", verbs
);
3338 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3340 error_report("rdma migration: error allocating pd and cq!");
3341 goto err_rdma_dest_wait
;
3344 ret
= qemu_rdma_alloc_qp(rdma
);
3346 error_report("rdma migration: error allocating qp!");
3347 goto err_rdma_dest_wait
;
3350 ret
= qemu_rdma_init_ram_blocks(rdma
);
3352 error_report("rdma migration: error initializing ram blocks!");
3353 goto err_rdma_dest_wait
;
3356 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3357 ret
= qemu_rdma_reg_control(rdma
, idx
);
3359 error_report("rdma: error registering %d control", idx
);
3360 goto err_rdma_dest_wait
;
3364 /* Accept the second connection request for return path */
3365 if (migrate_postcopy() && !rdma
->is_return_path
) {
3366 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3368 (void *)(intptr_t)rdma
->return_path
);
3370 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3374 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3376 error_report("rdma_accept returns %d", ret
);
3377 goto err_rdma_dest_wait
;
3380 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3382 error_report("rdma_accept get_cm_event failed %d", ret
);
3383 goto err_rdma_dest_wait
;
3386 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3387 error_report("rdma_accept not event established");
3388 rdma_ack_cm_event(cm_event
);
3389 goto err_rdma_dest_wait
;
3392 rdma_ack_cm_event(cm_event
);
3393 rdma
->connected
= true;
3395 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3397 error_report("rdma migration: error posting second control recv");
3398 goto err_rdma_dest_wait
;
3401 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3406 rdma
->error_state
= ret
;
3407 qemu_rdma_cleanup(rdma
);
3411 static int dest_ram_sort_func(const void *a
, const void *b
)
3413 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3414 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3416 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3420 * During each iteration of the migration, we listen for instructions
3421 * by the source VM to perform dynamic page registrations before they
3422 * can perform RDMA operations.
3424 * We respond with the 'rkey'.
3426 * Keep doing this until the source tells us to stop.
3428 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
)
3430 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3431 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3434 RDMAControlHeader unreg_resp
= { .len
= 0,
3435 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3438 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3440 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3442 RDMALocalBlocks
*local
;
3443 RDMAControlHeader head
;
3444 RDMARegister
*reg
, *registers
;
3446 RDMARegisterResult
*reg_result
;
3447 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3448 RDMALocalBlock
*block
;
3455 RCU_READ_LOCK_GUARD();
3456 rdma
= atomic_rcu_read(&rioc
->rdmain
);
3462 CHECK_ERROR_STATE();
3464 local
= &rdma
->local_ram_blocks
;
3466 trace_qemu_rdma_registration_handle_wait();
3468 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3474 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3475 error_report("rdma: Too many requests in this message (%d)."
3476 "Bailing.", head
.repeat
);
3481 switch (head
.type
) {
3482 case RDMA_CONTROL_COMPRESS
:
3483 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3484 network_to_compress(comp
);
3486 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3489 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3490 error_report("rdma: 'compress' bad block index %u (vs %d)",
3491 (unsigned int)comp
->block_idx
,
3492 rdma
->local_ram_blocks
.nb_blocks
);
3496 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3498 host_addr
= block
->local_host_addr
+
3499 (comp
->offset
- block
->offset
);
3501 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3504 case RDMA_CONTROL_REGISTER_FINISHED
:
3505 trace_qemu_rdma_registration_handle_finished();
3508 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3509 trace_qemu_rdma_registration_handle_ram_blocks();
3511 /* Sort our local RAM Block list so it's the same as the source,
3512 * we can do this since we've filled in a src_index in the list
3513 * as we received the RAMBlock list earlier.
3515 qsort(rdma
->local_ram_blocks
.block
,
3516 rdma
->local_ram_blocks
.nb_blocks
,
3517 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3518 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3519 local
->block
[i
].index
= i
;
3522 if (rdma
->pin_all
) {
3523 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3525 error_report("rdma migration: error dest "
3526 "registering ram blocks");
3532 * Dest uses this to prepare to transmit the RAMBlock descriptions
3533 * to the source VM after connection setup.
3534 * Both sides use the "remote" structure to communicate and update
3535 * their "local" descriptions with what was sent.
3537 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3538 rdma
->dest_blocks
[i
].remote_host_addr
=
3539 (uintptr_t)(local
->block
[i
].local_host_addr
);
3541 if (rdma
->pin_all
) {
3542 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3545 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3546 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3548 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3549 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3550 local
->block
[i
].block_name
,
3551 local
->block
[i
].offset
,
3552 local
->block
[i
].length
,
3553 local
->block
[i
].local_host_addr
,
3554 local
->block
[i
].src_index
);
3557 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3558 * sizeof(RDMADestBlock
);
3561 ret
= qemu_rdma_post_send_control(rdma
,
3562 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3565 error_report("rdma migration: error sending remote info");
3570 case RDMA_CONTROL_REGISTER_REQUEST
:
3571 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3573 reg_resp
.repeat
= head
.repeat
;
3574 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3576 for (count
= 0; count
< head
.repeat
; count
++) {
3578 uint8_t *chunk_start
, *chunk_end
;
3580 reg
= ®isters
[count
];
3581 network_to_register(reg
);
3583 reg_result
= &results
[count
];
3585 trace_qemu_rdma_registration_handle_register_loop(count
,
3586 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3588 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3589 error_report("rdma: 'register' bad block index %u (vs %d)",
3590 (unsigned int)reg
->current_index
,
3591 rdma
->local_ram_blocks
.nb_blocks
);
3595 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3596 if (block
->is_ram_block
) {
3597 if (block
->offset
> reg
->key
.current_addr
) {
3598 error_report("rdma: bad register address for block %s"
3599 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3600 block
->block_name
, block
->offset
,
3601 reg
->key
.current_addr
);
3605 host_addr
= (block
->local_host_addr
+
3606 (reg
->key
.current_addr
- block
->offset
));
3607 chunk
= ram_chunk_index(block
->local_host_addr
,
3608 (uint8_t *) host_addr
);
3610 chunk
= reg
->key
.chunk
;
3611 host_addr
= block
->local_host_addr
+
3612 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3613 /* Check for particularly bad chunk value */
3614 if (host_addr
< (void *)block
->local_host_addr
) {
3615 error_report("rdma: bad chunk for block %s"
3617 block
->block_name
, reg
->key
.chunk
);
3622 chunk_start
= ram_chunk_start(block
, chunk
);
3623 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3624 /* avoid "-Waddress-of-packed-member" warning */
3625 uint32_t tmp_rkey
= 0;
3626 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3627 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3628 chunk
, chunk_start
, chunk_end
)) {
3629 error_report("cannot get rkey");
3633 reg_result
->rkey
= tmp_rkey
;
3635 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3637 trace_qemu_rdma_registration_handle_register_rkey(
3640 result_to_network(reg_result
);
3643 ret
= qemu_rdma_post_send_control(rdma
,
3644 (uint8_t *) results
, ®_resp
);
3647 error_report("Failed to send control buffer");
3651 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3652 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3653 unreg_resp
.repeat
= head
.repeat
;
3654 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3656 for (count
= 0; count
< head
.repeat
; count
++) {
3657 reg
= ®isters
[count
];
3658 network_to_register(reg
);
3660 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3661 reg
->current_index
, reg
->key
.chunk
);
3663 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3665 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3666 block
->pmr
[reg
->key
.chunk
] = NULL
;
3669 perror("rdma unregistration chunk failed");
3674 rdma
->total_registrations
--;
3676 trace_qemu_rdma_registration_handle_unregister_success(
3680 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3683 error_report("Failed to send control buffer");
3687 case RDMA_CONTROL_REGISTER_RESULT
:
3688 error_report("Invalid RESULT message at dest.");
3692 error_report("Unknown control message %s", control_desc(head
.type
));
3699 rdma
->error_state
= ret
;
3705 * Called via a ram_control_load_hook during the initial RAM load section which
3706 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3708 * We've already built our local RAMBlock list, but not yet sent the list to
3712 rdma_block_notification_handle(QIOChannelRDMA
*rioc
, const char *name
)
3718 RCU_READ_LOCK_GUARD();
3719 rdma
= atomic_rcu_read(&rioc
->rdmain
);
3725 /* Find the matching RAMBlock in our local list */
3726 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3727 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3734 error_report("RAMBlock '%s' not found on destination", name
);
3738 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3739 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3740 rdma
->next_src_index
++;
3745 static int rdma_load_hook(QEMUFile
*f
, void *opaque
, uint64_t flags
, void *data
)
3748 case RAM_CONTROL_BLOCK_REG
:
3749 return rdma_block_notification_handle(opaque
, data
);
3751 case RAM_CONTROL_HOOK
:
3752 return qemu_rdma_registration_handle(f
, opaque
);
3755 /* Shouldn't be called with any other values */
3760 static int qemu_rdma_registration_start(QEMUFile
*f
, void *opaque
,
3761 uint64_t flags
, void *data
)
3763 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3766 RCU_READ_LOCK_GUARD();
3767 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3772 CHECK_ERROR_STATE();
3774 if (migration_in_postcopy()) {
3778 trace_qemu_rdma_registration_start(flags
);
3779 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3786 * Inform dest that dynamic registrations are done for now.
3787 * First, flush writes, if any.
3789 static int qemu_rdma_registration_stop(QEMUFile
*f
, void *opaque
,
3790 uint64_t flags
, void *data
)
3792 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3794 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3797 RCU_READ_LOCK_GUARD();
3798 rdma
= atomic_rcu_read(&rioc
->rdmaout
);
3803 CHECK_ERROR_STATE();
3805 if (migration_in_postcopy()) {
3810 ret
= qemu_rdma_drain_cq(f
, rdma
);
3816 if (flags
== RAM_CONTROL_SETUP
) {
3817 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3818 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3819 int reg_result_idx
, i
, nb_dest_blocks
;
3821 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3822 trace_qemu_rdma_registration_stop_ram();
3825 * Make sure that we parallelize the pinning on both sides.
3826 * For very large guests, doing this serially takes a really
3827 * long time, so we have to 'interleave' the pinning locally
3828 * with the control messages by performing the pinning on this
3829 * side before we receive the control response from the other
3830 * side that the pinning has completed.
3832 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3833 ®_result_idx
, rdma
->pin_all
?
3834 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3836 fprintf(stderr
, "receiving remote info!");
3840 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3843 * The protocol uses two different sets of rkeys (mutually exclusive):
3844 * 1. One key to represent the virtual address of the entire ram block.
3845 * (dynamic chunk registration disabled - pin everything with one rkey.)
3846 * 2. One to represent individual chunks within a ram block.
3847 * (dynamic chunk registration enabled - pin individual chunks.)
3849 * Once the capability is successfully negotiated, the destination transmits
3850 * the keys to use (or sends them later) including the virtual addresses
3851 * and then propagates the remote ram block descriptions to his local copy.
3854 if (local
->nb_blocks
!= nb_dest_blocks
) {
3855 fprintf(stderr
, "ram blocks mismatch (Number of blocks %d vs %d) "
3856 "Your QEMU command line parameters are probably "
3857 "not identical on both the source and destination.",
3858 local
->nb_blocks
, nb_dest_blocks
);
3859 rdma
->error_state
= -EINVAL
;
3863 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3864 memcpy(rdma
->dest_blocks
,
3865 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3866 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3867 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3869 /* We require that the blocks are in the same order */
3870 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3871 fprintf(stderr
, "Block %s/%d has a different length %" PRIu64
3872 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3873 local
->block
[i
].length
,
3874 rdma
->dest_blocks
[i
].length
);
3875 rdma
->error_state
= -EINVAL
;
3878 local
->block
[i
].remote_host_addr
=
3879 rdma
->dest_blocks
[i
].remote_host_addr
;
3880 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3884 trace_qemu_rdma_registration_stop(flags
);
3886 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3887 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3895 rdma
->error_state
= ret
;
3899 static const QEMUFileHooks rdma_read_hooks
= {
3900 .hook_ram_load
= rdma_load_hook
,
3903 static const QEMUFileHooks rdma_write_hooks
= {
3904 .before_ram_iterate
= qemu_rdma_registration_start
,
3905 .after_ram_iterate
= qemu_rdma_registration_stop
,
3906 .save_page
= qemu_rdma_save_page
,
3910 static void qio_channel_rdma_finalize(Object
*obj
)
3912 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
3914 qemu_rdma_cleanup(rioc
->rdmain
);
3915 g_free(rioc
->rdmain
);
3916 rioc
->rdmain
= NULL
;
3918 if (rioc
->rdmaout
) {
3919 qemu_rdma_cleanup(rioc
->rdmaout
);
3920 g_free(rioc
->rdmaout
);
3921 rioc
->rdmaout
= NULL
;
3925 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
3926 void *class_data G_GNUC_UNUSED
)
3928 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
3930 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
3931 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
3932 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
3933 ioc_klass
->io_close
= qio_channel_rdma_close
;
3934 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
3935 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
3936 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
3939 static const TypeInfo qio_channel_rdma_info
= {
3940 .parent
= TYPE_QIO_CHANNEL
,
3941 .name
= TYPE_QIO_CHANNEL_RDMA
,
3942 .instance_size
= sizeof(QIOChannelRDMA
),
3943 .instance_finalize
= qio_channel_rdma_finalize
,
3944 .class_init
= qio_channel_rdma_class_init
,
3947 static void qio_channel_rdma_register_types(void)
3949 type_register_static(&qio_channel_rdma_info
);
3952 type_init(qio_channel_rdma_register_types
);
3954 static QEMUFile
*qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
3956 QIOChannelRDMA
*rioc
;
3958 if (qemu_file_mode_is_not_valid(mode
)) {
3962 rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
3964 if (mode
[0] == 'w') {
3965 rioc
->file
= qemu_fopen_channel_output(QIO_CHANNEL(rioc
));
3966 rioc
->rdmaout
= rdma
;
3967 rioc
->rdmain
= rdma
->return_path
;
3968 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
3970 rioc
->file
= qemu_fopen_channel_input(QIO_CHANNEL(rioc
));
3971 rioc
->rdmain
= rdma
;
3972 rioc
->rdmaout
= rdma
->return_path
;
3973 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
3979 static void rdma_accept_incoming_migration(void *opaque
)
3981 RDMAContext
*rdma
= opaque
;
3984 Error
*local_err
= NULL
;
3986 trace_qemu_rdma_accept_incoming_migration();
3987 ret
= qemu_rdma_accept(rdma
);
3990 fprintf(stderr
, "RDMA ERROR: Migration initialization failed\n");
3994 trace_qemu_rdma_accept_incoming_migration_accepted();
3996 if (rdma
->is_return_path
) {
4000 f
= qemu_fopen_rdma(rdma
, "rb");
4002 fprintf(stderr
, "RDMA ERROR: could not qemu_fopen_rdma\n");
4003 qemu_rdma_cleanup(rdma
);
4007 rdma
->migration_started_on_destination
= 1;
4008 migration_fd_process_incoming(f
, &local_err
);
4010 error_reportf_err(local_err
, "RDMA ERROR:");
4014 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4017 RDMAContext
*rdma
, *rdma_return_path
= NULL
;
4018 Error
*local_err
= NULL
;
4020 trace_rdma_start_incoming_migration();
4022 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4023 if (ram_block_discard_is_required()) {
4024 error_setg(errp
, "RDMA: cannot disable RAM discard");
4028 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
4033 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
4039 trace_rdma_start_incoming_migration_after_dest_init();
4041 ret
= rdma_listen(rdma
->listen_id
, 5);
4044 ERROR(errp
, "listening on socket!");
4048 trace_rdma_start_incoming_migration_after_rdma_listen();
4050 /* initialize the RDMAContext for return path */
4051 if (migrate_postcopy()) {
4052 rdma_return_path
= qemu_rdma_data_init(host_port
, &local_err
);
4054 if (rdma_return_path
== NULL
) {
4058 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
4061 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4062 NULL
, (void *)(intptr_t)rdma
);
4065 error_propagate(errp
, local_err
);
4070 g_free(rdma_return_path
);
4073 void rdma_start_outgoing_migration(void *opaque
,
4074 const char *host_port
, Error
**errp
)
4076 MigrationState
*s
= opaque
;
4077 RDMAContext
*rdma_return_path
= NULL
;
4081 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4082 if (ram_block_discard_is_required()) {
4083 error_setg(errp
, "RDMA: cannot disable RAM discard");
4087 rdma
= qemu_rdma_data_init(host_port
, errp
);
4092 ret
= qemu_rdma_source_init(rdma
,
4093 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4099 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4100 ret
= qemu_rdma_connect(rdma
, errp
);
4106 /* RDMA postcopy need a seprate queue pair for return path */
4107 if (migrate_postcopy()) {
4108 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4110 if (rdma_return_path
== NULL
) {
4111 goto return_path_err
;
4114 ret
= qemu_rdma_source_init(rdma_return_path
,
4115 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4118 goto return_path_err
;
4121 ret
= qemu_rdma_connect(rdma_return_path
, errp
);
4124 goto return_path_err
;
4127 rdma
->return_path
= rdma_return_path
;
4128 rdma_return_path
->return_path
= rdma
;
4129 rdma_return_path
->is_return_path
= true;
4132 trace_rdma_start_outgoing_migration_after_rdma_connect();
4134 s
->to_dst_file
= qemu_fopen_rdma(rdma
, "wb");
4135 migrate_fd_connect(s
, NULL
);
4138 qemu_rdma_cleanup(rdma
);
4141 g_free(rdma_return_path
);