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/error-report.h"
25 #include "qemu/main-loop.h"
26 #include "qemu/module.h"
28 #include "qemu/sockets.h"
29 #include "qemu/bitmap.h"
30 #include "qemu/coroutine.h"
31 #include "exec/memory.h"
32 #include <sys/socket.h>
34 #include <arpa/inet.h>
35 #include <rdma/rdma_cma.h>
37 #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
*recv_comp_channel
; /* recv completion channel */
362 struct ibv_comp_channel
*send_comp_channel
; /* send completion channel */
363 struct ibv_pd
*pd
; /* protection domain */
364 struct ibv_cq
*recv_cq
; /* recvieve completion queue */
365 struct ibv_cq
*send_cq
; /* send completion queue */
368 * If a previous write failed (perhaps because of a failed
369 * memory registration, then do not attempt any future work
370 * and remember the error state.
377 * Description of ram blocks used throughout the code.
379 RDMALocalBlocks local_ram_blocks
;
380 RDMADestBlock
*dest_blocks
;
382 /* Index of the next RAMBlock received during block registration */
383 unsigned int next_src_index
;
386 * Migration on *destination* started.
387 * Then use coroutine yield function.
388 * Source runs in a thread, so we don't care.
390 int migration_started_on_destination
;
392 int total_registrations
;
395 int unregister_current
, unregister_next
;
396 uint64_t unregistrations
[RDMA_SIGNALED_SEND_MAX
];
398 GHashTable
*blockmap
;
400 /* the RDMAContext for return path */
401 struct RDMAContext
*return_path
;
405 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
406 OBJECT_DECLARE_SIMPLE_TYPE(QIOChannelRDMA
, QIO_CHANNEL_RDMA
)
410 struct QIOChannelRDMA
{
413 RDMAContext
*rdmaout
;
415 bool blocking
; /* XXX we don't actually honour this yet */
419 * Main structure for IB Send/Recv control messages.
420 * This gets prepended at the beginning of every Send/Recv.
422 typedef struct QEMU_PACKED
{
423 uint32_t len
; /* Total length of data portion */
424 uint32_t type
; /* which control command to perform */
425 uint32_t repeat
; /* number of commands in data portion of same type */
429 static void control_to_network(RDMAControlHeader
*control
)
431 control
->type
= htonl(control
->type
);
432 control
->len
= htonl(control
->len
);
433 control
->repeat
= htonl(control
->repeat
);
436 static void network_to_control(RDMAControlHeader
*control
)
438 control
->type
= ntohl(control
->type
);
439 control
->len
= ntohl(control
->len
);
440 control
->repeat
= ntohl(control
->repeat
);
444 * Register a single Chunk.
445 * Information sent by the source VM to inform the dest
446 * to register an single chunk of memory before we can perform
447 * the actual RDMA operation.
449 typedef struct QEMU_PACKED
{
451 uint64_t current_addr
; /* offset into the ram_addr_t space */
452 uint64_t chunk
; /* chunk to lookup if unregistering */
454 uint32_t current_index
; /* which ramblock the chunk belongs to */
456 uint64_t chunks
; /* how many sequential chunks to register */
459 static void register_to_network(RDMAContext
*rdma
, RDMARegister
*reg
)
461 RDMALocalBlock
*local_block
;
462 local_block
= &rdma
->local_ram_blocks
.block
[reg
->current_index
];
464 if (local_block
->is_ram_block
) {
466 * current_addr as passed in is an address in the local ram_addr_t
467 * space, we need to translate this for the destination
469 reg
->key
.current_addr
-= local_block
->offset
;
470 reg
->key
.current_addr
+= rdma
->dest_blocks
[reg
->current_index
].offset
;
472 reg
->key
.current_addr
= htonll(reg
->key
.current_addr
);
473 reg
->current_index
= htonl(reg
->current_index
);
474 reg
->chunks
= htonll(reg
->chunks
);
477 static void network_to_register(RDMARegister
*reg
)
479 reg
->key
.current_addr
= ntohll(reg
->key
.current_addr
);
480 reg
->current_index
= ntohl(reg
->current_index
);
481 reg
->chunks
= ntohll(reg
->chunks
);
484 typedef struct QEMU_PACKED
{
485 uint32_t value
; /* if zero, we will madvise() */
486 uint32_t block_idx
; /* which ram block index */
487 uint64_t offset
; /* Address in remote ram_addr_t space */
488 uint64_t length
; /* length of the chunk */
491 static void compress_to_network(RDMAContext
*rdma
, RDMACompress
*comp
)
493 comp
->value
= htonl(comp
->value
);
495 * comp->offset as passed in is an address in the local ram_addr_t
496 * space, we need to translate this for the destination
498 comp
->offset
-= rdma
->local_ram_blocks
.block
[comp
->block_idx
].offset
;
499 comp
->offset
+= rdma
->dest_blocks
[comp
->block_idx
].offset
;
500 comp
->block_idx
= htonl(comp
->block_idx
);
501 comp
->offset
= htonll(comp
->offset
);
502 comp
->length
= htonll(comp
->length
);
505 static void network_to_compress(RDMACompress
*comp
)
507 comp
->value
= ntohl(comp
->value
);
508 comp
->block_idx
= ntohl(comp
->block_idx
);
509 comp
->offset
= ntohll(comp
->offset
);
510 comp
->length
= ntohll(comp
->length
);
514 * The result of the dest's memory registration produces an "rkey"
515 * which the source VM must reference in order to perform
516 * the RDMA operation.
518 typedef struct QEMU_PACKED
{
522 } RDMARegisterResult
;
524 static void result_to_network(RDMARegisterResult
*result
)
526 result
->rkey
= htonl(result
->rkey
);
527 result
->host_addr
= htonll(result
->host_addr
);
530 static void network_to_result(RDMARegisterResult
*result
)
532 result
->rkey
= ntohl(result
->rkey
);
533 result
->host_addr
= ntohll(result
->host_addr
);
536 const char *print_wrid(int wrid
);
537 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
538 uint8_t *data
, RDMAControlHeader
*resp
,
540 int (*callback
)(RDMAContext
*rdma
));
542 static inline uint64_t ram_chunk_index(const uint8_t *start
,
545 return ((uintptr_t) host
- (uintptr_t) start
) >> RDMA_REG_CHUNK_SHIFT
;
548 static inline uint8_t *ram_chunk_start(const RDMALocalBlock
*rdma_ram_block
,
551 return (uint8_t *)(uintptr_t)(rdma_ram_block
->local_host_addr
+
552 (i
<< RDMA_REG_CHUNK_SHIFT
));
555 static inline uint8_t *ram_chunk_end(const RDMALocalBlock
*rdma_ram_block
,
558 uint8_t *result
= ram_chunk_start(rdma_ram_block
, i
) +
559 (1UL << RDMA_REG_CHUNK_SHIFT
);
561 if (result
> (rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
)) {
562 result
= rdma_ram_block
->local_host_addr
+ rdma_ram_block
->length
;
568 static int rdma_add_block(RDMAContext
*rdma
, const char *block_name
,
570 ram_addr_t block_offset
, uint64_t length
)
572 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
573 RDMALocalBlock
*block
;
574 RDMALocalBlock
*old
= local
->block
;
576 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
+ 1);
578 if (local
->nb_blocks
) {
581 if (rdma
->blockmap
) {
582 for (x
= 0; x
< local
->nb_blocks
; x
++) {
583 g_hash_table_remove(rdma
->blockmap
,
584 (void *)(uintptr_t)old
[x
].offset
);
585 g_hash_table_insert(rdma
->blockmap
,
586 (void *)(uintptr_t)old
[x
].offset
,
590 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * local
->nb_blocks
);
594 block
= &local
->block
[local
->nb_blocks
];
596 block
->block_name
= g_strdup(block_name
);
597 block
->local_host_addr
= host_addr
;
598 block
->offset
= block_offset
;
599 block
->length
= length
;
600 block
->index
= local
->nb_blocks
;
601 block
->src_index
= ~0U; /* Filled in by the receipt of the block list */
602 block
->nb_chunks
= ram_chunk_index(host_addr
, host_addr
+ length
) + 1UL;
603 block
->transit_bitmap
= bitmap_new(block
->nb_chunks
);
604 bitmap_clear(block
->transit_bitmap
, 0, block
->nb_chunks
);
605 block
->unregister_bitmap
= bitmap_new(block
->nb_chunks
);
606 bitmap_clear(block
->unregister_bitmap
, 0, block
->nb_chunks
);
607 block
->remote_keys
= g_new0(uint32_t, block
->nb_chunks
);
609 block
->is_ram_block
= local
->init
? false : true;
611 if (rdma
->blockmap
) {
612 g_hash_table_insert(rdma
->blockmap
, (void *)(uintptr_t)block_offset
, block
);
615 trace_rdma_add_block(block_name
, local
->nb_blocks
,
616 (uintptr_t) block
->local_host_addr
,
617 block
->offset
, block
->length
,
618 (uintptr_t) (block
->local_host_addr
+ block
->length
),
619 BITS_TO_LONGS(block
->nb_chunks
) *
620 sizeof(unsigned long) * 8,
629 * Memory regions need to be registered with the device and queue pairs setup
630 * in advanced before the migration starts. This tells us where the RAM blocks
631 * are so that we can register them individually.
633 static int qemu_rdma_init_one_block(RAMBlock
*rb
, void *opaque
)
635 const char *block_name
= qemu_ram_get_idstr(rb
);
636 void *host_addr
= qemu_ram_get_host_addr(rb
);
637 ram_addr_t block_offset
= qemu_ram_get_offset(rb
);
638 ram_addr_t length
= qemu_ram_get_used_length(rb
);
639 return rdma_add_block(opaque
, block_name
, host_addr
, block_offset
, length
);
643 * Identify the RAMBlocks and their quantity. They will be references to
644 * identify chunk boundaries inside each RAMBlock and also be referenced
645 * during dynamic page registration.
647 static int qemu_rdma_init_ram_blocks(RDMAContext
*rdma
)
649 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
652 assert(rdma
->blockmap
== NULL
);
653 memset(local
, 0, sizeof *local
);
654 ret
= foreach_not_ignored_block(qemu_rdma_init_one_block
, rdma
);
658 trace_qemu_rdma_init_ram_blocks(local
->nb_blocks
);
659 rdma
->dest_blocks
= g_new0(RDMADestBlock
,
660 rdma
->local_ram_blocks
.nb_blocks
);
666 * Note: If used outside of cleanup, the caller must ensure that the destination
667 * block structures are also updated
669 static int rdma_delete_block(RDMAContext
*rdma
, RDMALocalBlock
*block
)
671 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
672 RDMALocalBlock
*old
= local
->block
;
675 if (rdma
->blockmap
) {
676 g_hash_table_remove(rdma
->blockmap
, (void *)(uintptr_t)block
->offset
);
681 for (j
= 0; j
< block
->nb_chunks
; j
++) {
682 if (!block
->pmr
[j
]) {
685 ibv_dereg_mr(block
->pmr
[j
]);
686 rdma
->total_registrations
--;
693 ibv_dereg_mr(block
->mr
);
694 rdma
->total_registrations
--;
698 g_free(block
->transit_bitmap
);
699 block
->transit_bitmap
= NULL
;
701 g_free(block
->unregister_bitmap
);
702 block
->unregister_bitmap
= NULL
;
704 g_free(block
->remote_keys
);
705 block
->remote_keys
= NULL
;
707 g_free(block
->block_name
);
708 block
->block_name
= NULL
;
710 if (rdma
->blockmap
) {
711 for (x
= 0; x
< local
->nb_blocks
; x
++) {
712 g_hash_table_remove(rdma
->blockmap
,
713 (void *)(uintptr_t)old
[x
].offset
);
717 if (local
->nb_blocks
> 1) {
719 local
->block
= g_new0(RDMALocalBlock
, local
->nb_blocks
- 1);
722 memcpy(local
->block
, old
, sizeof(RDMALocalBlock
) * block
->index
);
725 if (block
->index
< (local
->nb_blocks
- 1)) {
726 memcpy(local
->block
+ block
->index
, old
+ (block
->index
+ 1),
727 sizeof(RDMALocalBlock
) *
728 (local
->nb_blocks
- (block
->index
+ 1)));
729 for (x
= block
->index
; x
< local
->nb_blocks
- 1; x
++) {
730 local
->block
[x
].index
--;
734 assert(block
== local
->block
);
738 trace_rdma_delete_block(block
, (uintptr_t)block
->local_host_addr
,
739 block
->offset
, block
->length
,
740 (uintptr_t)(block
->local_host_addr
+ block
->length
),
741 BITS_TO_LONGS(block
->nb_chunks
) *
742 sizeof(unsigned long) * 8, block
->nb_chunks
);
748 if (local
->nb_blocks
&& rdma
->blockmap
) {
749 for (x
= 0; x
< local
->nb_blocks
; x
++) {
750 g_hash_table_insert(rdma
->blockmap
,
751 (void *)(uintptr_t)local
->block
[x
].offset
,
760 * Put in the log file which RDMA device was opened and the details
761 * associated with that device.
763 static void qemu_rdma_dump_id(const char *who
, struct ibv_context
*verbs
)
765 struct ibv_port_attr port
;
767 if (ibv_query_port(verbs
, 1, &port
)) {
768 error_report("Failed to query port information");
772 printf("%s RDMA Device opened: kernel name %s "
773 "uverbs device name %s, "
774 "infiniband_verbs class device path %s, "
775 "infiniband class device path %s, "
776 "transport: (%d) %s\n",
779 verbs
->device
->dev_name
,
780 verbs
->device
->dev_path
,
781 verbs
->device
->ibdev_path
,
783 (port
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) ? "Infiniband" :
784 ((port
.link_layer
== IBV_LINK_LAYER_ETHERNET
)
785 ? "Ethernet" : "Unknown"));
789 * Put in the log file the RDMA gid addressing information,
790 * useful for folks who have trouble understanding the
791 * RDMA device hierarchy in the kernel.
793 static void qemu_rdma_dump_gid(const char *who
, struct rdma_cm_id
*id
)
797 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.sgid
, sgid
, sizeof sgid
);
798 inet_ntop(AF_INET6
, &id
->route
.addr
.addr
.ibaddr
.dgid
, dgid
, sizeof dgid
);
799 trace_qemu_rdma_dump_gid(who
, sgid
, dgid
);
803 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
804 * We will try the next addrinfo struct, and fail if there are
805 * no other valid addresses to bind against.
807 * If user is listening on '[::]', then we will not have a opened a device
808 * yet and have no way of verifying if the device is RoCE or not.
810 * In this case, the source VM will throw an error for ALL types of
811 * connections (both IPv4 and IPv6) if the destination machine does not have
812 * a regular infiniband network available for use.
814 * The only way to guarantee that an error is thrown for broken kernels is
815 * for the management software to choose a *specific* interface at bind time
816 * and validate what time of hardware it is.
818 * Unfortunately, this puts the user in a fix:
820 * If the source VM connects with an IPv4 address without knowing that the
821 * destination has bound to '[::]' the migration will unconditionally fail
822 * unless the management software is explicitly listening on the IPv4
823 * address while using a RoCE-based device.
825 * If the source VM connects with an IPv6 address, then we're OK because we can
826 * throw an error on the source (and similarly on the destination).
828 * But in mixed environments, this will be broken for a while until it is fixed
831 * We do provide a *tiny* bit of help in this function: We can list all of the
832 * devices in the system and check to see if all the devices are RoCE or
835 * If we detect that we have a *pure* RoCE environment, then we can safely
836 * thrown an error even if the management software has specified '[::]' as the
839 * However, if there is are multiple hetergeneous devices, then we cannot make
840 * this assumption and the user just has to be sure they know what they are
843 * Patches are being reviewed on linux-rdma.
845 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context
*verbs
, Error
**errp
)
847 /* This bug only exists in linux, to our knowledge. */
849 struct ibv_port_attr port_attr
;
852 * Verbs are only NULL if management has bound to '[::]'.
854 * Let's iterate through all the devices and see if there any pure IB
855 * devices (non-ethernet).
857 * If not, then we can safely proceed with the migration.
858 * Otherwise, there are no guarantees until the bug is fixed in linux.
862 struct ibv_device
**dev_list
= ibv_get_device_list(&num_devices
);
863 bool roce_found
= false;
864 bool ib_found
= false;
866 for (x
= 0; x
< num_devices
; x
++) {
867 verbs
= ibv_open_device(dev_list
[x
]);
869 if (errno
== EPERM
) {
876 if (ibv_query_port(verbs
, 1, &port_attr
)) {
877 ibv_close_device(verbs
);
878 ERROR(errp
, "Could not query initial IB port");
882 if (port_attr
.link_layer
== IBV_LINK_LAYER_INFINIBAND
) {
884 } else if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
888 ibv_close_device(verbs
);
894 fprintf(stderr
, "WARN: migrations may fail:"
895 " IPv6 over RoCE / iWARP in linux"
896 " is broken. But since you appear to have a"
897 " mixed RoCE / IB environment, be sure to only"
898 " migrate over the IB fabric until the kernel "
899 " fixes the bug.\n");
901 ERROR(errp
, "You only have RoCE / iWARP devices in your systems"
902 " and your management software has specified '[::]'"
903 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
912 * If we have a verbs context, that means that some other than '[::]' was
913 * used by the management software for binding. In which case we can
914 * actually warn the user about a potentially broken kernel.
917 /* IB ports start with 1, not 0 */
918 if (ibv_query_port(verbs
, 1, &port_attr
)) {
919 ERROR(errp
, "Could not query initial IB port");
923 if (port_attr
.link_layer
== IBV_LINK_LAYER_ETHERNET
) {
924 ERROR(errp
, "Linux kernel's RoCE / iWARP does not support IPv6 "
925 "(but patches on linux-rdma in progress)");
935 * Figure out which RDMA device corresponds to the requested IP hostname
936 * Also create the initial connection manager identifiers for opening
939 static int qemu_rdma_resolve_host(RDMAContext
*rdma
, Error
**errp
)
942 struct rdma_addrinfo
*res
;
944 struct rdma_cm_event
*cm_event
;
945 char ip
[40] = "unknown";
946 struct rdma_addrinfo
*e
;
948 if (rdma
->host
== NULL
|| !strcmp(rdma
->host
, "")) {
949 ERROR(errp
, "RDMA hostname has not been set");
953 /* create CM channel */
954 rdma
->channel
= rdma_create_event_channel();
955 if (!rdma
->channel
) {
956 ERROR(errp
, "could not create CM channel");
961 ret
= rdma_create_id(rdma
->channel
, &rdma
->cm_id
, NULL
, RDMA_PS_TCP
);
963 ERROR(errp
, "could not create channel id");
964 goto err_resolve_create_id
;
967 snprintf(port_str
, 16, "%d", rdma
->port
);
970 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
972 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
973 goto err_resolve_get_addr
;
976 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
977 inet_ntop(e
->ai_family
,
978 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
979 trace_qemu_rdma_resolve_host_trying(rdma
->host
, ip
);
981 ret
= rdma_resolve_addr(rdma
->cm_id
, NULL
, e
->ai_dst_addr
,
982 RDMA_RESOLVE_TIMEOUT_MS
);
984 if (e
->ai_family
== AF_INET6
) {
985 ret
= qemu_rdma_broken_ipv6_kernel(rdma
->cm_id
->verbs
, errp
);
994 rdma_freeaddrinfo(res
);
995 ERROR(errp
, "could not resolve address %s", rdma
->host
);
996 goto err_resolve_get_addr
;
999 rdma_freeaddrinfo(res
);
1000 qemu_rdma_dump_gid("source_resolve_addr", rdma
->cm_id
);
1002 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1004 ERROR(errp
, "could not perform event_addr_resolved");
1005 goto err_resolve_get_addr
;
1008 if (cm_event
->event
!= RDMA_CM_EVENT_ADDR_RESOLVED
) {
1009 ERROR(errp
, "result not equal to event_addr_resolved %s",
1010 rdma_event_str(cm_event
->event
));
1011 error_report("rdma_resolve_addr");
1012 rdma_ack_cm_event(cm_event
);
1014 goto err_resolve_get_addr
;
1016 rdma_ack_cm_event(cm_event
);
1019 ret
= rdma_resolve_route(rdma
->cm_id
, RDMA_RESOLVE_TIMEOUT_MS
);
1021 ERROR(errp
, "could not resolve rdma route");
1022 goto err_resolve_get_addr
;
1025 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1027 ERROR(errp
, "could not perform event_route_resolved");
1028 goto err_resolve_get_addr
;
1030 if (cm_event
->event
!= RDMA_CM_EVENT_ROUTE_RESOLVED
) {
1031 ERROR(errp
, "result not equal to event_route_resolved: %s",
1032 rdma_event_str(cm_event
->event
));
1033 rdma_ack_cm_event(cm_event
);
1035 goto err_resolve_get_addr
;
1037 rdma_ack_cm_event(cm_event
);
1038 rdma
->verbs
= rdma
->cm_id
->verbs
;
1039 qemu_rdma_dump_id("source_resolve_host", rdma
->cm_id
->verbs
);
1040 qemu_rdma_dump_gid("source_resolve_host", rdma
->cm_id
);
1043 err_resolve_get_addr
:
1044 rdma_destroy_id(rdma
->cm_id
);
1046 err_resolve_create_id
:
1047 rdma_destroy_event_channel(rdma
->channel
);
1048 rdma
->channel
= NULL
;
1053 * Create protection domain and completion queues
1055 static int qemu_rdma_alloc_pd_cq(RDMAContext
*rdma
)
1058 rdma
->pd
= ibv_alloc_pd(rdma
->verbs
);
1060 error_report("failed to allocate protection domain");
1064 /* create receive completion channel */
1065 rdma
->recv_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1066 if (!rdma
->recv_comp_channel
) {
1067 error_report("failed to allocate receive completion channel");
1068 goto err_alloc_pd_cq
;
1072 * Completion queue can be filled by read work requests.
1074 rdma
->recv_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1075 NULL
, rdma
->recv_comp_channel
, 0);
1076 if (!rdma
->recv_cq
) {
1077 error_report("failed to allocate receive completion queue");
1078 goto err_alloc_pd_cq
;
1081 /* create send completion channel */
1082 rdma
->send_comp_channel
= ibv_create_comp_channel(rdma
->verbs
);
1083 if (!rdma
->send_comp_channel
) {
1084 error_report("failed to allocate send completion channel");
1085 goto err_alloc_pd_cq
;
1088 rdma
->send_cq
= ibv_create_cq(rdma
->verbs
, (RDMA_SIGNALED_SEND_MAX
* 3),
1089 NULL
, rdma
->send_comp_channel
, 0);
1090 if (!rdma
->send_cq
) {
1091 error_report("failed to allocate send completion queue");
1092 goto err_alloc_pd_cq
;
1099 ibv_dealloc_pd(rdma
->pd
);
1101 if (rdma
->recv_comp_channel
) {
1102 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
1104 if (rdma
->send_comp_channel
) {
1105 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
1107 if (rdma
->recv_cq
) {
1108 ibv_destroy_cq(rdma
->recv_cq
);
1109 rdma
->recv_cq
= NULL
;
1112 rdma
->recv_comp_channel
= NULL
;
1113 rdma
->send_comp_channel
= NULL
;
1119 * Create queue pairs.
1121 static int qemu_rdma_alloc_qp(RDMAContext
*rdma
)
1123 struct ibv_qp_init_attr attr
= { 0 };
1126 attr
.cap
.max_send_wr
= RDMA_SIGNALED_SEND_MAX
;
1127 attr
.cap
.max_recv_wr
= 3;
1128 attr
.cap
.max_send_sge
= 1;
1129 attr
.cap
.max_recv_sge
= 1;
1130 attr
.send_cq
= rdma
->send_cq
;
1131 attr
.recv_cq
= rdma
->recv_cq
;
1132 attr
.qp_type
= IBV_QPT_RC
;
1134 ret
= rdma_create_qp(rdma
->cm_id
, rdma
->pd
, &attr
);
1139 rdma
->qp
= rdma
->cm_id
->qp
;
1143 /* Check whether On-Demand Paging is supported by RDAM device */
1144 static bool rdma_support_odp(struct ibv_context
*dev
)
1146 struct ibv_device_attr_ex attr
= {0};
1147 int ret
= ibv_query_device_ex(dev
, NULL
, &attr
);
1152 if (attr
.odp_caps
.general_caps
& IBV_ODP_SUPPORT
) {
1160 * ibv_advise_mr to avoid RNR NAK error as far as possible.
1161 * The responder mr registering with ODP will sent RNR NAK back to
1162 * the requester in the face of the page fault.
1164 static void qemu_rdma_advise_prefetch_mr(struct ibv_pd
*pd
, uint64_t addr
,
1165 uint32_t len
, uint32_t lkey
,
1166 const char *name
, bool wr
)
1168 #ifdef HAVE_IBV_ADVISE_MR
1170 int advice
= wr
? IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE
:
1171 IBV_ADVISE_MR_ADVICE_PREFETCH
;
1172 struct ibv_sge sg_list
= {.lkey
= lkey
, .addr
= addr
, .length
= len
};
1174 ret
= ibv_advise_mr(pd
, advice
,
1175 IBV_ADVISE_MR_FLAG_FLUSH
, &sg_list
, 1);
1176 /* ignore the error */
1178 trace_qemu_rdma_advise_mr(name
, len
, addr
, strerror(errno
));
1180 trace_qemu_rdma_advise_mr(name
, len
, addr
, "successed");
1185 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext
*rdma
)
1188 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
1190 for (i
= 0; i
< local
->nb_blocks
; i
++) {
1191 int access
= IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
;
1193 local
->block
[i
].mr
=
1194 ibv_reg_mr(rdma
->pd
,
1195 local
->block
[i
].local_host_addr
,
1196 local
->block
[i
].length
, access
1199 if (!local
->block
[i
].mr
&&
1200 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1201 access
|= IBV_ACCESS_ON_DEMAND
;
1202 /* register ODP mr */
1203 local
->block
[i
].mr
=
1204 ibv_reg_mr(rdma
->pd
,
1205 local
->block
[i
].local_host_addr
,
1206 local
->block
[i
].length
, access
);
1207 trace_qemu_rdma_register_odp_mr(local
->block
[i
].block_name
);
1209 if (local
->block
[i
].mr
) {
1210 qemu_rdma_advise_prefetch_mr(rdma
->pd
,
1211 (uintptr_t)local
->block
[i
].local_host_addr
,
1212 local
->block
[i
].length
,
1213 local
->block
[i
].mr
->lkey
,
1214 local
->block
[i
].block_name
,
1219 if (!local
->block
[i
].mr
) {
1220 perror("Failed to register local dest ram block!");
1223 rdma
->total_registrations
++;
1226 if (i
>= local
->nb_blocks
) {
1230 for (i
--; i
>= 0; i
--) {
1231 ibv_dereg_mr(local
->block
[i
].mr
);
1232 local
->block
[i
].mr
= NULL
;
1233 rdma
->total_registrations
--;
1241 * Find the ram block that corresponds to the page requested to be
1242 * transmitted by QEMU.
1244 * Once the block is found, also identify which 'chunk' within that
1245 * block that the page belongs to.
1247 * This search cannot fail or the migration will fail.
1249 static int qemu_rdma_search_ram_block(RDMAContext
*rdma
,
1250 uintptr_t block_offset
,
1253 uint64_t *block_index
,
1254 uint64_t *chunk_index
)
1256 uint64_t current_addr
= block_offset
+ offset
;
1257 RDMALocalBlock
*block
= g_hash_table_lookup(rdma
->blockmap
,
1258 (void *) block_offset
);
1260 assert(current_addr
>= block
->offset
);
1261 assert((current_addr
+ length
) <= (block
->offset
+ block
->length
));
1263 *block_index
= block
->index
;
1264 *chunk_index
= ram_chunk_index(block
->local_host_addr
,
1265 block
->local_host_addr
+ (current_addr
- block
->offset
));
1271 * Register a chunk with IB. If the chunk was already registered
1272 * previously, then skip.
1274 * Also return the keys associated with the registration needed
1275 * to perform the actual RDMA operation.
1277 static int qemu_rdma_register_and_get_keys(RDMAContext
*rdma
,
1278 RDMALocalBlock
*block
, uintptr_t host_addr
,
1279 uint32_t *lkey
, uint32_t *rkey
, int chunk
,
1280 uint8_t *chunk_start
, uint8_t *chunk_end
)
1284 *lkey
= block
->mr
->lkey
;
1287 *rkey
= block
->mr
->rkey
;
1292 /* allocate memory to store chunk MRs */
1294 block
->pmr
= g_new0(struct ibv_mr
*, block
->nb_chunks
);
1298 * If 'rkey', then we're the destination, so grant access to the source.
1300 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1302 if (!block
->pmr
[chunk
]) {
1303 uint64_t len
= chunk_end
- chunk_start
;
1304 int access
= rkey
? IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
:
1307 trace_qemu_rdma_register_and_get_keys(len
, chunk_start
);
1309 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1310 if (!block
->pmr
[chunk
] &&
1311 errno
== ENOTSUP
&& rdma_support_odp(rdma
->verbs
)) {
1312 access
|= IBV_ACCESS_ON_DEMAND
;
1313 /* register ODP mr */
1314 block
->pmr
[chunk
] = ibv_reg_mr(rdma
->pd
, chunk_start
, len
, access
);
1315 trace_qemu_rdma_register_odp_mr(block
->block_name
);
1317 if (block
->pmr
[chunk
]) {
1318 qemu_rdma_advise_prefetch_mr(rdma
->pd
, (uintptr_t)chunk_start
,
1319 len
, block
->pmr
[chunk
]->lkey
,
1320 block
->block_name
, rkey
);
1325 if (!block
->pmr
[chunk
]) {
1326 perror("Failed to register chunk!");
1327 fprintf(stderr
, "Chunk details: block: %d chunk index %d"
1328 " start %" PRIuPTR
" end %" PRIuPTR
1330 " local %" PRIuPTR
" registrations: %d\n",
1331 block
->index
, chunk
, (uintptr_t)chunk_start
,
1332 (uintptr_t)chunk_end
, host_addr
,
1333 (uintptr_t)block
->local_host_addr
,
1334 rdma
->total_registrations
);
1337 rdma
->total_registrations
++;
1340 *lkey
= block
->pmr
[chunk
]->lkey
;
1343 *rkey
= block
->pmr
[chunk
]->rkey
;
1349 * Register (at connection time) the memory used for control
1352 static int qemu_rdma_reg_control(RDMAContext
*rdma
, int idx
)
1354 rdma
->wr_data
[idx
].control_mr
= ibv_reg_mr(rdma
->pd
,
1355 rdma
->wr_data
[idx
].control
, RDMA_CONTROL_MAX_BUFFER
,
1356 IBV_ACCESS_LOCAL_WRITE
| IBV_ACCESS_REMOTE_WRITE
);
1357 if (rdma
->wr_data
[idx
].control_mr
) {
1358 rdma
->total_registrations
++;
1361 error_report("qemu_rdma_reg_control failed");
1365 const char *print_wrid(int wrid
)
1367 if (wrid
>= RDMA_WRID_RECV_CONTROL
) {
1368 return wrid_desc
[RDMA_WRID_RECV_CONTROL
];
1370 return wrid_desc
[wrid
];
1374 * Perform a non-optimized memory unregistration after every transfer
1375 * for demonstration purposes, only if pin-all is not requested.
1377 * Potential optimizations:
1378 * 1. Start a new thread to run this function continuously
1380 - and for receipt of unregister messages
1382 * 3. Use workload hints.
1384 static int qemu_rdma_unregister_waiting(RDMAContext
*rdma
)
1386 while (rdma
->unregistrations
[rdma
->unregister_current
]) {
1388 uint64_t wr_id
= rdma
->unregistrations
[rdma
->unregister_current
];
1390 (wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1392 (wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1393 RDMALocalBlock
*block
=
1394 &(rdma
->local_ram_blocks
.block
[index
]);
1395 RDMARegister reg
= { .current_index
= index
};
1396 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
1398 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
1399 .type
= RDMA_CONTROL_UNREGISTER_REQUEST
,
1403 trace_qemu_rdma_unregister_waiting_proc(chunk
,
1404 rdma
->unregister_current
);
1406 rdma
->unregistrations
[rdma
->unregister_current
] = 0;
1407 rdma
->unregister_current
++;
1409 if (rdma
->unregister_current
== RDMA_SIGNALED_SEND_MAX
) {
1410 rdma
->unregister_current
= 0;
1415 * Unregistration is speculative (because migration is single-threaded
1416 * and we cannot break the protocol's inifinband message ordering).
1417 * Thus, if the memory is currently being used for transmission,
1418 * then abort the attempt to unregister and try again
1419 * later the next time a completion is received for this memory.
1421 clear_bit(chunk
, block
->unregister_bitmap
);
1423 if (test_bit(chunk
, block
->transit_bitmap
)) {
1424 trace_qemu_rdma_unregister_waiting_inflight(chunk
);
1428 trace_qemu_rdma_unregister_waiting_send(chunk
);
1430 ret
= ibv_dereg_mr(block
->pmr
[chunk
]);
1431 block
->pmr
[chunk
] = NULL
;
1432 block
->remote_keys
[chunk
] = 0;
1435 perror("unregistration chunk failed");
1438 rdma
->total_registrations
--;
1440 reg
.key
.chunk
= chunk
;
1441 register_to_network(rdma
, ®
);
1442 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
1448 trace_qemu_rdma_unregister_waiting_complete(chunk
);
1454 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id
, uint64_t index
,
1457 uint64_t result
= wr_id
& RDMA_WRID_TYPE_MASK
;
1459 result
|= (index
<< RDMA_WRID_BLOCK_SHIFT
);
1460 result
|= (chunk
<< RDMA_WRID_CHUNK_SHIFT
);
1466 * Consult the connection manager to see a work request
1467 * (of any kind) has completed.
1468 * Return the work request ID that completed.
1470 static uint64_t qemu_rdma_poll(RDMAContext
*rdma
, struct ibv_cq
*cq
,
1471 uint64_t *wr_id_out
, uint32_t *byte_len
)
1477 ret
= ibv_poll_cq(cq
, 1, &wc
);
1480 *wr_id_out
= RDMA_WRID_NONE
;
1485 error_report("ibv_poll_cq return %d", ret
);
1489 wr_id
= wc
.wr_id
& RDMA_WRID_TYPE_MASK
;
1491 if (wc
.status
!= IBV_WC_SUCCESS
) {
1492 fprintf(stderr
, "ibv_poll_cq wc.status=%d %s!\n",
1493 wc
.status
, ibv_wc_status_str(wc
.status
));
1494 fprintf(stderr
, "ibv_poll_cq wrid=%s!\n", wrid_desc
[wr_id
]);
1499 if (rdma
->control_ready_expected
&&
1500 (wr_id
>= RDMA_WRID_RECV_CONTROL
)) {
1501 trace_qemu_rdma_poll_recv(wrid_desc
[RDMA_WRID_RECV_CONTROL
],
1502 wr_id
- RDMA_WRID_RECV_CONTROL
, wr_id
, rdma
->nb_sent
);
1503 rdma
->control_ready_expected
= 0;
1506 if (wr_id
== RDMA_WRID_RDMA_WRITE
) {
1508 (wc
.wr_id
& RDMA_WRID_CHUNK_MASK
) >> RDMA_WRID_CHUNK_SHIFT
;
1510 (wc
.wr_id
& RDMA_WRID_BLOCK_MASK
) >> RDMA_WRID_BLOCK_SHIFT
;
1511 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[index
]);
1513 trace_qemu_rdma_poll_write(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
,
1514 index
, chunk
, block
->local_host_addr
,
1515 (void *)(uintptr_t)block
->remote_host_addr
);
1517 clear_bit(chunk
, block
->transit_bitmap
);
1519 if (rdma
->nb_sent
> 0) {
1523 trace_qemu_rdma_poll_other(print_wrid(wr_id
), wr_id
, rdma
->nb_sent
);
1526 *wr_id_out
= wc
.wr_id
;
1528 *byte_len
= wc
.byte_len
;
1534 /* Wait for activity on the completion channel.
1535 * Returns 0 on success, none-0 on error.
1537 static int qemu_rdma_wait_comp_channel(RDMAContext
*rdma
,
1538 struct ibv_comp_channel
*comp_channel
)
1540 struct rdma_cm_event
*cm_event
;
1544 * Coroutine doesn't start until migration_fd_process_incoming()
1545 * so don't yield unless we know we're running inside of a coroutine.
1547 if (rdma
->migration_started_on_destination
&&
1548 migration_incoming_get_current()->state
== MIGRATION_STATUS_ACTIVE
) {
1549 yield_until_fd_readable(comp_channel
->fd
);
1551 /* This is the source side, we're in a separate thread
1552 * or destination prior to migration_fd_process_incoming()
1553 * after postcopy, the destination also in a separate thread.
1554 * we can't yield; so we have to poll the fd.
1555 * But we need to be able to handle 'cancel' or an error
1556 * without hanging forever.
1558 while (!rdma
->error_state
&& !rdma
->received_error
) {
1560 pfds
[0].fd
= comp_channel
->fd
;
1561 pfds
[0].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1562 pfds
[0].revents
= 0;
1564 pfds
[1].fd
= rdma
->channel
->fd
;
1565 pfds
[1].events
= G_IO_IN
| G_IO_HUP
| G_IO_ERR
;
1566 pfds
[1].revents
= 0;
1568 /* 0.1s timeout, should be fine for a 'cancel' */
1569 switch (qemu_poll_ns(pfds
, 2, 100 * 1000 * 1000)) {
1571 case 1: /* fd active */
1572 if (pfds
[0].revents
) {
1576 if (pfds
[1].revents
) {
1577 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
1579 error_report("failed to get cm event while wait "
1580 "completion channel");
1584 error_report("receive cm event while wait comp channel,"
1585 "cm event is %d", cm_event
->event
);
1586 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
1587 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
1588 rdma_ack_cm_event(cm_event
);
1591 rdma_ack_cm_event(cm_event
);
1595 case 0: /* Timeout, go around again */
1598 default: /* Error of some type -
1599 * I don't trust errno from qemu_poll_ns
1601 error_report("%s: poll failed", __func__
);
1605 if (migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) {
1606 /* Bail out and let the cancellation happen */
1612 if (rdma
->received_error
) {
1615 return rdma
->error_state
;
1618 static struct ibv_comp_channel
*to_channel(RDMAContext
*rdma
, int wrid
)
1620 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_comp_channel
:
1621 rdma
->recv_comp_channel
;
1624 static struct ibv_cq
*to_cq(RDMAContext
*rdma
, int wrid
)
1626 return wrid
< RDMA_WRID_RECV_CONTROL
? rdma
->send_cq
: rdma
->recv_cq
;
1630 * Block until the next work request has completed.
1632 * First poll to see if a work request has already completed,
1635 * If we encounter completed work requests for IDs other than
1636 * the one we're interested in, then that's generally an error.
1638 * The only exception is actual RDMA Write completions. These
1639 * completions only need to be recorded, but do not actually
1640 * need further processing.
1642 static int qemu_rdma_block_for_wrid(RDMAContext
*rdma
, int wrid_requested
,
1645 int num_cq_events
= 0, ret
= 0;
1648 uint64_t wr_id
= RDMA_WRID_NONE
, wr_id_in
;
1649 struct ibv_comp_channel
*ch
= to_channel(rdma
, wrid_requested
);
1650 struct ibv_cq
*poll_cq
= to_cq(rdma
, wrid_requested
);
1652 if (ibv_req_notify_cq(poll_cq
, 0)) {
1656 while (wr_id
!= wrid_requested
) {
1657 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1662 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1664 if (wr_id
== RDMA_WRID_NONE
) {
1667 if (wr_id
!= wrid_requested
) {
1668 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1669 wrid_requested
, print_wrid(wr_id
), wr_id
);
1673 if (wr_id
== wrid_requested
) {
1678 ret
= qemu_rdma_wait_comp_channel(rdma
, ch
);
1680 goto err_block_for_wrid
;
1683 ret
= ibv_get_cq_event(ch
, &cq
, &cq_ctx
);
1685 perror("ibv_get_cq_event");
1686 goto err_block_for_wrid
;
1691 ret
= -ibv_req_notify_cq(cq
, 0);
1693 goto err_block_for_wrid
;
1696 while (wr_id
!= wrid_requested
) {
1697 ret
= qemu_rdma_poll(rdma
, poll_cq
, &wr_id_in
, byte_len
);
1699 goto err_block_for_wrid
;
1702 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
1704 if (wr_id
== RDMA_WRID_NONE
) {
1707 if (wr_id
!= wrid_requested
) {
1708 trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested
),
1709 wrid_requested
, print_wrid(wr_id
), wr_id
);
1713 if (wr_id
== wrid_requested
) {
1714 goto success_block_for_wrid
;
1718 success_block_for_wrid
:
1719 if (num_cq_events
) {
1720 ibv_ack_cq_events(cq
, num_cq_events
);
1725 if (num_cq_events
) {
1726 ibv_ack_cq_events(cq
, num_cq_events
);
1729 rdma
->error_state
= ret
;
1734 * Post a SEND message work request for the control channel
1735 * containing some data and block until the post completes.
1737 static int qemu_rdma_post_send_control(RDMAContext
*rdma
, uint8_t *buf
,
1738 RDMAControlHeader
*head
)
1741 RDMAWorkRequestData
*wr
= &rdma
->wr_data
[RDMA_WRID_CONTROL
];
1742 struct ibv_send_wr
*bad_wr
;
1743 struct ibv_sge sge
= {
1744 .addr
= (uintptr_t)(wr
->control
),
1745 .length
= head
->len
+ sizeof(RDMAControlHeader
),
1746 .lkey
= wr
->control_mr
->lkey
,
1748 struct ibv_send_wr send_wr
= {
1749 .wr_id
= RDMA_WRID_SEND_CONTROL
,
1750 .opcode
= IBV_WR_SEND
,
1751 .send_flags
= IBV_SEND_SIGNALED
,
1756 trace_qemu_rdma_post_send_control(control_desc(head
->type
));
1759 * We don't actually need to do a memcpy() in here if we used
1760 * the "sge" properly, but since we're only sending control messages
1761 * (not RAM in a performance-critical path), then its OK for now.
1763 * The copy makes the RDMAControlHeader simpler to manipulate
1764 * for the time being.
1766 assert(head
->len
<= RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
));
1767 memcpy(wr
->control
, head
, sizeof(RDMAControlHeader
));
1768 control_to_network((void *) wr
->control
);
1771 memcpy(wr
->control
+ sizeof(RDMAControlHeader
), buf
, head
->len
);
1775 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
1778 error_report("Failed to use post IB SEND for control");
1782 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_SEND_CONTROL
, NULL
);
1784 error_report("rdma migration: send polling control error");
1791 * Post a RECV work request in anticipation of some future receipt
1792 * of data on the control channel.
1794 static int qemu_rdma_post_recv_control(RDMAContext
*rdma
, int idx
)
1796 struct ibv_recv_wr
*bad_wr
;
1797 struct ibv_sge sge
= {
1798 .addr
= (uintptr_t)(rdma
->wr_data
[idx
].control
),
1799 .length
= RDMA_CONTROL_MAX_BUFFER
,
1800 .lkey
= rdma
->wr_data
[idx
].control_mr
->lkey
,
1803 struct ibv_recv_wr recv_wr
= {
1804 .wr_id
= RDMA_WRID_RECV_CONTROL
+ idx
,
1810 if (ibv_post_recv(rdma
->qp
, &recv_wr
, &bad_wr
)) {
1818 * Block and wait for a RECV control channel message to arrive.
1820 static int qemu_rdma_exchange_get_response(RDMAContext
*rdma
,
1821 RDMAControlHeader
*head
, int expecting
, int idx
)
1824 int ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RECV_CONTROL
+ idx
,
1828 error_report("rdma migration: recv polling control error!");
1832 network_to_control((void *) rdma
->wr_data
[idx
].control
);
1833 memcpy(head
, rdma
->wr_data
[idx
].control
, sizeof(RDMAControlHeader
));
1835 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting
));
1837 if (expecting
== RDMA_CONTROL_NONE
) {
1838 trace_qemu_rdma_exchange_get_response_none(control_desc(head
->type
),
1840 } else if (head
->type
!= expecting
|| head
->type
== RDMA_CONTROL_ERROR
) {
1841 error_report("Was expecting a %s (%d) control message"
1842 ", but got: %s (%d), length: %d",
1843 control_desc(expecting
), expecting
,
1844 control_desc(head
->type
), head
->type
, head
->len
);
1845 if (head
->type
== RDMA_CONTROL_ERROR
) {
1846 rdma
->received_error
= true;
1850 if (head
->len
> RDMA_CONTROL_MAX_BUFFER
- sizeof(*head
)) {
1851 error_report("too long length: %d", head
->len
);
1854 if (sizeof(*head
) + head
->len
!= byte_len
) {
1855 error_report("Malformed length: %d byte_len %d", head
->len
, byte_len
);
1863 * When a RECV work request has completed, the work request's
1864 * buffer is pointed at the header.
1866 * This will advance the pointer to the data portion
1867 * of the control message of the work request's buffer that
1868 * was populated after the work request finished.
1870 static void qemu_rdma_move_header(RDMAContext
*rdma
, int idx
,
1871 RDMAControlHeader
*head
)
1873 rdma
->wr_data
[idx
].control_len
= head
->len
;
1874 rdma
->wr_data
[idx
].control_curr
=
1875 rdma
->wr_data
[idx
].control
+ sizeof(RDMAControlHeader
);
1879 * This is an 'atomic' high-level operation to deliver a single, unified
1880 * control-channel message.
1882 * Additionally, if the user is expecting some kind of reply to this message,
1883 * they can request a 'resp' response message be filled in by posting an
1884 * additional work request on behalf of the user and waiting for an additional
1887 * The extra (optional) response is used during registration to us from having
1888 * to perform an *additional* exchange of message just to provide a response by
1889 * instead piggy-backing on the acknowledgement.
1891 static int qemu_rdma_exchange_send(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1892 uint8_t *data
, RDMAControlHeader
*resp
,
1894 int (*callback
)(RDMAContext
*rdma
))
1899 * Wait until the dest is ready before attempting to deliver the message
1900 * by waiting for a READY message.
1902 if (rdma
->control_ready_expected
) {
1903 RDMAControlHeader resp
;
1904 ret
= qemu_rdma_exchange_get_response(rdma
,
1905 &resp
, RDMA_CONTROL_READY
, RDMA_WRID_READY
);
1912 * If the user is expecting a response, post a WR in anticipation of it.
1915 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_DATA
);
1917 error_report("rdma migration: error posting"
1918 " extra control recv for anticipated result!");
1924 * Post a WR to replace the one we just consumed for the READY message.
1926 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
1928 error_report("rdma migration: error posting first control recv!");
1933 * Deliver the control message that was requested.
1935 ret
= qemu_rdma_post_send_control(rdma
, data
, head
);
1938 error_report("Failed to send control buffer!");
1943 * If we're expecting a response, block and wait for it.
1947 trace_qemu_rdma_exchange_send_issue_callback();
1948 ret
= callback(rdma
);
1954 trace_qemu_rdma_exchange_send_waiting(control_desc(resp
->type
));
1955 ret
= qemu_rdma_exchange_get_response(rdma
, resp
,
1956 resp
->type
, RDMA_WRID_DATA
);
1962 qemu_rdma_move_header(rdma
, RDMA_WRID_DATA
, resp
);
1964 *resp_idx
= RDMA_WRID_DATA
;
1966 trace_qemu_rdma_exchange_send_received(control_desc(resp
->type
));
1969 rdma
->control_ready_expected
= 1;
1975 * This is an 'atomic' high-level operation to receive a single, unified
1976 * control-channel message.
1978 static int qemu_rdma_exchange_recv(RDMAContext
*rdma
, RDMAControlHeader
*head
,
1981 RDMAControlHeader ready
= {
1983 .type
= RDMA_CONTROL_READY
,
1989 * Inform the source that we're ready to receive a message.
1991 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &ready
);
1994 error_report("Failed to send control buffer!");
1999 * Block and wait for the message.
2001 ret
= qemu_rdma_exchange_get_response(rdma
, head
,
2002 expecting
, RDMA_WRID_READY
);
2008 qemu_rdma_move_header(rdma
, RDMA_WRID_READY
, head
);
2011 * Post a new RECV work request to replace the one we just consumed.
2013 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2015 error_report("rdma migration: error posting second control recv!");
2023 * Write an actual chunk of memory using RDMA.
2025 * If we're using dynamic registration on the dest-side, we have to
2026 * send a registration command first.
2028 static int qemu_rdma_write_one(QEMUFile
*f
, RDMAContext
*rdma
,
2029 int current_index
, uint64_t current_addr
,
2033 struct ibv_send_wr send_wr
= { 0 };
2034 struct ibv_send_wr
*bad_wr
;
2035 int reg_result_idx
, ret
, count
= 0;
2036 uint64_t chunk
, chunks
;
2037 uint8_t *chunk_start
, *chunk_end
;
2038 RDMALocalBlock
*block
= &(rdma
->local_ram_blocks
.block
[current_index
]);
2040 RDMARegisterResult
*reg_result
;
2041 RDMAControlHeader resp
= { .type
= RDMA_CONTROL_REGISTER_RESULT
};
2042 RDMAControlHeader head
= { .len
= sizeof(RDMARegister
),
2043 .type
= RDMA_CONTROL_REGISTER_REQUEST
,
2048 sge
.addr
= (uintptr_t)(block
->local_host_addr
+
2049 (current_addr
- block
->offset
));
2050 sge
.length
= length
;
2052 chunk
= ram_chunk_index(block
->local_host_addr
,
2053 (uint8_t *)(uintptr_t)sge
.addr
);
2054 chunk_start
= ram_chunk_start(block
, chunk
);
2056 if (block
->is_ram_block
) {
2057 chunks
= length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2059 if (chunks
&& ((length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2063 chunks
= block
->length
/ (1UL << RDMA_REG_CHUNK_SHIFT
);
2065 if (chunks
&& ((block
->length
% (1UL << RDMA_REG_CHUNK_SHIFT
)) == 0)) {
2070 trace_qemu_rdma_write_one_top(chunks
+ 1,
2072 (1UL << RDMA_REG_CHUNK_SHIFT
) / 1024 / 1024);
2074 chunk_end
= ram_chunk_end(block
, chunk
+ chunks
);
2077 while (test_bit(chunk
, block
->transit_bitmap
)) {
2079 trace_qemu_rdma_write_one_block(count
++, current_index
, chunk
,
2080 sge
.addr
, length
, rdma
->nb_sent
, block
->nb_chunks
);
2082 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2085 error_report("Failed to Wait for previous write to complete "
2086 "block %d chunk %" PRIu64
2087 " current %" PRIu64
" len %" PRIu64
" %d",
2088 current_index
, chunk
, sge
.addr
, length
, rdma
->nb_sent
);
2093 if (!rdma
->pin_all
|| !block
->is_ram_block
) {
2094 if (!block
->remote_keys
[chunk
]) {
2096 * This chunk has not yet been registered, so first check to see
2097 * if the entire chunk is zero. If so, tell the other size to
2098 * memset() + madvise() the entire chunk without RDMA.
2101 if (buffer_is_zero((void *)(uintptr_t)sge
.addr
, length
)) {
2102 RDMACompress comp
= {
2103 .offset
= current_addr
,
2105 .block_idx
= current_index
,
2109 head
.len
= sizeof(comp
);
2110 head
.type
= RDMA_CONTROL_COMPRESS
;
2112 trace_qemu_rdma_write_one_zero(chunk
, sge
.length
,
2113 current_index
, current_addr
);
2115 compress_to_network(rdma
, &comp
);
2116 ret
= qemu_rdma_exchange_send(rdma
, &head
,
2117 (uint8_t *) &comp
, NULL
, NULL
, NULL
);
2123 acct_update_position(f
, sge
.length
, true);
2129 * Otherwise, tell other side to register.
2131 reg
.current_index
= current_index
;
2132 if (block
->is_ram_block
) {
2133 reg
.key
.current_addr
= current_addr
;
2135 reg
.key
.chunk
= chunk
;
2137 reg
.chunks
= chunks
;
2139 trace_qemu_rdma_write_one_sendreg(chunk
, sge
.length
, current_index
,
2142 register_to_network(rdma
, ®
);
2143 ret
= qemu_rdma_exchange_send(rdma
, &head
, (uint8_t *) ®
,
2144 &resp
, ®_result_idx
, NULL
);
2149 /* try to overlap this single registration with the one we sent. */
2150 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2151 &sge
.lkey
, NULL
, chunk
,
2152 chunk_start
, chunk_end
)) {
2153 error_report("cannot get lkey");
2157 reg_result
= (RDMARegisterResult
*)
2158 rdma
->wr_data
[reg_result_idx
].control_curr
;
2160 network_to_result(reg_result
);
2162 trace_qemu_rdma_write_one_recvregres(block
->remote_keys
[chunk
],
2163 reg_result
->rkey
, chunk
);
2165 block
->remote_keys
[chunk
] = reg_result
->rkey
;
2166 block
->remote_host_addr
= reg_result
->host_addr
;
2168 /* already registered before */
2169 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2170 &sge
.lkey
, NULL
, chunk
,
2171 chunk_start
, chunk_end
)) {
2172 error_report("cannot get lkey!");
2177 send_wr
.wr
.rdma
.rkey
= block
->remote_keys
[chunk
];
2179 send_wr
.wr
.rdma
.rkey
= block
->remote_rkey
;
2181 if (qemu_rdma_register_and_get_keys(rdma
, block
, sge
.addr
,
2182 &sge
.lkey
, NULL
, chunk
,
2183 chunk_start
, chunk_end
)) {
2184 error_report("cannot get lkey!");
2190 * Encode the ram block index and chunk within this wrid.
2191 * We will use this information at the time of completion
2192 * to figure out which bitmap to check against and then which
2193 * chunk in the bitmap to look for.
2195 send_wr
.wr_id
= qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE
,
2196 current_index
, chunk
);
2198 send_wr
.opcode
= IBV_WR_RDMA_WRITE
;
2199 send_wr
.send_flags
= IBV_SEND_SIGNALED
;
2200 send_wr
.sg_list
= &sge
;
2201 send_wr
.num_sge
= 1;
2202 send_wr
.wr
.rdma
.remote_addr
= block
->remote_host_addr
+
2203 (current_addr
- block
->offset
);
2205 trace_qemu_rdma_write_one_post(chunk
, sge
.addr
, send_wr
.wr
.rdma
.remote_addr
,
2209 * ibv_post_send() does not return negative error numbers,
2210 * per the specification they are positive - no idea why.
2212 ret
= ibv_post_send(rdma
->qp
, &send_wr
, &bad_wr
);
2214 if (ret
== ENOMEM
) {
2215 trace_qemu_rdma_write_one_queue_full();
2216 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2218 error_report("rdma migration: failed to make "
2219 "room in full send queue! %d", ret
);
2225 } else if (ret
> 0) {
2226 perror("rdma migration: post rdma write failed");
2230 set_bit(chunk
, block
->transit_bitmap
);
2231 acct_update_position(f
, sge
.length
, false);
2232 rdma
->total_writes
++;
2238 * Push out any unwritten RDMA operations.
2240 * We support sending out multiple chunks at the same time.
2241 * Not all of them need to get signaled in the completion queue.
2243 static int qemu_rdma_write_flush(QEMUFile
*f
, RDMAContext
*rdma
)
2247 if (!rdma
->current_length
) {
2251 ret
= qemu_rdma_write_one(f
, rdma
,
2252 rdma
->current_index
, rdma
->current_addr
, rdma
->current_length
);
2260 trace_qemu_rdma_write_flush(rdma
->nb_sent
);
2263 rdma
->current_length
= 0;
2264 rdma
->current_addr
= 0;
2269 static inline int qemu_rdma_buffer_mergable(RDMAContext
*rdma
,
2270 uint64_t offset
, uint64_t len
)
2272 RDMALocalBlock
*block
;
2276 if (rdma
->current_index
< 0) {
2280 if (rdma
->current_chunk
< 0) {
2284 block
= &(rdma
->local_ram_blocks
.block
[rdma
->current_index
]);
2285 host_addr
= block
->local_host_addr
+ (offset
- block
->offset
);
2286 chunk_end
= ram_chunk_end(block
, rdma
->current_chunk
);
2288 if (rdma
->current_length
== 0) {
2293 * Only merge into chunk sequentially.
2295 if (offset
!= (rdma
->current_addr
+ rdma
->current_length
)) {
2299 if (offset
< block
->offset
) {
2303 if ((offset
+ len
) > (block
->offset
+ block
->length
)) {
2307 if ((host_addr
+ len
) > chunk_end
) {
2315 * We're not actually writing here, but doing three things:
2317 * 1. Identify the chunk the buffer belongs to.
2318 * 2. If the chunk is full or the buffer doesn't belong to the current
2319 * chunk, then start a new chunk and flush() the old chunk.
2320 * 3. To keep the hardware busy, we also group chunks into batches
2321 * and only require that a batch gets acknowledged in the completion
2322 * queue instead of each individual chunk.
2324 static int qemu_rdma_write(QEMUFile
*f
, RDMAContext
*rdma
,
2325 uint64_t block_offset
, uint64_t offset
,
2328 uint64_t current_addr
= block_offset
+ offset
;
2329 uint64_t index
= rdma
->current_index
;
2330 uint64_t chunk
= rdma
->current_chunk
;
2333 /* If we cannot merge it, we flush the current buffer first. */
2334 if (!qemu_rdma_buffer_mergable(rdma
, current_addr
, len
)) {
2335 ret
= qemu_rdma_write_flush(f
, rdma
);
2339 rdma
->current_length
= 0;
2340 rdma
->current_addr
= current_addr
;
2342 ret
= qemu_rdma_search_ram_block(rdma
, block_offset
,
2343 offset
, len
, &index
, &chunk
);
2345 error_report("ram block search failed");
2348 rdma
->current_index
= index
;
2349 rdma
->current_chunk
= chunk
;
2353 rdma
->current_length
+= len
;
2355 /* flush it if buffer is too large */
2356 if (rdma
->current_length
>= RDMA_MERGE_MAX
) {
2357 return qemu_rdma_write_flush(f
, rdma
);
2363 static void qemu_rdma_cleanup(RDMAContext
*rdma
)
2367 if (rdma
->cm_id
&& rdma
->connected
) {
2368 if ((rdma
->error_state
||
2369 migrate_get_current()->state
== MIGRATION_STATUS_CANCELLING
) &&
2370 !rdma
->received_error
) {
2371 RDMAControlHeader head
= { .len
= 0,
2372 .type
= RDMA_CONTROL_ERROR
,
2375 error_report("Early error. Sending error.");
2376 qemu_rdma_post_send_control(rdma
, NULL
, &head
);
2379 rdma_disconnect(rdma
->cm_id
);
2380 trace_qemu_rdma_cleanup_disconnect();
2381 rdma
->connected
= false;
2384 if (rdma
->channel
) {
2385 qemu_set_fd_handler(rdma
->channel
->fd
, NULL
, NULL
, NULL
);
2387 g_free(rdma
->dest_blocks
);
2388 rdma
->dest_blocks
= NULL
;
2390 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2391 if (rdma
->wr_data
[idx
].control_mr
) {
2392 rdma
->total_registrations
--;
2393 ibv_dereg_mr(rdma
->wr_data
[idx
].control_mr
);
2395 rdma
->wr_data
[idx
].control_mr
= NULL
;
2398 if (rdma
->local_ram_blocks
.block
) {
2399 while (rdma
->local_ram_blocks
.nb_blocks
) {
2400 rdma_delete_block(rdma
, &rdma
->local_ram_blocks
.block
[0]);
2405 rdma_destroy_qp(rdma
->cm_id
);
2408 if (rdma
->recv_cq
) {
2409 ibv_destroy_cq(rdma
->recv_cq
);
2410 rdma
->recv_cq
= NULL
;
2412 if (rdma
->send_cq
) {
2413 ibv_destroy_cq(rdma
->send_cq
);
2414 rdma
->send_cq
= NULL
;
2416 if (rdma
->recv_comp_channel
) {
2417 ibv_destroy_comp_channel(rdma
->recv_comp_channel
);
2418 rdma
->recv_comp_channel
= NULL
;
2420 if (rdma
->send_comp_channel
) {
2421 ibv_destroy_comp_channel(rdma
->send_comp_channel
);
2422 rdma
->send_comp_channel
= NULL
;
2425 ibv_dealloc_pd(rdma
->pd
);
2429 rdma_destroy_id(rdma
->cm_id
);
2433 /* the destination side, listen_id and channel is shared */
2434 if (rdma
->listen_id
) {
2435 if (!rdma
->is_return_path
) {
2436 rdma_destroy_id(rdma
->listen_id
);
2438 rdma
->listen_id
= NULL
;
2440 if (rdma
->channel
) {
2441 if (!rdma
->is_return_path
) {
2442 rdma_destroy_event_channel(rdma
->channel
);
2444 rdma
->channel
= NULL
;
2448 if (rdma
->channel
) {
2449 rdma_destroy_event_channel(rdma
->channel
);
2450 rdma
->channel
= NULL
;
2453 g_free(rdma
->host_port
);
2455 rdma
->host_port
= NULL
;
2459 static int qemu_rdma_source_init(RDMAContext
*rdma
, bool pin_all
, Error
**errp
)
2462 Error
*local_err
= NULL
, **temp
= &local_err
;
2465 * Will be validated against destination's actual capabilities
2466 * after the connect() completes.
2468 rdma
->pin_all
= pin_all
;
2470 ret
= qemu_rdma_resolve_host(rdma
, temp
);
2472 goto err_rdma_source_init
;
2475 ret
= qemu_rdma_alloc_pd_cq(rdma
);
2477 ERROR(temp
, "rdma migration: error allocating pd and cq! Your mlock()"
2478 " limits may be too low. Please check $ ulimit -a # and "
2479 "search for 'ulimit -l' in the output");
2480 goto err_rdma_source_init
;
2483 ret
= qemu_rdma_alloc_qp(rdma
);
2485 ERROR(temp
, "rdma migration: error allocating qp!");
2486 goto err_rdma_source_init
;
2489 ret
= qemu_rdma_init_ram_blocks(rdma
);
2491 ERROR(temp
, "rdma migration: error initializing ram blocks!");
2492 goto err_rdma_source_init
;
2495 /* Build the hash that maps from offset to RAMBlock */
2496 rdma
->blockmap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2497 for (idx
= 0; idx
< rdma
->local_ram_blocks
.nb_blocks
; idx
++) {
2498 g_hash_table_insert(rdma
->blockmap
,
2499 (void *)(uintptr_t)rdma
->local_ram_blocks
.block
[idx
].offset
,
2500 &rdma
->local_ram_blocks
.block
[idx
]);
2503 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2504 ret
= qemu_rdma_reg_control(rdma
, idx
);
2506 ERROR(temp
, "rdma migration: error registering %d control!",
2508 goto err_rdma_source_init
;
2514 err_rdma_source_init
:
2515 error_propagate(errp
, local_err
);
2516 qemu_rdma_cleanup(rdma
);
2520 static int qemu_get_cm_event_timeout(RDMAContext
*rdma
,
2521 struct rdma_cm_event
**cm_event
,
2522 long msec
, Error
**errp
)
2525 struct pollfd poll_fd
= {
2526 .fd
= rdma
->channel
->fd
,
2532 ret
= poll(&poll_fd
, 1, msec
);
2533 } while (ret
< 0 && errno
== EINTR
);
2536 ERROR(errp
, "poll cm event timeout");
2538 } else if (ret
< 0) {
2539 ERROR(errp
, "failed to poll cm event, errno=%i", errno
);
2541 } else if (poll_fd
.revents
& POLLIN
) {
2542 return rdma_get_cm_event(rdma
->channel
, cm_event
);
2544 ERROR(errp
, "no POLLIN event, revent=%x", poll_fd
.revents
);
2549 static int qemu_rdma_connect(RDMAContext
*rdma
, Error
**errp
, bool return_path
)
2551 RDMACapabilities cap
= {
2552 .version
= RDMA_CONTROL_VERSION_CURRENT
,
2555 struct rdma_conn_param conn_param
= { .initiator_depth
= 2,
2557 .private_data
= &cap
,
2558 .private_data_len
= sizeof(cap
),
2560 struct rdma_cm_event
*cm_event
;
2564 * Only negotiate the capability with destination if the user
2565 * on the source first requested the capability.
2567 if (rdma
->pin_all
) {
2568 trace_qemu_rdma_connect_pin_all_requested();
2569 cap
.flags
|= RDMA_CAPABILITY_PIN_ALL
;
2572 caps_to_network(&cap
);
2574 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
2576 ERROR(errp
, "posting second control recv");
2577 goto err_rdma_source_connect
;
2580 ret
= rdma_connect(rdma
->cm_id
, &conn_param
);
2582 perror("rdma_connect");
2583 ERROR(errp
, "connecting to destination!");
2584 goto err_rdma_source_connect
;
2588 ret
= qemu_get_cm_event_timeout(rdma
, &cm_event
, 5000, errp
);
2590 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
2593 perror("rdma_get_cm_event after rdma_connect");
2594 ERROR(errp
, "connecting to destination!");
2595 goto err_rdma_source_connect
;
2598 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
2599 error_report("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect");
2600 ERROR(errp
, "connecting to destination!");
2601 rdma_ack_cm_event(cm_event
);
2602 goto err_rdma_source_connect
;
2604 rdma
->connected
= true;
2606 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
2607 network_to_caps(&cap
);
2610 * Verify that the *requested* capabilities are supported by the destination
2611 * and disable them otherwise.
2613 if (rdma
->pin_all
&& !(cap
.flags
& RDMA_CAPABILITY_PIN_ALL
)) {
2614 ERROR(errp
, "Server cannot support pinning all memory. "
2615 "Will register memory dynamically.");
2616 rdma
->pin_all
= false;
2619 trace_qemu_rdma_connect_pin_all_outcome(rdma
->pin_all
);
2621 rdma_ack_cm_event(cm_event
);
2623 rdma
->control_ready_expected
= 1;
2627 err_rdma_source_connect
:
2628 qemu_rdma_cleanup(rdma
);
2632 static int qemu_rdma_dest_init(RDMAContext
*rdma
, Error
**errp
)
2635 struct rdma_cm_id
*listen_id
;
2636 char ip
[40] = "unknown";
2637 struct rdma_addrinfo
*res
, *e
;
2641 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2642 rdma
->wr_data
[idx
].control_len
= 0;
2643 rdma
->wr_data
[idx
].control_curr
= NULL
;
2646 if (!rdma
->host
|| !rdma
->host
[0]) {
2647 ERROR(errp
, "RDMA host is not set!");
2648 rdma
->error_state
= -EINVAL
;
2651 /* create CM channel */
2652 rdma
->channel
= rdma_create_event_channel();
2653 if (!rdma
->channel
) {
2654 ERROR(errp
, "could not create rdma event channel");
2655 rdma
->error_state
= -EINVAL
;
2660 ret
= rdma_create_id(rdma
->channel
, &listen_id
, NULL
, RDMA_PS_TCP
);
2662 ERROR(errp
, "could not create cm_id!");
2663 goto err_dest_init_create_listen_id
;
2666 snprintf(port_str
, 16, "%d", rdma
->port
);
2667 port_str
[15] = '\0';
2669 ret
= rdma_getaddrinfo(rdma
->host
, port_str
, NULL
, &res
);
2671 ERROR(errp
, "could not rdma_getaddrinfo address %s", rdma
->host
);
2672 goto err_dest_init_bind_addr
;
2675 ret
= rdma_set_option(listen_id
, RDMA_OPTION_ID
, RDMA_OPTION_ID_REUSEADDR
,
2676 &reuse
, sizeof reuse
);
2678 ERROR(errp
, "Error: could not set REUSEADDR option");
2679 goto err_dest_init_bind_addr
;
2681 for (e
= res
; e
!= NULL
; e
= e
->ai_next
) {
2682 inet_ntop(e
->ai_family
,
2683 &((struct sockaddr_in
*) e
->ai_dst_addr
)->sin_addr
, ip
, sizeof ip
);
2684 trace_qemu_rdma_dest_init_trying(rdma
->host
, ip
);
2685 ret
= rdma_bind_addr(listen_id
, e
->ai_dst_addr
);
2689 if (e
->ai_family
== AF_INET6
) {
2690 ret
= qemu_rdma_broken_ipv6_kernel(listen_id
->verbs
, errp
);
2698 rdma_freeaddrinfo(res
);
2700 ERROR(errp
, "Error: could not rdma_bind_addr!");
2701 goto err_dest_init_bind_addr
;
2704 rdma
->listen_id
= listen_id
;
2705 qemu_rdma_dump_gid("dest_init", listen_id
);
2708 err_dest_init_bind_addr
:
2709 rdma_destroy_id(listen_id
);
2710 err_dest_init_create_listen_id
:
2711 rdma_destroy_event_channel(rdma
->channel
);
2712 rdma
->channel
= NULL
;
2713 rdma
->error_state
= ret
;
2718 static void qemu_rdma_return_path_dest_init(RDMAContext
*rdma_return_path
,
2723 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
2724 rdma_return_path
->wr_data
[idx
].control_len
= 0;
2725 rdma_return_path
->wr_data
[idx
].control_curr
= NULL
;
2728 /*the CM channel and CM id is shared*/
2729 rdma_return_path
->channel
= rdma
->channel
;
2730 rdma_return_path
->listen_id
= rdma
->listen_id
;
2732 rdma
->return_path
= rdma_return_path
;
2733 rdma_return_path
->return_path
= rdma
;
2734 rdma_return_path
->is_return_path
= true;
2737 static void *qemu_rdma_data_init(const char *host_port
, Error
**errp
)
2739 RDMAContext
*rdma
= NULL
;
2740 InetSocketAddress
*addr
;
2743 rdma
= g_new0(RDMAContext
, 1);
2744 rdma
->current_index
= -1;
2745 rdma
->current_chunk
= -1;
2747 addr
= g_new(InetSocketAddress
, 1);
2748 if (!inet_parse(addr
, host_port
, NULL
)) {
2749 rdma
->port
= atoi(addr
->port
);
2750 rdma
->host
= g_strdup(addr
->host
);
2751 rdma
->host_port
= g_strdup(host_port
);
2753 ERROR(errp
, "bad RDMA migration address '%s'", host_port
);
2758 qapi_free_InetSocketAddress(addr
);
2765 * QEMUFile interface to the control channel.
2766 * SEND messages for control only.
2767 * VM's ram is handled with regular RDMA messages.
2769 static ssize_t
qio_channel_rdma_writev(QIOChannel
*ioc
,
2770 const struct iovec
*iov
,
2777 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2778 QEMUFile
*f
= rioc
->file
;
2785 RCU_READ_LOCK_GUARD();
2786 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
2789 error_setg(errp
, "RDMA control channel output is not set");
2793 CHECK_ERROR_STATE();
2796 * Push out any writes that
2797 * we're queued up for VM's ram.
2799 ret
= qemu_rdma_write_flush(f
, rdma
);
2801 rdma
->error_state
= ret
;
2802 error_setg(errp
, "qemu_rdma_write_flush returned %d", ret
);
2806 for (i
= 0; i
< niov
; i
++) {
2807 size_t remaining
= iov
[i
].iov_len
;
2808 uint8_t * data
= (void *)iov
[i
].iov_base
;
2810 RDMAControlHeader head
;
2812 len
= MIN(remaining
, RDMA_SEND_INCREMENT
);
2816 head
.type
= RDMA_CONTROL_QEMU_FILE
;
2818 ret
= qemu_rdma_exchange_send(rdma
, &head
, data
, NULL
, NULL
, NULL
);
2821 rdma
->error_state
= ret
;
2822 error_setg(errp
, "qemu_rdma_exchange_send returned %d", ret
);
2834 static size_t qemu_rdma_fill(RDMAContext
*rdma
, uint8_t *buf
,
2835 size_t size
, int idx
)
2839 if (rdma
->wr_data
[idx
].control_len
) {
2840 trace_qemu_rdma_fill(rdma
->wr_data
[idx
].control_len
, size
);
2842 len
= MIN(size
, rdma
->wr_data
[idx
].control_len
);
2843 memcpy(buf
, rdma
->wr_data
[idx
].control_curr
, len
);
2844 rdma
->wr_data
[idx
].control_curr
+= len
;
2845 rdma
->wr_data
[idx
].control_len
-= len
;
2852 * QEMUFile interface to the control channel.
2853 * RDMA links don't use bytestreams, so we have to
2854 * return bytes to QEMUFile opportunistically.
2856 static ssize_t
qio_channel_rdma_readv(QIOChannel
*ioc
,
2857 const struct iovec
*iov
,
2864 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2866 RDMAControlHeader head
;
2871 RCU_READ_LOCK_GUARD();
2872 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
2875 error_setg(errp
, "RDMA control channel input is not set");
2879 CHECK_ERROR_STATE();
2881 for (i
= 0; i
< niov
; i
++) {
2882 size_t want
= iov
[i
].iov_len
;
2883 uint8_t *data
= (void *)iov
[i
].iov_base
;
2886 * First, we hold on to the last SEND message we
2887 * were given and dish out the bytes until we run
2890 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2893 /* Got what we needed, so go to next iovec */
2898 /* If we got any data so far, then don't wait
2899 * for more, just return what we have */
2905 /* We've got nothing at all, so lets wait for
2908 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_QEMU_FILE
);
2911 rdma
->error_state
= ret
;
2912 error_setg(errp
, "qemu_rdma_exchange_recv returned %d", ret
);
2917 * SEND was received with new bytes, now try again.
2919 ret
= qemu_rdma_fill(rdma
, data
, want
, 0);
2923 /* Still didn't get enough, so lets just return */
2926 return QIO_CHANNEL_ERR_BLOCK
;
2936 * Block until all the outstanding chunks have been delivered by the hardware.
2938 static int qemu_rdma_drain_cq(QEMUFile
*f
, RDMAContext
*rdma
)
2942 if (qemu_rdma_write_flush(f
, rdma
) < 0) {
2946 while (rdma
->nb_sent
) {
2947 ret
= qemu_rdma_block_for_wrid(rdma
, RDMA_WRID_RDMA_WRITE
, NULL
);
2949 error_report("rdma migration: complete polling error!");
2954 qemu_rdma_unregister_waiting(rdma
);
2960 static int qio_channel_rdma_set_blocking(QIOChannel
*ioc
,
2964 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
2965 /* XXX we should make readv/writev actually honour this :-) */
2966 rioc
->blocking
= blocking
;
2971 typedef struct QIOChannelRDMASource QIOChannelRDMASource
;
2972 struct QIOChannelRDMASource
{
2974 QIOChannelRDMA
*rioc
;
2975 GIOCondition condition
;
2979 qio_channel_rdma_source_prepare(GSource
*source
,
2982 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
2984 GIOCondition cond
= 0;
2987 RCU_READ_LOCK_GUARD();
2988 if (rsource
->condition
== G_IO_IN
) {
2989 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
2991 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
2995 error_report("RDMAContext is NULL when prepare Gsource");
2999 if (rdma
->wr_data
[0].control_len
) {
3004 return cond
& rsource
->condition
;
3008 qio_channel_rdma_source_check(GSource
*source
)
3010 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3012 GIOCondition cond
= 0;
3014 RCU_READ_LOCK_GUARD();
3015 if (rsource
->condition
== G_IO_IN
) {
3016 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3018 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3022 error_report("RDMAContext is NULL when check Gsource");
3026 if (rdma
->wr_data
[0].control_len
) {
3031 return cond
& rsource
->condition
;
3035 qio_channel_rdma_source_dispatch(GSource
*source
,
3036 GSourceFunc callback
,
3039 QIOChannelFunc func
= (QIOChannelFunc
)callback
;
3040 QIOChannelRDMASource
*rsource
= (QIOChannelRDMASource
*)source
;
3042 GIOCondition cond
= 0;
3044 RCU_READ_LOCK_GUARD();
3045 if (rsource
->condition
== G_IO_IN
) {
3046 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmain
);
3048 rdma
= qatomic_rcu_read(&rsource
->rioc
->rdmaout
);
3052 error_report("RDMAContext is NULL when dispatch Gsource");
3056 if (rdma
->wr_data
[0].control_len
) {
3061 return (*func
)(QIO_CHANNEL(rsource
->rioc
),
3062 (cond
& rsource
->condition
),
3067 qio_channel_rdma_source_finalize(GSource
*source
)
3069 QIOChannelRDMASource
*ssource
= (QIOChannelRDMASource
*)source
;
3071 object_unref(OBJECT(ssource
->rioc
));
3074 GSourceFuncs qio_channel_rdma_source_funcs
= {
3075 qio_channel_rdma_source_prepare
,
3076 qio_channel_rdma_source_check
,
3077 qio_channel_rdma_source_dispatch
,
3078 qio_channel_rdma_source_finalize
3081 static GSource
*qio_channel_rdma_create_watch(QIOChannel
*ioc
,
3082 GIOCondition condition
)
3084 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3085 QIOChannelRDMASource
*ssource
;
3088 source
= g_source_new(&qio_channel_rdma_source_funcs
,
3089 sizeof(QIOChannelRDMASource
));
3090 ssource
= (QIOChannelRDMASource
*)source
;
3092 ssource
->rioc
= rioc
;
3093 object_ref(OBJECT(rioc
));
3095 ssource
->condition
= condition
;
3100 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel
*ioc
,
3103 IOHandler
*io_write
,
3106 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3108 aio_set_fd_handler(ctx
, rioc
->rdmain
->recv_comp_channel
->fd
,
3109 false, io_read
, io_write
, NULL
, NULL
, opaque
);
3110 aio_set_fd_handler(ctx
, rioc
->rdmain
->send_comp_channel
->fd
,
3111 false, io_read
, io_write
, NULL
, NULL
, opaque
);
3113 aio_set_fd_handler(ctx
, rioc
->rdmaout
->recv_comp_channel
->fd
,
3114 false, io_read
, io_write
, NULL
, NULL
, opaque
);
3115 aio_set_fd_handler(ctx
, rioc
->rdmaout
->send_comp_channel
->fd
,
3116 false, io_read
, io_write
, NULL
, NULL
, opaque
);
3120 struct rdma_close_rcu
{
3121 struct rcu_head rcu
;
3122 RDMAContext
*rdmain
;
3123 RDMAContext
*rdmaout
;
3126 /* callback from qio_channel_rdma_close via call_rcu */
3127 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu
*rcu
)
3130 qemu_rdma_cleanup(rcu
->rdmain
);
3134 qemu_rdma_cleanup(rcu
->rdmaout
);
3137 g_free(rcu
->rdmain
);
3138 g_free(rcu
->rdmaout
);
3142 static int qio_channel_rdma_close(QIOChannel
*ioc
,
3145 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3146 RDMAContext
*rdmain
, *rdmaout
;
3147 struct rdma_close_rcu
*rcu
= g_new(struct rdma_close_rcu
, 1);
3149 trace_qemu_rdma_close();
3151 rdmain
= rioc
->rdmain
;
3153 qatomic_rcu_set(&rioc
->rdmain
, NULL
);
3156 rdmaout
= rioc
->rdmaout
;
3158 qatomic_rcu_set(&rioc
->rdmaout
, NULL
);
3161 rcu
->rdmain
= rdmain
;
3162 rcu
->rdmaout
= rdmaout
;
3163 call_rcu(rcu
, qio_channel_rdma_close_rcu
, rcu
);
3169 qio_channel_rdma_shutdown(QIOChannel
*ioc
,
3170 QIOChannelShutdown how
,
3173 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(ioc
);
3174 RDMAContext
*rdmain
, *rdmaout
;
3176 RCU_READ_LOCK_GUARD();
3178 rdmain
= qatomic_rcu_read(&rioc
->rdmain
);
3179 rdmaout
= qatomic_rcu_read(&rioc
->rdmain
);
3182 case QIO_CHANNEL_SHUTDOWN_READ
:
3184 rdmain
->error_state
= -1;
3187 case QIO_CHANNEL_SHUTDOWN_WRITE
:
3189 rdmaout
->error_state
= -1;
3192 case QIO_CHANNEL_SHUTDOWN_BOTH
:
3195 rdmain
->error_state
= -1;
3198 rdmaout
->error_state
= -1;
3209 * This means that 'block_offset' is a full virtual address that does not
3210 * belong to a RAMBlock of the virtual machine and instead
3211 * represents a private malloc'd memory area that the caller wishes to
3215 * Offset is an offset to be added to block_offset and used
3216 * to also lookup the corresponding RAMBlock.
3218 * @size : Number of bytes to transfer
3220 * @bytes_sent : User-specificed pointer to indicate how many bytes were
3221 * sent. Usually, this will not be more than a few bytes of
3222 * the protocol because most transfers are sent asynchronously.
3224 static size_t qemu_rdma_save_page(QEMUFile
*f
,
3225 ram_addr_t block_offset
, ram_addr_t offset
,
3226 size_t size
, uint64_t *bytes_sent
)
3228 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3232 RCU_READ_LOCK_GUARD();
3233 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3239 CHECK_ERROR_STATE();
3241 if (migration_in_postcopy()) {
3242 return RAM_SAVE_CONTROL_NOT_SUPP
;
3248 * Add this page to the current 'chunk'. If the chunk
3249 * is full, or the page doesn't belong to the current chunk,
3250 * an actual RDMA write will occur and a new chunk will be formed.
3252 ret
= qemu_rdma_write(f
, rdma
, block_offset
, offset
, size
);
3254 error_report("rdma migration: write error! %d", ret
);
3259 * We always return 1 bytes because the RDMA
3260 * protocol is completely asynchronous. We do not yet know
3261 * whether an identified chunk is zero or not because we're
3262 * waiting for other pages to potentially be merged with
3263 * the current chunk. So, we have to call qemu_update_position()
3264 * later on when the actual write occurs.
3271 * Drain the Completion Queue if possible, but do not block,
3274 * If nothing to poll, the end of the iteration will do this
3275 * again to make sure we don't overflow the request queue.
3278 uint64_t wr_id
, wr_id_in
;
3279 int ret
= qemu_rdma_poll(rdma
, rdma
->recv_cq
, &wr_id_in
, NULL
);
3281 error_report("rdma migration: polling error! %d", ret
);
3285 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3287 if (wr_id
== RDMA_WRID_NONE
) {
3293 uint64_t wr_id
, wr_id_in
;
3294 int ret
= qemu_rdma_poll(rdma
, rdma
->send_cq
, &wr_id_in
, NULL
);
3296 error_report("rdma migration: polling error! %d", ret
);
3300 wr_id
= wr_id_in
& RDMA_WRID_TYPE_MASK
;
3302 if (wr_id
== RDMA_WRID_NONE
) {
3307 return RAM_SAVE_CONTROL_DELAYED
;
3309 rdma
->error_state
= ret
;
3313 static void rdma_accept_incoming_migration(void *opaque
);
3315 static void rdma_cm_poll_handler(void *opaque
)
3317 RDMAContext
*rdma
= opaque
;
3319 struct rdma_cm_event
*cm_event
;
3320 MigrationIncomingState
*mis
= migration_incoming_get_current();
3322 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3324 error_report("get_cm_event failed %d", errno
);
3328 if (cm_event
->event
== RDMA_CM_EVENT_DISCONNECTED
||
3329 cm_event
->event
== RDMA_CM_EVENT_DEVICE_REMOVAL
) {
3330 if (!rdma
->error_state
&&
3331 migration_incoming_get_current()->state
!=
3332 MIGRATION_STATUS_COMPLETED
) {
3333 error_report("receive cm event, cm event is %d", cm_event
->event
);
3334 rdma
->error_state
= -EPIPE
;
3335 if (rdma
->return_path
) {
3336 rdma
->return_path
->error_state
= -EPIPE
;
3339 rdma_ack_cm_event(cm_event
);
3341 if (mis
->migration_incoming_co
) {
3342 qemu_coroutine_enter(mis
->migration_incoming_co
);
3346 rdma_ack_cm_event(cm_event
);
3349 static int qemu_rdma_accept(RDMAContext
*rdma
)
3351 RDMACapabilities cap
;
3352 struct rdma_conn_param conn_param
= {
3353 .responder_resources
= 2,
3354 .private_data
= &cap
,
3355 .private_data_len
= sizeof(cap
),
3357 RDMAContext
*rdma_return_path
= NULL
;
3358 struct rdma_cm_event
*cm_event
;
3359 struct ibv_context
*verbs
;
3363 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3365 goto err_rdma_dest_wait
;
3368 if (cm_event
->event
!= RDMA_CM_EVENT_CONNECT_REQUEST
) {
3369 rdma_ack_cm_event(cm_event
);
3370 goto err_rdma_dest_wait
;
3374 * initialize the RDMAContext for return path for postcopy after first
3375 * connection request reached.
3377 if ((migrate_postcopy() || migrate_return_path())
3378 && !rdma
->is_return_path
) {
3379 rdma_return_path
= qemu_rdma_data_init(rdma
->host_port
, NULL
);
3380 if (rdma_return_path
== NULL
) {
3381 rdma_ack_cm_event(cm_event
);
3382 goto err_rdma_dest_wait
;
3385 qemu_rdma_return_path_dest_init(rdma_return_path
, rdma
);
3388 memcpy(&cap
, cm_event
->param
.conn
.private_data
, sizeof(cap
));
3390 network_to_caps(&cap
);
3392 if (cap
.version
< 1 || cap
.version
> RDMA_CONTROL_VERSION_CURRENT
) {
3393 error_report("Unknown source RDMA version: %d, bailing...",
3395 rdma_ack_cm_event(cm_event
);
3396 goto err_rdma_dest_wait
;
3400 * Respond with only the capabilities this version of QEMU knows about.
3402 cap
.flags
&= known_capabilities
;
3405 * Enable the ones that we do know about.
3406 * Add other checks here as new ones are introduced.
3408 if (cap
.flags
& RDMA_CAPABILITY_PIN_ALL
) {
3409 rdma
->pin_all
= true;
3412 rdma
->cm_id
= cm_event
->id
;
3413 verbs
= cm_event
->id
->verbs
;
3415 rdma_ack_cm_event(cm_event
);
3417 trace_qemu_rdma_accept_pin_state(rdma
->pin_all
);
3419 caps_to_network(&cap
);
3421 trace_qemu_rdma_accept_pin_verbsc(verbs
);
3424 rdma
->verbs
= verbs
;
3425 } else if (rdma
->verbs
!= verbs
) {
3426 error_report("ibv context not matching %p, %p!", rdma
->verbs
,
3428 goto err_rdma_dest_wait
;
3431 qemu_rdma_dump_id("dest_init", verbs
);
3433 ret
= qemu_rdma_alloc_pd_cq(rdma
);
3435 error_report("rdma migration: error allocating pd and cq!");
3436 goto err_rdma_dest_wait
;
3439 ret
= qemu_rdma_alloc_qp(rdma
);
3441 error_report("rdma migration: error allocating qp!");
3442 goto err_rdma_dest_wait
;
3445 ret
= qemu_rdma_init_ram_blocks(rdma
);
3447 error_report("rdma migration: error initializing ram blocks!");
3448 goto err_rdma_dest_wait
;
3451 for (idx
= 0; idx
< RDMA_WRID_MAX
; idx
++) {
3452 ret
= qemu_rdma_reg_control(rdma
, idx
);
3454 error_report("rdma: error registering %d control", idx
);
3455 goto err_rdma_dest_wait
;
3459 /* Accept the second connection request for return path */
3460 if ((migrate_postcopy() || migrate_return_path())
3461 && !rdma
->is_return_path
) {
3462 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
3464 (void *)(intptr_t)rdma
->return_path
);
3466 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_cm_poll_handler
,
3470 ret
= rdma_accept(rdma
->cm_id
, &conn_param
);
3472 error_report("rdma_accept returns %d", ret
);
3473 goto err_rdma_dest_wait
;
3476 ret
= rdma_get_cm_event(rdma
->channel
, &cm_event
);
3478 error_report("rdma_accept get_cm_event failed %d", ret
);
3479 goto err_rdma_dest_wait
;
3482 if (cm_event
->event
!= RDMA_CM_EVENT_ESTABLISHED
) {
3483 error_report("rdma_accept not event established");
3484 rdma_ack_cm_event(cm_event
);
3485 goto err_rdma_dest_wait
;
3488 rdma_ack_cm_event(cm_event
);
3489 rdma
->connected
= true;
3491 ret
= qemu_rdma_post_recv_control(rdma
, RDMA_WRID_READY
);
3493 error_report("rdma migration: error posting second control recv");
3494 goto err_rdma_dest_wait
;
3497 qemu_rdma_dump_gid("dest_connect", rdma
->cm_id
);
3502 rdma
->error_state
= ret
;
3503 qemu_rdma_cleanup(rdma
);
3504 g_free(rdma_return_path
);
3508 static int dest_ram_sort_func(const void *a
, const void *b
)
3510 unsigned int a_index
= ((const RDMALocalBlock
*)a
)->src_index
;
3511 unsigned int b_index
= ((const RDMALocalBlock
*)b
)->src_index
;
3513 return (a_index
< b_index
) ? -1 : (a_index
!= b_index
);
3517 * During each iteration of the migration, we listen for instructions
3518 * by the source VM to perform dynamic page registrations before they
3519 * can perform RDMA operations.
3521 * We respond with the 'rkey'.
3523 * Keep doing this until the source tells us to stop.
3525 static int qemu_rdma_registration_handle(QEMUFile
*f
, void *opaque
)
3527 RDMAControlHeader reg_resp
= { .len
= sizeof(RDMARegisterResult
),
3528 .type
= RDMA_CONTROL_REGISTER_RESULT
,
3531 RDMAControlHeader unreg_resp
= { .len
= 0,
3532 .type
= RDMA_CONTROL_UNREGISTER_FINISHED
,
3535 RDMAControlHeader blocks
= { .type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
,
3537 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(opaque
);
3539 RDMALocalBlocks
*local
;
3540 RDMAControlHeader head
;
3541 RDMARegister
*reg
, *registers
;
3543 RDMARegisterResult
*reg_result
;
3544 static RDMARegisterResult results
[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
];
3545 RDMALocalBlock
*block
;
3552 RCU_READ_LOCK_GUARD();
3553 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3559 CHECK_ERROR_STATE();
3561 local
= &rdma
->local_ram_blocks
;
3563 trace_qemu_rdma_registration_handle_wait();
3565 ret
= qemu_rdma_exchange_recv(rdma
, &head
, RDMA_CONTROL_NONE
);
3571 if (head
.repeat
> RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE
) {
3572 error_report("rdma: Too many requests in this message (%d)."
3573 "Bailing.", head
.repeat
);
3578 switch (head
.type
) {
3579 case RDMA_CONTROL_COMPRESS
:
3580 comp
= (RDMACompress
*) rdma
->wr_data
[idx
].control_curr
;
3581 network_to_compress(comp
);
3583 trace_qemu_rdma_registration_handle_compress(comp
->length
,
3586 if (comp
->block_idx
>= rdma
->local_ram_blocks
.nb_blocks
) {
3587 error_report("rdma: 'compress' bad block index %u (vs %d)",
3588 (unsigned int)comp
->block_idx
,
3589 rdma
->local_ram_blocks
.nb_blocks
);
3593 block
= &(rdma
->local_ram_blocks
.block
[comp
->block_idx
]);
3595 host_addr
= block
->local_host_addr
+
3596 (comp
->offset
- block
->offset
);
3598 ram_handle_compressed(host_addr
, comp
->value
, comp
->length
);
3601 case RDMA_CONTROL_REGISTER_FINISHED
:
3602 trace_qemu_rdma_registration_handle_finished();
3605 case RDMA_CONTROL_RAM_BLOCKS_REQUEST
:
3606 trace_qemu_rdma_registration_handle_ram_blocks();
3608 /* Sort our local RAM Block list so it's the same as the source,
3609 * we can do this since we've filled in a src_index in the list
3610 * as we received the RAMBlock list earlier.
3612 qsort(rdma
->local_ram_blocks
.block
,
3613 rdma
->local_ram_blocks
.nb_blocks
,
3614 sizeof(RDMALocalBlock
), dest_ram_sort_func
);
3615 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3616 local
->block
[i
].index
= i
;
3619 if (rdma
->pin_all
) {
3620 ret
= qemu_rdma_reg_whole_ram_blocks(rdma
);
3622 error_report("rdma migration: error dest "
3623 "registering ram blocks");
3629 * Dest uses this to prepare to transmit the RAMBlock descriptions
3630 * to the source VM after connection setup.
3631 * Both sides use the "remote" structure to communicate and update
3632 * their "local" descriptions with what was sent.
3634 for (i
= 0; i
< local
->nb_blocks
; i
++) {
3635 rdma
->dest_blocks
[i
].remote_host_addr
=
3636 (uintptr_t)(local
->block
[i
].local_host_addr
);
3638 if (rdma
->pin_all
) {
3639 rdma
->dest_blocks
[i
].remote_rkey
= local
->block
[i
].mr
->rkey
;
3642 rdma
->dest_blocks
[i
].offset
= local
->block
[i
].offset
;
3643 rdma
->dest_blocks
[i
].length
= local
->block
[i
].length
;
3645 dest_block_to_network(&rdma
->dest_blocks
[i
]);
3646 trace_qemu_rdma_registration_handle_ram_blocks_loop(
3647 local
->block
[i
].block_name
,
3648 local
->block
[i
].offset
,
3649 local
->block
[i
].length
,
3650 local
->block
[i
].local_host_addr
,
3651 local
->block
[i
].src_index
);
3654 blocks
.len
= rdma
->local_ram_blocks
.nb_blocks
3655 * sizeof(RDMADestBlock
);
3658 ret
= qemu_rdma_post_send_control(rdma
,
3659 (uint8_t *) rdma
->dest_blocks
, &blocks
);
3662 error_report("rdma migration: error sending remote info");
3667 case RDMA_CONTROL_REGISTER_REQUEST
:
3668 trace_qemu_rdma_registration_handle_register(head
.repeat
);
3670 reg_resp
.repeat
= head
.repeat
;
3671 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3673 for (count
= 0; count
< head
.repeat
; count
++) {
3675 uint8_t *chunk_start
, *chunk_end
;
3677 reg
= ®isters
[count
];
3678 network_to_register(reg
);
3680 reg_result
= &results
[count
];
3682 trace_qemu_rdma_registration_handle_register_loop(count
,
3683 reg
->current_index
, reg
->key
.current_addr
, reg
->chunks
);
3685 if (reg
->current_index
>= rdma
->local_ram_blocks
.nb_blocks
) {
3686 error_report("rdma: 'register' bad block index %u (vs %d)",
3687 (unsigned int)reg
->current_index
,
3688 rdma
->local_ram_blocks
.nb_blocks
);
3692 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3693 if (block
->is_ram_block
) {
3694 if (block
->offset
> reg
->key
.current_addr
) {
3695 error_report("rdma: bad register address for block %s"
3696 " offset: %" PRIx64
" current_addr: %" PRIx64
,
3697 block
->block_name
, block
->offset
,
3698 reg
->key
.current_addr
);
3702 host_addr
= (block
->local_host_addr
+
3703 (reg
->key
.current_addr
- block
->offset
));
3704 chunk
= ram_chunk_index(block
->local_host_addr
,
3705 (uint8_t *) host_addr
);
3707 chunk
= reg
->key
.chunk
;
3708 host_addr
= block
->local_host_addr
+
3709 (reg
->key
.chunk
* (1UL << RDMA_REG_CHUNK_SHIFT
));
3710 /* Check for particularly bad chunk value */
3711 if (host_addr
< (void *)block
->local_host_addr
) {
3712 error_report("rdma: bad chunk for block %s"
3714 block
->block_name
, reg
->key
.chunk
);
3719 chunk_start
= ram_chunk_start(block
, chunk
);
3720 chunk_end
= ram_chunk_end(block
, chunk
+ reg
->chunks
);
3721 /* avoid "-Waddress-of-packed-member" warning */
3722 uint32_t tmp_rkey
= 0;
3723 if (qemu_rdma_register_and_get_keys(rdma
, block
,
3724 (uintptr_t)host_addr
, NULL
, &tmp_rkey
,
3725 chunk
, chunk_start
, chunk_end
)) {
3726 error_report("cannot get rkey");
3730 reg_result
->rkey
= tmp_rkey
;
3732 reg_result
->host_addr
= (uintptr_t)block
->local_host_addr
;
3734 trace_qemu_rdma_registration_handle_register_rkey(
3737 result_to_network(reg_result
);
3740 ret
= qemu_rdma_post_send_control(rdma
,
3741 (uint8_t *) results
, ®_resp
);
3744 error_report("Failed to send control buffer");
3748 case RDMA_CONTROL_UNREGISTER_REQUEST
:
3749 trace_qemu_rdma_registration_handle_unregister(head
.repeat
);
3750 unreg_resp
.repeat
= head
.repeat
;
3751 registers
= (RDMARegister
*) rdma
->wr_data
[idx
].control_curr
;
3753 for (count
= 0; count
< head
.repeat
; count
++) {
3754 reg
= ®isters
[count
];
3755 network_to_register(reg
);
3757 trace_qemu_rdma_registration_handle_unregister_loop(count
,
3758 reg
->current_index
, reg
->key
.chunk
);
3760 block
= &(rdma
->local_ram_blocks
.block
[reg
->current_index
]);
3762 ret
= ibv_dereg_mr(block
->pmr
[reg
->key
.chunk
]);
3763 block
->pmr
[reg
->key
.chunk
] = NULL
;
3766 perror("rdma unregistration chunk failed");
3771 rdma
->total_registrations
--;
3773 trace_qemu_rdma_registration_handle_unregister_success(
3777 ret
= qemu_rdma_post_send_control(rdma
, NULL
, &unreg_resp
);
3780 error_report("Failed to send control buffer");
3784 case RDMA_CONTROL_REGISTER_RESULT
:
3785 error_report("Invalid RESULT message at dest.");
3789 error_report("Unknown control message %s", control_desc(head
.type
));
3796 rdma
->error_state
= ret
;
3802 * Called via a ram_control_load_hook during the initial RAM load section which
3803 * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks
3805 * We've already built our local RAMBlock list, but not yet sent the list to
3809 rdma_block_notification_handle(QIOChannelRDMA
*rioc
, const char *name
)
3815 RCU_READ_LOCK_GUARD();
3816 rdma
= qatomic_rcu_read(&rioc
->rdmain
);
3822 /* Find the matching RAMBlock in our local list */
3823 for (curr
= 0; curr
< rdma
->local_ram_blocks
.nb_blocks
; curr
++) {
3824 if (!strcmp(rdma
->local_ram_blocks
.block
[curr
].block_name
, name
)) {
3831 error_report("RAMBlock '%s' not found on destination", name
);
3835 rdma
->local_ram_blocks
.block
[curr
].src_index
= rdma
->next_src_index
;
3836 trace_rdma_block_notification_handle(name
, rdma
->next_src_index
);
3837 rdma
->next_src_index
++;
3842 static int rdma_load_hook(QEMUFile
*f
, uint64_t flags
, void *data
)
3844 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3846 case RAM_CONTROL_BLOCK_REG
:
3847 return rdma_block_notification_handle(rioc
, data
);
3849 case RAM_CONTROL_HOOK
:
3850 return qemu_rdma_registration_handle(f
, rioc
);
3853 /* Shouldn't be called with any other values */
3858 static int qemu_rdma_registration_start(QEMUFile
*f
,
3859 uint64_t flags
, void *data
)
3861 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3864 RCU_READ_LOCK_GUARD();
3865 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3870 CHECK_ERROR_STATE();
3872 if (migration_in_postcopy()) {
3876 trace_qemu_rdma_registration_start(flags
);
3877 qemu_put_be64(f
, RAM_SAVE_FLAG_HOOK
);
3884 * Inform dest that dynamic registrations are done for now.
3885 * First, flush writes, if any.
3887 static int qemu_rdma_registration_stop(QEMUFile
*f
,
3888 uint64_t flags
, void *data
)
3890 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(qemu_file_get_ioc(f
));
3892 RDMAControlHeader head
= { .len
= 0, .repeat
= 1 };
3895 RCU_READ_LOCK_GUARD();
3896 rdma
= qatomic_rcu_read(&rioc
->rdmaout
);
3901 CHECK_ERROR_STATE();
3903 if (migration_in_postcopy()) {
3908 ret
= qemu_rdma_drain_cq(f
, rdma
);
3914 if (flags
== RAM_CONTROL_SETUP
) {
3915 RDMAControlHeader resp
= {.type
= RDMA_CONTROL_RAM_BLOCKS_RESULT
};
3916 RDMALocalBlocks
*local
= &rdma
->local_ram_blocks
;
3917 int reg_result_idx
, i
, nb_dest_blocks
;
3919 head
.type
= RDMA_CONTROL_RAM_BLOCKS_REQUEST
;
3920 trace_qemu_rdma_registration_stop_ram();
3923 * Make sure that we parallelize the pinning on both sides.
3924 * For very large guests, doing this serially takes a really
3925 * long time, so we have to 'interleave' the pinning locally
3926 * with the control messages by performing the pinning on this
3927 * side before we receive the control response from the other
3928 * side that the pinning has completed.
3930 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, &resp
,
3931 ®_result_idx
, rdma
->pin_all
?
3932 qemu_rdma_reg_whole_ram_blocks
: NULL
);
3934 fprintf(stderr
, "receiving remote info!");
3938 nb_dest_blocks
= resp
.len
/ sizeof(RDMADestBlock
);
3941 * The protocol uses two different sets of rkeys (mutually exclusive):
3942 * 1. One key to represent the virtual address of the entire ram block.
3943 * (dynamic chunk registration disabled - pin everything with one rkey.)
3944 * 2. One to represent individual chunks within a ram block.
3945 * (dynamic chunk registration enabled - pin individual chunks.)
3947 * Once the capability is successfully negotiated, the destination transmits
3948 * the keys to use (or sends them later) including the virtual addresses
3949 * and then propagates the remote ram block descriptions to his local copy.
3952 if (local
->nb_blocks
!= nb_dest_blocks
) {
3953 fprintf(stderr
, "ram blocks mismatch (Number of blocks %d vs %d) "
3954 "Your QEMU command line parameters are probably "
3955 "not identical on both the source and destination.",
3956 local
->nb_blocks
, nb_dest_blocks
);
3957 rdma
->error_state
= -EINVAL
;
3961 qemu_rdma_move_header(rdma
, reg_result_idx
, &resp
);
3962 memcpy(rdma
->dest_blocks
,
3963 rdma
->wr_data
[reg_result_idx
].control_curr
, resp
.len
);
3964 for (i
= 0; i
< nb_dest_blocks
; i
++) {
3965 network_to_dest_block(&rdma
->dest_blocks
[i
]);
3967 /* We require that the blocks are in the same order */
3968 if (rdma
->dest_blocks
[i
].length
!= local
->block
[i
].length
) {
3969 fprintf(stderr
, "Block %s/%d has a different length %" PRIu64
3970 "vs %" PRIu64
, local
->block
[i
].block_name
, i
,
3971 local
->block
[i
].length
,
3972 rdma
->dest_blocks
[i
].length
);
3973 rdma
->error_state
= -EINVAL
;
3976 local
->block
[i
].remote_host_addr
=
3977 rdma
->dest_blocks
[i
].remote_host_addr
;
3978 local
->block
[i
].remote_rkey
= rdma
->dest_blocks
[i
].remote_rkey
;
3982 trace_qemu_rdma_registration_stop(flags
);
3984 head
.type
= RDMA_CONTROL_REGISTER_FINISHED
;
3985 ret
= qemu_rdma_exchange_send(rdma
, &head
, NULL
, NULL
, NULL
, NULL
);
3993 rdma
->error_state
= ret
;
3997 static const QEMUFileHooks rdma_read_hooks
= {
3998 .hook_ram_load
= rdma_load_hook
,
4001 static const QEMUFileHooks rdma_write_hooks
= {
4002 .before_ram_iterate
= qemu_rdma_registration_start
,
4003 .after_ram_iterate
= qemu_rdma_registration_stop
,
4004 .save_page
= qemu_rdma_save_page
,
4008 static void qio_channel_rdma_finalize(Object
*obj
)
4010 QIOChannelRDMA
*rioc
= QIO_CHANNEL_RDMA(obj
);
4012 qemu_rdma_cleanup(rioc
->rdmain
);
4013 g_free(rioc
->rdmain
);
4014 rioc
->rdmain
= NULL
;
4016 if (rioc
->rdmaout
) {
4017 qemu_rdma_cleanup(rioc
->rdmaout
);
4018 g_free(rioc
->rdmaout
);
4019 rioc
->rdmaout
= NULL
;
4023 static void qio_channel_rdma_class_init(ObjectClass
*klass
,
4024 void *class_data G_GNUC_UNUSED
)
4026 QIOChannelClass
*ioc_klass
= QIO_CHANNEL_CLASS(klass
);
4028 ioc_klass
->io_writev
= qio_channel_rdma_writev
;
4029 ioc_klass
->io_readv
= qio_channel_rdma_readv
;
4030 ioc_klass
->io_set_blocking
= qio_channel_rdma_set_blocking
;
4031 ioc_klass
->io_close
= qio_channel_rdma_close
;
4032 ioc_klass
->io_create_watch
= qio_channel_rdma_create_watch
;
4033 ioc_klass
->io_set_aio_fd_handler
= qio_channel_rdma_set_aio_fd_handler
;
4034 ioc_klass
->io_shutdown
= qio_channel_rdma_shutdown
;
4037 static const TypeInfo qio_channel_rdma_info
= {
4038 .parent
= TYPE_QIO_CHANNEL
,
4039 .name
= TYPE_QIO_CHANNEL_RDMA
,
4040 .instance_size
= sizeof(QIOChannelRDMA
),
4041 .instance_finalize
= qio_channel_rdma_finalize
,
4042 .class_init
= qio_channel_rdma_class_init
,
4045 static void qio_channel_rdma_register_types(void)
4047 type_register_static(&qio_channel_rdma_info
);
4050 type_init(qio_channel_rdma_register_types
);
4052 static QEMUFile
*qemu_fopen_rdma(RDMAContext
*rdma
, const char *mode
)
4054 QIOChannelRDMA
*rioc
;
4056 if (qemu_file_mode_is_not_valid(mode
)) {
4060 rioc
= QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA
));
4062 if (mode
[0] == 'w') {
4063 rioc
->file
= qemu_file_new_output(QIO_CHANNEL(rioc
));
4064 rioc
->rdmaout
= rdma
;
4065 rioc
->rdmain
= rdma
->return_path
;
4066 qemu_file_set_hooks(rioc
->file
, &rdma_write_hooks
);
4068 rioc
->file
= qemu_file_new_input(QIO_CHANNEL(rioc
));
4069 rioc
->rdmain
= rdma
;
4070 rioc
->rdmaout
= rdma
->return_path
;
4071 qemu_file_set_hooks(rioc
->file
, &rdma_read_hooks
);
4077 static void rdma_accept_incoming_migration(void *opaque
)
4079 RDMAContext
*rdma
= opaque
;
4082 Error
*local_err
= NULL
;
4084 trace_qemu_rdma_accept_incoming_migration();
4085 ret
= qemu_rdma_accept(rdma
);
4088 fprintf(stderr
, "RDMA ERROR: Migration initialization failed\n");
4092 trace_qemu_rdma_accept_incoming_migration_accepted();
4094 if (rdma
->is_return_path
) {
4098 f
= qemu_fopen_rdma(rdma
, "rb");
4100 fprintf(stderr
, "RDMA ERROR: could not qemu_fopen_rdma\n");
4101 qemu_rdma_cleanup(rdma
);
4105 rdma
->migration_started_on_destination
= 1;
4106 migration_fd_process_incoming(f
, &local_err
);
4108 error_reportf_err(local_err
, "RDMA ERROR:");
4112 void rdma_start_incoming_migration(const char *host_port
, Error
**errp
)
4116 Error
*local_err
= NULL
;
4118 trace_rdma_start_incoming_migration();
4120 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4121 if (ram_block_discard_is_required()) {
4122 error_setg(errp
, "RDMA: cannot disable RAM discard");
4126 rdma
= qemu_rdma_data_init(host_port
, &local_err
);
4131 ret
= qemu_rdma_dest_init(rdma
, &local_err
);
4137 trace_rdma_start_incoming_migration_after_dest_init();
4139 ret
= rdma_listen(rdma
->listen_id
, 5);
4142 ERROR(errp
, "listening on socket!");
4146 trace_rdma_start_incoming_migration_after_rdma_listen();
4148 qemu_set_fd_handler(rdma
->channel
->fd
, rdma_accept_incoming_migration
,
4149 NULL
, (void *)(intptr_t)rdma
);
4153 qemu_rdma_cleanup(rdma
);
4155 error_propagate(errp
, local_err
);
4158 g_free(rdma
->host_port
);
4163 void rdma_start_outgoing_migration(void *opaque
,
4164 const char *host_port
, Error
**errp
)
4166 MigrationState
*s
= opaque
;
4167 RDMAContext
*rdma_return_path
= NULL
;
4171 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4172 if (ram_block_discard_is_required()) {
4173 error_setg(errp
, "RDMA: cannot disable RAM discard");
4177 rdma
= qemu_rdma_data_init(host_port
, errp
);
4182 ret
= qemu_rdma_source_init(rdma
, migrate_rdma_pin_all(), errp
);
4188 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4189 ret
= qemu_rdma_connect(rdma
, errp
, false);
4195 /* RDMA postcopy need a separate queue pair for return path */
4196 if (migrate_postcopy() || migrate_return_path()) {
4197 rdma_return_path
= qemu_rdma_data_init(host_port
, errp
);
4199 if (rdma_return_path
== NULL
) {
4200 goto return_path_err
;
4203 ret
= qemu_rdma_source_init(rdma_return_path
,
4204 migrate_rdma_pin_all(), errp
);
4207 goto return_path_err
;
4210 ret
= qemu_rdma_connect(rdma_return_path
, errp
, true);
4213 goto return_path_err
;
4216 rdma
->return_path
= rdma_return_path
;
4217 rdma_return_path
->return_path
= rdma
;
4218 rdma_return_path
->is_return_path
= true;
4221 trace_rdma_start_outgoing_migration_after_rdma_connect();
4223 s
->to_dst_file
= qemu_fopen_rdma(rdma
, "wb");
4224 migrate_fd_connect(s
, NULL
);
4227 qemu_rdma_cleanup(rdma
);
4230 g_free(rdma_return_path
);