2 * RDMA protocol and interfaces
4 * Copyright IBM, Corp. 2010-2013
5 * Copyright Red Hat, Inc. 2015-2016
8 * Michael R. Hines <mrhines@us.ibm.com>
9 * Jiuxing Liu <jl@us.ibm.com>
10 * Daniel P. Berrange <berrange@redhat.com>
12 * This work is licensed under the terms of the GNU GPL, version 2 or
13 * later. See the COPYING file in the top-level directory.
17 #include "qemu/osdep.h"
18 #include "qapi/error.h"
19 #include "qemu/cutils.h"
20 #include "exec/target_page.h"
22 #include "migration.h"
23 #include "migration-stats.h"
24 #include "qemu-file.h"
26 #include "qemu/error-report.h"
27 #include "qemu/main-loop.h"
28 #include "qemu/module.h"
30 #include "qemu/sockets.h"
31 #include "qemu/bitmap.h"
32 #include "qemu/coroutine.h"
33 #include "exec/memory.h"
34 #include <sys/socket.h>
36 #include <arpa/inet.h>
37 #include <rdma/rdma_cma.h>
39 #include "qom/object.h"
43 #define RDMA_RESOLVE_TIMEOUT_MS 10000
45 /* Do not merge data if larger than this. */
46 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
47 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
49 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
52 * This is only for non-live state being migrated.
53 * Instead of RDMA_WRITE messages, we use RDMA_SEND
54 * messages for that state, which requires a different
55 * delivery design than main memory.
57 #define RDMA_SEND_INCREMENT 32768
60 * Maximum size infiniband SEND message
62 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
63 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
65 #define RDMA_CONTROL_VERSION_CURRENT 1
67 * Capabilities for negotiation.
69 #define RDMA_CAPABILITY_PIN_ALL 0x01
72 * Add the other flags above to this list of known capabilities
73 * as they are introduced.
75 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
78 * A work request ID is 64-bits and we split up these bits
81 * bits 0-15 : type of control message, 2^16
82 * bits 16-29: ram block index, 2^14
83 * bits 30-63: ram block chunk number, 2^34
85 * The last two bit ranges are only used for RDMA writes,
86 * in order to track their completion and potentially
87 * also track unregistration status of the message.
89 #define RDMA_WRID_TYPE_SHIFT 0UL
90 #define RDMA_WRID_BLOCK_SHIFT 16UL
91 #define RDMA_WRID_CHUNK_SHIFT 30UL
93 #define RDMA_WRID_TYPE_MASK \
94 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
96 #define RDMA_WRID_BLOCK_MASK \
97 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
99 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
102 * RDMA migration protocol:
103 * 1. RDMA Writes (data messages, i.e. RAM)
104 * 2. IB Send/Recv (control channel messages)
108 RDMA_WRID_RDMA_WRITE
= 1,
109 RDMA_WRID_SEND_CONTROL
= 2000,
110 RDMA_WRID_RECV_CONTROL
= 4000,
114 * Work request IDs for IB SEND messages only (not RDMA writes).
115 * This is used by the migration protocol to transmit
116 * control messages (such as device state and registration commands)
118 * We could use more WRs, but we have enough for now.
128 * SEND/RECV IB Control Messages.
131 RDMA_CONTROL_NONE
= 0,
133 RDMA_CONTROL_READY
, /* ready to receive */
134 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
135 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
136 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
137 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
138 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
139 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
140 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
141 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
142 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
147 * Memory and MR structures used to represent an IB Send/Recv work request.
148 * This is *not* used for RDMA writes, only IB Send/Recv.
151 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
152 struct ibv_mr
*control_mr
; /* registration metadata */
153 size_t control_len
; /* length of the message */
154 uint8_t *control_curr
; /* start of unconsumed bytes */
155 } RDMAWorkRequestData
;
158 * Negotiate RDMA capabilities during connection-setup time.
165 static void caps_to_network(RDMACapabilities
*cap
)
167 cap
->version
= htonl(cap
->version
);
168 cap
->flags
= htonl(cap
->flags
);
171 static void network_to_caps(RDMACapabilities
*cap
)
173 cap
->version
= ntohl(cap
->version
);
174 cap
->flags
= ntohl(cap
->flags
);
178 * Representation of a RAMBlock from an RDMA perspective.
179 * This is not transmitted, only local.
180 * This and subsequent structures cannot be linked lists
181 * because we're using a single IB message to transmit
182 * the information. It's small anyway, so a list is overkill.
184 typedef struct RDMALocalBlock
{
186 uint8_t *local_host_addr
; /* local virtual address */
187 uint64_t remote_host_addr
; /* remote virtual address */
190 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
191 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
192 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
193 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
194 int index
; /* which block are we */
195 unsigned int src_index
; /* (Only used on dest) */
198 unsigned long *transit_bitmap
;
199 unsigned long *unregister_bitmap
;
203 * Also represents a RAMblock, but only on the dest.
204 * This gets transmitted by the dest during connection-time
205 * to the source VM and then is used to populate the
206 * corresponding RDMALocalBlock with
207 * the information needed to perform the actual RDMA.
209 typedef struct QEMU_PACKED RDMADestBlock
{
210 uint64_t remote_host_addr
;
213 uint32_t remote_rkey
;
217 static const char *control_desc(unsigned int rdma_control
)
219 static const char *strs
[] = {
220 [RDMA_CONTROL_NONE
] = "NONE",
221 [RDMA_CONTROL_ERROR
] = "ERROR",
222 [RDMA_CONTROL_READY
] = "READY",
223 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
224 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
225 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
226 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
227 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
228 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
229 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
230 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
231 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
234 if (rdma_control
> RDMA_CONTROL_UNREGISTER_FINISHED
) {
235 return "??BAD CONTROL VALUE??";
238 return strs
[rdma_control
];
241 static uint64_t htonll(uint64_t v
)
243 union { uint32_t lv
[2]; uint64_t llv
; } u
;
244 u
.lv
[0] = htonl(v
>> 32);
245 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
249 static uint64_t ntohll(uint64_t v
)
251 union { uint32_t lv
[2]; uint64_t llv
; } u
;
253 return ((uint64_t)ntohl(u
.lv
[0]) << 32) | (uint64_t) ntohl(u
.lv
[1]);
256 static void dest_block_to_network(RDMADestBlock
*db
)
258 db
->remote_host_addr
= htonll(db
->remote_host_addr
);
259 db
->offset
= htonll(db
->offset
);
260 db
->length
= htonll(db
->length
);
261 db
->remote_rkey
= htonl(db
->remote_rkey
);
264 static void network_to_dest_block(RDMADestBlock
*db
)
266 db
->remote_host_addr
= ntohll(db
->remote_host_addr
);
267 db
->offset
= ntohll(db
->offset
);
268 db
->length
= ntohll(db
->length
);
269 db
->remote_rkey
= ntohl(db
->remote_rkey
);
273 * Virtual address of the above structures used for transmitting
274 * the RAMBlock descriptions at connection-time.
275 * This structure is *not* transmitted.
277 typedef struct RDMALocalBlocks
{
279 bool init
; /* main memory init complete */
280 RDMALocalBlock
*block
;
284 * Main data structure for RDMA state.
285 * While there is only one copy of this structure being allocated right now,
286 * this is the place where one would start if you wanted to consider
287 * having more than one RDMA connection open at the same time.
289 typedef struct RDMAContext
{
294 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
297 * This is used by *_exchange_send() to figure out whether or not
298 * the initial "READY" message has already been received or not.
299 * This is because other functions may potentially poll() and detect
300 * the READY message before send() does, in which case we need to
301 * know if it completed.
303 int control_ready_expected
;
305 /* number of outstanding writes */
308 /* store info about current buffer so that we can
309 merge it with future sends */
310 uint64_t current_addr
;
311 uint64_t current_length
;
312 /* index of ram block the current buffer belongs to */
314 /* index of the chunk in the current ram block */
320 * infiniband-specific variables for opening the device
321 * and maintaining connection state and so forth.
323 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
324 * cm_id->verbs, cm_id->channel, and cm_id->qp.
326 struct rdma_cm_id
*cm_id
; /* connection manager ID */
327 struct rdma_cm_id
*listen_id
;
330 struct ibv_context
*verbs
;
331 struct rdma_event_channel
*channel
;
332 struct ibv_qp
*qp
; /* queue pair */
333 struct ibv_comp_channel
*recv_comp_channel
; /* recv completion channel */
334 struct ibv_comp_channel
*send_comp_channel
; /* send completion channel */
335 struct ibv_pd
*pd
; /* protection domain */
336 struct ibv_cq
*recv_cq
; /* recvieve completion queue */
337 struct ibv_cq
*send_cq
; /* send completion queue */
340 * If a previous write failed (perhaps because of a failed
341 * memory registration, then do not attempt any future work
342 * and remember the error state.
349 * Description of ram blocks used throughout the code.
351 RDMALocalBlocks local_ram_blocks
;
352 RDMADestBlock
*dest_blocks
;
354 /* Index of the next RAMBlock received during block registration */
355 unsigned int next_src_index
;
358 * Migration on *destination* started.
359 * Then use coroutine yield function.
360 * Source runs in a thread, so we don't care.
362 int migration_started_on_destination
;
364 int total_registrations
;
367 int unregister_current
, unregister_next
;
368 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
370 GHashTable
*blockmap
;
372 /* the RDMAContext for return path */
373 struct RDMAContext
*return_path
;
377 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
378 OBJECT_DECLARE_SIMPLE_TYPE(QIOChannelRDMA
, QIO_CHANNEL_RDMA
)
382 struct QIOChannelRDMA
{
385 RDMAContext
*rdmaout
;
387 bool blocking
; /* XXX we don't actually honour this yet */
391 * Main structure for IB Send/Recv control messages.
392 * This gets prepended at the beginning of every Send/Recv.
394 typedef struct QEMU_PACKED
{
395 uint32_t len
; /* Total length of data portion */
396 uint32_t type
; /* which control command to perform */
397 uint32_t repeat
; /* number of commands in data portion of same type */
401 static void control_to_network(RDMAControlHeader
*control
)
403 control
->type
= htonl(control
->type
);
404 control
->len
= htonl(control
->len
);
405 control
->repeat
= htonl(control
->repeat
);
408 static void network_to_control(RDMAControlHeader
*control
)
410 control
->type
= ntohl(control
->type
);
411 control
->len
= ntohl(control
->len
);
412 control
->repeat
= ntohl(control
->repeat
);
416 * Register a single Chunk.
417 * Information sent by the source VM to inform the dest
418 * to register an single chunk of memory before we can perform
419 * the actual RDMA operation.
421 typedef struct QEMU_PACKED
{
423 uint64_t current_addr
; /* offset into the ram_addr_t space */
424 uint64_t chunk
; /* chunk to lookup if unregistering */
426 uint32_t current_index
; /* which ramblock the chunk belongs to */
428 uint64_t chunks
; /* how many sequential chunks to register */
431 static bool rdma_errored(RDMAContext
*rdma
)
433 if (rdma
->errored
&& !rdma
->error_reported
) {
434 error_report("RDMA is in an error state waiting migration"
436 rdma
->error_reported
= true;
438 return rdma
->errored
;
441 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
443 RDMALocalBlock
*local_block
;
444 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
446 if (local_block
->is_ram_block
) {
448 * current_addr as passed in is an address in the local ram_addr_t
449 * space, we need to translate this for the destination
451 reg
->key
.current_addr
-= local_block
->offset
;
452 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
454 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
455 reg
->current_index
= htonl(reg
->current_index
);
456 reg
->chunks
= htonll(reg
->chunks
);
459 static void network_to_register(RDMARegister
*reg
)
461 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
462 reg
->current_index
= ntohl(reg
->current_index
);
463 reg
->chunks
= ntohll(reg
->chunks
);
466 typedef struct QEMU_PACKED
{
467 uint32_t value
; /* if zero, we will madvise() */
468 uint32_t block_idx
; /* which ram block index */
469 uint64_t offset
; /* Address in remote ram_addr_t space */
470 uint64_t length
; /* length of the chunk */
473 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
475 comp
->value
= htonl(comp
->value
);
477 * comp->offset as passed in is an address in the local ram_addr_t
478 * space, we need to translate this for the destination
480 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
481 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
482 comp
->block_idx
= htonl(comp
->block_idx
);
483 comp
->offset
= htonll(comp
->offset
);
484 comp
->length
= htonll(comp
->length
);
487 static void network_to_compress(RDMACompress
*comp
)
489 comp
->value
= ntohl(comp
->value
);
490 comp
->block_idx
= ntohl(comp
->block_idx
);
491 comp
->offset
= ntohll(comp
->offset
);
492 comp
->length
= ntohll(comp
->length
);
496 * The result of the dest's memory registration produces an "rkey"
497 * which the source VM must reference in order to perform
498 * the RDMA operation.
500 typedef struct QEMU_PACKED
{
504 } RDMARegisterResult
;
506 static void result_to_network(RDMARegisterResult
*result
)
508 result
->rkey
= htonl(result
->rkey
);
509 result
->host_addr
= htonll(result
->host_addr
);
512 static void network_to_result(RDMARegisterResult
*result
)
514 result
->rkey
= ntohl(result
->rkey
);
515 result
->host_addr
= ntohll(result
->host_addr
);
518 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
519 uint8_t *data
, RDMAControlHeader
*resp
,
521 int (*callback
)(RDMAContext
*rdma
,
525 static inline uint64_t ram_chunk_index(const uint8_t *start
,
528 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
531 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
534 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
535 (i
<< RDMA_REG_CHUNK_SHIFT
));
538 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
541 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
542 (1UL << RDMA_REG_CHUNK_SHIFT
);
544 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
545 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
551 static void rdma_add_block(RDMAContext
*rdma
, const char *block_name
,
553 ram_addr_t block_offset
, uint64_t length
)
555 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
556 RDMALocalBlock
*block
;
557 RDMALocalBlock
*old
= local
->block
;
559 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
+ 1);
561 if (local
->nb_blocks
) {
564 if (rdma
->blockmap
) {
565 for (x
= 0; x
< local
->nb_blocks
; x
++) {
566 g_hash_table_remove(rdma
->blockmap
,
567 (void *)(uintptr_t)old
[x
].offset
);
568 g_hash_table_insert(rdma
->blockmap
,
569 (void *)(uintptr_t)old
[x
].offset
,
573 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
577 block
= &local
->block
[local
->nb_blocks
];
579 block
->block_name
= g_strdup(block_name
);
580 block
->local_host_addr
= host_addr
;
581 block
->offset
= block_offset
;
582 block
->length
= length
;
583 block
->index
= local
->nb_blocks
;
584 block
->src_index
= ~0U; /* Filled in by the receipt of the block list */
585 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
586 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
587 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
588 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
589 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
590 block
->remote_keys
= g_new0(uint32_t, block
->nb_chunks
);
592 block
->is_ram_block
= local
->init
? false : true;
594 if (rdma
->blockmap
) {
595 g_hash_table_insert(rdma
->blockmap
, (void *)(uintptr_t)block_offset
, block
);
598 trace_rdma_add_block(block_name
, local
->nb_blocks
,
599 (uintptr_t) block
->local_host_addr
,
600 block
->offset
, block
->length
,
601 (uintptr_t) (block
->local_host_addr
+ block
->length
),
602 BITS_TO_LONGS(block
->nb_chunks
) *
603 sizeof(unsigned long) * 8,
610 * Memory regions need to be registered with the device and queue pairs setup
611 * in advanced before the migration starts. This tells us where the RAM blocks
612 * are so that we can register them individually.
614 static int qemu_rdma_init_one_block(RAMBlock
*rb
, void *opaque
)
616 const char *block_name
= qemu_ram_get_idstr(rb
);
617 void *host_addr
= qemu_ram_get_host_addr(rb
);
618 ram_addr_t block_offset
= qemu_ram_get_offset(rb
);
619 ram_addr_t length
= qemu_ram_get_used_length(rb
);
620 rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
625 * Identify the RAMBlocks and their quantity. They will be references to
626 * identify chunk boundaries inside each RAMBlock and also be referenced
627 * during dynamic page registration.
629 static void qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
631 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
634 assert(rdma
->blockmap
== NULL
);
635 memset(local
, 0, sizeof *local
);
636 ret
= foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
638 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
639 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
640 rdma
->local_ram_blocks
.nb_blocks
);
645 * Note: If used outside of cleanup, the caller must ensure that the destination
646 * block structures are also updated
648 static void rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
650 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
651 RDMALocalBlock
*old
= local
->block
;
654 if (rdma
->blockmap
) {
655 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
660 for (j
= 0; j
< block
->nb_chunks
; j
++) {
661 if (!block
->pmr
[j
]) {
664 ibv_dereg_mr(block
->pmr
[j
]);
665 rdma
->total_registrations
--;
672 ibv_dereg_mr(block
->mr
);
673 rdma
->total_registrations
--;
677 g_free(block
->transit_bitmap
);
678 block
->transit_bitmap
= NULL
;
680 g_free(block
->unregister_bitmap
);
681 block
->unregister_bitmap
= NULL
;
683 g_free(block
->remote_keys
);
684 block
->remote_keys
= NULL
;
686 g_free(block
->block_name
);
687 block
->block_name
= NULL
;
689 if (rdma
->blockmap
) {
690 for (x
= 0; x
< local
->nb_blocks
; x
++) {
691 g_hash_table_remove(rdma
->blockmap
,
692 (void *)(uintptr_t)old
[x
].offset
);
696 if (local
->nb_blocks
> 1) {
698 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
701 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
704 if (block
->index
< (local
->nb_blocks
- 1)) {
705 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
706 sizeof(RDMALocalBlock
) *
707 (local
->nb_blocks
- (block
->index
+ 1)));
708 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
709 local
->block
[x
].index
--;
713 assert(block
== local
->block
);
717 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
718 block
->offset
, block
->length
,
719 (uintptr_t)(block
->local_host_addr
+ block
->length
),
720 BITS_TO_LONGS(block
->nb_chunks
) *
721 sizeof(unsigned long) * 8, block
->nb_chunks
);
727 if (local
->nb_blocks
&& rdma
->blockmap
) {
728 for (x
= 0; x
< local
->nb_blocks
; x
++) {
729 g_hash_table_insert(rdma
->blockmap
,
730 (void *)(uintptr_t)local
->block
[x
].offset
,
737 * Trace RDMA device open, with device details.
739 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
741 struct ibv_port_attr port
;
743 if (ibv_query_port(verbs
, 1, &port
)) {
744 trace_qemu_rdma_dump_id_failed(who
);
748 trace_qemu_rdma_dump_id(who
,
750 verbs
->device
->dev_name
,
751 verbs
->device
->dev_path
,
752 verbs
->device
->ibdev_path
,
754 port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
? "Infiniband"
755 : port
.link_layer
== IBV_LINK_LAYER_ETHERNET
? "Ethernet"
760 * Trace RDMA gid addressing information.
761 * Useful for understanding the RDMA device hierarchy in the kernel.
763 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
767 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
768 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
769 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
773 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
774 * We will try the next addrinfo struct, and fail if there are
775 * no other valid addresses to bind against.
777 * If user is listening on '[::]', then we will not have a opened a device
778 * yet and have no way of verifying if the device is RoCE or not.
780 * In this case, the source VM will throw an error for ALL types of
781 * connections (both IPv4 and IPv6) if the destination machine does not have
782 * a regular infiniband network available for use.
784 * The only way to guarantee that an error is thrown for broken kernels is
785 * for the management software to choose a *specific* interface at bind time
786 * and validate what time of hardware it is.
788 * Unfortunately, this puts the user in a fix:
790 * If the source VM connects with an IPv4 address without knowing that the
791 * destination has bound to '[::]' the migration will unconditionally fail
792 * unless the management software is explicitly listening on the IPv4
793 * address while using a RoCE-based device.
795 * If the source VM connects with an IPv6 address, then we're OK because we can
796 * throw an error on the source (and similarly on the destination).
798 * But in mixed environments, this will be broken for a while until it is fixed
801 * We do provide a *tiny* bit of help in this function: We can list all of the
802 * devices in the system and check to see if all the devices are RoCE or
805 * If we detect that we have a *pure* RoCE environment, then we can safely
806 * thrown an error even if the management software has specified '[::]' as the
809 * However, if there is are multiple hetergeneous devices, then we cannot make
810 * this assumption and the user just has to be sure they know what they are
813 * Patches are being reviewed on linux-rdma.
815 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
817 /* This bug only exists in linux, to our knowledge. */
819 struct ibv_port_attr port_attr
;
822 * Verbs are only NULL if management has bound to '[::]'.
824 * Let's iterate through all the devices and see if there any pure IB
825 * devices (non-ethernet).
827 * If not, then we can safely proceed with the migration.
828 * Otherwise, there are no guarantees until the bug is fixed in linux.
832 struct ibv_device
**dev_list
= ibv_get_device_list(&num_devices
);
833 bool roce_found
= false;
834 bool ib_found
= false;
836 for (x
= 0; x
< num_devices
; x
++) {
837 verbs
= ibv_open_device(dev_list
[x
]);
839 * ibv_open_device() is not documented to set errno. If
840 * it does, it's somebody else's doc bug. If it doesn't,
841 * the use of errno below is wrong.
842 * TODO Find out whether ibv_open_device() sets errno.
845 if (errno
== EPERM
) {
848 error_setg_errno(errp
, errno
,
849 "could not open RDMA device context");
854 if (ibv_query_port(verbs
, 1, &port_attr
)) {
855 ibv_close_device(verbs
);
857 "RDMA ERROR: Could not query initial IB port");
861 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
863 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
867 ibv_close_device(verbs
);
873 warn_report("migrations may fail:"
874 " IPv6 over RoCE / iWARP in linux"
875 " is broken. But since you appear to have a"
876 " mixed RoCE / IB environment, be sure to only"
877 " migrate over the IB fabric until the kernel "
880 error_setg(errp
, "RDMA ERROR: "
881 "You only have RoCE / iWARP devices in your systems"
882 " and your management software has specified '[::]'"
883 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
892 * If we have a verbs context, that means that some other than '[::]' was
893 * used by the management software for binding. In which case we can
894 * actually warn the user about a potentially broken kernel.
897 /* IB ports start with 1, not 0 */
898 if (ibv_query_port(verbs
, 1, &port_attr
)) {
899 error_setg(errp
, "RDMA ERROR: Could not query initial IB port");
903 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
904 error_setg(errp
, "RDMA ERROR: "
905 "Linux kernel's RoCE / iWARP does not support IPv6 "
906 "(but patches on linux-rdma in progress)");
916 * Figure out which RDMA device corresponds to the requested IP hostname
917 * Also create the initial connection manager identifiers for opening
920 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
924 struct rdma_addrinfo
*res
;
926 struct rdma_cm_event
*cm_event
;
927 char ip
[40] = "unknown";
928 struct rdma_addrinfo
*e
;
930 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
931 error_setg(errp
, "RDMA ERROR: RDMA hostname has not been set");
935 /* create CM channel */
936 rdma
->channel
= rdma_create_event_channel();
937 if (!rdma
->channel
) {
938 error_setg(errp
, "RDMA ERROR: could not create CM channel");
943 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
945 error_setg(errp
, "RDMA ERROR: could not create channel id");
946 goto err_resolve_create_id
;
949 snprintf(port_str
, 16, "%d", rdma
->port
);
952 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
954 error_setg(errp
, "RDMA ERROR: could not rdma_getaddrinfo address %s",
956 goto err_resolve_get_addr
;
959 /* Try all addresses, saving the first error in @err */
960 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
961 Error
**local_errp
= err
? NULL
: &err
;
963 inet_ntop(e
->ai_family
,
964 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
965 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
967 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
968 RDMA_RESOLVE_TIMEOUT_MS
);
970 if (e
->ai_family
== AF_INET6
) {
971 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
,
982 rdma_freeaddrinfo(res
);
984 error_propagate(errp
, err
);
986 error_setg(errp
, "RDMA ERROR: could not resolve address %s",
989 goto err_resolve_get_addr
;
992 rdma_freeaddrinfo(res
);
993 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
995 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
997 error_setg(errp
, "RDMA ERROR: could not perform event_addr_resolved");
998 goto err_resolve_get_addr
;
1001 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
1003 "RDMA ERROR: result not equal to event_addr_resolved %s",
1004 rdma_event_str(cm_event
->event
));
1005 rdma_ack_cm_event(cm_event
);
1006 goto err_resolve_get_addr
;
1008 rdma_ack_cm_event(cm_event
);
1011 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1013 error_setg(errp
, "RDMA ERROR: could not resolve rdma route");
1014 goto err_resolve_get_addr
;
1017 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1019 error_setg(errp
, "RDMA ERROR: could not perform event_route_resolved");
1020 goto err_resolve_get_addr
;
1022 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1023 error_setg(errp
, "RDMA ERROR: "
1024 "result not equal to event_route_resolved: %s",
1025 rdma_event_str(cm_event
->event
));
1026 rdma_ack_cm_event(cm_event
);
1027 goto err_resolve_get_addr
;
1029 rdma_ack_cm_event(cm_event
);
1030 rdma
->verbs
= rdma
->cm_id
->verbs
;
1031 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1032 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1035 err_resolve_get_addr
:
1036 rdma_destroy_id(rdma
->cm_id
);
1038 err_resolve_create_id
:
1039 rdma_destroy_event_channel(rdma
->channel
);
1040 rdma
->channel
= NULL
;
1045 * Create protection domain and completion queues
1047 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
, Error
**errp
)
1050 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1052 error_setg(errp
, "failed to allocate protection domain");
1056 /* create receive completion channel */
1057 rdma
->recv_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1058 if (!rdma
->recv_comp_channel
) {
1059 error_setg(errp
, "failed to allocate receive completion channel");
1060 goto err_alloc_pd_cq
;
1064 * Completion queue can be filled by read work requests.
1066 rdma
->recv_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1067 NULL
, rdma
->recv_comp_channel
, 0);
1068 if (!rdma
->recv_cq
) {
1069 error_setg(errp
, "failed to allocate receive completion queue");
1070 goto err_alloc_pd_cq
;
1073 /* create send completion channel */
1074 rdma
->send_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1075 if (!rdma
->send_comp_channel
) {
1076 error_setg(errp
, "failed to allocate send completion channel");
1077 goto err_alloc_pd_cq
;
1080 rdma
->send_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1081 NULL
, rdma
->send_comp_channel
, 0);
1082 if (!rdma
->send_cq
) {
1083 error_setg(errp
, "failed to allocate send completion queue");
1084 goto err_alloc_pd_cq
;
1091 ibv_dealloc_pd(rdma
->pd
);
1093 if (rdma
->recv_comp_channel
) {
1094 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
1096 if (rdma
->send_comp_channel
) {
1097 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
1099 if (rdma
->recv_cq
) {
1100 ibv_destroy_cq(rdma
->recv_cq
);
1101 rdma
->recv_cq
= NULL
;
1104 rdma
->recv_comp_channel
= NULL
;
1105 rdma
->send_comp_channel
= NULL
;
1111 * Create queue pairs.
1113 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1115 struct ibv_qp_init_attr attr
= { 0 };
1118 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1119 attr
.cap
.max_recv_wr
= 3;
1120 attr
.cap
.max_send_sge
= 1;
1121 attr
.cap
.max_recv_sge
= 1;
1122 attr
.send_cq
= rdma
->send_cq
;
1123 attr
.recv_cq
= rdma
->recv_cq
;
1124 attr
.qp_type
= IBV_QPT_RC
;
1126 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1131 rdma
->qp
= rdma
->cm_id
->qp
;
1135 /* Check whether On-Demand Paging is supported by RDAM device */
1136 static bool rdma_support_odp(struct ibv_context
*dev
)
1138 struct ibv_device_attr_ex attr
= {0};
1139 int ret
= ibv_query_device_ex(dev
, NULL
, &attr
);
1144 if (attr
.odp_caps
.general_caps
& IBV_ODP_SUPPORT
) {
1152 * ibv_advise_mr to avoid RNR NAK error as far as possible.
1153 * The responder mr registering with ODP will sent RNR NAK back to
1154 * the requester in the face of the page fault.
1156 static void qemu_rdma_advise_prefetch_mr(struct ibv_pd
*pd
, uint64_t addr
,
1157 uint32_t len
, uint32_t lkey
,
1158 const char *name
, bool wr
)
1160 #ifdef HAVE_IBV_ADVISE_MR
1162 int advice
= wr
? IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE
:
1163 IBV_ADVISE_MR_ADVICE_PREFETCH
;
1164 struct ibv_sge sg_list
= {.lkey
= lkey
, .addr
= addr
, .length
= len
};
1166 ret
= ibv_advise_mr(pd
, advice
,
1167 IBV_ADVISE_MR_FLAG_FLUSH
, &sg_list
, 1);
1168 /* ignore the error */
1169 trace_qemu_rdma_advise_mr(name
, len
, addr
, strerror(ret
));
1173 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
, Error
**errp
)
1176 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1178 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1179 int access
= IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
;
1181 local
->block
[i
].mr
=
1182 ibv_reg_mr(rdma
->pd
,
1183 local
->block
[i
].local_host_addr
,
1184 local
->block
[i
].length
, access
1187 * ibv_reg_mr() is not documented to set errno. If it does,
1188 * it's somebody else's doc bug. If it doesn't, the use of
1189 * errno below is wrong.
1190 * TODO Find out whether ibv_reg_mr() sets errno.
1192 if (!local
->block
[i
].mr
&&
1193 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1194 access
|= IBV_ACCESS_ON_DEMAND
;
1195 /* register ODP mr */
1196 local
->block
[i
].mr
=
1197 ibv_reg_mr(rdma
->pd
,
1198 local
->block
[i
].local_host_addr
,
1199 local
->block
[i
].length
, access
);
1200 trace_qemu_rdma_register_odp_mr(local
->block
[i
].block_name
);
1202 if (local
->block
[i
].mr
) {
1203 qemu_rdma_advise_prefetch_mr(rdma
->pd
,
1204 (uintptr_t)local
->block
[i
].local_host_addr
,
1205 local
->block
[i
].length
,
1206 local
->block
[i
].mr
->lkey
,
1207 local
->block
[i
].block_name
,
1212 if (!local
->block
[i
].mr
) {
1213 error_setg_errno(errp
, errno
,
1214 "Failed to register local dest ram block!");
1217 rdma
->total_registrations
++;
1223 for (i
--; i
>= 0; i
--) {
1224 ibv_dereg_mr(local
->block
[i
].mr
);
1225 local
->block
[i
].mr
= NULL
;
1226 rdma
->total_registrations
--;
1234 * Find the ram block that corresponds to the page requested to be
1235 * transmitted by QEMU.
1237 * Once the block is found, also identify which 'chunk' within that
1238 * block that the page belongs to.
1240 static void qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1241 uintptr_t block_offset
,
1244 uint64_t *block_index
,
1245 uint64_t *chunk_index
)
1247 uint64_t current_addr
= block_offset
+ offset
;
1248 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1249 (void *) block_offset
);
1251 assert(current_addr
>= block
->offset
);
1252 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1254 *block_index
= block
->index
;
1255 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1256 block
->local_host_addr
+ (current_addr
- block
->offset
));
1260 * Register a chunk with IB. If the chunk was already registered
1261 * previously, then skip.
1263 * Also return the keys associated with the registration needed
1264 * to perform the actual RDMA operation.
1266 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1267 RDMALocalBlock
*block
, uintptr_t host_addr
,
1268 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1269 uint8_t *chunk_start
, uint8_t *chunk_end
)
1273 *lkey
= block
->mr
->lkey
;
1276 *rkey
= block
->mr
->rkey
;
1281 /* allocate memory to store chunk MRs */
1283 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1287 * If 'rkey', then we're the destination, so grant access to the source.
1289 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1291 if (!block
->pmr
[chunk
]) {
1292 uint64_t len
= chunk_end
- chunk_start
;
1293 int access
= rkey
? IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
:
1296 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1298 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1300 * ibv_reg_mr() is not documented to set errno. If it does,
1301 * it's somebody else's doc bug. If it doesn't, the use of
1302 * errno below is wrong.
1303 * TODO Find out whether ibv_reg_mr() sets errno.
1305 if (!block
->pmr
[chunk
] &&
1306 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1307 access
|= IBV_ACCESS_ON_DEMAND
;
1308 /* register ODP mr */
1309 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1310 trace_qemu_rdma_register_odp_mr(block
->block_name
);
1312 if (block
->pmr
[chunk
]) {
1313 qemu_rdma_advise_prefetch_mr(rdma
->pd
, (uintptr_t)chunk_start
,
1314 len
, block
->pmr
[chunk
]->lkey
,
1315 block
->block_name
, rkey
);
1320 if (!block
->pmr
[chunk
]) {
1323 rdma
->total_registrations
++;
1326 *lkey
= block
->pmr
[chunk
]->lkey
;
1329 *rkey
= block
->pmr
[chunk
]->rkey
;
1335 * Register (at connection time) the memory used for control
1338 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1340 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1341 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1342 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1343 if (rdma
->wr_data
[idx
].control_mr
) {
1344 rdma
->total_registrations
++;
1351 * Perform a non-optimized memory unregistration after every transfer
1352 * for demonstration purposes, only if pin-all is not requested.
1354 * Potential optimizations:
1355 * 1. Start a new thread to run this function continuously
1357 - and for receipt of unregister messages
1359 * 3. Use workload hints.
1361 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1365 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1367 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1369 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1371 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1372 RDMALocalBlock
*block
=
1373 &(rdma
->local_ram_blocks
.block
[index
]);
1374 RDMARegister reg
= { .current_index
= index
};
1375 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1377 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1378 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1382 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1383 rdma
->unregister_current
);
1385 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1386 rdma
->unregister_current
++;
1388 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1389 rdma
->unregister_current
= 0;
1394 * Unregistration is speculative (because migration is single-threaded
1395 * and we cannot break the protocol's inifinband message ordering).
1396 * Thus, if the memory is currently being used for transmission,
1397 * then abort the attempt to unregister and try again
1398 * later the next time a completion is received for this memory.
1400 clear_bit(chunk
, block
->unregister_bitmap
);
1402 if (test_bit(chunk
, block
->transit_bitmap
)) {
1403 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1407 trace_qemu_rdma_unregister_waiting_send(chunk
);
1409 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1410 block
->pmr
[chunk
] = NULL
;
1411 block
->remote_keys
[chunk
] = 0;
1414 error_report("unregistration chunk failed: %s",
1418 rdma
->total_registrations
--;
1420 reg
.key
.chunk
= chunk
;
1421 register_to_network(rdma
, ®
);
1422 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1423 &resp
, NULL
, NULL
, &err
);
1425 error_report_err(err
);
1429 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1435 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1438 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1440 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1441 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1447 * Consult the connection manager to see a work request
1448 * (of any kind) has completed.
1449 * Return the work request ID that completed.
1451 static int qemu_rdma_poll(RDMAContext
*rdma
, struct ibv_cq
*cq
,
1452 uint64_t *wr_id_out
, uint32_t *byte_len
)
1458 ret
= ibv_poll_cq(cq
, 1, &wc
);
1461 *wr_id_out
= RDMA_WRID_NONE
;
1469 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1471 if (wc
.status
!= IBV_WC_SUCCESS
) {
1475 if (rdma
->control_ready_expected
&&
1476 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1477 trace_qemu_rdma_poll_recv(wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
,
1479 rdma
->control_ready_expected
= 0;
1482 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1484 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1486 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1487 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1489 trace_qemu_rdma_poll_write(wr_id
, rdma
->nb_sent
,
1490 index
, chunk
, block
->local_host_addr
,
1491 (void *)(uintptr_t)block
->remote_host_addr
);
1493 clear_bit(chunk
, block
->transit_bitmap
);
1495 if (rdma
->nb_sent
> 0) {
1499 trace_qemu_rdma_poll_other(wr_id
, rdma
->nb_sent
);
1502 *wr_id_out
= wc
.wr_id
;
1504 *byte_len
= wc
.byte_len
;
1510 /* Wait for activity on the completion channel.
1511 * Returns 0 on success, none-0 on error.
1513 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
,
1514 struct ibv_comp_channel
*comp_channel
)
1516 struct rdma_cm_event
*cm_event
;
1520 * Coroutine doesn't start until migration_fd_process_incoming()
1521 * so don't yield unless we know we're running inside of a coroutine.
1523 if (rdma
->migration_started_on_destination
&&
1524 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1525 yield_until_fd_readable(comp_channel
->fd
);
1527 /* This is the source side, we're in a separate thread
1528 * or destination prior to migration_fd_process_incoming()
1529 * after postcopy, the destination also in a separate thread.
1530 * we can't yield; so we have to poll the fd.
1531 * But we need to be able to handle 'cancel' or an error
1532 * without hanging forever.
1534 while (!rdma
->errored
&& !rdma
->received_error
) {
1536 pfds
[0].fd
= comp_channel
->fd
;
1537 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1538 pfds
[0].revents
= 0;
1540 pfds
[1].fd
= rdma
->channel
->fd
;
1541 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1542 pfds
[1].revents
= 0;
1544 /* 0.1s timeout, should be fine for a 'cancel' */
1545 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1547 case 1: /* fd active */
1548 if (pfds
[0].revents
) {
1552 if (pfds
[1].revents
) {
1553 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1558 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1559 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1560 rdma_ack_cm_event(cm_event
);
1563 rdma_ack_cm_event(cm_event
);
1567 case 0: /* Timeout, go around again */
1570 default: /* Error of some type -
1571 * I don't trust errno from qemu_poll_ns
1576 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1577 /* Bail out and let the cancellation happen */
1583 if (rdma
->received_error
) {
1586 return -rdma
->errored
;
1589 static struct ibv_comp_channel
*to_channel(RDMAContext
*rdma
, uint64_t wrid
)
1591 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_comp_channel
:
1592 rdma
->recv_comp_channel
;
1595 static struct ibv_cq
*to_cq(RDMAContext
*rdma
, uint64_t wrid
)
1597 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_cq
: rdma
->recv_cq
;
1601 * Block until the next work request has completed.
1603 * First poll to see if a work request has already completed,
1606 * If we encounter completed work requests for IDs other than
1607 * the one we're interested in, then that's generally an error.
1609 * The only exception is actual RDMA Write completions. These
1610 * completions only need to be recorded, but do not actually
1611 * need further processing.
1613 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
,
1614 uint64_t wrid_requested
,
1617 int num_cq_events
= 0, ret
;
1620 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1621 struct ibv_comp_channel
*ch
= to_channel(rdma
, wrid_requested
);
1622 struct ibv_cq
*poll_cq
= to_cq(rdma
, wrid_requested
);
1624 if (ibv_req_notify_cq(poll_cq
, 0)) {
1628 while (wr_id
!= wrid_requested
) {
1629 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1634 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1636 if (wr_id
== RDMA_WRID_NONE
) {
1639 if (wr_id
!= wrid_requested
) {
1640 trace_qemu_rdma_block_for_wrid_miss(wrid_requested
, wr_id
);
1644 if (wr_id
== wrid_requested
) {
1649 ret
= qemu_rdma_wait_comp_channel(rdma
, ch
);
1651 goto err_block_for_wrid
;
1654 ret
= ibv_get_cq_event(ch
, &cq
, &cq_ctx
);
1656 goto err_block_for_wrid
;
1661 if (ibv_req_notify_cq(cq
, 0)) {
1662 goto err_block_for_wrid
;
1665 while (wr_id
!= wrid_requested
) {
1666 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1668 goto err_block_for_wrid
;
1671 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1673 if (wr_id
== RDMA_WRID_NONE
) {
1676 if (wr_id
!= wrid_requested
) {
1677 trace_qemu_rdma_block_for_wrid_miss(wrid_requested
, wr_id
);
1681 if (wr_id
== wrid_requested
) {
1682 goto success_block_for_wrid
;
1686 success_block_for_wrid
:
1687 if (num_cq_events
) {
1688 ibv_ack_cq_events(cq
, num_cq_events
);
1693 if (num_cq_events
) {
1694 ibv_ack_cq_events(cq
, num_cq_events
);
1697 rdma
->errored
= true;
1702 * Post a SEND message work request for the control channel
1703 * containing some data and block until the post completes.
1705 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1706 RDMAControlHeader
*head
,
1710 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1711 struct ibv_send_wr
*bad_wr
;
1712 struct ibv_sge sge
= {
1713 .addr
= (uintptr_t)(wr
->control
),
1714 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1715 .lkey
= wr
->control_mr
->lkey
,
1717 struct ibv_send_wr send_wr
= {
1718 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1719 .opcode
= IBV_WR_SEND
,
1720 .send_flags
= IBV_SEND_SIGNALED
,
1725 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1728 * We don't actually need to do a memcpy() in here if we used
1729 * the "sge" properly, but since we're only sending control messages
1730 * (not RAM in a performance-critical path), then its OK for now.
1732 * The copy makes the RDMAControlHeader simpler to manipulate
1733 * for the time being.
1735 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1736 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1737 control_to_network((void *) wr
->control
);
1740 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1744 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1747 error_setg(errp
, "Failed to use post IB SEND for control");
1751 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1753 error_setg(errp
, "rdma migration: send polling control error");
1761 * Post a RECV work request in anticipation of some future receipt
1762 * of data on the control channel.
1764 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
,
1767 struct ibv_recv_wr
*bad_wr
;
1768 struct ibv_sge sge
= {
1769 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1770 .length
= RDMA_CONTROL_MAX_BUFFER
,
1771 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1774 struct ibv_recv_wr recv_wr
= {
1775 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1781 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1782 error_setg(errp
, "error posting control recv");
1790 * Block and wait for a RECV control channel message to arrive.
1792 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1793 RDMAControlHeader
*head
, uint32_t expecting
, int idx
,
1797 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1801 error_setg(errp
, "rdma migration: recv polling control error!");
1805 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1806 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1808 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1810 if (expecting
== RDMA_CONTROL_NONE
) {
1811 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1813 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1814 error_setg(errp
, "Was expecting a %s (%d) control message"
1815 ", but got: %s (%d), length: %d",
1816 control_desc(expecting
), expecting
,
1817 control_desc(head
->type
), head
->type
, head
->len
);
1818 if (head
->type
== RDMA_CONTROL_ERROR
) {
1819 rdma
->received_error
= true;
1823 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1824 error_setg(errp
, "too long length: %d", head
->len
);
1827 if (sizeof(*head
) + head
->len
!= byte_len
) {
1828 error_setg(errp
, "Malformed length: %d byte_len %d",
1829 head
->len
, byte_len
);
1837 * When a RECV work request has completed, the work request's
1838 * buffer is pointed at the header.
1840 * This will advance the pointer to the data portion
1841 * of the control message of the work request's buffer that
1842 * was populated after the work request finished.
1844 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1845 RDMAControlHeader
*head
)
1847 rdma
->wr_data
[idx
].control_len
= head
->len
;
1848 rdma
->wr_data
[idx
].control_curr
=
1849 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1853 * This is an 'atomic' high-level operation to deliver a single, unified
1854 * control-channel message.
1856 * Additionally, if the user is expecting some kind of reply to this message,
1857 * they can request a 'resp' response message be filled in by posting an
1858 * additional work request on behalf of the user and waiting for an additional
1861 * The extra (optional) response is used during registration to us from having
1862 * to perform an *additional* exchange of message just to provide a response by
1863 * instead piggy-backing on the acknowledgement.
1865 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1866 uint8_t *data
, RDMAControlHeader
*resp
,
1868 int (*callback
)(RDMAContext
*rdma
,
1875 * Wait until the dest is ready before attempting to deliver the message
1876 * by waiting for a READY message.
1878 if (rdma
->control_ready_expected
) {
1879 RDMAControlHeader resp_ignored
;
1881 ret
= qemu_rdma_exchange_get_response(rdma
, &resp_ignored
,
1883 RDMA_WRID_READY
, errp
);
1890 * If the user is expecting a response, post a WR in anticipation of it.
1893 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
, errp
);
1900 * Post a WR to replace the one we just consumed for the READY message.
1902 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
, errp
);
1908 * Deliver the control message that was requested.
1910 ret
= qemu_rdma_post_send_control(rdma
, data
, head
, errp
);
1917 * If we're expecting a response, block and wait for it.
1921 trace_qemu_rdma_exchange_send_issue_callback();
1922 ret
= callback(rdma
, errp
);
1928 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1929 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1930 resp
->type
, RDMA_WRID_DATA
,
1937 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1939 *resp_idx
= RDMA_WRID_DATA
;
1941 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1944 rdma
->control_ready_expected
= 1;
1950 * This is an 'atomic' high-level operation to receive a single, unified
1951 * control-channel message.
1953 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1954 uint32_t expecting
, Error
**errp
)
1956 RDMAControlHeader ready
= {
1958 .type
= RDMA_CONTROL_READY
,
1964 * Inform the source that we're ready to receive a message.
1966 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
, errp
);
1973 * Block and wait for the message.
1975 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1976 expecting
, RDMA_WRID_READY
, errp
);
1982 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1985 * Post a new RECV work request to replace the one we just consumed.
1987 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
, errp
);
1996 * Write an actual chunk of memory using RDMA.
1998 * If we're using dynamic registration on the dest-side, we have to
1999 * send a registration command first.
2001 static int qemu_rdma_write_one(RDMAContext
*rdma
,
2002 int current_index
, uint64_t current_addr
,
2003 uint64_t length
, Error
**errp
)
2006 struct ibv_send_wr send_wr
= { 0 };
2007 struct ibv_send_wr
*bad_wr
;
2008 int reg_result_idx
, ret
, count
= 0;
2009 uint64_t chunk
, chunks
;
2010 uint8_t *chunk_start
, *chunk_end
;
2011 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
2013 RDMARegisterResult
*reg_result
;
2014 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
2015 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
2016 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
2021 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
2022 (current_addr
- block
->offset
));
2023 sge
.length
= length
;
2025 chunk
= ram_chunk_index(block
->local_host_addr
,
2026 (uint8_t *)(uintptr_t)sge
.addr
);
2027 chunk_start
= ram_chunk_start(block
, chunk
);
2029 if (block
->is_ram_block
) {
2030 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2032 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2036 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2038 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2043 trace_qemu_rdma_write_one_top(chunks
+ 1,
2045 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2047 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2050 while (test_bit(chunk
, block
->transit_bitmap
)) {
2052 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2053 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2055 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2058 error_setg(errp
, "Failed to Wait for previous write to complete "
2059 "block %d chunk %" PRIu64
2060 " current %" PRIu64
" len %" PRIu64
" %d",
2061 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2066 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2067 if (!block
->remote_keys
[chunk
]) {
2069 * This chunk has not yet been registered, so first check to see
2070 * if the entire chunk is zero. If so, tell the other size to
2071 * memset() + madvise() the entire chunk without RDMA.
2074 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2075 RDMACompress comp
= {
2076 .offset
= current_addr
,
2078 .block_idx
= current_index
,
2082 head
.len
= sizeof(comp
);
2083 head
.type
= RDMA_CONTROL_COMPRESS
;
2085 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2086 current_index
, current_addr
);
2088 compress_to_network(rdma
, &comp
);
2089 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2090 (uint8_t *) &comp
, NULL
, NULL
, NULL
, errp
);
2097 * TODO: Here we are sending something, but we are not
2098 * accounting for anything transferred. The following is wrong:
2100 * stat64_add(&mig_stats.rdma_bytes, sge.length);
2102 * because we are using some kind of compression. I
2103 * would think that head.len would be the more similar
2104 * thing to a correct value.
2106 stat64_add(&mig_stats
.zero_pages
,
2107 sge
.length
/ qemu_target_page_size());
2112 * Otherwise, tell other side to register.
2114 reg
.current_index
= current_index
;
2115 if (block
->is_ram_block
) {
2116 reg
.key
.current_addr
= current_addr
;
2118 reg
.key
.chunk
= chunk
;
2120 reg
.chunks
= chunks
;
2122 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2125 register_to_network(rdma
, ®
);
2126 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2127 &resp
, ®_result_idx
, NULL
, errp
);
2132 /* try to overlap this single registration with the one we sent. */
2133 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2134 &sge
.lkey
, NULL
, chunk
,
2135 chunk_start
, chunk_end
)) {
2136 error_setg(errp
, "cannot get lkey");
2140 reg_result
= (RDMARegisterResult
*)
2141 rdma
->wr_data
[reg_result_idx
].control_curr
;
2143 network_to_result(reg_result
);
2145 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2146 reg_result
->rkey
, chunk
);
2148 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2149 block
->remote_host_addr
= reg_result
->host_addr
;
2151 /* already registered before */
2152 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2153 &sge
.lkey
, NULL
, chunk
,
2154 chunk_start
, chunk_end
)) {
2155 error_setg(errp
, "cannot get lkey!");
2160 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2162 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2164 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2165 &sge
.lkey
, NULL
, chunk
,
2166 chunk_start
, chunk_end
)) {
2167 error_setg(errp
, "cannot get lkey!");
2173 * Encode the ram block index and chunk within this wrid.
2174 * We will use this information at the time of completion
2175 * to figure out which bitmap to check against and then which
2176 * chunk in the bitmap to look for.
2178 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2179 current_index
, chunk
);
2181 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2182 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2183 send_wr
.sg_list
= &sge
;
2184 send_wr
.num_sge
= 1;
2185 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2186 (current_addr
- block
->offset
);
2188 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2192 * ibv_post_send() does not return negative error numbers,
2193 * per the specification they are positive - no idea why.
2195 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2197 if (ret
== ENOMEM
) {
2198 trace_qemu_rdma_write_one_queue_full();
2199 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2201 error_setg(errp
, "rdma migration: failed to make "
2202 "room in full send queue!");
2208 } else if (ret
> 0) {
2209 error_setg_errno(errp
, ret
,
2210 "rdma migration: post rdma write failed");
2214 set_bit(chunk
, block
->transit_bitmap
);
2215 stat64_add(&mig_stats
.normal_pages
, sge
.length
/ qemu_target_page_size());
2217 * We are adding to transferred the amount of data written, but no
2218 * overhead at all. I will asume that RDMA is magicaly and don't
2219 * need to transfer (at least) the addresses where it wants to
2220 * write the pages. Here it looks like it should be something
2222 * sizeof(send_wr) + sge.length
2223 * but this being RDMA, who knows.
2225 stat64_add(&mig_stats
.rdma_bytes
, sge
.length
);
2226 ram_transferred_add(sge
.length
);
2227 rdma
->total_writes
++;
2233 * Push out any unwritten RDMA operations.
2235 * We support sending out multiple chunks at the same time.
2236 * Not all of them need to get signaled in the completion queue.
2238 static int qemu_rdma_write_flush(RDMAContext
*rdma
, Error
**errp
)
2242 if (!rdma
->current_length
) {
2246 ret
= qemu_rdma_write_one(rdma
, rdma
->current_index
, rdma
->current_addr
,
2247 rdma
->current_length
, errp
);
2255 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2258 rdma
->current_length
= 0;
2259 rdma
->current_addr
= 0;
2264 static inline bool qemu_rdma_buffer_mergeable(RDMAContext
*rdma
,
2265 uint64_t offset
, uint64_t len
)
2267 RDMALocalBlock
*block
;
2271 if (rdma
->current_index
< 0) {
2275 if (rdma
->current_chunk
< 0) {
2279 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2280 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2281 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2283 if (rdma
->current_length
== 0) {
2288 * Only merge into chunk sequentially.
2290 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2294 if (offset
< block
->offset
) {
2298 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2302 if ((host_addr
+ len
) > chunk_end
) {
2310 * We're not actually writing here, but doing three things:
2312 * 1. Identify the chunk the buffer belongs to.
2313 * 2. If the chunk is full or the buffer doesn't belong to the current
2314 * chunk, then start a new chunk and flush() the old chunk.
2315 * 3. To keep the hardware busy, we also group chunks into batches
2316 * and only require that a batch gets acknowledged in the completion
2317 * queue instead of each individual chunk.
2319 static int qemu_rdma_write(RDMAContext
*rdma
,
2320 uint64_t block_offset
, uint64_t offset
,
2321 uint64_t len
, Error
**errp
)
2323 uint64_t current_addr
= block_offset
+ offset
;
2324 uint64_t index
= rdma
->current_index
;
2325 uint64_t chunk
= rdma
->current_chunk
;
2328 /* If we cannot merge it, we flush the current buffer first. */
2329 if (!qemu_rdma_buffer_mergeable(rdma
, current_addr
, len
)) {
2330 ret
= qemu_rdma_write_flush(rdma
, errp
);
2334 rdma
->current_length
= 0;
2335 rdma
->current_addr
= current_addr
;
2337 qemu_rdma_search_ram_block(rdma
, block_offset
,
2338 offset
, len
, &index
, &chunk
);
2339 rdma
->current_index
= index
;
2340 rdma
->current_chunk
= chunk
;
2344 rdma
->current_length
+= len
;
2346 /* flush it if buffer is too large */
2347 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2348 return qemu_rdma_write_flush(rdma
, errp
);
2354 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2358 if (rdma
->cm_id
&& rdma
->connected
) {
2359 if ((rdma
->errored
||
2360 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2361 !rdma
->received_error
) {
2362 RDMAControlHeader head
= { .len
= 0,
2363 .type
= RDMA_CONTROL_ERROR
,
2366 warn_report("Early error. Sending error.");
2367 if (qemu_rdma_post_send_control(rdma
, NULL
, &head
, &err
) < 0) {
2368 warn_report_err(err
);
2372 rdma_disconnect(rdma
->cm_id
);
2373 trace_qemu_rdma_cleanup_disconnect();
2374 rdma
->connected
= false;
2377 if (rdma
->channel
) {
2378 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2380 g_free(rdma
->dest_blocks
);
2381 rdma
->dest_blocks
= NULL
;
2383 for (int i
= 0; i
< RDMA_WRID_MAX
; i
++) {
2384 if (rdma
->wr_data
[i
].control_mr
) {
2385 rdma
->total_registrations
--;
2386 ibv_dereg_mr(rdma
->wr_data
[i
].control_mr
);
2388 rdma
->wr_data
[i
].control_mr
= NULL
;
2391 if (rdma
->local_ram_blocks
.block
) {
2392 while (rdma
->local_ram_blocks
.nb_blocks
) {
2393 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2398 rdma_destroy_qp(rdma
->cm_id
);
2401 if (rdma
->recv_cq
) {
2402 ibv_destroy_cq(rdma
->recv_cq
);
2403 rdma
->recv_cq
= NULL
;
2405 if (rdma
->send_cq
) {
2406 ibv_destroy_cq(rdma
->send_cq
);
2407 rdma
->send_cq
= NULL
;
2409 if (rdma
->recv_comp_channel
) {
2410 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
2411 rdma
->recv_comp_channel
= NULL
;
2413 if (rdma
->send_comp_channel
) {
2414 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
2415 rdma
->send_comp_channel
= NULL
;
2418 ibv_dealloc_pd(rdma
->pd
);
2422 rdma_destroy_id(rdma
->cm_id
);
2426 /* the destination side, listen_id and channel is shared */
2427 if (rdma
->listen_id
) {
2428 if (!rdma
->is_return_path
) {
2429 rdma_destroy_id(rdma
->listen_id
);
2431 rdma
->listen_id
= NULL
;
2433 if (rdma
->channel
) {
2434 if (!rdma
->is_return_path
) {
2435 rdma_destroy_event_channel(rdma
->channel
);
2437 rdma
->channel
= NULL
;
2441 if (rdma
->channel
) {
2442 rdma_destroy_event_channel(rdma
->channel
);
2443 rdma
->channel
= NULL
;
2446 g_free(rdma
->host_port
);
2448 rdma
->host_port
= NULL
;
2452 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2457 * Will be validated against destination's actual capabilities
2458 * after the connect() completes.
2460 rdma
->pin_all
= pin_all
;
2462 ret
= qemu_rdma_resolve_host(rdma
, errp
);
2464 goto err_rdma_source_init
;
2467 ret
= qemu_rdma_alloc_pd_cq(rdma
, errp
);
2469 goto err_rdma_source_init
;
2472 ret
= qemu_rdma_alloc_qp(rdma
);
2474 error_setg(errp
, "RDMA ERROR: rdma migration: error allocating qp!");
2475 goto err_rdma_source_init
;
2478 qemu_rdma_init_ram_blocks(rdma
);
2480 /* Build the hash that maps from offset to RAMBlock */
2481 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2482 for (int i
= 0; i
< rdma
->local_ram_blocks
.nb_blocks
; i
++) {
2483 g_hash_table_insert(rdma
->blockmap
,
2484 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[i
].offset
,
2485 &rdma
->local_ram_blocks
.block
[i
]);
2488 for (int i
= 0; i
< RDMA_WRID_MAX
; i
++) {
2489 ret
= qemu_rdma_reg_control(rdma
, i
);
2491 error_setg(errp
, "RDMA ERROR: rdma migration: error "
2492 "registering %d control!", i
);
2493 goto err_rdma_source_init
;
2499 err_rdma_source_init
:
2500 qemu_rdma_cleanup(rdma
);
2504 static int qemu_get_cm_event_timeout(RDMAContext
*rdma
,
2505 struct rdma_cm_event
**cm_event
,
2506 long msec
, Error
**errp
)
2509 struct pollfd poll_fd
= {
2510 .fd
= rdma
->channel
->fd
,
2516 ret
= poll(&poll_fd
, 1, msec
);
2517 } while (ret
< 0 && errno
== EINTR
);
2520 error_setg(errp
, "RDMA ERROR: poll cm event timeout");
2522 } else if (ret
< 0) {
2523 error_setg(errp
, "RDMA ERROR: failed to poll cm event, errno=%i",
2526 } else if (poll_fd
.revents
& POLLIN
) {
2527 if (rdma_get_cm_event(rdma
->channel
, cm_event
) < 0) {
2528 error_setg(errp
, "RDMA ERROR: failed to get cm event");
2533 error_setg(errp
, "RDMA ERROR: no POLLIN event, revent=%x",
2539 static int qemu_rdma_connect(RDMAContext
*rdma
, bool return_path
,
2542 RDMACapabilities cap
= {
2543 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2546 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2548 .private_data
= &cap
,
2549 .private_data_len
= sizeof(cap
),
2551 struct rdma_cm_event
*cm_event
;
2555 * Only negotiate the capability with destination if the user
2556 * on the source first requested the capability.
2558 if (rdma
->pin_all
) {
2559 trace_qemu_rdma_connect_pin_all_requested();
2560 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2563 caps_to_network(&cap
);
2565 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
, errp
);
2567 goto err_rdma_source_connect
;
2570 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2572 error_setg_errno(errp
, errno
,
2573 "RDMA ERROR: connecting to destination!");
2574 goto err_rdma_source_connect
;
2578 ret
= qemu_get_cm_event_timeout(rdma
, &cm_event
, 5000, errp
);
2580 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2582 error_setg_errno(errp
, errno
,
2583 "RDMA ERROR: failed to get cm event");
2587 goto err_rdma_source_connect
;
2590 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2591 error_setg(errp
, "RDMA ERROR: connecting to destination!");
2592 rdma_ack_cm_event(cm_event
);
2593 goto err_rdma_source_connect
;
2595 rdma
->connected
= true;
2597 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2598 network_to_caps(&cap
);
2601 * Verify that the *requested* capabilities are supported by the destination
2602 * and disable them otherwise.
2604 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2605 warn_report("RDMA: Server cannot support pinning all memory. "
2606 "Will register memory dynamically.");
2607 rdma
->pin_all
= false;
2610 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2612 rdma_ack_cm_event(cm_event
);
2614 rdma
->control_ready_expected
= 1;
2618 err_rdma_source_connect
:
2619 qemu_rdma_cleanup(rdma
);
2623 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2627 struct rdma_cm_id
*listen_id
;
2628 char ip
[40] = "unknown";
2629 struct rdma_addrinfo
*res
, *e
;
2633 for (int i
= 0; i
< RDMA_WRID_MAX
; i
++) {
2634 rdma
->wr_data
[i
].control_len
= 0;
2635 rdma
->wr_data
[i
].control_curr
= NULL
;
2638 if (!rdma
->host
|| !rdma
->host
[0]) {
2639 error_setg(errp
, "RDMA ERROR: RDMA host is not set!");
2640 rdma
->errored
= true;
2643 /* create CM channel */
2644 rdma
->channel
= rdma_create_event_channel();
2645 if (!rdma
->channel
) {
2646 error_setg(errp
, "RDMA ERROR: could not create rdma event channel");
2647 rdma
->errored
= true;
2652 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2654 error_setg(errp
, "RDMA ERROR: could not create cm_id!");
2655 goto err_dest_init_create_listen_id
;
2658 snprintf(port_str
, 16, "%d", rdma
->port
);
2659 port_str
[15] = '\0';
2661 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2663 error_setg(errp
, "RDMA ERROR: could not rdma_getaddrinfo address %s",
2665 goto err_dest_init_bind_addr
;
2668 ret
= rdma_set_option(listen_id
, RDMA_OPTION_ID
, RDMA_OPTION_ID_REUSEADDR
,
2669 &reuse
, sizeof reuse
);
2671 error_setg(errp
, "RDMA ERROR: Error: could not set REUSEADDR option");
2672 goto err_dest_init_bind_addr
;
2675 /* Try all addresses, saving the first error in @err */
2676 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2677 Error
**local_errp
= err
? NULL
: &err
;
2679 inet_ntop(e
->ai_family
,
2680 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2681 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2682 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2686 if (e
->ai_family
== AF_INET6
) {
2687 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
,
2697 rdma_freeaddrinfo(res
);
2700 error_propagate(errp
, err
);
2702 error_setg(errp
, "RDMA ERROR: Error: could not rdma_bind_addr!");
2704 goto err_dest_init_bind_addr
;
2707 rdma
->listen_id
= listen_id
;
2708 qemu_rdma_dump_gid("dest_init", listen_id
);
2711 err_dest_init_bind_addr
:
2712 rdma_destroy_id(listen_id
);
2713 err_dest_init_create_listen_id
:
2714 rdma_destroy_event_channel(rdma
->channel
);
2715 rdma
->channel
= NULL
;
2716 rdma
->errored
= true;
2721 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2724 for (int i
= 0; i
< RDMA_WRID_MAX
; i
++) {
2725 rdma_return_path
->wr_data
[i
].control_len
= 0;
2726 rdma_return_path
->wr_data
[i
].control_curr
= NULL
;
2729 /*the CM channel and CM id is shared*/
2730 rdma_return_path
->channel
= rdma
->channel
;
2731 rdma_return_path
->listen_id
= rdma
->listen_id
;
2733 rdma
->return_path
= rdma_return_path
;
2734 rdma_return_path
->return_path
= rdma
;
2735 rdma_return_path
->is_return_path
= true;
2738 static RDMAContext
*qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2740 RDMAContext
*rdma
= NULL
;
2741 InetSocketAddress
*addr
;
2743 rdma
= g_new0(RDMAContext
, 1);
2744 rdma
->current_index
= -1;
2745 rdma
->current_chunk
= -1;
2747 addr
= g_new(InetSocketAddress
, 1);
2748 if (!inet_parse(addr
, host_port
, NULL
)) {
2749 rdma
->port
= atoi(addr
->port
);
2750 rdma
->host
= g_strdup(addr
->host
);
2751 rdma
->host_port
= g_strdup(host_port
);
2753 error_setg(errp
, "RDMA ERROR: bad RDMA migration address '%s'",
2759 qapi_free_InetSocketAddress(addr
);
2764 * QEMUFile interface to the control channel.
2765 * SEND messages for control only.
2766 * VM's ram is handled with regular RDMA messages.
2768 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2769 const struct iovec
*iov
,
2776 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2782 RCU_READ_LOCK_GUARD();
2783 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
2786 error_setg(errp
, "RDMA control channel output is not set");
2790 if (rdma
->errored
) {
2792 "RDMA is in an error state waiting migration to abort!");
2797 * Push out any writes that
2798 * we're queued up for VM's ram.
2800 ret
= qemu_rdma_write_flush(rdma
, errp
);
2802 rdma
->errored
= true;
2806 for (i
= 0; i
< niov
; i
++) {
2807 size_t remaining
= iov
[i
].iov_len
;
2808 uint8_t * data
= (void *)iov
[i
].iov_base
;
2810 RDMAControlHeader head
= {};
2812 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2816 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2818 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2819 data
, NULL
, NULL
, NULL
, errp
);
2822 rdma
->errored
= true;
2834 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2835 size_t size
, int idx
)
2839 if (rdma
->wr_data
[idx
].control_len
) {
2840 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2842 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2843 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2844 rdma
->wr_data
[idx
].control_curr
+= len
;
2845 rdma
->wr_data
[idx
].control_len
-= len
;
2852 * QEMUFile interface to the control channel.
2853 * RDMA links don't use bytestreams, so we have to
2854 * return bytes to QEMUFile opportunistically.
2856 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2857 const struct iovec
*iov
,
2864 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2866 RDMAControlHeader head
;
2871 RCU_READ_LOCK_GUARD();
2872 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
2875 error_setg(errp
, "RDMA control channel input is not set");
2879 if (rdma
->errored
) {
2881 "RDMA is in an error state waiting migration to abort!");
2885 for (i
= 0; i
< niov
; i
++) {
2886 size_t want
= iov
[i
].iov_len
;
2887 uint8_t *data
= (void *)iov
[i
].iov_base
;
2890 * First, we hold on to the last SEND message we
2891 * were given and dish out the bytes until we run
2894 len
= qemu_rdma_fill(rdma
, data
, want
, 0);
2897 /* Got what we needed, so go to next iovec */
2902 /* If we got any data so far, then don't wait
2903 * for more, just return what we have */
2909 /* We've got nothing at all, so lets wait for
2912 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
,
2916 rdma
->errored
= true;
2921 * SEND was received with new bytes, now try again.
2923 len
= qemu_rdma_fill(rdma
, data
, want
, 0);
2927 /* Still didn't get enough, so lets just return */
2930 return QIO_CHANNEL_ERR_BLOCK
;
2940 * Block until all the outstanding chunks have been delivered by the hardware.
2942 static int qemu_rdma_drain_cq(RDMAContext
*rdma
)
2947 if (qemu_rdma_write_flush(rdma
, &err
) < 0) {
2948 error_report_err(err
);
2952 while (rdma
->nb_sent
) {
2953 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2955 error_report("rdma migration: complete polling error!");
2960 qemu_rdma_unregister_waiting(rdma
);
2966 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2970 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2971 /* XXX we should make readv/writev actually honour this :-) */
2972 rioc
->blocking
= blocking
;
2977 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2978 struct QIOChannelRDMASource
{
2980 QIOChannelRDMA
*rioc
;
2981 GIOCondition condition
;
2985 qio_channel_rdma_source_prepare(GSource
*source
,
2988 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2990 GIOCondition cond
= 0;
2993 RCU_READ_LOCK_GUARD();
2994 if (rsource
->condition
== G_IO_IN
) {
2995 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
2997 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3001 error_report("RDMAContext is NULL when prepare Gsource");
3005 if (rdma
->wr_data
[0].control_len
) {
3010 return cond
& rsource
->condition
;
3014 qio_channel_rdma_source_check(GSource
*source
)
3016 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3018 GIOCondition cond
= 0;
3020 RCU_READ_LOCK_GUARD();
3021 if (rsource
->condition
== G_IO_IN
) {
3022 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3024 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3028 error_report("RDMAContext is NULL when check Gsource");
3032 if (rdma
->wr_data
[0].control_len
) {
3037 return cond
& rsource
->condition
;
3041 qio_channel_rdma_source_dispatch(GSource
*source
,
3042 GSourceFunc callback
,
3045 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
3046 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3048 GIOCondition cond
= 0;
3050 RCU_READ_LOCK_GUARD();
3051 if (rsource
->condition
== G_IO_IN
) {
3052 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3054 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3058 error_report("RDMAContext is NULL when dispatch Gsource");
3062 if (rdma
->wr_data
[0].control_len
) {
3067 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
3068 (cond
& rsource
->condition
),
3073 qio_channel_rdma_source_finalize(GSource
*source
)
3075 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
3077 object_unref(OBJECT(ssource
->rioc
));
3080 static GSourceFuncs qio_channel_rdma_source_funcs
= {
3081 qio_channel_rdma_source_prepare
,
3082 qio_channel_rdma_source_check
,
3083 qio_channel_rdma_source_dispatch
,
3084 qio_channel_rdma_source_finalize
3087 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
3088 GIOCondition condition
)
3090 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3091 QIOChannelRDMASource
*ssource
;
3094 source
= g_source_new(&qio_channel_rdma_source_funcs
,
3095 sizeof(QIOChannelRDMASource
));
3096 ssource
= (QIOChannelRDMASource
*)source
;
3098 ssource
->rioc
= rioc
;
3099 object_ref(OBJECT(rioc
));
3101 ssource
->condition
= condition
;
3106 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
3107 AioContext
*read_ctx
,
3109 AioContext
*write_ctx
,
3110 IOHandler
*io_write
,
3113 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3115 aio_set_fd_handler(read_ctx
, rioc
->rdmain
->recv_comp_channel
->fd
,
3116 io_read
, io_write
, NULL
, NULL
, opaque
);
3117 aio_set_fd_handler(read_ctx
, rioc
->rdmain
->send_comp_channel
->fd
,
3118 io_read
, io_write
, NULL
, NULL
, opaque
);
3120 aio_set_fd_handler(write_ctx
, rioc
->rdmaout
->recv_comp_channel
->fd
,
3121 io_read
, io_write
, NULL
, NULL
, opaque
);
3122 aio_set_fd_handler(write_ctx
, rioc
->rdmaout
->send_comp_channel
->fd
,
3123 io_read
, io_write
, NULL
, NULL
, opaque
);
3127 struct rdma_close_rcu
{
3128 struct rcu_head rcu
;
3129 RDMAContext
*rdmain
;
3130 RDMAContext
*rdmaout
;
3133 /* callback from qio_channel_rdma_close via call_rcu */
3134 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu
*rcu
)
3137 qemu_rdma_cleanup(rcu
->rdmain
);
3141 qemu_rdma_cleanup(rcu
->rdmaout
);
3144 g_free(rcu
->rdmain
);
3145 g_free(rcu
->rdmaout
);
3149 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3152 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3153 RDMAContext
*rdmain
, *rdmaout
;
3154 struct rdma_close_rcu
*rcu
= g_new(struct rdma_close_rcu
, 1);
3156 trace_qemu_rdma_close();
3158 rdmain
= rioc
->rdmain
;
3160 qatomic_rcu_set(&rioc
->rdmain
, NULL
);
3163 rdmaout
= rioc
->rdmaout
;
3165 qatomic_rcu_set(&rioc
->rdmaout
, NULL
);
3168 rcu
->rdmain
= rdmain
;
3169 rcu
->rdmaout
= rdmaout
;
3170 call_rcu(rcu
, qio_channel_rdma_close_rcu
, rcu
);
3176 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3177 QIOChannelShutdown how
,
3180 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3181 RDMAContext
*rdmain
, *rdmaout
;
3183 RCU_READ_LOCK_GUARD();
3185 rdmain
= qatomic_rcu_read(&rioc
->rdmain
);
3186 rdmaout
= qatomic_rcu_read(&rioc
->rdmain
);
3189 case QIO_CHANNEL_SHUTDOWN_READ
:
3191 rdmain
->errored
= true;
3194 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3196 rdmaout
->errored
= true;
3199 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3202 rdmain
->errored
= true;
3205 rdmaout
->errored
= true;
3216 * This means that 'block_offset' is a full virtual address that does not
3217 * belong to a RAMBlock of the virtual machine and instead
3218 * represents a private malloc'd memory area that the caller wishes to
3222 * Offset is an offset to be added to block_offset and used
3223 * to also lookup the corresponding RAMBlock.
3225 * @size : Number of bytes to transfer
3227 * @pages_sent : User-specificed pointer to indicate how many pages were
3228 * sent. Usually, this will not be more than a few bytes of
3229 * the protocol because most transfers are sent asynchronously.
3231 static int qemu_rdma_save_page(QEMUFile
*f
, ram_addr_t block_offset
,
3232 ram_addr_t offset
, size_t size
)
3234 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3239 RCU_READ_LOCK_GUARD();
3240 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3246 if (rdma_errored(rdma
)) {
3253 * Add this page to the current 'chunk'. If the chunk
3254 * is full, or the page doesn't belong to the current chunk,
3255 * an actual RDMA write will occur and a new chunk will be formed.
3257 ret
= qemu_rdma_write(rdma
, block_offset
, offset
, size
, &err
);
3259 error_report_err(err
);
3264 * Drain the Completion Queue if possible, but do not block,
3267 * If nothing to poll, the end of the iteration will do this
3268 * again to make sure we don't overflow the request queue.
3271 uint64_t wr_id
, wr_id_in
;
3272 ret
= qemu_rdma_poll(rdma
, rdma
->recv_cq
, &wr_id_in
, NULL
);
3275 error_report("rdma migration: polling error");
3279 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3281 if (wr_id
== RDMA_WRID_NONE
) {
3287 uint64_t wr_id
, wr_id_in
;
3288 ret
= qemu_rdma_poll(rdma
, rdma
->send_cq
, &wr_id_in
, NULL
);
3291 error_report("rdma migration: polling error");
3295 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3297 if (wr_id
== RDMA_WRID_NONE
) {
3302 return RAM_SAVE_CONTROL_DELAYED
;
3305 rdma
->errored
= true;
3309 int rdma_control_save_page(QEMUFile
*f
, ram_addr_t block_offset
,
3310 ram_addr_t offset
, size_t size
)
3312 if (!migrate_rdma() || migration_in_postcopy()) {
3313 return RAM_SAVE_CONTROL_NOT_SUPP
;
3316 int ret
= qemu_rdma_save_page(f
, block_offset
, offset
, size
);
3318 if (ret
!= RAM_SAVE_CONTROL_DELAYED
&&
3319 ret
!= RAM_SAVE_CONTROL_NOT_SUPP
) {
3321 qemu_file_set_error(f
, ret
);
3327 static void rdma_accept_incoming_migration(void *opaque
);
3329 static void rdma_cm_poll_handler(void *opaque
)
3331 RDMAContext
*rdma
= opaque
;
3333 struct rdma_cm_event
*cm_event
;
3334 MigrationIncomingState
*mis
= migration_incoming_get_current();
3336 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3338 error_report("get_cm_event failed %d", errno
);
3342 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3343 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3344 if (!rdma
->errored
&&
3345 migration_incoming_get_current()->state
!=
3346 MIGRATION_STATUS_COMPLETED
) {
3347 error_report("receive cm event, cm event is %d", cm_event
->event
);
3348 rdma
->errored
= true;
3349 if (rdma
->return_path
) {
3350 rdma
->return_path
->errored
= true;
3353 rdma_ack_cm_event(cm_event
);
3354 if (mis
->loadvm_co
) {
3355 qemu_coroutine_enter(mis
->loadvm_co
);
3359 rdma_ack_cm_event(cm_event
);
3362 static int qemu_rdma_accept(RDMAContext
*rdma
)
3365 RDMACapabilities cap
;
3366 struct rdma_conn_param conn_param
= {
3367 .responder_resources
= 2,
3368 .private_data
= &cap
,
3369 .private_data_len
= sizeof(cap
),
3371 RDMAContext
*rdma_return_path
= NULL
;
3372 struct rdma_cm_event
*cm_event
;
3373 struct ibv_context
*verbs
;
3376 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3378 goto err_rdma_dest_wait
;
3381 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3382 rdma_ack_cm_event(cm_event
);
3383 goto err_rdma_dest_wait
;
3387 * initialize the RDMAContext for return path for postcopy after first
3388 * connection request reached.
3390 if ((migrate_postcopy() || migrate_return_path())
3391 && !rdma
->is_return_path
) {
3392 rdma_return_path
= qemu_rdma_data_init(rdma
->host_port
, NULL
);
3393 if (rdma_return_path
== NULL
) {
3394 rdma_ack_cm_event(cm_event
);
3395 goto err_rdma_dest_wait
;
3398 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
3401 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3403 network_to_caps(&cap
);
3405 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3406 error_report("Unknown source RDMA version: %d, bailing...",
3408 rdma_ack_cm_event(cm_event
);
3409 goto err_rdma_dest_wait
;
3413 * Respond with only the capabilities this version of QEMU knows about.
3415 cap
.flags
&= known_capabilities
;
3418 * Enable the ones that we do know about.
3419 * Add other checks here as new ones are introduced.
3421 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3422 rdma
->pin_all
= true;
3425 rdma
->cm_id
= cm_event
->id
;
3426 verbs
= cm_event
->id
->verbs
;
3428 rdma_ack_cm_event(cm_event
);
3430 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3432 caps_to_network(&cap
);
3434 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3437 rdma
->verbs
= verbs
;
3438 } else if (rdma
->verbs
!= verbs
) {
3439 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3441 goto err_rdma_dest_wait
;
3444 qemu_rdma_dump_id("dest_init", verbs
);
3446 ret
= qemu_rdma_alloc_pd_cq(rdma
, &err
);
3448 error_report_err(err
);
3449 goto err_rdma_dest_wait
;
3452 ret
= qemu_rdma_alloc_qp(rdma
);
3454 error_report("rdma migration: error allocating qp!");
3455 goto err_rdma_dest_wait
;
3458 qemu_rdma_init_ram_blocks(rdma
);
3460 for (int i
= 0; i
< RDMA_WRID_MAX
; i
++) {
3461 ret
= qemu_rdma_reg_control(rdma
, i
);
3463 error_report("rdma: error registering %d control", i
);
3464 goto err_rdma_dest_wait
;
3468 /* Accept the second connection request for return path */
3469 if ((migrate_postcopy() || migrate_return_path())
3470 && !rdma
->is_return_path
) {
3471 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3473 (void *)(intptr_t)rdma
->return_path
);
3475 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3479 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3481 error_report("rdma_accept failed");
3482 goto err_rdma_dest_wait
;
3485 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3487 error_report("rdma_accept get_cm_event failed");
3488 goto err_rdma_dest_wait
;
3491 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3492 error_report("rdma_accept not event established");
3493 rdma_ack_cm_event(cm_event
);
3494 goto err_rdma_dest_wait
;
3497 rdma_ack_cm_event(cm_event
);
3498 rdma
->connected
= true;
3500 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
, &err
);
3502 error_report_err(err
);
3503 goto err_rdma_dest_wait
;
3506 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3511 rdma
->errored
= true;
3512 qemu_rdma_cleanup(rdma
);
3513 g_free(rdma_return_path
);
3517 static int dest_ram_sort_func(const void *a
, const void *b
)
3519 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3520 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3522 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3526 * During each iteration of the migration, we listen for instructions
3527 * by the source VM to perform dynamic page registrations before they
3528 * can perform RDMA operations.
3530 * We respond with the 'rkey'.
3532 * Keep doing this until the source tells us to stop.
3534 int rdma_registration_handle(QEMUFile
*f
)
3536 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3537 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3540 RDMAControlHeader unreg_resp
= { .len
= 0,
3541 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3544 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3546 QIOChannelRDMA
*rioc
;
3549 RDMALocalBlocks
*local
;
3550 RDMAControlHeader head
;
3551 RDMARegister
*reg
, *registers
;
3553 RDMARegisterResult
*reg_result
;
3554 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3555 RDMALocalBlock
*block
;
3562 if (!migrate_rdma()) {
3566 RCU_READ_LOCK_GUARD();
3567 rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3568 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3574 if (rdma_errored(rdma
)) {
3578 local
= &rdma
->local_ram_blocks
;
3580 trace_rdma_registration_handle_wait();
3582 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
, &err
);
3585 error_report_err(err
);
3589 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3590 error_report("rdma: Too many requests in this message (%d)."
3591 "Bailing.", head
.repeat
);
3595 switch (head
.type
) {
3596 case RDMA_CONTROL_COMPRESS
:
3597 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3598 network_to_compress(comp
);
3600 trace_rdma_registration_handle_compress(comp
->length
,
3603 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3604 error_report("rdma: 'compress' bad block index %u (vs %d)",
3605 (unsigned int)comp
->block_idx
,
3606 rdma
->local_ram_blocks
.nb_blocks
);
3609 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3611 host_addr
= block
->local_host_addr
+
3612 (comp
->offset
- block
->offset
);
3614 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3617 case RDMA_CONTROL_REGISTER_FINISHED
:
3618 trace_rdma_registration_handle_finished();
3621 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3622 trace_rdma_registration_handle_ram_blocks();
3624 /* Sort our local RAM Block list so it's the same as the source,
3625 * we can do this since we've filled in a src_index in the list
3626 * as we received the RAMBlock list earlier.
3628 qsort(rdma
->local_ram_blocks
.block
,
3629 rdma
->local_ram_blocks
.nb_blocks
,
3630 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3631 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3632 local
->block
[i
].index
= i
;
3635 if (rdma
->pin_all
) {
3636 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
, &err
);
3638 error_report_err(err
);
3644 * Dest uses this to prepare to transmit the RAMBlock descriptions
3645 * to the source VM after connection setup.
3646 * Both sides use the "remote" structure to communicate and update
3647 * their "local" descriptions with what was sent.
3649 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3650 rdma
->dest_blocks
[i
].remote_host_addr
=
3651 (uintptr_t)(local
->block
[i
].local_host_addr
);
3653 if (rdma
->pin_all
) {
3654 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3657 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3658 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3660 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3661 trace_rdma_registration_handle_ram_blocks_loop(
3662 local
->block
[i
].block_name
,
3663 local
->block
[i
].offset
,
3664 local
->block
[i
].length
,
3665 local
->block
[i
].local_host_addr
,
3666 local
->block
[i
].src_index
);
3669 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3670 * sizeof(RDMADestBlock
);
3673 ret
= qemu_rdma_post_send_control(rdma
,
3674 (uint8_t *) rdma
->dest_blocks
, &blocks
,
3678 error_report_err(err
);
3683 case RDMA_CONTROL_REGISTER_REQUEST
:
3684 trace_rdma_registration_handle_register(head
.repeat
);
3686 reg_resp
.repeat
= head
.repeat
;
3687 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3689 for (count
= 0; count
< head
.repeat
; count
++) {
3691 uint8_t *chunk_start
, *chunk_end
;
3693 reg
= ®isters
[count
];
3694 network_to_register(reg
);
3696 reg_result
= &results
[count
];
3698 trace_rdma_registration_handle_register_loop(count
,
3699 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3701 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3702 error_report("rdma: 'register' bad block index %u (vs %d)",
3703 (unsigned int)reg
->current_index
,
3704 rdma
->local_ram_blocks
.nb_blocks
);
3707 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3708 if (block
->is_ram_block
) {
3709 if (block
->offset
> reg
->key
.current_addr
) {
3710 error_report("rdma: bad register address for block %s"
3711 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3712 block
->block_name
, block
->offset
,
3713 reg
->key
.current_addr
);
3716 host_addr
= (block
->local_host_addr
+
3717 (reg
->key
.current_addr
- block
->offset
));
3718 chunk
= ram_chunk_index(block
->local_host_addr
,
3719 (uint8_t *) host_addr
);
3721 chunk
= reg
->key
.chunk
;
3722 host_addr
= block
->local_host_addr
+
3723 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3724 /* Check for particularly bad chunk value */
3725 if (host_addr
< (void *)block
->local_host_addr
) {
3726 error_report("rdma: bad chunk for block %s"
3728 block
->block_name
, reg
->key
.chunk
);
3732 chunk_start
= ram_chunk_start(block
, chunk
);
3733 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3734 /* avoid "-Waddress-of-packed-member" warning */
3735 uint32_t tmp_rkey
= 0;
3736 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3737 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3738 chunk
, chunk_start
, chunk_end
)) {
3739 error_report("cannot get rkey");
3742 reg_result
->rkey
= tmp_rkey
;
3744 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3746 trace_rdma_registration_handle_register_rkey(reg_result
->rkey
);
3748 result_to_network(reg_result
);
3751 ret
= qemu_rdma_post_send_control(rdma
,
3752 (uint8_t *) results
, ®_resp
, &err
);
3755 error_report_err(err
);
3759 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3760 trace_rdma_registration_handle_unregister(head
.repeat
);
3761 unreg_resp
.repeat
= head
.repeat
;
3762 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3764 for (count
= 0; count
< head
.repeat
; count
++) {
3765 reg
= ®isters
[count
];
3766 network_to_register(reg
);
3768 trace_rdma_registration_handle_unregister_loop(count
,
3769 reg
->current_index
, reg
->key
.chunk
);
3771 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3773 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3774 block
->pmr
[reg
->key
.chunk
] = NULL
;
3777 error_report("rdma unregistration chunk failed: %s",
3782 rdma
->total_registrations
--;
3784 trace_rdma_registration_handle_unregister_success(reg
->key
.chunk
);
3787 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
, &err
);
3790 error_report_err(err
);
3794 case RDMA_CONTROL_REGISTER_RESULT
:
3795 error_report("Invalid RESULT message at dest.");
3798 error_report("Unknown control message %s", control_desc(head
.type
));
3804 rdma
->errored
= true;
3809 * Called during the initial RAM load section which lists the
3810 * RAMBlocks by name. This lets us know the order of the RAMBlocks on
3811 * the source. We've already built our local RAMBlock list, but not
3812 * yet sent the list to the source.
3814 int rdma_block_notification_handle(QEMUFile
*f
, const char *name
)
3819 if (!migrate_rdma()) {
3823 RCU_READ_LOCK_GUARD();
3824 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3825 RDMAContext
*rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3831 /* Find the matching RAMBlock in our local list */
3832 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3833 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3840 error_report("RAMBlock '%s' not found on destination", name
);
3844 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3845 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3846 rdma
->next_src_index
++;
3851 int rdma_registration_start(QEMUFile
*f
, uint64_t flags
)
3853 if (!migrate_rdma() || migration_in_postcopy()) {
3857 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3858 RCU_READ_LOCK_GUARD();
3859 RDMAContext
*rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3864 if (rdma_errored(rdma
)) {
3868 trace_rdma_registration_start(flags
);
3869 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3876 * Inform dest that dynamic registrations are done for now.
3877 * First, flush writes, if any.
3879 int rdma_registration_stop(QEMUFile
*f
, uint64_t flags
)
3881 QIOChannelRDMA
*rioc
;
3884 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3887 if (!migrate_rdma() || migration_in_postcopy()) {
3891 RCU_READ_LOCK_GUARD();
3892 rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3893 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3898 if (rdma_errored(rdma
)) {
3903 ret
= qemu_rdma_drain_cq(rdma
);
3909 if (flags
== RAM_CONTROL_SETUP
) {
3910 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3911 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3912 int reg_result_idx
, i
, nb_dest_blocks
;
3914 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3915 trace_rdma_registration_stop_ram();
3918 * Make sure that we parallelize the pinning on both sides.
3919 * For very large guests, doing this serially takes a really
3920 * long time, so we have to 'interleave' the pinning locally
3921 * with the control messages by performing the pinning on this
3922 * side before we receive the control response from the other
3923 * side that the pinning has completed.
3925 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3926 ®_result_idx
, rdma
->pin_all
?
3927 qemu_rdma_reg_whole_ram_blocks
: NULL
,
3930 error_report_err(err
);
3934 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3937 * The protocol uses two different sets of rkeys (mutually exclusive):
3938 * 1. One key to represent the virtual address of the entire ram block.
3939 * (dynamic chunk registration disabled - pin everything with one rkey.)
3940 * 2. One to represent individual chunks within a ram block.
3941 * (dynamic chunk registration enabled - pin individual chunks.)
3943 * Once the capability is successfully negotiated, the destination transmits
3944 * the keys to use (or sends them later) including the virtual addresses
3945 * and then propagates the remote ram block descriptions to his local copy.
3948 if (local
->nb_blocks
!= nb_dest_blocks
) {
3949 error_report("ram blocks mismatch (Number of blocks %d vs %d)",
3950 local
->nb_blocks
, nb_dest_blocks
);
3951 error_printf("Your QEMU command line parameters are probably "
3952 "not identical on both the source and destination.");
3953 rdma
->errored
= true;
3957 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3958 memcpy(rdma
->dest_blocks
,
3959 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3960 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3961 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3963 /* We require that the blocks are in the same order */
3964 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3965 error_report("Block %s/%d has a different length %" PRIu64
3967 local
->block
[i
].block_name
, i
,
3968 local
->block
[i
].length
,
3969 rdma
->dest_blocks
[i
].length
);
3970 rdma
->errored
= true;
3973 local
->block
[i
].remote_host_addr
=
3974 rdma
->dest_blocks
[i
].remote_host_addr
;
3975 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3979 trace_rdma_registration_stop(flags
);
3981 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3982 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
, &err
);
3985 error_report_err(err
);
3991 rdma
->errored
= true;
3995 static void qio_channel_rdma_finalize(Object
*obj
)
3997 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
3999 qemu_rdma_cleanup(rioc
->rdmain
);
4000 g_free(rioc
->rdmain
);
4001 rioc
->rdmain
= NULL
;
4003 if (rioc
->rdmaout
) {
4004 qemu_rdma_cleanup(rioc
->rdmaout
);
4005 g_free(rioc
->rdmaout
);
4006 rioc
->rdmaout
= NULL
;
4010 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
4011 void *class_data G_GNUC_UNUSED
)
4013 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
4015 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
4016 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
4017 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
4018 ioc_klass
->io_close
= qio_channel_rdma_close
;
4019 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
4020 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
4021 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
4024 static const TypeInfo qio_channel_rdma_info
= {
4025 .parent
= TYPE_QIO_CHANNEL
,
4026 .name
= TYPE_QIO_CHANNEL_RDMA
,
4027 .instance_size
= sizeof(QIOChannelRDMA
),
4028 .instance_finalize
= qio_channel_rdma_finalize
,
4029 .class_init
= qio_channel_rdma_class_init
,
4032 static void qio_channel_rdma_register_types(void)
4034 type_register_static(&qio_channel_rdma_info
);
4037 type_init(qio_channel_rdma_register_types
);
4039 static QEMUFile
*rdma_new_input(RDMAContext
*rdma
)
4041 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4043 rioc
->file
= qemu_file_new_input(QIO_CHANNEL(rioc
));
4044 rioc
->rdmain
= rdma
;
4045 rioc
->rdmaout
= rdma
->return_path
;
4050 static QEMUFile
*rdma_new_output(RDMAContext
*rdma
)
4052 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4054 rioc
->file
= qemu_file_new_output(QIO_CHANNEL(rioc
));
4055 rioc
->rdmaout
= rdma
;
4056 rioc
->rdmain
= rdma
->return_path
;
4061 static void rdma_accept_incoming_migration(void *opaque
)
4063 RDMAContext
*rdma
= opaque
;
4066 Error
*local_err
= NULL
;
4068 trace_qemu_rdma_accept_incoming_migration();
4069 ret
= qemu_rdma_accept(rdma
);
4072 error_report("RDMA ERROR: Migration initialization failed");
4076 trace_qemu_rdma_accept_incoming_migration_accepted();
4078 if (rdma
->is_return_path
) {
4082 f
= rdma_new_input(rdma
);
4084 error_report("RDMA ERROR: could not open RDMA for input");
4085 qemu_rdma_cleanup(rdma
);
4089 rdma
->migration_started_on_destination
= 1;
4090 migration_fd_process_incoming(f
, &local_err
);
4092 error_reportf_err(local_err
, "RDMA ERROR:");
4096 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4098 MigrationState
*s
= migrate_get_current();
4102 trace_rdma_start_incoming_migration();
4104 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4105 if (ram_block_discard_is_required()) {
4106 error_setg(errp
, "RDMA: cannot disable RAM discard");
4110 rdma
= qemu_rdma_data_init(host_port
, errp
);
4115 ret
= qemu_rdma_dest_init(rdma
, errp
);
4120 trace_rdma_start_incoming_migration_after_dest_init();
4122 ret
= rdma_listen(rdma
->listen_id
, 5);
4125 error_setg(errp
, "RDMA ERROR: listening on socket!");
4129 trace_rdma_start_incoming_migration_after_rdma_listen();
4130 s
->rdma_migration
= true;
4131 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4132 NULL
, (void *)(intptr_t)rdma
);
4136 qemu_rdma_cleanup(rdma
);
4140 g_free(rdma
->host_port
);
4145 void rdma_start_outgoing_migration(void *opaque
,
4146 const char *host_port
, Error
**errp
)
4148 MigrationState
*s
= opaque
;
4149 RDMAContext
*rdma_return_path
= NULL
;
4153 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4154 if (ram_block_discard_is_required()) {
4155 error_setg(errp
, "RDMA: cannot disable RAM discard");
4159 rdma
= qemu_rdma_data_init(host_port
, errp
);
4164 ret
= qemu_rdma_source_init(rdma
, migrate_rdma_pin_all(), errp
);
4170 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4171 ret
= qemu_rdma_connect(rdma
, false, errp
);
4177 /* RDMA postcopy need a separate queue pair for return path */
4178 if (migrate_postcopy() || migrate_return_path()) {
4179 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4181 if (rdma_return_path
== NULL
) {
4182 goto return_path_err
;
4185 ret
= qemu_rdma_source_init(rdma_return_path
,
4186 migrate_rdma_pin_all(), errp
);
4189 goto return_path_err
;
4192 ret
= qemu_rdma_connect(rdma_return_path
, true, errp
);
4195 goto return_path_err
;
4198 rdma
->return_path
= rdma_return_path
;
4199 rdma_return_path
->return_path
= rdma
;
4200 rdma_return_path
->is_return_path
= true;
4203 trace_rdma_start_outgoing_migration_after_rdma_connect();
4205 s
->to_dst_file
= rdma_new_output(rdma
);
4206 s
->rdma_migration
= true;
4207 migrate_fd_connect(s
, NULL
);
4210 qemu_rdma_cleanup(rdma
);
4213 g_free(rdma_return_path
);