2 * RDMA protocol and interfaces
4 * Copyright IBM, Corp. 2010-2013
5 * Copyright Red Hat, Inc. 2015-2016
8 * Michael R. Hines <mrhines@us.ibm.com>
9 * Jiuxing Liu <jl@us.ibm.com>
10 * Daniel P. Berrange <berrange@redhat.com>
12 * This work is licensed under the terms of the GNU GPL, version 2 or
13 * later. See the COPYING file in the top-level directory.
17 #include "qemu/osdep.h"
18 #include "qapi/error.h"
19 #include "qemu/cutils.h"
21 #include "migration.h"
22 #include "qemu-file.h"
24 #include "qemu-file-channel.h"
25 #include "qemu/error-report.h"
26 #include "qemu/main-loop.h"
27 #include "qemu/module.h"
29 #include "qemu/sockets.h"
30 #include "qemu/bitmap.h"
31 #include "qemu/coroutine.h"
32 #include "exec/memory.h"
33 #include <sys/socket.h>
35 #include <arpa/inet.h>
36 #include <rdma/rdma_cma.h>
38 #include "qom/object.h"
42 * Print and error on both the Monitor and the Log file.
44 #define ERROR(errp, fmt, ...) \
46 fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \
47 if (errp && (*(errp) == NULL)) { \
48 error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \
52 #define RDMA_RESOLVE_TIMEOUT_MS 10000
54 /* Do not merge data if larger than this. */
55 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
56 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
58 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
61 * This is only for non-live state being migrated.
62 * Instead of RDMA_WRITE messages, we use RDMA_SEND
63 * messages for that state, which requires a different
64 * delivery design than main memory.
66 #define RDMA_SEND_INCREMENT 32768
69 * Maximum size infiniband SEND message
71 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
72 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
74 #define RDMA_CONTROL_VERSION_CURRENT 1
76 * Capabilities for negotiation.
78 #define RDMA_CAPABILITY_PIN_ALL 0x01
81 * Add the other flags above to this list of known capabilities
82 * as they are introduced.
84 static uint32_t known_capabilities
= RDMA_CAPABILITY_PIN_ALL
;
86 #define CHECK_ERROR_STATE() \
88 if (rdma->error_state) { \
89 if (!rdma->error_reported) { \
90 error_report("RDMA is in an error state waiting migration" \
92 rdma->error_reported = 1; \
94 return rdma->error_state; \
99 * A work request ID is 64-bits and we split up these bits
102 * bits 0-15 : type of control message, 2^16
103 * bits 16-29: ram block index, 2^14
104 * bits 30-63: ram block chunk number, 2^34
106 * The last two bit ranges are only used for RDMA writes,
107 * in order to track their completion and potentially
108 * also track unregistration status of the message.
110 #define RDMA_WRID_TYPE_SHIFT 0UL
111 #define RDMA_WRID_BLOCK_SHIFT 16UL
112 #define RDMA_WRID_CHUNK_SHIFT 30UL
114 #define RDMA_WRID_TYPE_MASK \
115 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
117 #define RDMA_WRID_BLOCK_MASK \
118 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
120 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
123 * RDMA migration protocol:
124 * 1. RDMA Writes (data messages, i.e. RAM)
125 * 2. IB Send/Recv (control channel messages)
129 RDMA_WRID_RDMA_WRITE
= 1,
130 RDMA_WRID_SEND_CONTROL
= 2000,
131 RDMA_WRID_RECV_CONTROL
= 4000,
134 static const char *wrid_desc
[] = {
135 [RDMA_WRID_NONE
] = "NONE",
136 [RDMA_WRID_RDMA_WRITE
] = "WRITE RDMA",
137 [RDMA_WRID_SEND_CONTROL
] = "CONTROL SEND",
138 [RDMA_WRID_RECV_CONTROL
] = "CONTROL RECV",
142 * Work request IDs for IB SEND messages only (not RDMA writes).
143 * This is used by the migration protocol to transmit
144 * control messages (such as device state and registration commands)
146 * We could use more WRs, but we have enough for now.
156 * SEND/RECV IB Control Messages.
159 RDMA_CONTROL_NONE
= 0,
161 RDMA_CONTROL_READY
, /* ready to receive */
162 RDMA_CONTROL_QEMU_FILE
, /* QEMUFile-transmitted bytes */
163 RDMA_CONTROL_RAM_BLOCKS_REQUEST
, /* RAMBlock synchronization */
164 RDMA_CONTROL_RAM_BLOCKS_RESULT
, /* RAMBlock synchronization */
165 RDMA_CONTROL_COMPRESS
, /* page contains repeat values */
166 RDMA_CONTROL_REGISTER_REQUEST
, /* dynamic page registration */
167 RDMA_CONTROL_REGISTER_RESULT
, /* key to use after registration */
168 RDMA_CONTROL_REGISTER_FINISHED
, /* current iteration finished */
169 RDMA_CONTROL_UNREGISTER_REQUEST
, /* dynamic UN-registration */
170 RDMA_CONTROL_UNREGISTER_FINISHED
, /* unpinning finished */
175 * Memory and MR structures used to represent an IB Send/Recv work request.
176 * This is *not* used for RDMA writes, only IB Send/Recv.
179 uint8_t control
[RDMA_CONTROL_MAX_BUFFER
]; /* actual buffer to register */
180 struct ibv_mr
*control_mr
; /* registration metadata */
181 size_t control_len
; /* length of the message */
182 uint8_t *control_curr
; /* start of unconsumed bytes */
183 } RDMAWorkRequestData
;
186 * Negotiate RDMA capabilities during connection-setup time.
193 static void caps_to_network(RDMACapabilities
*cap
)
195 cap
->version
= htonl(cap
->version
);
196 cap
->flags
= htonl(cap
->flags
);
199 static void network_to_caps(RDMACapabilities
*cap
)
201 cap
->version
= ntohl(cap
->version
);
202 cap
->flags
= ntohl(cap
->flags
);
206 * Representation of a RAMBlock from an RDMA perspective.
207 * This is not transmitted, only local.
208 * This and subsequent structures cannot be linked lists
209 * because we're using a single IB message to transmit
210 * the information. It's small anyway, so a list is overkill.
212 typedef struct RDMALocalBlock
{
214 uint8_t *local_host_addr
; /* local virtual address */
215 uint64_t remote_host_addr
; /* remote virtual address */
218 struct ibv_mr
**pmr
; /* MRs for chunk-level registration */
219 struct ibv_mr
*mr
; /* MR for non-chunk-level registration */
220 uint32_t *remote_keys
; /* rkeys for chunk-level registration */
221 uint32_t remote_rkey
; /* rkeys for non-chunk-level registration */
222 int index
; /* which block are we */
223 unsigned int src_index
; /* (Only used on dest) */
226 unsigned long *transit_bitmap
;
227 unsigned long *unregister_bitmap
;
231 * Also represents a RAMblock, but only on the dest.
232 * This gets transmitted by the dest during connection-time
233 * to the source VM and then is used to populate the
234 * corresponding RDMALocalBlock with
235 * the information needed to perform the actual RDMA.
237 typedef struct QEMU_PACKED RDMADestBlock
{
238 uint64_t remote_host_addr
;
241 uint32_t remote_rkey
;
245 static const char *control_desc(unsigned int rdma_control
)
247 static const char *strs
[] = {
248 [RDMA_CONTROL_NONE
] = "NONE",
249 [RDMA_CONTROL_ERROR
] = "ERROR",
250 [RDMA_CONTROL_READY
] = "READY",
251 [RDMA_CONTROL_QEMU_FILE
] = "QEMU FILE",
252 [RDMA_CONTROL_RAM_BLOCKS_REQUEST
] = "RAM BLOCKS REQUEST",
253 [RDMA_CONTROL_RAM_BLOCKS_RESULT
] = "RAM BLOCKS RESULT",
254 [RDMA_CONTROL_COMPRESS
] = "COMPRESS",
255 [RDMA_CONTROL_REGISTER_REQUEST
] = "REGISTER REQUEST",
256 [RDMA_CONTROL_REGISTER_RESULT
] = "REGISTER RESULT",
257 [RDMA_CONTROL_REGISTER_FINISHED
] = "REGISTER FINISHED",
258 [RDMA_CONTROL_UNREGISTER_REQUEST
] = "UNREGISTER REQUEST",
259 [RDMA_CONTROL_UNREGISTER_FINISHED
] = "UNREGISTER FINISHED",
262 if (rdma_control
> RDMA_CONTROL_UNREGISTER_FINISHED
) {
263 return "??BAD CONTROL VALUE??";
266 return strs
[rdma_control
];
269 static uint64_t htonll(uint64_t v
)
271 union { uint32_t lv
[2]; uint64_t llv
; } u
;
272 u
.lv
[0] = htonl(v
>> 32);
273 u
.lv
[1] = htonl(v
& 0xFFFFFFFFULL
);
277 static uint64_t ntohll(uint64_t v
)
279 union { uint32_t lv
[2]; uint64_t llv
; } u
;
281 return ((uint64_t)ntohl(u
.lv
[0]) << 32) | (uint64_t) ntohl(u
.lv
[1]);
284 static void dest_block_to_network(RDMADestBlock
*db
)
286 db
->remote_host_addr
= htonll(db
->remote_host_addr
);
287 db
->offset
= htonll(db
->offset
);
288 db
->length
= htonll(db
->length
);
289 db
->remote_rkey
= htonl(db
->remote_rkey
);
292 static void network_to_dest_block(RDMADestBlock
*db
)
294 db
->remote_host_addr
= ntohll(db
->remote_host_addr
);
295 db
->offset
= ntohll(db
->offset
);
296 db
->length
= ntohll(db
->length
);
297 db
->remote_rkey
= ntohl(db
->remote_rkey
);
301 * Virtual address of the above structures used for transmitting
302 * the RAMBlock descriptions at connection-time.
303 * This structure is *not* transmitted.
305 typedef struct RDMALocalBlocks
{
307 bool init
; /* main memory init complete */
308 RDMALocalBlock
*block
;
312 * Main data structure for RDMA state.
313 * While there is only one copy of this structure being allocated right now,
314 * this is the place where one would start if you wanted to consider
315 * having more than one RDMA connection open at the same time.
317 typedef struct RDMAContext
{
322 RDMAWorkRequestData wr_data
[RDMA_WRID_MAX
];
325 * This is used by *_exchange_send() to figure out whether or not
326 * the initial "READY" message has already been received or not.
327 * This is because other functions may potentially poll() and detect
328 * the READY message before send() does, in which case we need to
329 * know if it completed.
331 int control_ready_expected
;
333 /* number of outstanding writes */
336 /* store info about current buffer so that we can
337 merge it with future sends */
338 uint64_t current_addr
;
339 uint64_t current_length
;
340 /* index of ram block the current buffer belongs to */
342 /* index of the chunk in the current ram block */
348 * infiniband-specific variables for opening the device
349 * and maintaining connection state and so forth.
351 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
352 * cm_id->verbs, cm_id->channel, and cm_id->qp.
354 struct rdma_cm_id
*cm_id
; /* connection manager ID */
355 struct rdma_cm_id
*listen_id
;
358 struct ibv_context
*verbs
;
359 struct rdma_event_channel
*channel
;
360 struct ibv_qp
*qp
; /* queue pair */
361 struct ibv_comp_channel
*comp_channel
; /* completion channel */
362 struct ibv_pd
*pd
; /* protection domain */
363 struct ibv_cq
*cq
; /* completion queue */
366 * If a previous write failed (perhaps because of a failed
367 * memory registration, then do not attempt any future work
368 * and remember the error state.
375 * Description of ram blocks used throughout the code.
377 RDMALocalBlocks local_ram_blocks
;
378 RDMADestBlock
*dest_blocks
;
380 /* Index of the next RAMBlock received during block registration */
381 unsigned int next_src_index
;
384 * Migration on *destination* started.
385 * Then use coroutine yield function.
386 * Source runs in a thread, so we don't care.
388 int migration_started_on_destination
;
390 int total_registrations
;
393 int unregister_current
, unregister_next
;
394 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
396 GHashTable
*blockmap
;
398 /* the RDMAContext for return path */
399 struct RDMAContext
*return_path
;
403 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
404 OBJECT_DECLARE_SIMPLE_TYPE(QIOChannelRDMA
, QIO_CHANNEL_RDMA
)
408 struct QIOChannelRDMA
{
411 RDMAContext
*rdmaout
;
413 bool blocking
; /* XXX we don't actually honour this yet */
417 * Main structure for IB Send/Recv control messages.
418 * This gets prepended at the beginning of every Send/Recv.
420 typedef struct QEMU_PACKED
{
421 uint32_t len
; /* Total length of data portion */
422 uint32_t type
; /* which control command to perform */
423 uint32_t repeat
; /* number of commands in data portion of same type */
427 static void control_to_network(RDMAControlHeader
*control
)
429 control
->type
= htonl(control
->type
);
430 control
->len
= htonl(control
->len
);
431 control
->repeat
= htonl(control
->repeat
);
434 static void network_to_control(RDMAControlHeader
*control
)
436 control
->type
= ntohl(control
->type
);
437 control
->len
= ntohl(control
->len
);
438 control
->repeat
= ntohl(control
->repeat
);
442 * Register a single Chunk.
443 * Information sent by the source VM to inform the dest
444 * to register an single chunk of memory before we can perform
445 * the actual RDMA operation.
447 typedef struct QEMU_PACKED
{
449 uint64_t current_addr
; /* offset into the ram_addr_t space */
450 uint64_t chunk
; /* chunk to lookup if unregistering */
452 uint32_t current_index
; /* which ramblock the chunk belongs to */
454 uint64_t chunks
; /* how many sequential chunks to register */
457 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
459 RDMALocalBlock
*local_block
;
460 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
462 if (local_block
->is_ram_block
) {
464 * current_addr as passed in is an address in the local ram_addr_t
465 * space, we need to translate this for the destination
467 reg
->key
.current_addr
-= local_block
->offset
;
468 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
470 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
471 reg
->current_index
= htonl(reg
->current_index
);
472 reg
->chunks
= htonll(reg
->chunks
);
475 static void network_to_register(RDMARegister
*reg
)
477 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
478 reg
->current_index
= ntohl(reg
->current_index
);
479 reg
->chunks
= ntohll(reg
->chunks
);
482 typedef struct QEMU_PACKED
{
483 uint32_t value
; /* if zero, we will madvise() */
484 uint32_t block_idx
; /* which ram block index */
485 uint64_t offset
; /* Address in remote ram_addr_t space */
486 uint64_t length
; /* length of the chunk */
489 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
491 comp
->value
= htonl(comp
->value
);
493 * comp->offset as passed in is an address in the local ram_addr_t
494 * space, we need to translate this for the destination
496 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
497 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
498 comp
->block_idx
= htonl(comp
->block_idx
);
499 comp
->offset
= htonll(comp
->offset
);
500 comp
->length
= htonll(comp
->length
);
503 static void network_to_compress(RDMACompress
*comp
)
505 comp
->value
= ntohl(comp
->value
);
506 comp
->block_idx
= ntohl(comp
->block_idx
);
507 comp
->offset
= ntohll(comp
->offset
);
508 comp
->length
= ntohll(comp
->length
);
512 * The result of the dest's memory registration produces an "rkey"
513 * which the source VM must reference in order to perform
514 * the RDMA operation.
516 typedef struct QEMU_PACKED
{
520 } RDMARegisterResult
;
522 static void result_to_network(RDMARegisterResult
*result
)
524 result
->rkey
= htonl(result
->rkey
);
525 result
->host_addr
= htonll(result
->host_addr
);
528 static void network_to_result(RDMARegisterResult
*result
)
530 result
->rkey
= ntohl(result
->rkey
);
531 result
->host_addr
= ntohll(result
->host_addr
);
534 const char *print_wrid(int wrid
);
535 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
536 uint8_t *data
, RDMAControlHeader
*resp
,
538 int (*callback
)(RDMAContext
*rdma
));
540 static inline uint64_t ram_chunk_index(const uint8_t *start
,
543 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
546 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
549 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
550 (i
<< RDMA_REG_CHUNK_SHIFT
));
553 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
556 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
557 (1UL << RDMA_REG_CHUNK_SHIFT
);
559 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
560 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
566 static int rdma_add_block(RDMAContext
*rdma
, const char *block_name
,
568 ram_addr_t block_offset
, uint64_t length
)
570 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
571 RDMALocalBlock
*block
;
572 RDMALocalBlock
*old
= local
->block
;
574 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
+ 1);
576 if (local
->nb_blocks
) {
579 if (rdma
->blockmap
) {
580 for (x
= 0; x
< local
->nb_blocks
; x
++) {
581 g_hash_table_remove(rdma
->blockmap
,
582 (void *)(uintptr_t)old
[x
].offset
);
583 g_hash_table_insert(rdma
->blockmap
,
584 (void *)(uintptr_t)old
[x
].offset
,
588 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
592 block
= &local
->block
[local
->nb_blocks
];
594 block
->block_name
= g_strdup(block_name
);
595 block
->local_host_addr
= host_addr
;
596 block
->offset
= block_offset
;
597 block
->length
= length
;
598 block
->index
= local
->nb_blocks
;
599 block
->src_index
= ~0U; /* Filled in by the receipt of the block list */
600 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
601 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
602 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
603 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
604 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
605 block
->remote_keys
= g_new0(uint32_t, block
->nb_chunks
);
607 block
->is_ram_block
= local
->init
? false : true;
609 if (rdma
->blockmap
) {
610 g_hash_table_insert(rdma
->blockmap
, (void *)(uintptr_t)block_offset
, block
);
613 trace_rdma_add_block(block_name
, local
->nb_blocks
,
614 (uintptr_t) block
->local_host_addr
,
615 block
->offset
, block
->length
,
616 (uintptr_t) (block
->local_host_addr
+ block
->length
),
617 BITS_TO_LONGS(block
->nb_chunks
) *
618 sizeof(unsigned long) * 8,
627 * Memory regions need to be registered with the device and queue pairs setup
628 * in advanced before the migration starts. This tells us where the RAM blocks
629 * are so that we can register them individually.
631 static int qemu_rdma_init_one_block(RAMBlock
*rb
, void *opaque
)
633 const char *block_name
= qemu_ram_get_idstr(rb
);
634 void *host_addr
= qemu_ram_get_host_addr(rb
);
635 ram_addr_t block_offset
= qemu_ram_get_offset(rb
);
636 ram_addr_t length
= qemu_ram_get_used_length(rb
);
637 return rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
641 * Identify the RAMBlocks and their quantity. They will be references to
642 * identify chunk boundaries inside each RAMBlock and also be referenced
643 * during dynamic page registration.
645 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
647 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
650 assert(rdma
->blockmap
== NULL
);
651 memset(local
, 0, sizeof *local
);
652 ret
= foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
656 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
657 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
658 rdma
->local_ram_blocks
.nb_blocks
);
664 * Note: If used outside of cleanup, the caller must ensure that the destination
665 * block structures are also updated
667 static int rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
669 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
670 RDMALocalBlock
*old
= local
->block
;
673 if (rdma
->blockmap
) {
674 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
679 for (j
= 0; j
< block
->nb_chunks
; j
++) {
680 if (!block
->pmr
[j
]) {
683 ibv_dereg_mr(block
->pmr
[j
]);
684 rdma
->total_registrations
--;
691 ibv_dereg_mr(block
->mr
);
692 rdma
->total_registrations
--;
696 g_free(block
->transit_bitmap
);
697 block
->transit_bitmap
= NULL
;
699 g_free(block
->unregister_bitmap
);
700 block
->unregister_bitmap
= NULL
;
702 g_free(block
->remote_keys
);
703 block
->remote_keys
= NULL
;
705 g_free(block
->block_name
);
706 block
->block_name
= NULL
;
708 if (rdma
->blockmap
) {
709 for (x
= 0; x
< local
->nb_blocks
; x
++) {
710 g_hash_table_remove(rdma
->blockmap
,
711 (void *)(uintptr_t)old
[x
].offset
);
715 if (local
->nb_blocks
> 1) {
717 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
720 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
723 if (block
->index
< (local
->nb_blocks
- 1)) {
724 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
725 sizeof(RDMALocalBlock
) *
726 (local
->nb_blocks
- (block
->index
+ 1)));
727 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
728 local
->block
[x
].index
--;
732 assert(block
== local
->block
);
736 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
737 block
->offset
, block
->length
,
738 (uintptr_t)(block
->local_host_addr
+ block
->length
),
739 BITS_TO_LONGS(block
->nb_chunks
) *
740 sizeof(unsigned long) * 8, block
->nb_chunks
);
746 if (local
->nb_blocks
&& rdma
->blockmap
) {
747 for (x
= 0; x
< local
->nb_blocks
; x
++) {
748 g_hash_table_insert(rdma
->blockmap
,
749 (void *)(uintptr_t)local
->block
[x
].offset
,
758 * Put in the log file which RDMA device was opened and the details
759 * associated with that device.
761 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
763 struct ibv_port_attr port
;
765 if (ibv_query_port(verbs
, 1, &port
)) {
766 error_report("Failed to query port information");
770 printf("%s RDMA Device opened: kernel name %s "
771 "uverbs device name %s, "
772 "infiniband_verbs class device path %s, "
773 "infiniband class device path %s, "
774 "transport: (%d) %s\n",
777 verbs
->device
->dev_name
,
778 verbs
->device
->dev_path
,
779 verbs
->device
->ibdev_path
,
781 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
782 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
783 ? "Ethernet" : "Unknown"));
787 * Put in the log file the RDMA gid addressing information,
788 * useful for folks who have trouble understanding the
789 * RDMA device hierarchy in the kernel.
791 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
795 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
796 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
797 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
801 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
802 * We will try the next addrinfo struct, and fail if there are
803 * no other valid addresses to bind against.
805 * If user is listening on '[::]', then we will not have a opened a device
806 * yet and have no way of verifying if the device is RoCE or not.
808 * In this case, the source VM will throw an error for ALL types of
809 * connections (both IPv4 and IPv6) if the destination machine does not have
810 * a regular infiniband network available for use.
812 * The only way to guarantee that an error is thrown for broken kernels is
813 * for the management software to choose a *specific* interface at bind time
814 * and validate what time of hardware it is.
816 * Unfortunately, this puts the user in a fix:
818 * If the source VM connects with an IPv4 address without knowing that the
819 * destination has bound to '[::]' the migration will unconditionally fail
820 * unless the management software is explicitly listening on the IPv4
821 * address while using a RoCE-based device.
823 * If the source VM connects with an IPv6 address, then we're OK because we can
824 * throw an error on the source (and similarly on the destination).
826 * But in mixed environments, this will be broken for a while until it is fixed
829 * We do provide a *tiny* bit of help in this function: We can list all of the
830 * devices in the system and check to see if all the devices are RoCE or
833 * If we detect that we have a *pure* RoCE environment, then we can safely
834 * thrown an error even if the management software has specified '[::]' as the
837 * However, if there is are multiple hetergeneous devices, then we cannot make
838 * this assumption and the user just has to be sure they know what they are
841 * Patches are being reviewed on linux-rdma.
843 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
845 /* This bug only exists in linux, to our knowledge. */
847 struct ibv_port_attr port_attr
;
850 * Verbs are only NULL if management has bound to '[::]'.
852 * Let's iterate through all the devices and see if there any pure IB
853 * devices (non-ethernet).
855 * If not, then we can safely proceed with the migration.
856 * Otherwise, there are no guarantees until the bug is fixed in linux.
860 struct ibv_device
**dev_list
= ibv_get_device_list(&num_devices
);
861 bool roce_found
= false;
862 bool ib_found
= false;
864 for (x
= 0; x
< num_devices
; x
++) {
865 verbs
= ibv_open_device(dev_list
[x
]);
867 if (errno
== EPERM
) {
874 if (ibv_query_port(verbs
, 1, &port_attr
)) {
875 ibv_close_device(verbs
);
876 ERROR(errp
, "Could not query initial IB port");
880 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
882 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
886 ibv_close_device(verbs
);
892 fprintf(stderr
, "WARN: migrations may fail:"
893 " IPv6 over RoCE / iWARP in linux"
894 " is broken. But since you appear to have a"
895 " mixed RoCE / IB environment, be sure to only"
896 " migrate over the IB fabric until the kernel "
897 " fixes the bug.\n");
899 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
900 " and your management software has specified '[::]'"
901 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
910 * If we have a verbs context, that means that some other than '[::]' was
911 * used by the management software for binding. In which case we can
912 * actually warn the user about a potentially broken kernel.
915 /* IB ports start with 1, not 0 */
916 if (ibv_query_port(verbs
, 1, &port_attr
)) {
917 ERROR(errp
, "Could not query initial IB port");
921 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
922 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
923 "(but patches on linux-rdma in progress)");
933 * Figure out which RDMA device corresponds to the requested IP hostname
934 * Also create the initial connection manager identifiers for opening
937 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
940 struct rdma_addrinfo
*res
;
942 struct rdma_cm_event
*cm_event
;
943 char ip
[40] = "unknown";
944 struct rdma_addrinfo
*e
;
946 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
947 ERROR(errp
, "RDMA hostname has not been set");
951 /* create CM channel */
952 rdma
->channel
= rdma_create_event_channel();
953 if (!rdma
->channel
) {
954 ERROR(errp
, "could not create CM channel");
959 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
961 ERROR(errp
, "could not create channel id");
962 goto err_resolve_create_id
;
965 snprintf(port_str
, 16, "%d", rdma
->port
);
968 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
970 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
971 goto err_resolve_get_addr
;
974 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
975 inet_ntop(e
->ai_family
,
976 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
977 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
979 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
980 RDMA_RESOLVE_TIMEOUT_MS
);
982 if (e
->ai_family
== AF_INET6
) {
983 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
992 rdma_freeaddrinfo(res
);
993 ERROR(errp
, "could not resolve address %s", rdma
->host
);
994 goto err_resolve_get_addr
;
997 rdma_freeaddrinfo(res
);
998 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
1000 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1002 ERROR(errp
, "could not perform event_addr_resolved");
1003 goto err_resolve_get_addr
;
1006 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
1007 ERROR(errp
, "result not equal to event_addr_resolved %s",
1008 rdma_event_str(cm_event
->event
));
1009 error_report("rdma_resolve_addr");
1010 rdma_ack_cm_event(cm_event
);
1012 goto err_resolve_get_addr
;
1014 rdma_ack_cm_event(cm_event
);
1017 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1019 ERROR(errp
, "could not resolve rdma route");
1020 goto err_resolve_get_addr
;
1023 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1025 ERROR(errp
, "could not perform event_route_resolved");
1026 goto err_resolve_get_addr
;
1028 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1029 ERROR(errp
, "result not equal to event_route_resolved: %s",
1030 rdma_event_str(cm_event
->event
));
1031 rdma_ack_cm_event(cm_event
);
1033 goto err_resolve_get_addr
;
1035 rdma_ack_cm_event(cm_event
);
1036 rdma
->verbs
= rdma
->cm_id
->verbs
;
1037 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1038 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1041 err_resolve_get_addr
:
1042 rdma_destroy_id(rdma
->cm_id
);
1044 err_resolve_create_id
:
1045 rdma_destroy_event_channel(rdma
->channel
);
1046 rdma
->channel
= NULL
;
1051 * Create protection domain and completion queues
1053 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1056 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1058 error_report("failed to allocate protection domain");
1062 /* create completion channel */
1063 rdma
->comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1064 if (!rdma
->comp_channel
) {
1065 error_report("failed to allocate completion channel");
1066 goto err_alloc_pd_cq
;
1070 * Completion queue can be filled by both read and write work requests,
1071 * so must reflect the sum of both possible queue sizes.
1073 rdma
->cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1074 NULL
, rdma
->comp_channel
, 0);
1076 error_report("failed to allocate completion queue");
1077 goto err_alloc_pd_cq
;
1084 ibv_dealloc_pd(rdma
->pd
);
1086 if (rdma
->comp_channel
) {
1087 ibv_destroy_comp_channel(rdma
->comp_channel
);
1090 rdma
->comp_channel
= NULL
;
1096 * Create queue pairs.
1098 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1100 struct ibv_qp_init_attr attr
= { 0 };
1103 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1104 attr
.cap
.max_recv_wr
= 3;
1105 attr
.cap
.max_send_sge
= 1;
1106 attr
.cap
.max_recv_sge
= 1;
1107 attr
.send_cq
= rdma
->cq
;
1108 attr
.recv_cq
= rdma
->cq
;
1109 attr
.qp_type
= IBV_QPT_RC
;
1111 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1116 rdma
->qp
= rdma
->cm_id
->qp
;
1120 /* Check whether On-Demand Paging is supported by RDAM device */
1121 static bool rdma_support_odp(struct ibv_context
*dev
)
1123 struct ibv_device_attr_ex attr
= {0};
1124 int ret
= ibv_query_device_ex(dev
, NULL
, &attr
);
1129 if (attr
.odp_caps
.general_caps
& IBV_ODP_SUPPORT
) {
1136 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1139 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1141 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1142 int access
= IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
;
1144 local
->block
[i
].mr
=
1145 ibv_reg_mr(rdma
->pd
,
1146 local
->block
[i
].local_host_addr
,
1147 local
->block
[i
].length
, access
1150 if (!local
->block
[i
].mr
&&
1151 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1152 access
|= IBV_ACCESS_ON_DEMAND
;
1153 /* register ODP mr */
1154 local
->block
[i
].mr
=
1155 ibv_reg_mr(rdma
->pd
,
1156 local
->block
[i
].local_host_addr
,
1157 local
->block
[i
].length
, access
);
1158 trace_qemu_rdma_register_odp_mr(local
->block
[i
].block_name
);
1161 if (!local
->block
[i
].mr
) {
1162 perror("Failed to register local dest ram block!");
1165 rdma
->total_registrations
++;
1168 if (i
>= local
->nb_blocks
) {
1172 for (i
--; i
>= 0; i
--) {
1173 ibv_dereg_mr(local
->block
[i
].mr
);
1174 local
->block
[i
].mr
= NULL
;
1175 rdma
->total_registrations
--;
1183 * Find the ram block that corresponds to the page requested to be
1184 * transmitted by QEMU.
1186 * Once the block is found, also identify which 'chunk' within that
1187 * block that the page belongs to.
1189 * This search cannot fail or the migration will fail.
1191 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1192 uintptr_t block_offset
,
1195 uint64_t *block_index
,
1196 uint64_t *chunk_index
)
1198 uint64_t current_addr
= block_offset
+ offset
;
1199 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1200 (void *) block_offset
);
1202 assert(current_addr
>= block
->offset
);
1203 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1205 *block_index
= block
->index
;
1206 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1207 block
->local_host_addr
+ (current_addr
- block
->offset
));
1213 * Register a chunk with IB. If the chunk was already registered
1214 * previously, then skip.
1216 * Also return the keys associated with the registration needed
1217 * to perform the actual RDMA operation.
1219 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1220 RDMALocalBlock
*block
, uintptr_t host_addr
,
1221 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1222 uint8_t *chunk_start
, uint8_t *chunk_end
)
1226 *lkey
= block
->mr
->lkey
;
1229 *rkey
= block
->mr
->rkey
;
1234 /* allocate memory to store chunk MRs */
1236 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1240 * If 'rkey', then we're the destination, so grant access to the source.
1242 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1244 if (!block
->pmr
[chunk
]) {
1245 uint64_t len
= chunk_end
- chunk_start
;
1246 int access
= rkey
? IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
:
1249 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1251 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1252 if (!block
->pmr
[chunk
] &&
1253 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1254 access
|= IBV_ACCESS_ON_DEMAND
;
1255 /* register ODP mr */
1256 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1257 trace_qemu_rdma_register_odp_mr(block
->block_name
);
1260 if (!block
->pmr
[chunk
]) {
1261 perror("Failed to register chunk!");
1262 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1263 " start %" PRIuPTR
" end %" PRIuPTR
1265 " local %" PRIuPTR
" registrations: %d\n",
1266 block
->index
, chunk
, (uintptr_t)chunk_start
,
1267 (uintptr_t)chunk_end
, host_addr
,
1268 (uintptr_t)block
->local_host_addr
,
1269 rdma
->total_registrations
);
1272 rdma
->total_registrations
++;
1275 *lkey
= block
->pmr
[chunk
]->lkey
;
1278 *rkey
= block
->pmr
[chunk
]->rkey
;
1284 * Register (at connection time) the memory used for control
1287 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1289 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1290 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1291 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1292 if (rdma
->wr_data
[idx
].control_mr
) {
1293 rdma
->total_registrations
++;
1296 error_report("qemu_rdma_reg_control failed");
1300 const char *print_wrid(int wrid
)
1302 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1303 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1305 return wrid_desc
[wrid
];
1309 * RDMA requires memory registration (mlock/pinning), but this is not good for
1312 * In preparation for the future where LRU information or workload-specific
1313 * writable writable working set memory access behavior is available to QEMU
1314 * it would be nice to have in place the ability to UN-register/UN-pin
1315 * particular memory regions from the RDMA hardware when it is determine that
1316 * those regions of memory will likely not be accessed again in the near future.
1318 * While we do not yet have such information right now, the following
1319 * compile-time option allows us to perform a non-optimized version of this
1322 * By uncommenting this option, you will cause *all* RDMA transfers to be
1323 * unregistered immediately after the transfer completes on both sides of the
1324 * connection. This has no effect in 'rdma-pin-all' mode, only regular mode.
1326 * This will have a terrible impact on migration performance, so until future
1327 * workload information or LRU information is available, do not attempt to use
1328 * this feature except for basic testing.
1330 /* #define RDMA_UNREGISTRATION_EXAMPLE */
1333 * Perform a non-optimized memory unregistration after every transfer
1334 * for demonstration purposes, only if pin-all is not requested.
1336 * Potential optimizations:
1337 * 1. Start a new thread to run this function continuously
1339 - and for receipt of unregister messages
1341 * 3. Use workload hints.
1343 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1345 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1347 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1349 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1351 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1352 RDMALocalBlock
*block
=
1353 &(rdma
->local_ram_blocks
.block
[index
]);
1354 RDMARegister reg
= { .current_index
= index
};
1355 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1357 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1358 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1362 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1363 rdma
->unregister_current
);
1365 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1366 rdma
->unregister_current
++;
1368 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1369 rdma
->unregister_current
= 0;
1374 * Unregistration is speculative (because migration is single-threaded
1375 * and we cannot break the protocol's inifinband message ordering).
1376 * Thus, if the memory is currently being used for transmission,
1377 * then abort the attempt to unregister and try again
1378 * later the next time a completion is received for this memory.
1380 clear_bit(chunk
, block
->unregister_bitmap
);
1382 if (test_bit(chunk
, block
->transit_bitmap
)) {
1383 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1387 trace_qemu_rdma_unregister_waiting_send(chunk
);
1389 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1390 block
->pmr
[chunk
] = NULL
;
1391 block
->remote_keys
[chunk
] = 0;
1394 perror("unregistration chunk failed");
1397 rdma
->total_registrations
--;
1399 reg
.key
.chunk
= chunk
;
1400 register_to_network(rdma
, ®
);
1401 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1407 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1413 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1416 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1418 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1419 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1425 * Set bit for unregistration in the next iteration.
1426 * We cannot transmit right here, but will unpin later.
1428 static void qemu_rdma_signal_unregister(RDMAContext
*rdma
, uint64_t index
,
1429 uint64_t chunk
, uint64_t wr_id
)
1431 if (rdma
->unregistrations
[rdma
->unregister_next
] != 0) {
1432 error_report("rdma migration: queue is full");
1434 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1436 if (!test_and_set_bit(chunk
, block
->unregister_bitmap
)) {
1437 trace_qemu_rdma_signal_unregister_append(chunk
,
1438 rdma
->unregister_next
);
1440 rdma
->unregistrations
[rdma
->unregister_next
++] =
1441 qemu_rdma_make_wrid(wr_id
, index
, chunk
);
1443 if (rdma
->unregister_next
== RDMA_SIGNALED_SEND_MAX
) {
1444 rdma
->unregister_next
= 0;
1447 trace_qemu_rdma_signal_unregister_already(chunk
);
1453 * Consult the connection manager to see a work request
1454 * (of any kind) has completed.
1455 * Return the work request ID that completed.
1457 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, uint64_t *wr_id_out
,
1464 ret
= ibv_poll_cq(rdma
->cq
, 1, &wc
);
1467 *wr_id_out
= RDMA_WRID_NONE
;
1472 error_report("ibv_poll_cq return %d", ret
);
1476 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1478 if (wc
.status
!= IBV_WC_SUCCESS
) {
1479 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1480 wc
.status
, ibv_wc_status_str(wc
.status
));
1481 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1486 if (rdma
->control_ready_expected
&&
1487 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1488 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1489 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1490 rdma
->control_ready_expected
= 0;
1493 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1495 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1497 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1498 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1500 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1501 index
, chunk
, block
->local_host_addr
,
1502 (void *)(uintptr_t)block
->remote_host_addr
);
1504 clear_bit(chunk
, block
->transit_bitmap
);
1506 if (rdma
->nb_sent
> 0) {
1510 if (!rdma
->pin_all
) {
1512 * FYI: If one wanted to signal a specific chunk to be unregistered
1513 * using LRU or workload-specific information, this is the function
1514 * you would call to do so. That chunk would then get asynchronously
1515 * unregistered later.
1517 #ifdef RDMA_UNREGISTRATION_EXAMPLE
1518 qemu_rdma_signal_unregister(rdma
, index
, chunk
, wc
.wr_id
);
1522 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1525 *wr_id_out
= wc
.wr_id
;
1527 *byte_len
= wc
.byte_len
;
1533 /* Wait for activity on the completion channel.
1534 * Returns 0 on success, none-0 on error.
1536 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
)
1538 struct rdma_cm_event
*cm_event
;
1542 * Coroutine doesn't start until migration_fd_process_incoming()
1543 * so don't yield unless we know we're running inside of a coroutine.
1545 if (rdma
->migration_started_on_destination
&&
1546 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1547 yield_until_fd_readable(rdma
->comp_channel
->fd
);
1549 /* This is the source side, we're in a separate thread
1550 * or destination prior to migration_fd_process_incoming()
1551 * after postcopy, the destination also in a separate thread.
1552 * we can't yield; so we have to poll the fd.
1553 * But we need to be able to handle 'cancel' or an error
1554 * without hanging forever.
1556 while (!rdma
->error_state
&& !rdma
->received_error
) {
1558 pfds
[0].fd
= rdma
->comp_channel
->fd
;
1559 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1560 pfds
[0].revents
= 0;
1562 pfds
[1].fd
= rdma
->channel
->fd
;
1563 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1564 pfds
[1].revents
= 0;
1566 /* 0.1s timeout, should be fine for a 'cancel' */
1567 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1569 case 1: /* fd active */
1570 if (pfds
[0].revents
) {
1574 if (pfds
[1].revents
) {
1575 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1577 error_report("failed to get cm event while wait "
1578 "completion channel");
1582 error_report("receive cm event while wait comp channel,"
1583 "cm event is %d", cm_event
->event
);
1584 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1585 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1586 rdma_ack_cm_event(cm_event
);
1589 rdma_ack_cm_event(cm_event
);
1593 case 0: /* Timeout, go around again */
1596 default: /* Error of some type -
1597 * I don't trust errno from qemu_poll_ns
1599 error_report("%s: poll failed", __func__
);
1603 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1604 /* Bail out and let the cancellation happen */
1610 if (rdma
->received_error
) {
1613 return rdma
->error_state
;
1617 * Block until the next work request has completed.
1619 * First poll to see if a work request has already completed,
1622 * If we encounter completed work requests for IDs other than
1623 * the one we're interested in, then that's generally an error.
1625 * The only exception is actual RDMA Write completions. These
1626 * completions only need to be recorded, but do not actually
1627 * need further processing.
1629 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1632 int num_cq_events
= 0, ret
= 0;
1635 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1637 if (ibv_req_notify_cq(rdma
->cq
, 0)) {
1641 while (wr_id
!= wrid_requested
) {
1642 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1647 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1649 if (wr_id
== RDMA_WRID_NONE
) {
1652 if (wr_id
!= wrid_requested
) {
1653 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1654 wrid_requested
, print_wrid(wr_id
), wr_id
);
1658 if (wr_id
== wrid_requested
) {
1663 ret
= qemu_rdma_wait_comp_channel(rdma
);
1665 goto err_block_for_wrid
;
1668 ret
= ibv_get_cq_event(rdma
->comp_channel
, &cq
, &cq_ctx
);
1670 perror("ibv_get_cq_event");
1671 goto err_block_for_wrid
;
1676 ret
= -ibv_req_notify_cq(cq
, 0);
1678 goto err_block_for_wrid
;
1681 while (wr_id
!= wrid_requested
) {
1682 ret
= qemu_rdma_poll(rdma
, &wr_id_in
, byte_len
);
1684 goto err_block_for_wrid
;
1687 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1689 if (wr_id
== RDMA_WRID_NONE
) {
1692 if (wr_id
!= wrid_requested
) {
1693 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1694 wrid_requested
, print_wrid(wr_id
), wr_id
);
1698 if (wr_id
== wrid_requested
) {
1699 goto success_block_for_wrid
;
1703 success_block_for_wrid
:
1704 if (num_cq_events
) {
1705 ibv_ack_cq_events(cq
, num_cq_events
);
1710 if (num_cq_events
) {
1711 ibv_ack_cq_events(cq
, num_cq_events
);
1714 rdma
->error_state
= ret
;
1719 * Post a SEND message work request for the control channel
1720 * containing some data and block until the post completes.
1722 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1723 RDMAControlHeader
*head
)
1726 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1727 struct ibv_send_wr
*bad_wr
;
1728 struct ibv_sge sge
= {
1729 .addr
= (uintptr_t)(wr
->control
),
1730 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1731 .lkey
= wr
->control_mr
->lkey
,
1733 struct ibv_send_wr send_wr
= {
1734 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1735 .opcode
= IBV_WR_SEND
,
1736 .send_flags
= IBV_SEND_SIGNALED
,
1741 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1744 * We don't actually need to do a memcpy() in here if we used
1745 * the "sge" properly, but since we're only sending control messages
1746 * (not RAM in a performance-critical path), then its OK for now.
1748 * The copy makes the RDMAControlHeader simpler to manipulate
1749 * for the time being.
1751 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1752 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1753 control_to_network((void *) wr
->control
);
1756 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1760 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1763 error_report("Failed to use post IB SEND for control");
1767 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1769 error_report("rdma migration: send polling control error");
1776 * Post a RECV work request in anticipation of some future receipt
1777 * of data on the control channel.
1779 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1781 struct ibv_recv_wr
*bad_wr
;
1782 struct ibv_sge sge
= {
1783 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1784 .length
= RDMA_CONTROL_MAX_BUFFER
,
1785 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1788 struct ibv_recv_wr recv_wr
= {
1789 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1795 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1803 * Block and wait for a RECV control channel message to arrive.
1805 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1806 RDMAControlHeader
*head
, int expecting
, int idx
)
1809 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1813 error_report("rdma migration: recv polling control error!");
1817 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1818 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1820 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1822 if (expecting
== RDMA_CONTROL_NONE
) {
1823 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1825 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1826 error_report("Was expecting a %s (%d) control message"
1827 ", but got: %s (%d), length: %d",
1828 control_desc(expecting
), expecting
,
1829 control_desc(head
->type
), head
->type
, head
->len
);
1830 if (head
->type
== RDMA_CONTROL_ERROR
) {
1831 rdma
->received_error
= true;
1835 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1836 error_report("too long length: %d", head
->len
);
1839 if (sizeof(*head
) + head
->len
!= byte_len
) {
1840 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1848 * When a RECV work request has completed, the work request's
1849 * buffer is pointed at the header.
1851 * This will advance the pointer to the data portion
1852 * of the control message of the work request's buffer that
1853 * was populated after the work request finished.
1855 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1856 RDMAControlHeader
*head
)
1858 rdma
->wr_data
[idx
].control_len
= head
->len
;
1859 rdma
->wr_data
[idx
].control_curr
=
1860 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1864 * This is an 'atomic' high-level operation to deliver a single, unified
1865 * control-channel message.
1867 * Additionally, if the user is expecting some kind of reply to this message,
1868 * they can request a 'resp' response message be filled in by posting an
1869 * additional work request on behalf of the user and waiting for an additional
1872 * The extra (optional) response is used during registration to us from having
1873 * to perform an *additional* exchange of message just to provide a response by
1874 * instead piggy-backing on the acknowledgement.
1876 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1877 uint8_t *data
, RDMAControlHeader
*resp
,
1879 int (*callback
)(RDMAContext
*rdma
))
1884 * Wait until the dest is ready before attempting to deliver the message
1885 * by waiting for a READY message.
1887 if (rdma
->control_ready_expected
) {
1888 RDMAControlHeader resp
;
1889 ret
= qemu_rdma_exchange_get_response(rdma
,
1890 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1897 * If the user is expecting a response, post a WR in anticipation of it.
1900 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1902 error_report("rdma migration: error posting"
1903 " extra control recv for anticipated result!");
1909 * Post a WR to replace the one we just consumed for the READY message.
1911 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1913 error_report("rdma migration: error posting first control recv!");
1918 * Deliver the control message that was requested.
1920 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1923 error_report("Failed to send control buffer!");
1928 * If we're expecting a response, block and wait for it.
1932 trace_qemu_rdma_exchange_send_issue_callback();
1933 ret
= callback(rdma
);
1939 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1940 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1941 resp
->type
, RDMA_WRID_DATA
);
1947 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1949 *resp_idx
= RDMA_WRID_DATA
;
1951 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1954 rdma
->control_ready_expected
= 1;
1960 * This is an 'atomic' high-level operation to receive a single, unified
1961 * control-channel message.
1963 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1966 RDMAControlHeader ready
= {
1968 .type
= RDMA_CONTROL_READY
,
1974 * Inform the source that we're ready to receive a message.
1976 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1979 error_report("Failed to send control buffer!");
1984 * Block and wait for the message.
1986 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
1987 expecting
, RDMA_WRID_READY
);
1993 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
1996 * Post a new RECV work request to replace the one we just consumed.
1998 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2000 error_report("rdma migration: error posting second control recv!");
2008 * Write an actual chunk of memory using RDMA.
2010 * If we're using dynamic registration on the dest-side, we have to
2011 * send a registration command first.
2013 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
2014 int current_index
, uint64_t current_addr
,
2018 struct ibv_send_wr send_wr
= { 0 };
2019 struct ibv_send_wr
*bad_wr
;
2020 int reg_result_idx
, ret
, count
= 0;
2021 uint64_t chunk
, chunks
;
2022 uint8_t *chunk_start
, *chunk_end
;
2023 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
2025 RDMARegisterResult
*reg_result
;
2026 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
2027 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
2028 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
2033 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
2034 (current_addr
- block
->offset
));
2035 sge
.length
= length
;
2037 chunk
= ram_chunk_index(block
->local_host_addr
,
2038 (uint8_t *)(uintptr_t)sge
.addr
);
2039 chunk_start
= ram_chunk_start(block
, chunk
);
2041 if (block
->is_ram_block
) {
2042 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2044 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2048 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2050 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2055 trace_qemu_rdma_write_one_top(chunks
+ 1,
2057 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2059 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2061 if (!rdma
->pin_all
) {
2062 #ifdef RDMA_UNREGISTRATION_EXAMPLE
2063 qemu_rdma_unregister_waiting(rdma
);
2067 while (test_bit(chunk
, block
->transit_bitmap
)) {
2069 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2070 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2072 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2075 error_report("Failed to Wait for previous write to complete "
2076 "block %d chunk %" PRIu64
2077 " current %" PRIu64
" len %" PRIu64
" %d",
2078 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2083 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2084 if (!block
->remote_keys
[chunk
]) {
2086 * This chunk has not yet been registered, so first check to see
2087 * if the entire chunk is zero. If so, tell the other size to
2088 * memset() + madvise() the entire chunk without RDMA.
2091 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2092 RDMACompress comp
= {
2093 .offset
= current_addr
,
2095 .block_idx
= current_index
,
2099 head
.len
= sizeof(comp
);
2100 head
.type
= RDMA_CONTROL_COMPRESS
;
2102 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2103 current_index
, current_addr
);
2105 compress_to_network(rdma
, &comp
);
2106 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2107 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2113 acct_update_position(f
, sge
.length
, true);
2119 * Otherwise, tell other side to register.
2121 reg
.current_index
= current_index
;
2122 if (block
->is_ram_block
) {
2123 reg
.key
.current_addr
= current_addr
;
2125 reg
.key
.chunk
= chunk
;
2127 reg
.chunks
= chunks
;
2129 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2132 register_to_network(rdma
, ®
);
2133 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2134 &resp
, ®_result_idx
, NULL
);
2139 /* try to overlap this single registration with the one we sent. */
2140 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2141 &sge
.lkey
, NULL
, chunk
,
2142 chunk_start
, chunk_end
)) {
2143 error_report("cannot get lkey");
2147 reg_result
= (RDMARegisterResult
*)
2148 rdma
->wr_data
[reg_result_idx
].control_curr
;
2150 network_to_result(reg_result
);
2152 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2153 reg_result
->rkey
, chunk
);
2155 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2156 block
->remote_host_addr
= reg_result
->host_addr
;
2158 /* already registered before */
2159 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2160 &sge
.lkey
, NULL
, chunk
,
2161 chunk_start
, chunk_end
)) {
2162 error_report("cannot get lkey!");
2167 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2169 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2171 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2172 &sge
.lkey
, NULL
, chunk
,
2173 chunk_start
, chunk_end
)) {
2174 error_report("cannot get lkey!");
2180 * Encode the ram block index and chunk within this wrid.
2181 * We will use this information at the time of completion
2182 * to figure out which bitmap to check against and then which
2183 * chunk in the bitmap to look for.
2185 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2186 current_index
, chunk
);
2188 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2189 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2190 send_wr
.sg_list
= &sge
;
2191 send_wr
.num_sge
= 1;
2192 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2193 (current_addr
- block
->offset
);
2195 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2199 * ibv_post_send() does not return negative error numbers,
2200 * per the specification they are positive - no idea why.
2202 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2204 if (ret
== ENOMEM
) {
2205 trace_qemu_rdma_write_one_queue_full();
2206 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2208 error_report("rdma migration: failed to make "
2209 "room in full send queue! %d", ret
);
2215 } else if (ret
> 0) {
2216 perror("rdma migration: post rdma write failed");
2220 set_bit(chunk
, block
->transit_bitmap
);
2221 acct_update_position(f
, sge
.length
, false);
2222 rdma
->total_writes
++;
2228 * Push out any unwritten RDMA operations.
2230 * We support sending out multiple chunks at the same time.
2231 * Not all of them need to get signaled in the completion queue.
2233 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2237 if (!rdma
->current_length
) {
2241 ret
= qemu_rdma_write_one(f
, rdma
,
2242 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2250 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2253 rdma
->current_length
= 0;
2254 rdma
->current_addr
= 0;
2259 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2260 uint64_t offset
, uint64_t len
)
2262 RDMALocalBlock
*block
;
2266 if (rdma
->current_index
< 0) {
2270 if (rdma
->current_chunk
< 0) {
2274 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2275 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2276 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2278 if (rdma
->current_length
== 0) {
2283 * Only merge into chunk sequentially.
2285 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2289 if (offset
< block
->offset
) {
2293 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2297 if ((host_addr
+ len
) > chunk_end
) {
2305 * We're not actually writing here, but doing three things:
2307 * 1. Identify the chunk the buffer belongs to.
2308 * 2. If the chunk is full or the buffer doesn't belong to the current
2309 * chunk, then start a new chunk and flush() the old chunk.
2310 * 3. To keep the hardware busy, we also group chunks into batches
2311 * and only require that a batch gets acknowledged in the completion
2312 * queue instead of each individual chunk.
2314 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2315 uint64_t block_offset
, uint64_t offset
,
2318 uint64_t current_addr
= block_offset
+ offset
;
2319 uint64_t index
= rdma
->current_index
;
2320 uint64_t chunk
= rdma
->current_chunk
;
2323 /* If we cannot merge it, we flush the current buffer first. */
2324 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2325 ret
= qemu_rdma_write_flush(f
, rdma
);
2329 rdma
->current_length
= 0;
2330 rdma
->current_addr
= current_addr
;
2332 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2333 offset
, len
, &index
, &chunk
);
2335 error_report("ram block search failed");
2338 rdma
->current_index
= index
;
2339 rdma
->current_chunk
= chunk
;
2343 rdma
->current_length
+= len
;
2345 /* flush it if buffer is too large */
2346 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2347 return qemu_rdma_write_flush(f
, rdma
);
2353 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2357 if (rdma
->cm_id
&& rdma
->connected
) {
2358 if ((rdma
->error_state
||
2359 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2360 !rdma
->received_error
) {
2361 RDMAControlHeader head
= { .len
= 0,
2362 .type
= RDMA_CONTROL_ERROR
,
2365 error_report("Early error. Sending error.");
2366 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2369 rdma_disconnect(rdma
->cm_id
);
2370 trace_qemu_rdma_cleanup_disconnect();
2371 rdma
->connected
= false;
2374 if (rdma
->channel
) {
2375 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2377 g_free(rdma
->dest_blocks
);
2378 rdma
->dest_blocks
= NULL
;
2380 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2381 if (rdma
->wr_data
[idx
].control_mr
) {
2382 rdma
->total_registrations
--;
2383 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2385 rdma
->wr_data
[idx
].control_mr
= NULL
;
2388 if (rdma
->local_ram_blocks
.block
) {
2389 while (rdma
->local_ram_blocks
.nb_blocks
) {
2390 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2395 rdma_destroy_qp(rdma
->cm_id
);
2399 ibv_destroy_cq(rdma
->cq
);
2402 if (rdma
->comp_channel
) {
2403 ibv_destroy_comp_channel(rdma
->comp_channel
);
2404 rdma
->comp_channel
= NULL
;
2407 ibv_dealloc_pd(rdma
->pd
);
2411 rdma_destroy_id(rdma
->cm_id
);
2415 /* the destination side, listen_id and channel is shared */
2416 if (rdma
->listen_id
) {
2417 if (!rdma
->is_return_path
) {
2418 rdma_destroy_id(rdma
->listen_id
);
2420 rdma
->listen_id
= NULL
;
2422 if (rdma
->channel
) {
2423 if (!rdma
->is_return_path
) {
2424 rdma_destroy_event_channel(rdma
->channel
);
2426 rdma
->channel
= NULL
;
2430 if (rdma
->channel
) {
2431 rdma_destroy_event_channel(rdma
->channel
);
2432 rdma
->channel
= NULL
;
2435 g_free(rdma
->host_port
);
2437 rdma
->host_port
= NULL
;
2441 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2444 Error
*local_err
= NULL
, **temp
= &local_err
;
2447 * Will be validated against destination's actual capabilities
2448 * after the connect() completes.
2450 rdma
->pin_all
= pin_all
;
2452 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2454 goto err_rdma_source_init
;
2457 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2459 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2460 " limits may be too low. Please check $ ulimit -a # and "
2461 "search for 'ulimit -l' in the output");
2462 goto err_rdma_source_init
;
2465 ret
= qemu_rdma_alloc_qp(rdma
);
2467 ERROR(temp
, "rdma migration: error allocating qp!");
2468 goto err_rdma_source_init
;
2471 ret
= qemu_rdma_init_ram_blocks(rdma
);
2473 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2474 goto err_rdma_source_init
;
2477 /* Build the hash that maps from offset to RAMBlock */
2478 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2479 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2480 g_hash_table_insert(rdma
->blockmap
,
2481 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2482 &rdma
->local_ram_blocks
.block
[idx
]);
2485 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2486 ret
= qemu_rdma_reg_control(rdma
, idx
);
2488 ERROR(temp
, "rdma migration: error registering %d control!",
2490 goto err_rdma_source_init
;
2496 err_rdma_source_init
:
2497 error_propagate(errp
, local_err
);
2498 qemu_rdma_cleanup(rdma
);
2502 static int qemu_get_cm_event_timeout(RDMAContext
*rdma
,
2503 struct rdma_cm_event
**cm_event
,
2504 long msec
, Error
**errp
)
2507 struct pollfd poll_fd
= {
2508 .fd
= rdma
->channel
->fd
,
2514 ret
= poll(&poll_fd
, 1, msec
);
2515 } while (ret
< 0 && errno
== EINTR
);
2518 ERROR(errp
, "poll cm event timeout");
2520 } else if (ret
< 0) {
2521 ERROR(errp
, "failed to poll cm event, errno=%i", errno
);
2523 } else if (poll_fd
.revents
& POLLIN
) {
2524 return rdma_get_cm_event(rdma
->channel
, cm_event
);
2526 ERROR(errp
, "no POLLIN event, revent=%x", poll_fd
.revents
);
2531 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
, bool return_path
)
2533 RDMACapabilities cap
= {
2534 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2537 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2539 .private_data
= &cap
,
2540 .private_data_len
= sizeof(cap
),
2542 struct rdma_cm_event
*cm_event
;
2546 * Only negotiate the capability with destination if the user
2547 * on the source first requested the capability.
2549 if (rdma
->pin_all
) {
2550 trace_qemu_rdma_connect_pin_all_requested();
2551 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2554 caps_to_network(&cap
);
2556 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2558 ERROR(errp
, "posting second control recv");
2559 goto err_rdma_source_connect
;
2562 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2564 perror("rdma_connect");
2565 ERROR(errp
, "connecting to destination!");
2566 goto err_rdma_source_connect
;
2570 ret
= qemu_get_cm_event_timeout(rdma
, &cm_event
, 5000, errp
);
2572 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2575 perror("rdma_get_cm_event after rdma_connect");
2576 ERROR(errp
, "connecting to destination!");
2577 goto err_rdma_source_connect
;
2580 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2581 error_report("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2582 ERROR(errp
, "connecting to destination!");
2583 rdma_ack_cm_event(cm_event
);
2584 goto err_rdma_source_connect
;
2586 rdma
->connected
= true;
2588 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2589 network_to_caps(&cap
);
2592 * Verify that the *requested* capabilities are supported by the destination
2593 * and disable them otherwise.
2595 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2596 ERROR(errp
, "Server cannot support pinning all memory. "
2597 "Will register memory dynamically.");
2598 rdma
->pin_all
= false;
2601 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2603 rdma_ack_cm_event(cm_event
);
2605 rdma
->control_ready_expected
= 1;
2609 err_rdma_source_connect
:
2610 qemu_rdma_cleanup(rdma
);
2614 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2617 struct rdma_cm_id
*listen_id
;
2618 char ip
[40] = "unknown";
2619 struct rdma_addrinfo
*res
, *e
;
2622 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2623 rdma
->wr_data
[idx
].control_len
= 0;
2624 rdma
->wr_data
[idx
].control_curr
= NULL
;
2627 if (!rdma
->host
|| !rdma
->host
[0]) {
2628 ERROR(errp
, "RDMA host is not set!");
2629 rdma
->error_state
= -EINVAL
;
2632 /* create CM channel */
2633 rdma
->channel
= rdma_create_event_channel();
2634 if (!rdma
->channel
) {
2635 ERROR(errp
, "could not create rdma event channel");
2636 rdma
->error_state
= -EINVAL
;
2641 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2643 ERROR(errp
, "could not create cm_id!");
2644 goto err_dest_init_create_listen_id
;
2647 snprintf(port_str
, 16, "%d", rdma
->port
);
2648 port_str
[15] = '\0';
2650 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2652 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2653 goto err_dest_init_bind_addr
;
2656 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2657 inet_ntop(e
->ai_family
,
2658 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2659 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2660 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2664 if (e
->ai_family
== AF_INET6
) {
2665 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2673 rdma_freeaddrinfo(res
);
2675 ERROR(errp
, "Error: could not rdma_bind_addr!");
2676 goto err_dest_init_bind_addr
;
2679 rdma
->listen_id
= listen_id
;
2680 qemu_rdma_dump_gid("dest_init", listen_id
);
2683 err_dest_init_bind_addr
:
2684 rdma_destroy_id(listen_id
);
2685 err_dest_init_create_listen_id
:
2686 rdma_destroy_event_channel(rdma
->channel
);
2687 rdma
->channel
= NULL
;
2688 rdma
->error_state
= ret
;
2693 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2698 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2699 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2700 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2703 /*the CM channel and CM id is shared*/
2704 rdma_return_path
->channel
= rdma
->channel
;
2705 rdma_return_path
->listen_id
= rdma
->listen_id
;
2707 rdma
->return_path
= rdma_return_path
;
2708 rdma_return_path
->return_path
= rdma
;
2709 rdma_return_path
->is_return_path
= true;
2712 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2714 RDMAContext
*rdma
= NULL
;
2715 InetSocketAddress
*addr
;
2718 rdma
= g_new0(RDMAContext
, 1);
2719 rdma
->current_index
= -1;
2720 rdma
->current_chunk
= -1;
2722 addr
= g_new(InetSocketAddress
, 1);
2723 if (!inet_parse(addr
, host_port
, NULL
)) {
2724 rdma
->port
= atoi(addr
->port
);
2725 rdma
->host
= g_strdup(addr
->host
);
2726 rdma
->host_port
= g_strdup(host_port
);
2728 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2733 qapi_free_InetSocketAddress(addr
);
2740 * QEMUFile interface to the control channel.
2741 * SEND messages for control only.
2742 * VM's ram is handled with regular RDMA messages.
2744 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2745 const struct iovec
*iov
,
2751 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2752 QEMUFile
*f
= rioc
->file
;
2759 RCU_READ_LOCK_GUARD();
2760 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
2766 CHECK_ERROR_STATE();
2769 * Push out any writes that
2770 * we're queued up for VM's ram.
2772 ret
= qemu_rdma_write_flush(f
, rdma
);
2774 rdma
->error_state
= ret
;
2778 for (i
= 0; i
< niov
; i
++) {
2779 size_t remaining
= iov
[i
].iov_len
;
2780 uint8_t * data
= (void *)iov
[i
].iov_base
;
2782 RDMAControlHeader head
;
2784 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2788 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2790 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2793 rdma
->error_state
= ret
;
2805 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2806 size_t size
, int idx
)
2810 if (rdma
->wr_data
[idx
].control_len
) {
2811 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2813 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2814 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2815 rdma
->wr_data
[idx
].control_curr
+= len
;
2816 rdma
->wr_data
[idx
].control_len
-= len
;
2823 * QEMUFile interface to the control channel.
2824 * RDMA links don't use bytestreams, so we have to
2825 * return bytes to QEMUFile opportunistically.
2827 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2828 const struct iovec
*iov
,
2834 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2836 RDMAControlHeader head
;
2841 RCU_READ_LOCK_GUARD();
2842 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
2848 CHECK_ERROR_STATE();
2850 for (i
= 0; i
< niov
; i
++) {
2851 size_t want
= iov
[i
].iov_len
;
2852 uint8_t *data
= (void *)iov
[i
].iov_base
;
2855 * First, we hold on to the last SEND message we
2856 * were given and dish out the bytes until we run
2859 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2862 /* Got what we needed, so go to next iovec */
2867 /* If we got any data so far, then don't wait
2868 * for more, just return what we have */
2874 /* We've got nothing at all, so lets wait for
2877 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2880 rdma
->error_state
= ret
;
2885 * SEND was received with new bytes, now try again.
2887 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2891 /* Still didn't get enough, so lets just return */
2894 return QIO_CHANNEL_ERR_BLOCK
;
2904 * Block until all the outstanding chunks have been delivered by the hardware.
2906 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2910 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2914 while (rdma
->nb_sent
) {
2915 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2917 error_report("rdma migration: complete polling error!");
2922 qemu_rdma_unregister_waiting(rdma
);
2928 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2932 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2933 /* XXX we should make readv/writev actually honour this :-) */
2934 rioc
->blocking
= blocking
;
2939 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2940 struct QIOChannelRDMASource
{
2942 QIOChannelRDMA
*rioc
;
2943 GIOCondition condition
;
2947 qio_channel_rdma_source_prepare(GSource
*source
,
2950 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2952 GIOCondition cond
= 0;
2955 RCU_READ_LOCK_GUARD();
2956 if (rsource
->condition
== G_IO_IN
) {
2957 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
2959 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
2963 error_report("RDMAContext is NULL when prepare Gsource");
2967 if (rdma
->wr_data
[0].control_len
) {
2972 return cond
& rsource
->condition
;
2976 qio_channel_rdma_source_check(GSource
*source
)
2978 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2980 GIOCondition cond
= 0;
2982 RCU_READ_LOCK_GUARD();
2983 if (rsource
->condition
== G_IO_IN
) {
2984 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
2986 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
2990 error_report("RDMAContext is NULL when check Gsource");
2994 if (rdma
->wr_data
[0].control_len
) {
2999 return cond
& rsource
->condition
;
3003 qio_channel_rdma_source_dispatch(GSource
*source
,
3004 GSourceFunc callback
,
3007 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
3008 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3010 GIOCondition cond
= 0;
3012 RCU_READ_LOCK_GUARD();
3013 if (rsource
->condition
== G_IO_IN
) {
3014 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3016 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3020 error_report("RDMAContext is NULL when dispatch Gsource");
3024 if (rdma
->wr_data
[0].control_len
) {
3029 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
3030 (cond
& rsource
->condition
),
3035 qio_channel_rdma_source_finalize(GSource
*source
)
3037 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
3039 object_unref(OBJECT(ssource
->rioc
));
3042 GSourceFuncs qio_channel_rdma_source_funcs
= {
3043 qio_channel_rdma_source_prepare
,
3044 qio_channel_rdma_source_check
,
3045 qio_channel_rdma_source_dispatch
,
3046 qio_channel_rdma_source_finalize
3049 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
3050 GIOCondition condition
)
3052 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3053 QIOChannelRDMASource
*ssource
;
3056 source
= g_source_new(&qio_channel_rdma_source_funcs
,
3057 sizeof(QIOChannelRDMASource
));
3058 ssource
= (QIOChannelRDMASource
*)source
;
3060 ssource
->rioc
= rioc
;
3061 object_ref(OBJECT(rioc
));
3063 ssource
->condition
= condition
;
3068 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
3071 IOHandler
*io_write
,
3074 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3076 aio_set_fd_handler(ctx
, rioc
->rdmain
->comp_channel
->fd
,
3077 false, io_read
, io_write
, NULL
, opaque
);
3079 aio_set_fd_handler(ctx
, rioc
->rdmaout
->comp_channel
->fd
,
3080 false, io_read
, io_write
, NULL
, opaque
);
3084 struct rdma_close_rcu
{
3085 struct rcu_head rcu
;
3086 RDMAContext
*rdmain
;
3087 RDMAContext
*rdmaout
;
3090 /* callback from qio_channel_rdma_close via call_rcu */
3091 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu
*rcu
)
3094 qemu_rdma_cleanup(rcu
->rdmain
);
3098 qemu_rdma_cleanup(rcu
->rdmaout
);
3101 g_free(rcu
->rdmain
);
3102 g_free(rcu
->rdmaout
);
3106 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3109 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3110 RDMAContext
*rdmain
, *rdmaout
;
3111 struct rdma_close_rcu
*rcu
= g_new(struct rdma_close_rcu
, 1);
3113 trace_qemu_rdma_close();
3115 rdmain
= rioc
->rdmain
;
3117 qatomic_rcu_set(&rioc
->rdmain
, NULL
);
3120 rdmaout
= rioc
->rdmaout
;
3122 qatomic_rcu_set(&rioc
->rdmaout
, NULL
);
3125 rcu
->rdmain
= rdmain
;
3126 rcu
->rdmaout
= rdmaout
;
3127 call_rcu(rcu
, qio_channel_rdma_close_rcu
, rcu
);
3133 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3134 QIOChannelShutdown how
,
3137 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3138 RDMAContext
*rdmain
, *rdmaout
;
3140 RCU_READ_LOCK_GUARD();
3142 rdmain
= qatomic_rcu_read(&rioc
->rdmain
);
3143 rdmaout
= qatomic_rcu_read(&rioc
->rdmain
);
3146 case QIO_CHANNEL_SHUTDOWN_READ
:
3148 rdmain
->error_state
= -1;
3151 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3153 rdmaout
->error_state
= -1;
3156 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3159 rdmain
->error_state
= -1;
3162 rdmaout
->error_state
= -1;
3173 * This means that 'block_offset' is a full virtual address that does not
3174 * belong to a RAMBlock of the virtual machine and instead
3175 * represents a private malloc'd memory area that the caller wishes to
3179 * Offset is an offset to be added to block_offset and used
3180 * to also lookup the corresponding RAMBlock.
3183 * Initiate an transfer this size.
3186 * A 'hint' or 'advice' that means that we wish to speculatively
3187 * and asynchronously unregister this memory. In this case, there is no
3188 * guarantee that the unregister will actually happen, for example,
3189 * if the memory is being actively transmitted. Additionally, the memory
3190 * may be re-registered at any future time if a write within the same
3191 * chunk was requested again, even if you attempted to unregister it
3194 * @size < 0 : TODO, not yet supported
3195 * Unregister the memory NOW. This means that the caller does not
3196 * expect there to be any future RDMA transfers and we just want to clean
3197 * things up. This is used in case the upper layer owns the memory and
3198 * cannot wait for qemu_fclose() to occur.
3200 * @bytes_sent : User-specificed pointer to indicate how many bytes were
3201 * sent. Usually, this will not be more than a few bytes of
3202 * the protocol because most transfers are sent asynchronously.
3204 static size_t qemu_rdma_save_page(QEMUFile
*f
, void *opaque
,
3205 ram_addr_t block_offset
, ram_addr_t offset
,
3206 size_t size
, uint64_t *bytes_sent
)
3208 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3212 RCU_READ_LOCK_GUARD();
3213 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3219 CHECK_ERROR_STATE();
3221 if (migration_in_postcopy()) {
3222 return RAM_SAVE_CONTROL_NOT_SUPP
;
3229 * Add this page to the current 'chunk'. If the chunk
3230 * is full, or the page doesn't belong to the current chunk,
3231 * an actual RDMA write will occur and a new chunk will be formed.
3233 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
3235 error_report("rdma migration: write error! %d", ret
);
3240 * We always return 1 bytes because the RDMA
3241 * protocol is completely asynchronous. We do not yet know
3242 * whether an identified chunk is zero or not because we're
3243 * waiting for other pages to potentially be merged with
3244 * the current chunk. So, we have to call qemu_update_position()
3245 * later on when the actual write occurs.
3251 uint64_t index
, chunk
;
3253 /* TODO: Change QEMUFileOps prototype to be signed: size_t => long
3255 ret = qemu_rdma_drain_cq(f, rdma);
3257 fprintf(stderr, "rdma: failed to synchronously drain"
3258 " completion queue before unregistration.\n");
3264 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
3265 offset
, size
, &index
, &chunk
);
3268 error_report("ram block search failed");
3272 qemu_rdma_signal_unregister(rdma
, index
, chunk
, 0);
3275 * TODO: Synchronous, guaranteed unregistration (should not occur during
3276 * fast-path). Otherwise, unregisters will process on the next call to
3277 * qemu_rdma_drain_cq()
3279 qemu_rdma_unregister_waiting(rdma);
3285 * Drain the Completion Queue if possible, but do not block,
3288 * If nothing to poll, the end of the iteration will do this
3289 * again to make sure we don't overflow the request queue.
3292 uint64_t wr_id
, wr_id_in
;
3293 int ret
= qemu_rdma_poll(rdma
, &wr_id_in
, NULL
);
3295 error_report("rdma migration: polling error! %d", ret
);
3299 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3301 if (wr_id
== RDMA_WRID_NONE
) {
3306 return RAM_SAVE_CONTROL_DELAYED
;
3308 rdma
->error_state
= ret
;
3312 static void rdma_accept_incoming_migration(void *opaque
);
3314 static void rdma_cm_poll_handler(void *opaque
)
3316 RDMAContext
*rdma
= opaque
;
3318 struct rdma_cm_event
*cm_event
;
3319 MigrationIncomingState
*mis
= migration_incoming_get_current();
3321 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3323 error_report("get_cm_event failed %d", errno
);
3327 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3328 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3329 if (!rdma
->error_state
&&
3330 migration_incoming_get_current()->state
!=
3331 MIGRATION_STATUS_COMPLETED
) {
3332 error_report("receive cm event, cm event is %d", cm_event
->event
);
3333 rdma
->error_state
= -EPIPE
;
3334 if (rdma
->return_path
) {
3335 rdma
->return_path
->error_state
= -EPIPE
;
3338 rdma_ack_cm_event(cm_event
);
3340 if (mis
->migration_incoming_co
) {
3341 qemu_coroutine_enter(mis
->migration_incoming_co
);
3345 rdma_ack_cm_event(cm_event
);
3348 static int qemu_rdma_accept(RDMAContext
*rdma
)
3350 RDMACapabilities cap
;
3351 struct rdma_conn_param conn_param
= {
3352 .responder_resources
= 2,
3353 .private_data
= &cap
,
3354 .private_data_len
= sizeof(cap
),
3356 RDMAContext
*rdma_return_path
= NULL
;
3357 struct rdma_cm_event
*cm_event
;
3358 struct ibv_context
*verbs
;
3362 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3364 goto err_rdma_dest_wait
;
3367 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3368 rdma_ack_cm_event(cm_event
);
3369 goto err_rdma_dest_wait
;
3373 * initialize the RDMAContext for return path for postcopy after first
3374 * connection request reached.
3376 if (migrate_postcopy() && !rdma
->is_return_path
) {
3377 rdma_return_path
= qemu_rdma_data_init(rdma
->host_port
, NULL
);
3378 if (rdma_return_path
== NULL
) {
3379 rdma_ack_cm_event(cm_event
);
3380 goto err_rdma_dest_wait
;
3383 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
3386 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3388 network_to_caps(&cap
);
3390 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3391 error_report("Unknown source RDMA version: %d, bailing...",
3393 rdma_ack_cm_event(cm_event
);
3394 goto err_rdma_dest_wait
;
3398 * Respond with only the capabilities this version of QEMU knows about.
3400 cap
.flags
&= known_capabilities
;
3403 * Enable the ones that we do know about.
3404 * Add other checks here as new ones are introduced.
3406 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3407 rdma
->pin_all
= true;
3410 rdma
->cm_id
= cm_event
->id
;
3411 verbs
= cm_event
->id
->verbs
;
3413 rdma_ack_cm_event(cm_event
);
3415 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3417 caps_to_network(&cap
);
3419 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3422 rdma
->verbs
= verbs
;
3423 } else if (rdma
->verbs
!= verbs
) {
3424 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3426 goto err_rdma_dest_wait
;
3429 qemu_rdma_dump_id("dest_init", verbs
);
3431 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3433 error_report("rdma migration: error allocating pd and cq!");
3434 goto err_rdma_dest_wait
;
3437 ret
= qemu_rdma_alloc_qp(rdma
);
3439 error_report("rdma migration: error allocating qp!");
3440 goto err_rdma_dest_wait
;
3443 ret
= qemu_rdma_init_ram_blocks(rdma
);
3445 error_report("rdma migration: error initializing ram blocks!");
3446 goto err_rdma_dest_wait
;
3449 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3450 ret
= qemu_rdma_reg_control(rdma
, idx
);
3452 error_report("rdma: error registering %d control", idx
);
3453 goto err_rdma_dest_wait
;
3457 /* Accept the second connection request for return path */
3458 if (migrate_postcopy() && !rdma
->is_return_path
) {
3459 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3461 (void *)(intptr_t)rdma
->return_path
);
3463 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3467 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3469 error_report("rdma_accept returns %d", ret
);
3470 goto err_rdma_dest_wait
;
3473 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3475 error_report("rdma_accept get_cm_event failed %d", ret
);
3476 goto err_rdma_dest_wait
;
3479 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3480 error_report("rdma_accept not event established");
3481 rdma_ack_cm_event(cm_event
);
3482 goto err_rdma_dest_wait
;
3485 rdma_ack_cm_event(cm_event
);
3486 rdma
->connected
= true;
3488 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3490 error_report("rdma migration: error posting second control recv");
3491 goto err_rdma_dest_wait
;
3494 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3499 rdma
->error_state
= ret
;
3500 qemu_rdma_cleanup(rdma
);
3501 g_free(rdma_return_path
);
3505 static int dest_ram_sort_func(const void *a
, const void *b
)
3507 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3508 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3510 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3514 * During each iteration of the migration, we listen for instructions
3515 * by the source VM to perform dynamic page registrations before they
3516 * can perform RDMA operations.
3518 * We respond with the 'rkey'.
3520 * Keep doing this until the source tells us to stop.
3522 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
)
3524 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3525 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3528 RDMAControlHeader unreg_resp
= { .len
= 0,
3529 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3532 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3534 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3536 RDMALocalBlocks
*local
;
3537 RDMAControlHeader head
;
3538 RDMARegister
*reg
, *registers
;
3540 RDMARegisterResult
*reg_result
;
3541 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3542 RDMALocalBlock
*block
;
3549 RCU_READ_LOCK_GUARD();
3550 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3556 CHECK_ERROR_STATE();
3558 local
= &rdma
->local_ram_blocks
;
3560 trace_qemu_rdma_registration_handle_wait();
3562 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3568 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3569 error_report("rdma: Too many requests in this message (%d)."
3570 "Bailing.", head
.repeat
);
3575 switch (head
.type
) {
3576 case RDMA_CONTROL_COMPRESS
:
3577 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3578 network_to_compress(comp
);
3580 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3583 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3584 error_report("rdma: 'compress' bad block index %u (vs %d)",
3585 (unsigned int)comp
->block_idx
,
3586 rdma
->local_ram_blocks
.nb_blocks
);
3590 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3592 host_addr
= block
->local_host_addr
+
3593 (comp
->offset
- block
->offset
);
3595 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3598 case RDMA_CONTROL_REGISTER_FINISHED
:
3599 trace_qemu_rdma_registration_handle_finished();
3602 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3603 trace_qemu_rdma_registration_handle_ram_blocks();
3605 /* Sort our local RAM Block list so it's the same as the source,
3606 * we can do this since we've filled in a src_index in the list
3607 * as we received the RAMBlock list earlier.
3609 qsort(rdma
->local_ram_blocks
.block
,
3610 rdma
->local_ram_blocks
.nb_blocks
,
3611 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3612 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3613 local
->block
[i
].index
= i
;
3616 if (rdma
->pin_all
) {
3617 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3619 error_report("rdma migration: error dest "
3620 "registering ram blocks");
3626 * Dest uses this to prepare to transmit the RAMBlock descriptions
3627 * to the source VM after connection setup.
3628 * Both sides use the "remote" structure to communicate and update
3629 * their "local" descriptions with what was sent.
3631 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3632 rdma
->dest_blocks
[i
].remote_host_addr
=
3633 (uintptr_t)(local
->block
[i
].local_host_addr
);
3635 if (rdma
->pin_all
) {
3636 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3639 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3640 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3642 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3643 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3644 local
->block
[i
].block_name
,
3645 local
->block
[i
].offset
,
3646 local
->block
[i
].length
,
3647 local
->block
[i
].local_host_addr
,
3648 local
->block
[i
].src_index
);
3651 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3652 * sizeof(RDMADestBlock
);
3655 ret
= qemu_rdma_post_send_control(rdma
,
3656 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3659 error_report("rdma migration: error sending remote info");
3664 case RDMA_CONTROL_REGISTER_REQUEST
:
3665 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3667 reg_resp
.repeat
= head
.repeat
;
3668 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3670 for (count
= 0; count
< head
.repeat
; count
++) {
3672 uint8_t *chunk_start
, *chunk_end
;
3674 reg
= ®isters
[count
];
3675 network_to_register(reg
);
3677 reg_result
= &results
[count
];
3679 trace_qemu_rdma_registration_handle_register_loop(count
,
3680 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3682 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3683 error_report("rdma: 'register' bad block index %u (vs %d)",
3684 (unsigned int)reg
->current_index
,
3685 rdma
->local_ram_blocks
.nb_blocks
);
3689 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3690 if (block
->is_ram_block
) {
3691 if (block
->offset
> reg
->key
.current_addr
) {
3692 error_report("rdma: bad register address for block %s"
3693 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3694 block
->block_name
, block
->offset
,
3695 reg
->key
.current_addr
);
3699 host_addr
= (block
->local_host_addr
+
3700 (reg
->key
.current_addr
- block
->offset
));
3701 chunk
= ram_chunk_index(block
->local_host_addr
,
3702 (uint8_t *) host_addr
);
3704 chunk
= reg
->key
.chunk
;
3705 host_addr
= block
->local_host_addr
+
3706 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3707 /* Check for particularly bad chunk value */
3708 if (host_addr
< (void *)block
->local_host_addr
) {
3709 error_report("rdma: bad chunk for block %s"
3711 block
->block_name
, reg
->key
.chunk
);
3716 chunk_start
= ram_chunk_start(block
, chunk
);
3717 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3718 /* avoid "-Waddress-of-packed-member" warning */
3719 uint32_t tmp_rkey
= 0;
3720 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3721 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3722 chunk
, chunk_start
, chunk_end
)) {
3723 error_report("cannot get rkey");
3727 reg_result
->rkey
= tmp_rkey
;
3729 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3731 trace_qemu_rdma_registration_handle_register_rkey(
3734 result_to_network(reg_result
);
3737 ret
= qemu_rdma_post_send_control(rdma
,
3738 (uint8_t *) results
, ®_resp
);
3741 error_report("Failed to send control buffer");
3745 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3746 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3747 unreg_resp
.repeat
= head
.repeat
;
3748 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3750 for (count
= 0; count
< head
.repeat
; count
++) {
3751 reg
= ®isters
[count
];
3752 network_to_register(reg
);
3754 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3755 reg
->current_index
, reg
->key
.chunk
);
3757 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3759 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3760 block
->pmr
[reg
->key
.chunk
] = NULL
;
3763 perror("rdma unregistration chunk failed");
3768 rdma
->total_registrations
--;
3770 trace_qemu_rdma_registration_handle_unregister_success(
3774 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3777 error_report("Failed to send control buffer");
3781 case RDMA_CONTROL_REGISTER_RESULT
:
3782 error_report("Invalid RESULT message at dest.");
3786 error_report("Unknown control message %s", control_desc(head
.type
));
3793 rdma
->error_state
= ret
;
3799 * Called via a ram_control_load_hook during the initial RAM load section which
3800 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3802 * We've already built our local RAMBlock list, but not yet sent the list to
3806 rdma_block_notification_handle(QIOChannelRDMA
*rioc
, const char *name
)
3812 RCU_READ_LOCK_GUARD();
3813 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3819 /* Find the matching RAMBlock in our local list */
3820 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3821 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3828 error_report("RAMBlock '%s' not found on destination", name
);
3832 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3833 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3834 rdma
->next_src_index
++;
3839 static int rdma_load_hook(QEMUFile
*f
, void *opaque
, uint64_t flags
, void *data
)
3842 case RAM_CONTROL_BLOCK_REG
:
3843 return rdma_block_notification_handle(opaque
, data
);
3845 case RAM_CONTROL_HOOK
:
3846 return qemu_rdma_registration_handle(f
, opaque
);
3849 /* Shouldn't be called with any other values */
3854 static int qemu_rdma_registration_start(QEMUFile
*f
, void *opaque
,
3855 uint64_t flags
, void *data
)
3857 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3860 RCU_READ_LOCK_GUARD();
3861 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3866 CHECK_ERROR_STATE();
3868 if (migration_in_postcopy()) {
3872 trace_qemu_rdma_registration_start(flags
);
3873 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3880 * Inform dest that dynamic registrations are done for now.
3881 * First, flush writes, if any.
3883 static int qemu_rdma_registration_stop(QEMUFile
*f
, void *opaque
,
3884 uint64_t flags
, void *data
)
3886 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3888 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3891 RCU_READ_LOCK_GUARD();
3892 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3897 CHECK_ERROR_STATE();
3899 if (migration_in_postcopy()) {
3904 ret
= qemu_rdma_drain_cq(f
, rdma
);
3910 if (flags
== RAM_CONTROL_SETUP
) {
3911 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3912 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3913 int reg_result_idx
, i
, nb_dest_blocks
;
3915 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3916 trace_qemu_rdma_registration_stop_ram();
3919 * Make sure that we parallelize the pinning on both sides.
3920 * For very large guests, doing this serially takes a really
3921 * long time, so we have to 'interleave' the pinning locally
3922 * with the control messages by performing the pinning on this
3923 * side before we receive the control response from the other
3924 * side that the pinning has completed.
3926 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3927 ®_result_idx
, rdma
->pin_all
?
3928 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3930 fprintf(stderr
, "receiving remote info!");
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 fprintf(stderr
, "ram blocks mismatch (Number of blocks %d vs %d) "
3950 "Your QEMU command line parameters are probably "
3951 "not identical on both the source and destination.",
3952 local
->nb_blocks
, nb_dest_blocks
);
3953 rdma
->error_state
= -EINVAL
;
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 fprintf(stderr
, "Block %s/%d has a different length %" PRIu64
3966 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3967 local
->block
[i
].length
,
3968 rdma
->dest_blocks
[i
].length
);
3969 rdma
->error_state
= -EINVAL
;
3972 local
->block
[i
].remote_host_addr
=
3973 rdma
->dest_blocks
[i
].remote_host_addr
;
3974 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3978 trace_qemu_rdma_registration_stop(flags
);
3980 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3981 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3989 rdma
->error_state
= ret
;
3993 static const QEMUFileHooks rdma_read_hooks
= {
3994 .hook_ram_load
= rdma_load_hook
,
3997 static const QEMUFileHooks rdma_write_hooks
= {
3998 .before_ram_iterate
= qemu_rdma_registration_start
,
3999 .after_ram_iterate
= qemu_rdma_registration_stop
,
4000 .save_page
= qemu_rdma_save_page
,
4004 static void qio_channel_rdma_finalize(Object
*obj
)
4006 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
4008 qemu_rdma_cleanup(rioc
->rdmain
);
4009 g_free(rioc
->rdmain
);
4010 rioc
->rdmain
= NULL
;
4012 if (rioc
->rdmaout
) {
4013 qemu_rdma_cleanup(rioc
->rdmaout
);
4014 g_free(rioc
->rdmaout
);
4015 rioc
->rdmaout
= NULL
;
4019 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
4020 void *class_data G_GNUC_UNUSED
)
4022 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
4024 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
4025 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
4026 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
4027 ioc_klass
->io_close
= qio_channel_rdma_close
;
4028 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
4029 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
4030 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
4033 static const TypeInfo qio_channel_rdma_info
= {
4034 .parent
= TYPE_QIO_CHANNEL
,
4035 .name
= TYPE_QIO_CHANNEL_RDMA
,
4036 .instance_size
= sizeof(QIOChannelRDMA
),
4037 .instance_finalize
= qio_channel_rdma_finalize
,
4038 .class_init
= qio_channel_rdma_class_init
,
4041 static void qio_channel_rdma_register_types(void)
4043 type_register_static(&qio_channel_rdma_info
);
4046 type_init(qio_channel_rdma_register_types
);
4048 static QEMUFile
*qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
4050 QIOChannelRDMA
*rioc
;
4052 if (qemu_file_mode_is_not_valid(mode
)) {
4056 rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4058 if (mode
[0] == 'w') {
4059 rioc
->file
= qemu_fopen_channel_output(QIO_CHANNEL(rioc
));
4060 rioc
->rdmaout
= rdma
;
4061 rioc
->rdmain
= rdma
->return_path
;
4062 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
4064 rioc
->file
= qemu_fopen_channel_input(QIO_CHANNEL(rioc
));
4065 rioc
->rdmain
= rdma
;
4066 rioc
->rdmaout
= rdma
->return_path
;
4067 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
4073 static void rdma_accept_incoming_migration(void *opaque
)
4075 RDMAContext
*rdma
= opaque
;
4078 Error
*local_err
= NULL
;
4080 trace_qemu_rdma_accept_incoming_migration();
4081 ret
= qemu_rdma_accept(rdma
);
4084 fprintf(stderr
, "RDMA ERROR: Migration initialization failed\n");
4088 trace_qemu_rdma_accept_incoming_migration_accepted();
4090 if (rdma
->is_return_path
) {
4094 f
= qemu_fopen_rdma(rdma
, "rb");
4096 fprintf(stderr
, "RDMA ERROR: could not qemu_fopen_rdma\n");
4097 qemu_rdma_cleanup(rdma
);
4101 rdma
->migration_started_on_destination
= 1;
4102 migration_fd_process_incoming(f
, &local_err
);
4104 error_reportf_err(local_err
, "RDMA ERROR:");
4108 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4111 RDMAContext
*rdma
, *rdma_return_path
= NULL
;
4112 Error
*local_err
= NULL
;
4114 trace_rdma_start_incoming_migration();
4116 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4117 if (ram_block_discard_is_required()) {
4118 error_setg(errp
, "RDMA: cannot disable RAM discard");
4122 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
4127 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
4133 trace_rdma_start_incoming_migration_after_dest_init();
4135 ret
= rdma_listen(rdma
->listen_id
, 5);
4138 ERROR(errp
, "listening on socket!");
4142 trace_rdma_start_incoming_migration_after_rdma_listen();
4144 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4145 NULL
, (void *)(intptr_t)rdma
);
4149 qemu_rdma_cleanup(rdma
);
4151 error_propagate(errp
, local_err
);
4154 g_free(rdma
->host_port
);
4157 g_free(rdma_return_path
);
4160 void rdma_start_outgoing_migration(void *opaque
,
4161 const char *host_port
, Error
**errp
)
4163 MigrationState
*s
= opaque
;
4164 RDMAContext
*rdma_return_path
= NULL
;
4168 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4169 if (ram_block_discard_is_required()) {
4170 error_setg(errp
, "RDMA: cannot disable RAM discard");
4174 rdma
= qemu_rdma_data_init(host_port
, errp
);
4179 ret
= qemu_rdma_source_init(rdma
,
4180 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4186 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4187 ret
= qemu_rdma_connect(rdma
, errp
, false);
4193 /* RDMA postcopy need a separate queue pair for return path */
4194 if (migrate_postcopy()) {
4195 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4197 if (rdma_return_path
== NULL
) {
4198 goto return_path_err
;
4201 ret
= qemu_rdma_source_init(rdma_return_path
,
4202 s
->enabled_capabilities
[MIGRATION_CAPABILITY_RDMA_PIN_ALL
], errp
);
4205 goto return_path_err
;
4208 ret
= qemu_rdma_connect(rdma_return_path
, errp
, true);
4211 goto return_path_err
;
4214 rdma
->return_path
= rdma_return_path
;
4215 rdma_return_path
->return_path
= rdma
;
4216 rdma_return_path
->is_return_path
= true;
4219 trace_rdma_start_outgoing_migration_after_rdma_connect();
4221 s
->to_dst_file
= qemu_fopen_rdma(rdma
, "wb");
4222 migrate_fd_connect(s
, NULL
);
4225 qemu_rdma_cleanup(rdma
);
4228 g_free(rdma_return_path
);