1 /*
2 * RDMA protocol and interfaces
3 *
4 * Copyright IBM, Corp. 2010-2013
5 * Copyright Red Hat, Inc. 2015-2016
6 *
7 * Authors:
8 * Michael R. Hines <mrhines@us.ibm.com>
9 * Jiuxing Liu <jl@us.ibm.com>
10 * Daniel P. Berrange <berrange@redhat.com>
11 *
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.
14 *
15 */
16
17 #include "qemu/osdep.h"
18 #include "qapi/error.h"
19 #include "qemu/cutils.h"
20 #include "exec/target_page.h"
21 #include "rdma.h"
22 #include "migration.h"
23 #include "migration-stats.h"
24 #include "qemu-file.h"
25 #include "ram.h"
26 #include "qemu/error-report.h"
27 #include "qemu/main-loop.h"
28 #include "qemu/module.h"
29 #include "qemu/rcu.h"
30 #include "qemu/sockets.h"
31 #include "qemu/bitmap.h"
32 #include "qemu/coroutine.h"
33 #include "exec/memory.h"
34 #include <sys/socket.h>
35 #include <netdb.h>
36 #include <arpa/inet.h>
37 #include <rdma/rdma_cma.h>
38 #include "trace.h"
39 #include "qom/object.h"
40 #include "options.h"
41 #include <poll.h>
42
43 #define RDMA_RESOLVE_TIMEOUT_MS 10000
44
45 /* Do not merge data if larger than this. */
46 #define RDMA_MERGE_MAX (2 * 1024 * 1024)
47 #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096)
48
49 #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */
50
51 /*
52 * This is only for non-live state being migrated.
53 * Instead of RDMA_WRITE messages, we use RDMA_SEND
54 * messages for that state, which requires a different
55 * delivery design than main memory.
56 */
57 #define RDMA_SEND_INCREMENT 32768
58
59 /*
60 * Maximum size infiniband SEND message
61 */
62 #define RDMA_CONTROL_MAX_BUFFER (512 * 1024)
63 #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096
64
65 #define RDMA_CONTROL_VERSION_CURRENT 1
66 /*
67 * Capabilities for negotiation.
68 */
69 #define RDMA_CAPABILITY_PIN_ALL 0x01
70
71 /*
72 * Add the other flags above to this list of known capabilities
73 * as they are introduced.
74 */
75 static uint32_t known_capabilities = RDMA_CAPABILITY_PIN_ALL;
76
77 /*
78 * A work request ID is 64-bits and we split up these bits
79 * into 3 parts:
80 *
81 * bits 0-15 : type of control message, 2^16
82 * bits 16-29: ram block index, 2^14
83 * bits 30-63: ram block chunk number, 2^34
84 *
85 * The last two bit ranges are only used for RDMA writes,
86 * in order to track their completion and potentially
87 * also track unregistration status of the message.
88 */
89 #define RDMA_WRID_TYPE_SHIFT 0UL
90 #define RDMA_WRID_BLOCK_SHIFT 16UL
91 #define RDMA_WRID_CHUNK_SHIFT 30UL
92
93 #define RDMA_WRID_TYPE_MASK \
94 ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL)
95
96 #define RDMA_WRID_BLOCK_MASK \
97 (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL))
98
99 #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK)
100
101 /*
102 * RDMA migration protocol:
103 * 1. RDMA Writes (data messages, i.e. RAM)
104 * 2. IB Send/Recv (control channel messages)
105 */
106 enum {
107 RDMA_WRID_NONE = 0,
108 RDMA_WRID_RDMA_WRITE = 1,
109 RDMA_WRID_SEND_CONTROL = 2000,
110 RDMA_WRID_RECV_CONTROL = 4000,
111 };
112
113 /*
114 * Work request IDs for IB SEND messages only (not RDMA writes).
115 * This is used by the migration protocol to transmit
116 * control messages (such as device state and registration commands)
117 *
118 * We could use more WRs, but we have enough for now.
119 */
120 enum {
121 RDMA_WRID_READY = 0,
122 RDMA_WRID_DATA,
123 RDMA_WRID_CONTROL,
124 RDMA_WRID_MAX,
125 };
126
127 /*
128 * SEND/RECV IB Control Messages.
129 */
130 enum {
131 RDMA_CONTROL_NONE = 0,
132 RDMA_CONTROL_ERROR,
133 RDMA_CONTROL_READY, /* ready to receive */
134 RDMA_CONTROL_QEMU_FILE, /* QEMUFile-transmitted bytes */
135 RDMA_CONTROL_RAM_BLOCKS_REQUEST, /* RAMBlock synchronization */
136 RDMA_CONTROL_RAM_BLOCKS_RESULT, /* RAMBlock synchronization */
137 RDMA_CONTROL_COMPRESS, /* page contains repeat values */
138 RDMA_CONTROL_REGISTER_REQUEST, /* dynamic page registration */
139 RDMA_CONTROL_REGISTER_RESULT, /* key to use after registration */
140 RDMA_CONTROL_REGISTER_FINISHED, /* current iteration finished */
141 RDMA_CONTROL_UNREGISTER_REQUEST, /* dynamic UN-registration */
142 RDMA_CONTROL_UNREGISTER_FINISHED, /* unpinning finished */
143 };
144
145
146 /*
147 * Memory and MR structures used to represent an IB Send/Recv work request.
148 * This is *not* used for RDMA writes, only IB Send/Recv.
149 */
150 typedef struct {
151 uint8_t control[RDMA_CONTROL_MAX_BUFFER]; /* actual buffer to register */
152 struct ibv_mr *control_mr; /* registration metadata */
153 size_t control_len; /* length of the message */
154 uint8_t *control_curr; /* start of unconsumed bytes */
155 } RDMAWorkRequestData;
156
157 /*
158 * Negotiate RDMA capabilities during connection-setup time.
159 */
160 typedef struct {
161 uint32_t version;
162 uint32_t flags;
163 } RDMACapabilities;
164
caps_to_network(RDMACapabilities * cap)165 static void caps_to_network(RDMACapabilities *cap)
166 {
167 cap->version = htonl(cap->version);
168 cap->flags = htonl(cap->flags);
169 }
170
network_to_caps(RDMACapabilities * cap)171 static void network_to_caps(RDMACapabilities *cap)
172 {
173 cap->version = ntohl(cap->version);
174 cap->flags = ntohl(cap->flags);
175 }
176
177 /*
178 * Representation of a RAMBlock from an RDMA perspective.
179 * This is not transmitted, only local.
180 * This and subsequent structures cannot be linked lists
181 * because we're using a single IB message to transmit
182 * the information. It's small anyway, so a list is overkill.
183 */
184 typedef struct RDMALocalBlock {
185 char *block_name;
186 uint8_t *local_host_addr; /* local virtual address */
187 uint64_t remote_host_addr; /* remote virtual address */
188 uint64_t offset;
189 uint64_t length;
190 struct ibv_mr **pmr; /* MRs for chunk-level registration */
191 struct ibv_mr *mr; /* MR for non-chunk-level registration */
192 uint32_t *remote_keys; /* rkeys for chunk-level registration */
193 uint32_t remote_rkey; /* rkeys for non-chunk-level registration */
194 int index; /* which block are we */
195 unsigned int src_index; /* (Only used on dest) */
196 bool is_ram_block;
197 int nb_chunks;
198 unsigned long *transit_bitmap;
199 unsigned long *unregister_bitmap;
200 } RDMALocalBlock;
201
202 /*
203 * Also represents a RAMblock, but only on the dest.
204 * This gets transmitted by the dest during connection-time
205 * to the source VM and then is used to populate the
206 * corresponding RDMALocalBlock with
207 * the information needed to perform the actual RDMA.
208 */
209 typedef struct QEMU_PACKED RDMADestBlock {
210 uint64_t remote_host_addr;
211 uint64_t offset;
212 uint64_t length;
213 uint32_t remote_rkey;
214 uint32_t padding;
215 } RDMADestBlock;
216
control_desc(unsigned int rdma_control)217 static const char *control_desc(unsigned int rdma_control)
218 {
219 static const char *strs[] = {
220 [RDMA_CONTROL_NONE] = "NONE",
221 [RDMA_CONTROL_ERROR] = "ERROR",
222 [RDMA_CONTROL_READY] = "READY",
223 [RDMA_CONTROL_QEMU_FILE] = "QEMU FILE",
224 [RDMA_CONTROL_RAM_BLOCKS_REQUEST] = "RAM BLOCKS REQUEST",
225 [RDMA_CONTROL_RAM_BLOCKS_RESULT] = "RAM BLOCKS RESULT",
226 [RDMA_CONTROL_COMPRESS] = "COMPRESS",
227 [RDMA_CONTROL_REGISTER_REQUEST] = "REGISTER REQUEST",
228 [RDMA_CONTROL_REGISTER_RESULT] = "REGISTER RESULT",
229 [RDMA_CONTROL_REGISTER_FINISHED] = "REGISTER FINISHED",
230 [RDMA_CONTROL_UNREGISTER_REQUEST] = "UNREGISTER REQUEST",
231 [RDMA_CONTROL_UNREGISTER_FINISHED] = "UNREGISTER FINISHED",
232 };
233
234 if (rdma_control > RDMA_CONTROL_UNREGISTER_FINISHED) {
235 return "??BAD CONTROL VALUE??";
236 }
237
238 return strs[rdma_control];
239 }
240
241 #if !defined(htonll)
htonll(uint64_t v)242 static uint64_t htonll(uint64_t v)
243 {
244 union { uint32_t lv[2]; uint64_t llv; } u;
245 u.lv[0] = htonl(v >> 32);
246 u.lv[1] = htonl(v & 0xFFFFFFFFULL);
247 return u.llv;
248 }
249 #endif
250
251 #if !defined(ntohll)
ntohll(uint64_t v)252 static uint64_t ntohll(uint64_t v)
253 {
254 union { uint32_t lv[2]; uint64_t llv; } u;
255 u.llv = v;
256 return ((uint64_t)ntohl(u.lv[0]) << 32) | (uint64_t) ntohl(u.lv[1]);
257 }
258 #endif
259
dest_block_to_network(RDMADestBlock * db)260 static void dest_block_to_network(RDMADestBlock *db)
261 {
262 db->remote_host_addr = htonll(db->remote_host_addr);
263 db->offset = htonll(db->offset);
264 db->length = htonll(db->length);
265 db->remote_rkey = htonl(db->remote_rkey);
266 }
267
network_to_dest_block(RDMADestBlock * db)268 static void network_to_dest_block(RDMADestBlock *db)
269 {
270 db->remote_host_addr = ntohll(db->remote_host_addr);
271 db->offset = ntohll(db->offset);
272 db->length = ntohll(db->length);
273 db->remote_rkey = ntohl(db->remote_rkey);
274 }
275
276 /*
277 * Virtual address of the above structures used for transmitting
278 * the RAMBlock descriptions at connection-time.
279 * This structure is *not* transmitted.
280 */
281 typedef struct RDMALocalBlocks {
282 int nb_blocks;
283 bool init; /* main memory init complete */
284 RDMALocalBlock *block;
285 } RDMALocalBlocks;
286
287 /*
288 * Main data structure for RDMA state.
289 * While there is only one copy of this structure being allocated right now,
290 * this is the place where one would start if you wanted to consider
291 * having more than one RDMA connection open at the same time.
292 */
293 typedef struct RDMAContext {
294 char *host;
295 int port;
296
297 RDMAWorkRequestData wr_data[RDMA_WRID_MAX];
298
299 /*
300 * This is used by *_exchange_send() to figure out whether or not
301 * the initial "READY" message has already been received or not.
302 * This is because other functions may potentially poll() and detect
303 * the READY message before send() does, in which case we need to
304 * know if it completed.
305 */
306 int control_ready_expected;
307
308 /* number of outstanding writes */
309 int nb_sent;
310
311 /* store info about current buffer so that we can
312 merge it with future sends */
313 uint64_t current_addr;
314 uint64_t current_length;
315 /* index of ram block the current buffer belongs to */
316 int current_index;
317 /* index of the chunk in the current ram block */
318 int current_chunk;
319
320 bool pin_all;
321
322 /*
323 * infiniband-specific variables for opening the device
324 * and maintaining connection state and so forth.
325 *
326 * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in
327 * cm_id->verbs, cm_id->channel, and cm_id->qp.
328 */
329 struct rdma_cm_id *cm_id; /* connection manager ID */
330 struct rdma_cm_id *listen_id;
331 bool connected;
332
333 struct ibv_context *verbs;
334 struct rdma_event_channel *channel;
335 struct ibv_qp *qp; /* queue pair */
336 struct ibv_comp_channel *recv_comp_channel; /* recv completion channel */
337 struct ibv_comp_channel *send_comp_channel; /* send completion channel */
338 struct ibv_pd *pd; /* protection domain */
339 struct ibv_cq *recv_cq; /* recvieve completion queue */
340 struct ibv_cq *send_cq; /* send completion queue */
341
342 /*
343 * If a previous write failed (perhaps because of a failed
344 * memory registration, then do not attempt any future work
345 * and remember the error state.
346 */
347 bool errored;
348 bool error_reported;
349 bool received_error;
350
351 /*
352 * Description of ram blocks used throughout the code.
353 */
354 RDMALocalBlocks local_ram_blocks;
355 RDMADestBlock *dest_blocks;
356
357 /* Index of the next RAMBlock received during block registration */
358 unsigned int next_src_index;
359
360 /*
361 * Migration on *destination* started.
362 * Then use coroutine yield function.
363 * Source runs in a thread, so we don't care.
364 */
365 int migration_started_on_destination;
366
367 int total_registrations;
368 int total_writes;
369
370 int unregister_current, unregister_next;
371 uint64_t unregistrations[RDMA_SIGNALED_SEND_MAX];
372
373 GHashTable *blockmap;
374
375 /* the RDMAContext for return path */
376 struct RDMAContext *return_path;
377 bool is_return_path;
378 } RDMAContext;
379
380 #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma"
381 OBJECT_DECLARE_SIMPLE_TYPE(QIOChannelRDMA, QIO_CHANNEL_RDMA)
382
383
384
385 struct QIOChannelRDMA {
386 QIOChannel parent;
387 RDMAContext *rdmain;
388 RDMAContext *rdmaout;
389 QEMUFile *file;
390 bool blocking; /* XXX we don't actually honour this yet */
391 };
392
393 /*
394 * Main structure for IB Send/Recv control messages.
395 * This gets prepended at the beginning of every Send/Recv.
396 */
397 typedef struct QEMU_PACKED {
398 uint32_t len; /* Total length of data portion */
399 uint32_t type; /* which control command to perform */
400 uint32_t repeat; /* number of commands in data portion of same type */
401 uint32_t padding;
402 } RDMAControlHeader;
403
control_to_network(RDMAControlHeader * control)404 static void control_to_network(RDMAControlHeader *control)
405 {
406 control->type = htonl(control->type);
407 control->len = htonl(control->len);
408 control->repeat = htonl(control->repeat);
409 }
410
network_to_control(RDMAControlHeader * control)411 static void network_to_control(RDMAControlHeader *control)
412 {
413 control->type = ntohl(control->type);
414 control->len = ntohl(control->len);
415 control->repeat = ntohl(control->repeat);
416 }
417
418 /*
419 * Register a single Chunk.
420 * Information sent by the source VM to inform the dest
421 * to register an single chunk of memory before we can perform
422 * the actual RDMA operation.
423 */
424 typedef struct QEMU_PACKED {
425 union QEMU_PACKED {
426 uint64_t current_addr; /* offset into the ram_addr_t space */
427 uint64_t chunk; /* chunk to lookup if unregistering */
428 } key;
429 uint32_t current_index; /* which ramblock the chunk belongs to */
430 uint32_t padding;
431 uint64_t chunks; /* how many sequential chunks to register */
432 } RDMARegister;
433
rdma_errored(RDMAContext * rdma)434 static bool rdma_errored(RDMAContext *rdma)
435 {
436 if (rdma->errored && !rdma->error_reported) {
437 error_report("RDMA is in an error state waiting migration"
438 " to abort!");
439 rdma->error_reported = true;
440 }
441 return rdma->errored;
442 }
443
register_to_network(RDMAContext * rdma,RDMARegister * reg)444 static void register_to_network(RDMAContext *rdma, RDMARegister *reg)
445 {
446 RDMALocalBlock *local_block;
447 local_block = &rdma->local_ram_blocks.block[reg->current_index];
448
449 if (local_block->is_ram_block) {
450 /*
451 * current_addr as passed in is an address in the local ram_addr_t
452 * space, we need to translate this for the destination
453 */
454 reg->key.current_addr -= local_block->offset;
455 reg->key.current_addr += rdma->dest_blocks[reg->current_index].offset;
456 }
457 reg->key.current_addr = htonll(reg->key.current_addr);
458 reg->current_index = htonl(reg->current_index);
459 reg->chunks = htonll(reg->chunks);
460 }
461
network_to_register(RDMARegister * reg)462 static void network_to_register(RDMARegister *reg)
463 {
464 reg->key.current_addr = ntohll(reg->key.current_addr);
465 reg->current_index = ntohl(reg->current_index);
466 reg->chunks = ntohll(reg->chunks);
467 }
468
469 typedef struct QEMU_PACKED {
470 uint32_t value; /* if zero, we will madvise() */
471 uint32_t block_idx; /* which ram block index */
472 uint64_t offset; /* Address in remote ram_addr_t space */
473 uint64_t length; /* length of the chunk */
474 } RDMACompress;
475
compress_to_network(RDMAContext * rdma,RDMACompress * comp)476 static void compress_to_network(RDMAContext *rdma, RDMACompress *comp)
477 {
478 comp->value = htonl(comp->value);
479 /*
480 * comp->offset as passed in is an address in the local ram_addr_t
481 * space, we need to translate this for the destination
482 */
483 comp->offset -= rdma->local_ram_blocks.block[comp->block_idx].offset;
484 comp->offset += rdma->dest_blocks[comp->block_idx].offset;
485 comp->block_idx = htonl(comp->block_idx);
486 comp->offset = htonll(comp->offset);
487 comp->length = htonll(comp->length);
488 }
489
network_to_compress(RDMACompress * comp)490 static void network_to_compress(RDMACompress *comp)
491 {
492 comp->value = ntohl(comp->value);
493 comp->block_idx = ntohl(comp->block_idx);
494 comp->offset = ntohll(comp->offset);
495 comp->length = ntohll(comp->length);
496 }
497
498 /*
499 * The result of the dest's memory registration produces an "rkey"
500 * which the source VM must reference in order to perform
501 * the RDMA operation.
502 */
503 typedef struct QEMU_PACKED {
504 uint32_t rkey;
505 uint32_t padding;
506 uint64_t host_addr;
507 } RDMARegisterResult;
508
result_to_network(RDMARegisterResult * result)509 static void result_to_network(RDMARegisterResult *result)
510 {
511 result->rkey = htonl(result->rkey);
512 result->host_addr = htonll(result->host_addr);
513 };
514
network_to_result(RDMARegisterResult * result)515 static void network_to_result(RDMARegisterResult *result)
516 {
517 result->rkey = ntohl(result->rkey);
518 result->host_addr = ntohll(result->host_addr);
519 };
520
521 static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head,
522 uint8_t *data, RDMAControlHeader *resp,
523 int *resp_idx,
524 int (*callback)(RDMAContext *rdma,
525 Error **errp),
526 Error **errp);
527
ram_chunk_index(const uint8_t * start,const uint8_t * host)528 static inline uint64_t ram_chunk_index(const uint8_t *start,
529 const uint8_t *host)
530 {
531 return ((uintptr_t) host - (uintptr_t) start) >> RDMA_REG_CHUNK_SHIFT;
532 }
533
ram_chunk_start(const RDMALocalBlock * rdma_ram_block,uint64_t i)534 static inline uint8_t *ram_chunk_start(const RDMALocalBlock *rdma_ram_block,
535 uint64_t i)
536 {
537 return (uint8_t *)(uintptr_t)(rdma_ram_block->local_host_addr +
538 (i << RDMA_REG_CHUNK_SHIFT));
539 }
540
ram_chunk_end(const RDMALocalBlock * rdma_ram_block,uint64_t i)541 static inline uint8_t *ram_chunk_end(const RDMALocalBlock *rdma_ram_block,
542 uint64_t i)
543 {
544 uint8_t *result = ram_chunk_start(rdma_ram_block, i) +
545 (1UL << RDMA_REG_CHUNK_SHIFT);
546
547 if (result > (rdma_ram_block->local_host_addr + rdma_ram_block->length)) {
548 result = rdma_ram_block->local_host_addr + rdma_ram_block->length;
549 }
550
551 return result;
552 }
553
rdma_add_block(RDMAContext * rdma,const char * block_name,void * host_addr,ram_addr_t block_offset,uint64_t length)554 static void rdma_add_block(RDMAContext *rdma, const char *block_name,
555 void *host_addr,
556 ram_addr_t block_offset, uint64_t length)
557 {
558 RDMALocalBlocks *local = &rdma->local_ram_blocks;
559 RDMALocalBlock *block;
560 RDMALocalBlock *old = local->block;
561
562 local->block = g_new0(RDMALocalBlock, local->nb_blocks + 1);
563
564 if (local->nb_blocks) {
565 if (rdma->blockmap) {
566 for (int x = 0; x < local->nb_blocks; x++) {
567 g_hash_table_remove(rdma->blockmap,
568 (void *)(uintptr_t)old[x].offset);
569 g_hash_table_insert(rdma->blockmap,
570 (void *)(uintptr_t)old[x].offset,
571 &local->block[x]);
572 }
573 }
574 memcpy(local->block, old, sizeof(RDMALocalBlock) * local->nb_blocks);
575 g_free(old);
576 }
577
578 block = &local->block[local->nb_blocks];
579
580 block->block_name = g_strdup(block_name);
581 block->local_host_addr = host_addr;
582 block->offset = block_offset;
583 block->length = length;
584 block->index = local->nb_blocks;
585 block->src_index = ~0U; /* Filled in by the receipt of the block list */
586 block->nb_chunks = ram_chunk_index(host_addr, host_addr + length) + 1UL;
587 block->transit_bitmap = bitmap_new(block->nb_chunks);
588 bitmap_clear(block->transit_bitmap, 0, block->nb_chunks);
589 block->unregister_bitmap = bitmap_new(block->nb_chunks);
590 bitmap_clear(block->unregister_bitmap, 0, block->nb_chunks);
591 block->remote_keys = g_new0(uint32_t, block->nb_chunks);
592
593 block->is_ram_block = local->init ? false : true;
594
595 if (rdma->blockmap) {
596 g_hash_table_insert(rdma->blockmap, (void *)(uintptr_t)block_offset, block);
597 }
598
599 trace_rdma_add_block(block_name, local->nb_blocks,
600 (uintptr_t) block->local_host_addr,
601 block->offset, block->length,
602 (uintptr_t) (block->local_host_addr + block->length),
603 BITS_TO_LONGS(block->nb_chunks) *
604 sizeof(unsigned long) * 8,
605 block->nb_chunks);
606
607 local->nb_blocks++;
608 }
609
610 /*
611 * Memory regions need to be registered with the device and queue pairs setup
612 * in advanced before the migration starts. This tells us where the RAM blocks
613 * are so that we can register them individually.
614 */
qemu_rdma_init_one_block(RAMBlock * rb,void * opaque)615 static int qemu_rdma_init_one_block(RAMBlock *rb, void *opaque)
616 {
617 const char *block_name = qemu_ram_get_idstr(rb);
618 void *host_addr = qemu_ram_get_host_addr(rb);
619 ram_addr_t block_offset = qemu_ram_get_offset(rb);
620 ram_addr_t length = qemu_ram_get_used_length(rb);
621 rdma_add_block(opaque, block_name, host_addr, block_offset, length);
622 return 0;
623 }
624
625 /*
626 * Identify the RAMBlocks and their quantity. They will be references to
627 * identify chunk boundaries inside each RAMBlock and also be referenced
628 * during dynamic page registration.
629 */
qemu_rdma_init_ram_blocks(RDMAContext * rdma)630 static void qemu_rdma_init_ram_blocks(RDMAContext *rdma)
631 {
632 RDMALocalBlocks *local = &rdma->local_ram_blocks;
633 int ret;
634
635 assert(rdma->blockmap == NULL);
636 memset(local, 0, sizeof *local);
637 ret = foreach_not_ignored_block(qemu_rdma_init_one_block, rdma);
638 assert(!ret);
639 trace_qemu_rdma_init_ram_blocks(local->nb_blocks);
640 rdma->dest_blocks = g_new0(RDMADestBlock,
641 rdma->local_ram_blocks.nb_blocks);
642 local->init = true;
643 }
644
645 /*
646 * Note: If used outside of cleanup, the caller must ensure that the destination
647 * block structures are also updated
648 */
rdma_delete_block(RDMAContext * rdma,RDMALocalBlock * block)649 static void rdma_delete_block(RDMAContext *rdma, RDMALocalBlock *block)
650 {
651 RDMALocalBlocks *local = &rdma->local_ram_blocks;
652 RDMALocalBlock *old = local->block;
653
654 if (rdma->blockmap) {
655 g_hash_table_remove(rdma->blockmap, (void *)(uintptr_t)block->offset);
656 }
657 if (block->pmr) {
658 for (int j = 0; j < block->nb_chunks; j++) {
659 if (!block->pmr[j]) {
660 continue;
661 }
662 ibv_dereg_mr(block->pmr[j]);
663 rdma->total_registrations--;
664 }
665 g_free(block->pmr);
666 block->pmr = NULL;
667 }
668
669 if (block->mr) {
670 ibv_dereg_mr(block->mr);
671 rdma->total_registrations--;
672 block->mr = NULL;
673 }
674
675 g_free(block->transit_bitmap);
676 block->transit_bitmap = NULL;
677
678 g_free(block->unregister_bitmap);
679 block->unregister_bitmap = NULL;
680
681 g_free(block->remote_keys);
682 block->remote_keys = NULL;
683
684 g_free(block->block_name);
685 block->block_name = NULL;
686
687 if (rdma->blockmap) {
688 for (int x = 0; x < local->nb_blocks; x++) {
689 g_hash_table_remove(rdma->blockmap,
690 (void *)(uintptr_t)old[x].offset);
691 }
692 }
693
694 if (local->nb_blocks > 1) {
695
696 local->block = g_new0(RDMALocalBlock, local->nb_blocks - 1);
697
698 if (block->index) {
699 memcpy(local->block, old, sizeof(RDMALocalBlock) * block->index);
700 }
701
702 if (block->index < (local->nb_blocks - 1)) {
703 memcpy(local->block + block->index, old + (block->index + 1),
704 sizeof(RDMALocalBlock) *
705 (local->nb_blocks - (block->index + 1)));
706 for (int x = block->index; x < local->nb_blocks - 1; x++) {
707 local->block[x].index--;
708 }
709 }
710 } else {
711 assert(block == local->block);
712 local->block = NULL;
713 }
714
715 trace_rdma_delete_block(block, (uintptr_t)block->local_host_addr,
716 block->offset, block->length,
717 (uintptr_t)(block->local_host_addr + block->length),
718 BITS_TO_LONGS(block->nb_chunks) *
719 sizeof(unsigned long) * 8, block->nb_chunks);
720
721 g_free(old);
722
723 local->nb_blocks--;
724
725 if (local->nb_blocks && rdma->blockmap) {
726 for (int x = 0; x < local->nb_blocks; x++) {
727 g_hash_table_insert(rdma->blockmap,
728 (void *)(uintptr_t)local->block[x].offset,
729 &local->block[x]);
730 }
731 }
732 }
733
734 /*
735 * Trace RDMA device open, with device details.
736 */
qemu_rdma_dump_id(const char * who,struct ibv_context * verbs)737 static void qemu_rdma_dump_id(const char *who, struct ibv_context *verbs)
738 {
739 struct ibv_port_attr port;
740
741 if (ibv_query_port(verbs, 1, &port)) {
742 trace_qemu_rdma_dump_id_failed(who);
743 return;
744 }
745
746 trace_qemu_rdma_dump_id(who,
747 verbs->device->name,
748 verbs->device->dev_name,
749 verbs->device->dev_path,
750 verbs->device->ibdev_path,
751 port.link_layer,
752 port.link_layer == IBV_LINK_LAYER_INFINIBAND ? "Infiniband"
753 : port.link_layer == IBV_LINK_LAYER_ETHERNET ? "Ethernet"
754 : "Unknown");
755 }
756
757 /*
758 * Trace RDMA gid addressing information.
759 * Useful for understanding the RDMA device hierarchy in the kernel.
760 */
qemu_rdma_dump_gid(const char * who,struct rdma_cm_id * id)761 static void qemu_rdma_dump_gid(const char *who, struct rdma_cm_id *id)
762 {
763 char sgid[33];
764 char dgid[33];
765 inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.sgid, sgid, sizeof sgid);
766 inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.dgid, dgid, sizeof dgid);
767 trace_qemu_rdma_dump_gid(who, sgid, dgid);
768 }
769
770 /*
771 * As of now, IPv6 over RoCE / iWARP is not supported by linux.
772 * We will try the next addrinfo struct, and fail if there are
773 * no other valid addresses to bind against.
774 *
775 * If user is listening on '[::]', then we will not have a opened a device
776 * yet and have no way of verifying if the device is RoCE or not.
777 *
778 * In this case, the source VM will throw an error for ALL types of
779 * connections (both IPv4 and IPv6) if the destination machine does not have
780 * a regular infiniband network available for use.
781 *
782 * The only way to guarantee that an error is thrown for broken kernels is
783 * for the management software to choose a *specific* interface at bind time
784 * and validate what time of hardware it is.
785 *
786 * Unfortunately, this puts the user in a fix:
787 *
788 * If the source VM connects with an IPv4 address without knowing that the
789 * destination has bound to '[::]' the migration will unconditionally fail
790 * unless the management software is explicitly listening on the IPv4
791 * address while using a RoCE-based device.
792 *
793 * If the source VM connects with an IPv6 address, then we're OK because we can
794 * throw an error on the source (and similarly on the destination).
795 *
796 * But in mixed environments, this will be broken for a while until it is fixed
797 * inside linux.
798 *
799 * We do provide a *tiny* bit of help in this function: We can list all of the
800 * devices in the system and check to see if all the devices are RoCE or
801 * Infiniband.
802 *
803 * If we detect that we have a *pure* RoCE environment, then we can safely
804 * thrown an error even if the management software has specified '[::]' as the
805 * bind address.
806 *
807 * However, if there is are multiple hetergeneous devices, then we cannot make
808 * this assumption and the user just has to be sure they know what they are
809 * doing.
810 *
811 * Patches are being reviewed on linux-rdma.
812 */
qemu_rdma_broken_ipv6_kernel(struct ibv_context * verbs,Error ** errp)813 static int qemu_rdma_broken_ipv6_kernel(struct ibv_context *verbs, Error **errp)
814 {
815 /* This bug only exists in linux, to our knowledge. */
816 #ifdef CONFIG_LINUX
817 struct ibv_port_attr port_attr;
818
819 /*
820 * Verbs are only NULL if management has bound to '[::]'.
821 *
822 * Let's iterate through all the devices and see if there any pure IB
823 * devices (non-ethernet).
824 *
825 * If not, then we can safely proceed with the migration.
826 * Otherwise, there are no guarantees until the bug is fixed in linux.
827 */
828 if (!verbs) {
829 int num_devices;
830 struct ibv_device **dev_list = ibv_get_device_list(&num_devices);
831 bool roce_found = false;
832 bool ib_found = false;
833
834 for (int x = 0; x < num_devices; x++) {
835 verbs = ibv_open_device(dev_list[x]);
836 /*
837 * ibv_open_device() is not documented to set errno. If
838 * it does, it's somebody else's doc bug. If it doesn't,
839 * the use of errno below is wrong.
840 * TODO Find out whether ibv_open_device() sets errno.
841 */
842 if (!verbs) {
843 if (errno == EPERM) {
844 continue;
845 } else {
846 error_setg_errno(errp, errno,
847 "could not open RDMA device context");
848 return -1;
849 }
850 }
851
852 if (ibv_query_port(verbs, 1, &port_attr)) {
853 ibv_close_device(verbs);
854 error_setg(errp,
855 "RDMA ERROR: Could not query initial IB port");
856 return -1;
857 }
858
859 if (port_attr.link_layer == IBV_LINK_LAYER_INFINIBAND) {
860 ib_found = true;
861 } else if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) {
862 roce_found = true;
863 }
864
865 ibv_close_device(verbs);
866
867 }
868
869 if (roce_found) {
870 if (ib_found) {
871 warn_report("migrations may fail:"
872 " IPv6 over RoCE / iWARP in linux"
873 " is broken. But since you appear to have a"
874 " mixed RoCE / IB environment, be sure to only"
875 " migrate over the IB fabric until the kernel "
876 " fixes the bug.");
877 } else {
878 error_setg(errp, "RDMA ERROR: "
879 "You only have RoCE / iWARP devices in your systems"
880 " and your management software has specified '[::]'"
881 ", but IPv6 over RoCE / iWARP is not supported in Linux.");
882 return -1;
883 }
884 }
885
886 return 0;
887 }
888
889 /*
890 * If we have a verbs context, that means that some other than '[::]' was
891 * used by the management software for binding. In which case we can
892 * actually warn the user about a potentially broken kernel.
893 */
894
895 /* IB ports start with 1, not 0 */
896 if (ibv_query_port(verbs, 1, &port_attr)) {
897 error_setg(errp, "RDMA ERROR: Could not query initial IB port");
898 return -1;
899 }
900
901 if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) {
902 error_setg(errp, "RDMA ERROR: "
903 "Linux kernel's RoCE / iWARP does not support IPv6 "
904 "(but patches on linux-rdma in progress)");
905 return -1;
906 }
907
908 #endif
909
910 return 0;
911 }
912
913 /*
914 * Figure out which RDMA device corresponds to the requested IP hostname
915 * Also create the initial connection manager identifiers for opening
916 * the connection.
917 */
qemu_rdma_resolve_host(RDMAContext * rdma,Error ** errp)918 static int qemu_rdma_resolve_host(RDMAContext *rdma, Error **errp)
919 {
920 Error *err = NULL;
921 int ret;
922 struct rdma_addrinfo *res;
923 char port_str[16];
924 struct rdma_cm_event *cm_event;
925 char ip[40] = "unknown";
926
927 if (rdma->host == NULL || !strcmp(rdma->host, "")) {
928 error_setg(errp, "RDMA ERROR: RDMA hostname has not been set");
929 return -1;
930 }
931
932 /* create CM channel */
933 rdma->channel = rdma_create_event_channel();
934 if (!rdma->channel) {
935 error_setg(errp, "RDMA ERROR: could not create CM channel");
936 return -1;
937 }
938
939 /* create CM id */
940 ret = rdma_create_id(rdma->channel, &rdma->cm_id, NULL, RDMA_PS_TCP);
941 if (ret < 0) {
942 error_setg(errp, "RDMA ERROR: could not create channel id");
943 goto err_resolve_create_id;
944 }
945
946 snprintf(port_str, 16, "%d", rdma->port);
947 port_str[15] = '\0';
948
949 ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res);
950 if (ret) {
951 error_setg(errp, "RDMA ERROR: could not rdma_getaddrinfo address %s",
952 rdma->host);
953 goto err_resolve_get_addr;
954 }
955
956 /* Try all addresses, saving the first error in @err */
957 for (struct rdma_addrinfo *e = res; e != NULL; e = e->ai_next) {
958 Error **local_errp = err ? NULL : &err;
959
960 inet_ntop(e->ai_family,
961 &((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip);
962 trace_qemu_rdma_resolve_host_trying(rdma->host, ip);
963
964 ret = rdma_resolve_addr(rdma->cm_id, NULL, e->ai_dst_addr,
965 RDMA_RESOLVE_TIMEOUT_MS);
966 if (ret >= 0) {
967 if (e->ai_family == AF_INET6) {
968 ret = qemu_rdma_broken_ipv6_kernel(rdma->cm_id->verbs,
969 local_errp);
970 if (ret < 0) {
971 continue;
972 }
973 }
974 error_free(err);
975 goto route;
976 }
977 }
978
979 rdma_freeaddrinfo(res);
980 if (err) {
981 error_propagate(errp, err);
982 } else {
983 error_setg(errp, "RDMA ERROR: could not resolve address %s",
984 rdma->host);
985 }
986 goto err_resolve_get_addr;
987
988 route:
989 rdma_freeaddrinfo(res);
990 qemu_rdma_dump_gid("source_resolve_addr", rdma->cm_id);
991
992 ret = rdma_get_cm_event(rdma->channel, &cm_event);
993 if (ret < 0) {
994 error_setg(errp, "RDMA ERROR: could not perform event_addr_resolved");
995 goto err_resolve_get_addr;
996 }
997
998 if (cm_event->event != RDMA_CM_EVENT_ADDR_RESOLVED) {
999 error_setg(errp,
1000 "RDMA ERROR: result not equal to event_addr_resolved %s",
1001 rdma_event_str(cm_event->event));
1002 rdma_ack_cm_event(cm_event);
1003 goto err_resolve_get_addr;
1004 }
1005 rdma_ack_cm_event(cm_event);
1006
1007 /* resolve route */
1008 ret = rdma_resolve_route(rdma->cm_id, RDMA_RESOLVE_TIMEOUT_MS);
1009 if (ret < 0) {
1010 error_setg(errp, "RDMA ERROR: could not resolve rdma route");
1011 goto err_resolve_get_addr;
1012 }
1013
1014 ret = rdma_get_cm_event(rdma->channel, &cm_event);
1015 if (ret < 0) {
1016 error_setg(errp, "RDMA ERROR: could not perform event_route_resolved");
1017 goto err_resolve_get_addr;
1018 }
1019 if (cm_event->event != RDMA_CM_EVENT_ROUTE_RESOLVED) {
1020 error_setg(errp, "RDMA ERROR: "
1021 "result not equal to event_route_resolved: %s",
1022 rdma_event_str(cm_event->event));
1023 rdma_ack_cm_event(cm_event);
1024 goto err_resolve_get_addr;
1025 }
1026 rdma_ack_cm_event(cm_event);
1027 rdma->verbs = rdma->cm_id->verbs;
1028 qemu_rdma_dump_id("source_resolve_host", rdma->cm_id->verbs);
1029 qemu_rdma_dump_gid("source_resolve_host", rdma->cm_id);
1030 return 0;
1031
1032 err_resolve_get_addr:
1033 rdma_destroy_id(rdma->cm_id);
1034 rdma->cm_id = NULL;
1035 err_resolve_create_id:
1036 rdma_destroy_event_channel(rdma->channel);
1037 rdma->channel = NULL;
1038 return -1;
1039 }
1040
1041 /*
1042 * Create protection domain and completion queues
1043 */
qemu_rdma_alloc_pd_cq(RDMAContext * rdma,Error ** errp)1044 static int qemu_rdma_alloc_pd_cq(RDMAContext *rdma, Error **errp)
1045 {
1046 /* allocate pd */
1047 rdma->pd = ibv_alloc_pd(rdma->verbs);
1048 if (!rdma->pd) {
1049 error_setg(errp, "failed to allocate protection domain");
1050 return -1;
1051 }
1052
1053 /* create receive completion channel */
1054 rdma->recv_comp_channel = ibv_create_comp_channel(rdma->verbs);
1055 if (!rdma->recv_comp_channel) {
1056 error_setg(errp, "failed to allocate receive completion channel");
1057 goto err_alloc_pd_cq;
1058 }
1059
1060 /*
1061 * Completion queue can be filled by read work requests.
1062 */
1063 rdma->recv_cq = ibv_create_cq(rdma->verbs, (RDMA_SIGNALED_SEND_MAX * 3),
1064 NULL, rdma->recv_comp_channel, 0);
1065 if (!rdma->recv_cq) {
1066 error_setg(errp, "failed to allocate receive completion queue");
1067 goto err_alloc_pd_cq;
1068 }
1069
1070 /* create send completion channel */
1071 rdma->send_comp_channel = ibv_create_comp_channel(rdma->verbs);
1072 if (!rdma->send_comp_channel) {
1073 error_setg(errp, "failed to allocate send completion channel");
1074 goto err_alloc_pd_cq;
1075 }
1076
1077 rdma->send_cq = ibv_create_cq(rdma->verbs, (RDMA_SIGNALED_SEND_MAX * 3),
1078 NULL, rdma->send_comp_channel, 0);
1079 if (!rdma->send_cq) {
1080 error_setg(errp, "failed to allocate send completion queue");
1081 goto err_alloc_pd_cq;
1082 }
1083
1084 return 0;
1085
1086 err_alloc_pd_cq:
1087 if (rdma->pd) {
1088 ibv_dealloc_pd(rdma->pd);
1089 }
1090 if (rdma->recv_comp_channel) {
1091 ibv_destroy_comp_channel(rdma->recv_comp_channel);
1092 }
1093 if (rdma->send_comp_channel) {
1094 ibv_destroy_comp_channel(rdma->send_comp_channel);
1095 }
1096 if (rdma->recv_cq) {
1097 ibv_destroy_cq(rdma->recv_cq);
1098 rdma->recv_cq = NULL;
1099 }
1100 rdma->pd = NULL;
1101 rdma->recv_comp_channel = NULL;
1102 rdma->send_comp_channel = NULL;
1103 return -1;
1104
1105 }
1106
1107 /*
1108 * Create queue pairs.
1109 */
qemu_rdma_alloc_qp(RDMAContext * rdma)1110 static int qemu_rdma_alloc_qp(RDMAContext *rdma)
1111 {
1112 struct ibv_qp_init_attr attr = { 0 };
1113
1114 attr.cap.max_send_wr = RDMA_SIGNALED_SEND_MAX;
1115 attr.cap.max_recv_wr = 3;
1116 attr.cap.max_send_sge = 1;
1117 attr.cap.max_recv_sge = 1;
1118 attr.send_cq = rdma->send_cq;
1119 attr.recv_cq = rdma->recv_cq;
1120 attr.qp_type = IBV_QPT_RC;
1121
1122 if (rdma_create_qp(rdma->cm_id, rdma->pd, &attr) < 0) {
1123 return -1;
1124 }
1125
1126 rdma->qp = rdma->cm_id->qp;
1127 return 0;
1128 }
1129
1130 /* Check whether On-Demand Paging is supported by RDAM device */
rdma_support_odp(struct ibv_context * dev)1131 static bool rdma_support_odp(struct ibv_context *dev)
1132 {
1133 struct ibv_device_attr_ex attr = {0};
1134
1135 if (ibv_query_device_ex(dev, NULL, &attr)) {
1136 return false;
1137 }
1138
1139 if (attr.odp_caps.general_caps & IBV_ODP_SUPPORT) {
1140 return true;
1141 }
1142
1143 return false;
1144 }
1145
1146 /*
1147 * ibv_advise_mr to avoid RNR NAK error as far as possible.
1148 * The responder mr registering with ODP will sent RNR NAK back to
1149 * the requester in the face of the page fault.
1150 */
qemu_rdma_advise_prefetch_mr(struct ibv_pd * pd,uint64_t addr,uint32_t len,uint32_t lkey,const char * name,bool wr)1151 static void qemu_rdma_advise_prefetch_mr(struct ibv_pd *pd, uint64_t addr,
1152 uint32_t len, uint32_t lkey,
1153 const char *name, bool wr)
1154 {
1155 #ifdef HAVE_IBV_ADVISE_MR
1156 int ret;
1157 int advice = wr ? IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE :
1158 IBV_ADVISE_MR_ADVICE_PREFETCH;
1159 struct ibv_sge sg_list = {.lkey = lkey, .addr = addr, .length = len};
1160
1161 ret = ibv_advise_mr(pd, advice,
1162 IBV_ADVISE_MR_FLAG_FLUSH, &sg_list, 1);
1163 /* ignore the error */
1164 trace_qemu_rdma_advise_mr(name, len, addr, strerror(ret));
1165 #endif
1166 }
1167
qemu_rdma_reg_whole_ram_blocks(RDMAContext * rdma,Error ** errp)1168 static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma, Error **errp)
1169 {
1170 int i;
1171 RDMALocalBlocks *local = &rdma->local_ram_blocks;
1172
1173 for (i = 0; i < local->nb_blocks; i++) {
1174 int access = IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE;
1175
1176 local->block[i].mr =
1177 ibv_reg_mr(rdma->pd,
1178 local->block[i].local_host_addr,
1179 local->block[i].length, access
1180 );
1181 /*
1182 * ibv_reg_mr() is not documented to set errno. If it does,
1183 * it's somebody else's doc bug. If it doesn't, the use of
1184 * errno below is wrong.
1185 * TODO Find out whether ibv_reg_mr() sets errno.
1186 */
1187 if (!local->block[i].mr &&
1188 errno == ENOTSUP && rdma_support_odp(rdma->verbs)) {
1189 access |= IBV_ACCESS_ON_DEMAND;
1190 /* register ODP mr */
1191 local->block[i].mr =
1192 ibv_reg_mr(rdma->pd,
1193 local->block[i].local_host_addr,
1194 local->block[i].length, access);
1195 trace_qemu_rdma_register_odp_mr(local->block[i].block_name);
1196
1197 if (local->block[i].mr) {
1198 qemu_rdma_advise_prefetch_mr(rdma->pd,
1199 (uintptr_t)local->block[i].local_host_addr,
1200 local->block[i].length,
1201 local->block[i].mr->lkey,
1202 local->block[i].block_name,
1203 true);
1204 }
1205 }
1206
1207 if (!local->block[i].mr) {
1208 error_setg_errno(errp, errno,
1209 "Failed to register local dest ram block!");
1210 goto err;
1211 }
1212 rdma->total_registrations++;
1213 }
1214
1215 return 0;
1216
1217 err:
1218 for (i--; i >= 0; i--) {
1219 ibv_dereg_mr(local->block[i].mr);
1220 local->block[i].mr = NULL;
1221 rdma->total_registrations--;
1222 }
1223
1224 return -1;
1225
1226 }
1227
1228 /*
1229 * Find the ram block that corresponds to the page requested to be
1230 * transmitted by QEMU.
1231 *
1232 * Once the block is found, also identify which 'chunk' within that
1233 * block that the page belongs to.
1234 */
qemu_rdma_search_ram_block(RDMAContext * rdma,uintptr_t block_offset,uint64_t offset,uint64_t length,uint64_t * block_index,uint64_t * chunk_index)1235 static void qemu_rdma_search_ram_block(RDMAContext *rdma,
1236 uintptr_t block_offset,
1237 uint64_t offset,
1238 uint64_t length,
1239 uint64_t *block_index,
1240 uint64_t *chunk_index)
1241 {
1242 uint64_t current_addr = block_offset + offset;
1243 RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap,
1244 (void *) block_offset);
1245 assert(block);
1246 assert(current_addr >= block->offset);
1247 assert((current_addr + length) <= (block->offset + block->length));
1248
1249 *block_index = block->index;
1250 *chunk_index = ram_chunk_index(block->local_host_addr,
1251 block->local_host_addr + (current_addr - block->offset));
1252 }
1253
1254 /*
1255 * Register a chunk with IB. If the chunk was already registered
1256 * previously, then skip.
1257 *
1258 * Also return the keys associated with the registration needed
1259 * to perform the actual RDMA operation.
1260 */
qemu_rdma_register_and_get_keys(RDMAContext * rdma,RDMALocalBlock * block,uintptr_t host_addr,uint32_t * lkey,uint32_t * rkey,int chunk,uint8_t * chunk_start,uint8_t * chunk_end)1261 static int qemu_rdma_register_and_get_keys(RDMAContext *rdma,
1262 RDMALocalBlock *block, uintptr_t host_addr,
1263 uint32_t *lkey, uint32_t *rkey, int chunk,
1264 uint8_t *chunk_start, uint8_t *chunk_end)
1265 {
1266 if (block->mr) {
1267 if (lkey) {
1268 *lkey = block->mr->lkey;
1269 }
1270 if (rkey) {
1271 *rkey = block->mr->rkey;
1272 }
1273 return 0;
1274 }
1275
1276 /* allocate memory to store chunk MRs */
1277 if (!block->pmr) {
1278 block->pmr = g_new0(struct ibv_mr *, block->nb_chunks);
1279 }
1280
1281 /*
1282 * If 'rkey', then we're the destination, so grant access to the source.
1283 *
1284 * If 'lkey', then we're the source VM, so grant access only to ourselves.
1285 */
1286 if (!block->pmr[chunk]) {
1287 uint64_t len = chunk_end - chunk_start;
1288 int access = rkey ? IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE :
1289 0;
1290
1291 trace_qemu_rdma_register_and_get_keys(len, chunk_start);
1292
1293 block->pmr[chunk] = ibv_reg_mr(rdma->pd, chunk_start, len, access);
1294 /*
1295 * ibv_reg_mr() is not documented to set errno. If it does,
1296 * it's somebody else's doc bug. If it doesn't, the use of
1297 * errno below is wrong.
1298 * TODO Find out whether ibv_reg_mr() sets errno.
1299 */
1300 if (!block->pmr[chunk] &&
1301 errno == ENOTSUP && rdma_support_odp(rdma->verbs)) {
1302 access |= IBV_ACCESS_ON_DEMAND;
1303 /* register ODP mr */
1304 block->pmr[chunk] = ibv_reg_mr(rdma->pd, chunk_start, len, access);
1305 trace_qemu_rdma_register_odp_mr(block->block_name);
1306
1307 if (block->pmr[chunk]) {
1308 qemu_rdma_advise_prefetch_mr(rdma->pd, (uintptr_t)chunk_start,
1309 len, block->pmr[chunk]->lkey,
1310 block->block_name, rkey);
1311
1312 }
1313 }
1314 }
1315 if (!block->pmr[chunk]) {
1316 return -1;
1317 }
1318 rdma->total_registrations++;
1319
1320 if (lkey) {
1321 *lkey = block->pmr[chunk]->lkey;
1322 }
1323 if (rkey) {
1324 *rkey = block->pmr[chunk]->rkey;
1325 }
1326 return 0;
1327 }
1328
1329 /*
1330 * Register (at connection time) the memory used for control
1331 * channel messages.
1332 */
qemu_rdma_reg_control(RDMAContext * rdma,int idx)1333 static int qemu_rdma_reg_control(RDMAContext *rdma, int idx)
1334 {
1335 rdma->wr_data[idx].control_mr = ibv_reg_mr(rdma->pd,
1336 rdma->wr_data[idx].control, RDMA_CONTROL_MAX_BUFFER,
1337 IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE);
1338 if (rdma->wr_data[idx].control_mr) {
1339 rdma->total_registrations++;
1340 return 0;
1341 }
1342 return -1;
1343 }
1344
1345 /*
1346 * Perform a non-optimized memory unregistration after every transfer
1347 * for demonstration purposes, only if pin-all is not requested.
1348 *
1349 * Potential optimizations:
1350 * 1. Start a new thread to run this function continuously
1351 - for bit clearing
1352 - and for receipt of unregister messages
1353 * 2. Use an LRU.
1354 * 3. Use workload hints.
1355 */
qemu_rdma_unregister_waiting(RDMAContext * rdma)1356 static int qemu_rdma_unregister_waiting(RDMAContext *rdma)
1357 {
1358 Error *err = NULL;
1359
1360 while (rdma->unregistrations[rdma->unregister_current]) {
1361 int ret;
1362 uint64_t wr_id = rdma->unregistrations[rdma->unregister_current];
1363 uint64_t chunk =
1364 (wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT;
1365 uint64_t index =
1366 (wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT;
1367 RDMALocalBlock *block =
1368 &(rdma->local_ram_blocks.block[index]);
1369 RDMARegister reg = { .current_index = index };
1370 RDMAControlHeader resp = { .type = RDMA_CONTROL_UNREGISTER_FINISHED,
1371 };
1372 RDMAControlHeader head = { .len = sizeof(RDMARegister),
1373 .type = RDMA_CONTROL_UNREGISTER_REQUEST,
1374 .repeat = 1,
1375 };
1376
1377 trace_qemu_rdma_unregister_waiting_proc(chunk,
1378 rdma->unregister_current);
1379
1380 rdma->unregistrations[rdma->unregister_current] = 0;
1381 rdma->unregister_current++;
1382
1383 if (rdma->unregister_current == RDMA_SIGNALED_SEND_MAX) {
1384 rdma->unregister_current = 0;
1385 }
1386
1387
1388 /*
1389 * Unregistration is speculative (because migration is single-threaded
1390 * and we cannot break the protocol's inifinband message ordering).
1391 * Thus, if the memory is currently being used for transmission,
1392 * then abort the attempt to unregister and try again
1393 * later the next time a completion is received for this memory.
1394 */
1395 clear_bit(chunk, block->unregister_bitmap);
1396
1397 if (test_bit(chunk, block->transit_bitmap)) {
1398 trace_qemu_rdma_unregister_waiting_inflight(chunk);
1399 continue;
1400 }
1401
1402 trace_qemu_rdma_unregister_waiting_send(chunk);
1403
1404 ret = ibv_dereg_mr(block->pmr[chunk]);
1405 block->pmr[chunk] = NULL;
1406 block->remote_keys[chunk] = 0;
1407
1408 if (ret != 0) {
1409 error_report("unregistration chunk failed: %s",
1410 strerror(ret));
1411 return -1;
1412 }
1413 rdma->total_registrations--;
1414
1415 reg.key.chunk = chunk;
1416 register_to_network(rdma, ®);
1417 ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®,
1418 &resp, NULL, NULL, &err);
1419 if (ret < 0) {
1420 error_report_err(err);
1421 return -1;
1422 }
1423
1424 trace_qemu_rdma_unregister_waiting_complete(chunk);
1425 }
1426
1427 return 0;
1428 }
1429
qemu_rdma_make_wrid(uint64_t wr_id,uint64_t index,uint64_t chunk)1430 static uint64_t qemu_rdma_make_wrid(uint64_t wr_id, uint64_t index,
1431 uint64_t chunk)
1432 {
1433 uint64_t result = wr_id & RDMA_WRID_TYPE_MASK;
1434
1435 result |= (index << RDMA_WRID_BLOCK_SHIFT);
1436 result |= (chunk << RDMA_WRID_CHUNK_SHIFT);
1437
1438 return result;
1439 }
1440
1441 /*
1442 * Consult the connection manager to see a work request
1443 * (of any kind) has completed.
1444 * Return the work request ID that completed.
1445 */
qemu_rdma_poll(RDMAContext * rdma,struct ibv_cq * cq,uint64_t * wr_id_out,uint32_t * byte_len)1446 static int qemu_rdma_poll(RDMAContext *rdma, struct ibv_cq *cq,
1447 uint64_t *wr_id_out, uint32_t *byte_len)
1448 {
1449 int ret;
1450 struct ibv_wc wc;
1451 uint64_t wr_id;
1452
1453 ret = ibv_poll_cq(cq, 1, &wc);
1454
1455 if (!ret) {
1456 *wr_id_out = RDMA_WRID_NONE;
1457 return 0;
1458 }
1459
1460 if (ret < 0) {
1461 return -1;
1462 }
1463
1464 wr_id = wc.wr_id & RDMA_WRID_TYPE_MASK;
1465
1466 if (wc.status != IBV_WC_SUCCESS) {
1467 return -1;
1468 }
1469
1470 if (rdma->control_ready_expected &&
1471 (wr_id >= RDMA_WRID_RECV_CONTROL)) {
1472 trace_qemu_rdma_poll_recv(wr_id - RDMA_WRID_RECV_CONTROL, wr_id,
1473 rdma->nb_sent);
1474 rdma->control_ready_expected = 0;
1475 }
1476
1477 if (wr_id == RDMA_WRID_RDMA_WRITE) {
1478 uint64_t chunk =
1479 (wc.wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT;
1480 uint64_t index =
1481 (wc.wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT;
1482 RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]);
1483
1484 trace_qemu_rdma_poll_write(wr_id, rdma->nb_sent,
1485 index, chunk, block->local_host_addr,
1486 (void *)(uintptr_t)block->remote_host_addr);
1487
1488 clear_bit(chunk, block->transit_bitmap);
1489
1490 if (rdma->nb_sent > 0) {
1491 rdma->nb_sent--;
1492 }
1493 } else {
1494 trace_qemu_rdma_poll_other(wr_id, rdma->nb_sent);
1495 }
1496
1497 *wr_id_out = wc.wr_id;
1498 if (byte_len) {
1499 *byte_len = wc.byte_len;
1500 }
1501
1502 return 0;
1503 }
1504
1505 /* Wait for activity on the completion channel.
1506 * Returns 0 on success, none-0 on error.
1507 */
qemu_rdma_wait_comp_channel(RDMAContext * rdma,struct ibv_comp_channel * comp_channel)1508 static int qemu_rdma_wait_comp_channel(RDMAContext *rdma,
1509 struct ibv_comp_channel *comp_channel)
1510 {
1511 struct rdma_cm_event *cm_event;
1512
1513 /*
1514 * Coroutine doesn't start until migration_fd_process_incoming()
1515 * so don't yield unless we know we're running inside of a coroutine.
1516 */
1517 if (rdma->migration_started_on_destination &&
1518 migration_incoming_get_current()->state == MIGRATION_STATUS_ACTIVE) {
1519 yield_until_fd_readable(comp_channel->fd);
1520 } else {
1521 /* This is the source side, we're in a separate thread
1522 * or destination prior to migration_fd_process_incoming()
1523 * after postcopy, the destination also in a separate thread.
1524 * we can't yield; so we have to poll the fd.
1525 * But we need to be able to handle 'cancel' or an error
1526 * without hanging forever.
1527 */
1528 while (!rdma->errored && !rdma->received_error) {
1529 GPollFD pfds[2];
1530 pfds[0].fd = comp_channel->fd;
1531 pfds[0].events = G_IO_IN | G_IO_HUP | G_IO_ERR;
1532 pfds[0].revents = 0;
1533
1534 pfds[1].fd = rdma->channel->fd;
1535 pfds[1].events = G_IO_IN | G_IO_HUP | G_IO_ERR;
1536 pfds[1].revents = 0;
1537
1538 /* 0.1s timeout, should be fine for a 'cancel' */
1539 switch (qemu_poll_ns(pfds, 2, 100 * 1000 * 1000)) {
1540 case 2:
1541 case 1: /* fd active */
1542 if (pfds[0].revents) {
1543 return 0;
1544 }
1545
1546 if (pfds[1].revents) {
1547 if (rdma_get_cm_event(rdma->channel, &cm_event) < 0) {
1548 return -1;
1549 }
1550
1551 if (cm_event->event == RDMA_CM_EVENT_DISCONNECTED ||
1552 cm_event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) {
1553 rdma_ack_cm_event(cm_event);
1554 return -1;
1555 }
1556 rdma_ack_cm_event(cm_event);
1557 }
1558 break;
1559
1560 case 0: /* Timeout, go around again */
1561 break;
1562
1563 default: /* Error of some type -
1564 * I don't trust errno from qemu_poll_ns
1565 */
1566 return -1;
1567 }
1568
1569 if (migrate_get_current()->state == MIGRATION_STATUS_CANCELLING) {
1570 /* Bail out and let the cancellation happen */
1571 return -1;
1572 }
1573 }
1574 }
1575
1576 if (rdma->received_error) {
1577 return -1;
1578 }
1579 return -rdma->errored;
1580 }
1581
to_channel(RDMAContext * rdma,uint64_t wrid)1582 static struct ibv_comp_channel *to_channel(RDMAContext *rdma, uint64_t wrid)
1583 {
1584 return wrid < RDMA_WRID_RECV_CONTROL ? rdma->send_comp_channel :
1585 rdma->recv_comp_channel;
1586 }
1587
to_cq(RDMAContext * rdma,uint64_t wrid)1588 static struct ibv_cq *to_cq(RDMAContext *rdma, uint64_t wrid)
1589 {
1590 return wrid < RDMA_WRID_RECV_CONTROL ? rdma->send_cq : rdma->recv_cq;
1591 }
1592
1593 /*
1594 * Block until the next work request has completed.
1595 *
1596 * First poll to see if a work request has already completed,
1597 * otherwise block.
1598 *
1599 * If we encounter completed work requests for IDs other than
1600 * the one we're interested in, then that's generally an error.
1601 *
1602 * The only exception is actual RDMA Write completions. These
1603 * completions only need to be recorded, but do not actually
1604 * need further processing.
1605 */
qemu_rdma_block_for_wrid(RDMAContext * rdma,uint64_t wrid_requested,uint32_t * byte_len)1606 static int qemu_rdma_block_for_wrid(RDMAContext *rdma,
1607 uint64_t wrid_requested,
1608 uint32_t *byte_len)
1609 {
1610 int num_cq_events = 0, ret;
1611 struct ibv_cq *cq;
1612 void *cq_ctx;
1613 uint64_t wr_id = RDMA_WRID_NONE, wr_id_in;
1614 struct ibv_comp_channel *ch = to_channel(rdma, wrid_requested);
1615 struct ibv_cq *poll_cq = to_cq(rdma, wrid_requested);
1616
1617 if (ibv_req_notify_cq(poll_cq, 0)) {
1618 return -1;
1619 }
1620 /* poll cq first */
1621 while (wr_id != wrid_requested) {
1622 ret = qemu_rdma_poll(rdma, poll_cq, &wr_id_in, byte_len);
1623 if (ret < 0) {
1624 return -1;
1625 }
1626
1627 wr_id = wr_id_in & RDMA_WRID_TYPE_MASK;
1628
1629 if (wr_id == RDMA_WRID_NONE) {
1630 break;
1631 }
1632 if (wr_id != wrid_requested) {
1633 trace_qemu_rdma_block_for_wrid_miss(wrid_requested, wr_id);
1634 }
1635 }
1636
1637 if (wr_id == wrid_requested) {
1638 return 0;
1639 }
1640
1641 while (1) {
1642 ret = qemu_rdma_wait_comp_channel(rdma, ch);
1643 if (ret < 0) {
1644 goto err_block_for_wrid;
1645 }
1646
1647 ret = ibv_get_cq_event(ch, &cq, &cq_ctx);
1648 if (ret < 0) {
1649 goto err_block_for_wrid;
1650 }
1651
1652 num_cq_events++;
1653
1654 if (ibv_req_notify_cq(cq, 0)) {
1655 goto err_block_for_wrid;
1656 }
1657
1658 while (wr_id != wrid_requested) {
1659 ret = qemu_rdma_poll(rdma, poll_cq, &wr_id_in, byte_len);
1660 if (ret < 0) {
1661 goto err_block_for_wrid;
1662 }
1663
1664 wr_id = wr_id_in & RDMA_WRID_TYPE_MASK;
1665
1666 if (wr_id == RDMA_WRID_NONE) {
1667 break;
1668 }
1669 if (wr_id != wrid_requested) {
1670 trace_qemu_rdma_block_for_wrid_miss(wrid_requested, wr_id);
1671 }
1672 }
1673
1674 if (wr_id == wrid_requested) {
1675 goto success_block_for_wrid;
1676 }
1677 }
1678
1679 success_block_for_wrid:
1680 if (num_cq_events) {
1681 ibv_ack_cq_events(cq, num_cq_events);
1682 }
1683 return 0;
1684
1685 err_block_for_wrid:
1686 if (num_cq_events) {
1687 ibv_ack_cq_events(cq, num_cq_events);
1688 }
1689
1690 rdma->errored = true;
1691 return -1;
1692 }
1693
1694 /*
1695 * Post a SEND message work request for the control channel
1696 * containing some data and block until the post completes.
1697 */
qemu_rdma_post_send_control(RDMAContext * rdma,uint8_t * buf,RDMAControlHeader * head,Error ** errp)1698 static int qemu_rdma_post_send_control(RDMAContext *rdma, uint8_t *buf,
1699 RDMAControlHeader *head,
1700 Error **errp)
1701 {
1702 int ret;
1703 RDMAWorkRequestData *wr = &rdma->wr_data[RDMA_WRID_CONTROL];
1704 struct ibv_send_wr *bad_wr;
1705 struct ibv_sge sge = {
1706 .addr = (uintptr_t)(wr->control),
1707 .length = head->len + sizeof(RDMAControlHeader),
1708 .lkey = wr->control_mr->lkey,
1709 };
1710 struct ibv_send_wr send_wr = {
1711 .wr_id = RDMA_WRID_SEND_CONTROL,
1712 .opcode = IBV_WR_SEND,
1713 .send_flags = IBV_SEND_SIGNALED,
1714 .sg_list = &sge,
1715 .num_sge = 1,
1716 };
1717
1718 trace_qemu_rdma_post_send_control(control_desc(head->type));
1719
1720 /*
1721 * We don't actually need to do a memcpy() in here if we used
1722 * the "sge" properly, but since we're only sending control messages
1723 * (not RAM in a performance-critical path), then its OK for now.
1724 *
1725 * The copy makes the RDMAControlHeader simpler to manipulate
1726 * for the time being.
1727 */
1728 assert(head->len <= RDMA_CONTROL_MAX_BUFFER - sizeof(*head));
1729 memcpy(wr->control, head, sizeof(RDMAControlHeader));
1730 control_to_network((void *) wr->control);
1731
1732 if (buf) {
1733 memcpy(wr->control + sizeof(RDMAControlHeader), buf, head->len);
1734 }
1735
1736
1737 ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr);
1738
1739 if (ret > 0) {
1740 error_setg(errp, "Failed to use post IB SEND for control");
1741 return -1;
1742 }
1743
1744 ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_SEND_CONTROL, NULL);
1745 if (ret < 0) {
1746 error_setg(errp, "rdma migration: send polling control error");
1747 return -1;
1748 }
1749
1750 return 0;
1751 }
1752
1753 /*
1754 * Post a RECV work request in anticipation of some future receipt
1755 * of data on the control channel.
1756 */
qemu_rdma_post_recv_control(RDMAContext * rdma,int idx,Error ** errp)1757 static int qemu_rdma_post_recv_control(RDMAContext *rdma, int idx,
1758 Error **errp)
1759 {
1760 struct ibv_recv_wr *bad_wr;
1761 struct ibv_sge sge = {
1762 .addr = (uintptr_t)(rdma->wr_data[idx].control),
1763 .length = RDMA_CONTROL_MAX_BUFFER,
1764 .lkey = rdma->wr_data[idx].control_mr->lkey,
1765 };
1766
1767 struct ibv_recv_wr recv_wr = {
1768 .wr_id = RDMA_WRID_RECV_CONTROL + idx,
1769 .sg_list = &sge,
1770 .num_sge = 1,
1771 };
1772
1773
1774 if (ibv_post_recv(rdma->qp, &recv_wr, &bad_wr)) {
1775 error_setg(errp, "error posting control recv");
1776 return -1;
1777 }
1778
1779 return 0;
1780 }
1781
1782 /*
1783 * Block and wait for a RECV control channel message to arrive.
1784 */
qemu_rdma_exchange_get_response(RDMAContext * rdma,RDMAControlHeader * head,uint32_t expecting,int idx,Error ** errp)1785 static int qemu_rdma_exchange_get_response(RDMAContext *rdma,
1786 RDMAControlHeader *head, uint32_t expecting, int idx,
1787 Error **errp)
1788 {
1789 uint32_t byte_len;
1790 int ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RECV_CONTROL + idx,
1791 &byte_len);
1792
1793 if (ret < 0) {
1794 error_setg(errp, "rdma migration: recv polling control error!");
1795 return -1;
1796 }
1797
1798 network_to_control((void *) rdma->wr_data[idx].control);
1799 memcpy(head, rdma->wr_data[idx].control, sizeof(RDMAControlHeader));
1800
1801 trace_qemu_rdma_exchange_get_response_start(control_desc(expecting));
1802
1803 if (expecting == RDMA_CONTROL_NONE) {
1804 trace_qemu_rdma_exchange_get_response_none(control_desc(head->type),
1805 head->type);
1806 } else if (head->type != expecting || head->type == RDMA_CONTROL_ERROR) {
1807 error_setg(errp, "Was expecting a %s (%d) control message"
1808 ", but got: %s (%d), length: %d",
1809 control_desc(expecting), expecting,
1810 control_desc(head->type), head->type, head->len);
1811 if (head->type == RDMA_CONTROL_ERROR) {
1812 rdma->received_error = true;
1813 }
1814 return -1;
1815 }
1816 if (head->len > RDMA_CONTROL_MAX_BUFFER - sizeof(*head)) {
1817 error_setg(errp, "too long length: %d", head->len);
1818 return -1;
1819 }
1820 if (sizeof(*head) + head->len != byte_len) {
1821 error_setg(errp, "Malformed length: %d byte_len %d",
1822 head->len, byte_len);
1823 return -1;
1824 }
1825
1826 return 0;
1827 }
1828
1829 /*
1830 * When a RECV work request has completed, the work request's
1831 * buffer is pointed at the header.
1832 *
1833 * This will advance the pointer to the data portion
1834 * of the control message of the work request's buffer that
1835 * was populated after the work request finished.
1836 */
qemu_rdma_move_header(RDMAContext * rdma,int idx,RDMAControlHeader * head)1837 static void qemu_rdma_move_header(RDMAContext *rdma, int idx,
1838 RDMAControlHeader *head)
1839 {
1840 rdma->wr_data[idx].control_len = head->len;
1841 rdma->wr_data[idx].control_curr =
1842 rdma->wr_data[idx].control + sizeof(RDMAControlHeader);
1843 }
1844
1845 /*
1846 * This is an 'atomic' high-level operation to deliver a single, unified
1847 * control-channel message.
1848 *
1849 * Additionally, if the user is expecting some kind of reply to this message,
1850 * they can request a 'resp' response message be filled in by posting an
1851 * additional work request on behalf of the user and waiting for an additional
1852 * completion.
1853 *
1854 * The extra (optional) response is used during registration to us from having
1855 * to perform an *additional* exchange of message just to provide a response by
1856 * instead piggy-backing on the acknowledgement.
1857 */
qemu_rdma_exchange_send(RDMAContext * rdma,RDMAControlHeader * head,uint8_t * data,RDMAControlHeader * resp,int * resp_idx,int (* callback)(RDMAContext * rdma,Error ** errp),Error ** errp)1858 static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head,
1859 uint8_t *data, RDMAControlHeader *resp,
1860 int *resp_idx,
1861 int (*callback)(RDMAContext *rdma,
1862 Error **errp),
1863 Error **errp)
1864 {
1865 int ret;
1866
1867 /*
1868 * Wait until the dest is ready before attempting to deliver the message
1869 * by waiting for a READY message.
1870 */
1871 if (rdma->control_ready_expected) {
1872 RDMAControlHeader resp_ignored;
1873
1874 ret = qemu_rdma_exchange_get_response(rdma, &resp_ignored,
1875 RDMA_CONTROL_READY,
1876 RDMA_WRID_READY, errp);
1877 if (ret < 0) {
1878 return -1;
1879 }
1880 }
1881
1882 /*
1883 * If the user is expecting a response, post a WR in anticipation of it.
1884 */
1885 if (resp) {
1886 ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_DATA, errp);
1887 if (ret < 0) {
1888 return -1;
1889 }
1890 }
1891
1892 /*
1893 * Post a WR to replace the one we just consumed for the READY message.
1894 */
1895 ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY, errp);
1896 if (ret < 0) {
1897 return -1;
1898 }
1899
1900 /*
1901 * Deliver the control message that was requested.
1902 */
1903 ret = qemu_rdma_post_send_control(rdma, data, head, errp);
1904
1905 if (ret < 0) {
1906 return -1;
1907 }
1908
1909 /*
1910 * If we're expecting a response, block and wait for it.
1911 */
1912 if (resp) {
1913 if (callback) {
1914 trace_qemu_rdma_exchange_send_issue_callback();
1915 ret = callback(rdma, errp);
1916 if (ret < 0) {
1917 return -1;
1918 }
1919 }
1920
1921 trace_qemu_rdma_exchange_send_waiting(control_desc(resp->type));
1922 ret = qemu_rdma_exchange_get_response(rdma, resp,
1923 resp->type, RDMA_WRID_DATA,
1924 errp);
1925
1926 if (ret < 0) {
1927 return -1;
1928 }
1929
1930 qemu_rdma_move_header(rdma, RDMA_WRID_DATA, resp);
1931 if (resp_idx) {
1932 *resp_idx = RDMA_WRID_DATA;
1933 }
1934 trace_qemu_rdma_exchange_send_received(control_desc(resp->type));
1935 }
1936
1937 rdma->control_ready_expected = 1;
1938
1939 return 0;
1940 }
1941
1942 /*
1943 * This is an 'atomic' high-level operation to receive a single, unified
1944 * control-channel message.
1945 */
qemu_rdma_exchange_recv(RDMAContext * rdma,RDMAControlHeader * head,uint32_t expecting,Error ** errp)1946 static int qemu_rdma_exchange_recv(RDMAContext *rdma, RDMAControlHeader *head,
1947 uint32_t expecting, Error **errp)
1948 {
1949 RDMAControlHeader ready = {
1950 .len = 0,
1951 .type = RDMA_CONTROL_READY,
1952 .repeat = 1,
1953 };
1954 int ret;
1955
1956 /*
1957 * Inform the source that we're ready to receive a message.
1958 */
1959 ret = qemu_rdma_post_send_control(rdma, NULL, &ready, errp);
1960
1961 if (ret < 0) {
1962 return -1;
1963 }
1964
1965 /*
1966 * Block and wait for the message.
1967 */
1968 ret = qemu_rdma_exchange_get_response(rdma, head,
1969 expecting, RDMA_WRID_READY, errp);
1970
1971 if (ret < 0) {
1972 return -1;
1973 }
1974
1975 qemu_rdma_move_header(rdma, RDMA_WRID_READY, head);
1976
1977 /*
1978 * Post a new RECV work request to replace the one we just consumed.
1979 */
1980 ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY, errp);
1981 if (ret < 0) {
1982 return -1;
1983 }
1984
1985 return 0;
1986 }
1987
1988 /*
1989 * Write an actual chunk of memory using RDMA.
1990 *
1991 * If we're using dynamic registration on the dest-side, we have to
1992 * send a registration command first.
1993 */
qemu_rdma_write_one(RDMAContext * rdma,int current_index,uint64_t current_addr,uint64_t length,Error ** errp)1994 static int qemu_rdma_write_one(RDMAContext *rdma,
1995 int current_index, uint64_t current_addr,
1996 uint64_t length, Error **errp)
1997 {
1998 struct ibv_sge sge;
1999 struct ibv_send_wr send_wr = { 0 };
2000 struct ibv_send_wr *bad_wr;
2001 int reg_result_idx, ret, count = 0;
2002 uint64_t chunk, chunks;
2003 uint8_t *chunk_start, *chunk_end;
2004 RDMALocalBlock *block = &(rdma->local_ram_blocks.block[current_index]);
2005 RDMARegister reg;
2006 RDMARegisterResult *reg_result;
2007 RDMAControlHeader resp = { .type = RDMA_CONTROL_REGISTER_RESULT };
2008 RDMAControlHeader head = { .len = sizeof(RDMARegister),
2009 .type = RDMA_CONTROL_REGISTER_REQUEST,
2010 .repeat = 1,
2011 };
2012
2013 retry:
2014 sge.addr = (uintptr_t)(block->local_host_addr +
2015 (current_addr - block->offset));
2016 sge.length = length;
2017
2018 chunk = ram_chunk_index(block->local_host_addr,
2019 (uint8_t *)(uintptr_t)sge.addr);
2020 chunk_start = ram_chunk_start(block, chunk);
2021
2022 if (block->is_ram_block) {
2023 chunks = length / (1UL << RDMA_REG_CHUNK_SHIFT);
2024
2025 if (chunks && ((length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) {
2026 chunks--;
2027 }
2028 } else {
2029 chunks = block->length / (1UL << RDMA_REG_CHUNK_SHIFT);
2030
2031 if (chunks && ((block->length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) {
2032 chunks--;
2033 }
2034 }
2035
2036 trace_qemu_rdma_write_one_top(chunks + 1,
2037 (chunks + 1) *
2038 (1UL << RDMA_REG_CHUNK_SHIFT) / 1024 / 1024);
2039
2040 chunk_end = ram_chunk_end(block, chunk + chunks);
2041
2042
2043 while (test_bit(chunk, block->transit_bitmap)) {
2044 (void)count;
2045 trace_qemu_rdma_write_one_block(count++, current_index, chunk,
2046 sge.addr, length, rdma->nb_sent, block->nb_chunks);
2047
2048 ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL);
2049
2050 if (ret < 0) {
2051 error_setg(errp, "Failed to Wait for previous write to complete "
2052 "block %d chunk %" PRIu64
2053 " current %" PRIu64 " len %" PRIu64 " %d",
2054 current_index, chunk, sge.addr, length, rdma->nb_sent);
2055 return -1;
2056 }
2057 }
2058
2059 if (!rdma->pin_all || !block->is_ram_block) {
2060 if (!block->remote_keys[chunk]) {
2061 /*
2062 * This chunk has not yet been registered, so first check to see
2063 * if the entire chunk is zero. If so, tell the other size to
2064 * memset() + madvise() the entire chunk without RDMA.
2065 */
2066
2067 if (buffer_is_zero((void *)(uintptr_t)sge.addr, length)) {
2068 RDMACompress comp = {
2069 .offset = current_addr,
2070 .value = 0,
2071 .block_idx = current_index,
2072 .length = length,
2073 };
2074
2075 head.len = sizeof(comp);
2076 head.type = RDMA_CONTROL_COMPRESS;
2077
2078 trace_qemu_rdma_write_one_zero(chunk, sge.length,
2079 current_index, current_addr);
2080
2081 compress_to_network(rdma, &comp);
2082 ret = qemu_rdma_exchange_send(rdma, &head,
2083 (uint8_t *) &comp, NULL, NULL, NULL, errp);
2084
2085 if (ret < 0) {
2086 return -1;
2087 }
2088
2089 /*
2090 * TODO: Here we are sending something, but we are not
2091 * accounting for anything transferred. The following is wrong:
2092 *
2093 * stat64_add(&mig_stats.rdma_bytes, sge.length);
2094 *
2095 * because we are using some kind of compression. I
2096 * would think that head.len would be the more similar
2097 * thing to a correct value.
2098 */
2099 stat64_add(&mig_stats.zero_pages,
2100 sge.length / qemu_target_page_size());
2101 return 1;
2102 }
2103
2104 /*
2105 * Otherwise, tell other side to register.
2106 */
2107 reg.current_index = current_index;
2108 if (block->is_ram_block) {
2109 reg.key.current_addr = current_addr;
2110 } else {
2111 reg.key.chunk = chunk;
2112 }
2113 reg.chunks = chunks;
2114
2115 trace_qemu_rdma_write_one_sendreg(chunk, sge.length, current_index,
2116 current_addr);
2117
2118 register_to_network(rdma, ®);
2119 ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®,
2120 &resp, ®_result_idx, NULL, errp);
2121 if (ret < 0) {
2122 return -1;
2123 }
2124
2125 /* try to overlap this single registration with the one we sent. */
2126 if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr,
2127 &sge.lkey, NULL, chunk,
2128 chunk_start, chunk_end)) {
2129 error_setg(errp, "cannot get lkey");
2130 return -1;
2131 }
2132
2133 reg_result = (RDMARegisterResult *)
2134 rdma->wr_data[reg_result_idx].control_curr;
2135
2136 network_to_result(reg_result);
2137
2138 trace_qemu_rdma_write_one_recvregres(block->remote_keys[chunk],
2139 reg_result->rkey, chunk);
2140
2141 block->remote_keys[chunk] = reg_result->rkey;
2142 block->remote_host_addr = reg_result->host_addr;
2143 } else {
2144 /* already registered before */
2145 if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr,
2146 &sge.lkey, NULL, chunk,
2147 chunk_start, chunk_end)) {
2148 error_setg(errp, "cannot get lkey!");
2149 return -1;
2150 }
2151 }
2152
2153 send_wr.wr.rdma.rkey = block->remote_keys[chunk];
2154 } else {
2155 send_wr.wr.rdma.rkey = block->remote_rkey;
2156
2157 if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr,
2158 &sge.lkey, NULL, chunk,
2159 chunk_start, chunk_end)) {
2160 error_setg(errp, "cannot get lkey!");
2161 return -1;
2162 }
2163 }
2164
2165 /*
2166 * Encode the ram block index and chunk within this wrid.
2167 * We will use this information at the time of completion
2168 * to figure out which bitmap to check against and then which
2169 * chunk in the bitmap to look for.
2170 */
2171 send_wr.wr_id = qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE,
2172 current_index, chunk);
2173
2174 send_wr.opcode = IBV_WR_RDMA_WRITE;
2175 send_wr.send_flags = IBV_SEND_SIGNALED;
2176 send_wr.sg_list = &sge;
2177 send_wr.num_sge = 1;
2178 send_wr.wr.rdma.remote_addr = block->remote_host_addr +
2179 (current_addr - block->offset);
2180
2181 trace_qemu_rdma_write_one_post(chunk, sge.addr, send_wr.wr.rdma.remote_addr,
2182 sge.length);
2183
2184 /*
2185 * ibv_post_send() does not return negative error numbers,
2186 * per the specification they are positive - no idea why.
2187 */
2188 ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr);
2189
2190 if (ret == ENOMEM) {
2191 trace_qemu_rdma_write_one_queue_full();
2192 ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL);
2193 if (ret < 0) {
2194 error_setg(errp, "rdma migration: failed to make "
2195 "room in full send queue!");
2196 return -1;
2197 }
2198
2199 goto retry;
2200
2201 } else if (ret > 0) {
2202 error_setg_errno(errp, ret,
2203 "rdma migration: post rdma write failed");
2204 return -1;
2205 }
2206
2207 set_bit(chunk, block->transit_bitmap);
2208 stat64_add(&mig_stats.normal_pages, sge.length / qemu_target_page_size());
2209 /*
2210 * We are adding to transferred the amount of data written, but no
2211 * overhead at all. I will assume that RDMA is magicaly and don't
2212 * need to transfer (at least) the addresses where it wants to
2213 * write the pages. Here it looks like it should be something
2214 * like:
2215 * sizeof(send_wr) + sge.length
2216 * but this being RDMA, who knows.
2217 */
2218 stat64_add(&mig_stats.rdma_bytes, sge.length);
2219 ram_transferred_add(sge.length);
2220 rdma->total_writes++;
2221
2222 return 0;
2223 }
2224
2225 /*
2226 * Push out any unwritten RDMA operations.
2227 *
2228 * We support sending out multiple chunks at the same time.
2229 * Not all of them need to get signaled in the completion queue.
2230 */
qemu_rdma_write_flush(RDMAContext * rdma,Error ** errp)2231 static int qemu_rdma_write_flush(RDMAContext *rdma, Error **errp)
2232 {
2233 int ret;
2234
2235 if (!rdma->current_length) {
2236 return 0;
2237 }
2238
2239 ret = qemu_rdma_write_one(rdma, rdma->current_index, rdma->current_addr,
2240 rdma->current_length, errp);
2241
2242 if (ret < 0) {
2243 return -1;
2244 }
2245
2246 if (ret == 0) {
2247 rdma->nb_sent++;
2248 trace_qemu_rdma_write_flush(rdma->nb_sent);
2249 }
2250
2251 rdma->current_length = 0;
2252 rdma->current_addr = 0;
2253
2254 return 0;
2255 }
2256
qemu_rdma_buffer_mergeable(RDMAContext * rdma,uint64_t offset,uint64_t len)2257 static inline bool qemu_rdma_buffer_mergeable(RDMAContext *rdma,
2258 uint64_t offset, uint64_t len)
2259 {
2260 RDMALocalBlock *block;
2261 uint8_t *host_addr;
2262 uint8_t *chunk_end;
2263
2264 if (rdma->current_index < 0) {
2265 return false;
2266 }
2267
2268 if (rdma->current_chunk < 0) {
2269 return false;
2270 }
2271
2272 block = &(rdma->local_ram_blocks.block[rdma->current_index]);
2273 host_addr = block->local_host_addr + (offset - block->offset);
2274 chunk_end = ram_chunk_end(block, rdma->current_chunk);
2275
2276 if (rdma->current_length == 0) {
2277 return false;
2278 }
2279
2280 /*
2281 * Only merge into chunk sequentially.
2282 */
2283 if (offset != (rdma->current_addr + rdma->current_length)) {
2284 return false;
2285 }
2286
2287 if (offset < block->offset) {
2288 return false;
2289 }
2290
2291 if ((offset + len) > (block->offset + block->length)) {
2292 return false;
2293 }
2294
2295 if ((host_addr + len) > chunk_end) {
2296 return false;
2297 }
2298
2299 return true;
2300 }
2301
2302 /*
2303 * We're not actually writing here, but doing three things:
2304 *
2305 * 1. Identify the chunk the buffer belongs to.
2306 * 2. If the chunk is full or the buffer doesn't belong to the current
2307 * chunk, then start a new chunk and flush() the old chunk.
2308 * 3. To keep the hardware busy, we also group chunks into batches
2309 * and only require that a batch gets acknowledged in the completion
2310 * queue instead of each individual chunk.
2311 */
qemu_rdma_write(RDMAContext * rdma,uint64_t block_offset,uint64_t offset,uint64_t len,Error ** errp)2312 static int qemu_rdma_write(RDMAContext *rdma,
2313 uint64_t block_offset, uint64_t offset,
2314 uint64_t len, Error **errp)
2315 {
2316 uint64_t current_addr = block_offset + offset;
2317 uint64_t index = rdma->current_index;
2318 uint64_t chunk = rdma->current_chunk;
2319
2320 /* If we cannot merge it, we flush the current buffer first. */
2321 if (!qemu_rdma_buffer_mergeable(rdma, current_addr, len)) {
2322 if (qemu_rdma_write_flush(rdma, errp) < 0) {
2323 return -1;
2324 }
2325 rdma->current_length = 0;
2326 rdma->current_addr = current_addr;
2327
2328 qemu_rdma_search_ram_block(rdma, block_offset,
2329 offset, len, &index, &chunk);
2330 rdma->current_index = index;
2331 rdma->current_chunk = chunk;
2332 }
2333
2334 /* merge it */
2335 rdma->current_length += len;
2336
2337 /* flush it if buffer is too large */
2338 if (rdma->current_length >= RDMA_MERGE_MAX) {
2339 return qemu_rdma_write_flush(rdma, errp);
2340 }
2341
2342 return 0;
2343 }
2344
qemu_rdma_cleanup(RDMAContext * rdma)2345 static void qemu_rdma_cleanup(RDMAContext *rdma)
2346 {
2347 Error *err = NULL;
2348
2349 if (rdma->cm_id && rdma->connected) {
2350 if ((rdma->errored ||
2351 migrate_get_current()->state == MIGRATION_STATUS_CANCELLING) &&
2352 !rdma->received_error) {
2353 RDMAControlHeader head = { .len = 0,
2354 .type = RDMA_CONTROL_ERROR,
2355 .repeat = 1,
2356 };
2357 warn_report("Early error. Sending error.");
2358 if (qemu_rdma_post_send_control(rdma, NULL, &head, &err) < 0) {
2359 warn_report_err(err);
2360 }
2361 }
2362
2363 rdma_disconnect(rdma->cm_id);
2364 trace_qemu_rdma_cleanup_disconnect();
2365 rdma->connected = false;
2366 }
2367
2368 if (rdma->channel) {
2369 qemu_set_fd_handler(rdma->channel->fd, NULL, NULL, NULL);
2370 }
2371 g_free(rdma->dest_blocks);
2372 rdma->dest_blocks = NULL;
2373
2374 for (int i = 0; i < RDMA_WRID_MAX; i++) {
2375 if (rdma->wr_data[i].control_mr) {
2376 rdma->total_registrations--;
2377 ibv_dereg_mr(rdma->wr_data[i].control_mr);
2378 }
2379 rdma->wr_data[i].control_mr = NULL;
2380 }
2381
2382 if (rdma->local_ram_blocks.block) {
2383 while (rdma->local_ram_blocks.nb_blocks) {
2384 rdma_delete_block(rdma, &rdma->local_ram_blocks.block[0]);
2385 }
2386 }
2387
2388 if (rdma->qp) {
2389 rdma_destroy_qp(rdma->cm_id);
2390 rdma->qp = NULL;
2391 }
2392 if (rdma->recv_cq) {
2393 ibv_destroy_cq(rdma->recv_cq);
2394 rdma->recv_cq = NULL;
2395 }
2396 if (rdma->send_cq) {
2397 ibv_destroy_cq(rdma->send_cq);
2398 rdma->send_cq = NULL;
2399 }
2400 if (rdma->recv_comp_channel) {
2401 ibv_destroy_comp_channel(rdma->recv_comp_channel);
2402 rdma->recv_comp_channel = NULL;
2403 }
2404 if (rdma->send_comp_channel) {
2405 ibv_destroy_comp_channel(rdma->send_comp_channel);
2406 rdma->send_comp_channel = NULL;
2407 }
2408 if (rdma->pd) {
2409 ibv_dealloc_pd(rdma->pd);
2410 rdma->pd = NULL;
2411 }
2412 if (rdma->cm_id) {
2413 rdma_destroy_id(rdma->cm_id);
2414 rdma->cm_id = NULL;
2415 }
2416
2417 /* the destination side, listen_id and channel is shared */
2418 if (rdma->listen_id) {
2419 if (!rdma->is_return_path) {
2420 rdma_destroy_id(rdma->listen_id);
2421 }
2422 rdma->listen_id = NULL;
2423
2424 if (rdma->channel) {
2425 if (!rdma->is_return_path) {
2426 rdma_destroy_event_channel(rdma->channel);
2427 }
2428 rdma->channel = NULL;
2429 }
2430 }
2431
2432 if (rdma->channel) {
2433 rdma_destroy_event_channel(rdma->channel);
2434 rdma->channel = NULL;
2435 }
2436 g_free(rdma->host);
2437 rdma->host = NULL;
2438 }
2439
2440
qemu_rdma_source_init(RDMAContext * rdma,bool pin_all,Error ** errp)2441 static int qemu_rdma_source_init(RDMAContext *rdma, bool pin_all, Error **errp)
2442 {
2443 int ret;
2444
2445 /*
2446 * Will be validated against destination's actual capabilities
2447 * after the connect() completes.
2448 */
2449 rdma->pin_all = pin_all;
2450
2451 ret = qemu_rdma_resolve_host(rdma, errp);
2452 if (ret < 0) {
2453 goto err_rdma_source_init;
2454 }
2455
2456 ret = qemu_rdma_alloc_pd_cq(rdma, errp);
2457 if (ret < 0) {
2458 goto err_rdma_source_init;
2459 }
2460
2461 ret = qemu_rdma_alloc_qp(rdma);
2462 if (ret < 0) {
2463 error_setg(errp, "RDMA ERROR: rdma migration: error allocating qp!");
2464 goto err_rdma_source_init;
2465 }
2466
2467 qemu_rdma_init_ram_blocks(rdma);
2468
2469 /* Build the hash that maps from offset to RAMBlock */
2470 rdma->blockmap = g_hash_table_new(g_direct_hash, g_direct_equal);
2471 for (int i = 0; i < rdma->local_ram_blocks.nb_blocks; i++) {
2472 g_hash_table_insert(rdma->blockmap,
2473 (void *)(uintptr_t)rdma->local_ram_blocks.block[i].offset,
2474 &rdma->local_ram_blocks.block[i]);
2475 }
2476
2477 for (int i = 0; i < RDMA_WRID_MAX; i++) {
2478 ret = qemu_rdma_reg_control(rdma, i);
2479 if (ret < 0) {
2480 error_setg(errp, "RDMA ERROR: rdma migration: error "
2481 "registering %d control!", i);
2482 goto err_rdma_source_init;
2483 }
2484 }
2485
2486 return 0;
2487
2488 err_rdma_source_init:
2489 qemu_rdma_cleanup(rdma);
2490 return -1;
2491 }
2492
qemu_get_cm_event_timeout(RDMAContext * rdma,struct rdma_cm_event ** cm_event,long msec,Error ** errp)2493 static int qemu_get_cm_event_timeout(RDMAContext *rdma,
2494 struct rdma_cm_event **cm_event,
2495 long msec, Error **errp)
2496 {
2497 int ret;
2498 struct pollfd poll_fd = {
2499 .fd = rdma->channel->fd,
2500 .events = POLLIN,
2501 .revents = 0
2502 };
2503
2504 do {
2505 ret = poll(&poll_fd, 1, msec);
2506 } while (ret < 0 && errno == EINTR);
2507
2508 if (ret == 0) {
2509 error_setg(errp, "RDMA ERROR: poll cm event timeout");
2510 return -1;
2511 } else if (ret < 0) {
2512 error_setg(errp, "RDMA ERROR: failed to poll cm event, errno=%i",
2513 errno);
2514 return -1;
2515 } else if (poll_fd.revents & POLLIN) {
2516 if (rdma_get_cm_event(rdma->channel, cm_event) < 0) {
2517 error_setg(errp, "RDMA ERROR: failed to get cm event");
2518 return -1;
2519 }
2520 return 0;
2521 } else {
2522 error_setg(errp, "RDMA ERROR: no POLLIN event, revent=%x",
2523 poll_fd.revents);
2524 return -1;
2525 }
2526 }
2527
qemu_rdma_connect(RDMAContext * rdma,bool return_path,Error ** errp)2528 static int qemu_rdma_connect(RDMAContext *rdma, bool return_path,
2529 Error **errp)
2530 {
2531 RDMACapabilities cap = {
2532 .version = RDMA_CONTROL_VERSION_CURRENT,
2533 .flags = 0,
2534 };
2535 struct rdma_conn_param conn_param = { .initiator_depth = 2,
2536 .retry_count = 5,
2537 .private_data = &cap,
2538 .private_data_len = sizeof(cap),
2539 };
2540 struct rdma_cm_event *cm_event;
2541 int ret;
2542
2543 /*
2544 * Only negotiate the capability with destination if the user
2545 * on the source first requested the capability.
2546 */
2547 if (rdma->pin_all) {
2548 trace_qemu_rdma_connect_pin_all_requested();
2549 cap.flags |= RDMA_CAPABILITY_PIN_ALL;
2550 }
2551
2552 caps_to_network(&cap);
2553
2554 ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY, errp);
2555 if (ret < 0) {
2556 goto err_rdma_source_connect;
2557 }
2558
2559 ret = rdma_connect(rdma->cm_id, &conn_param);
2560 if (ret < 0) {
2561 error_setg_errno(errp, errno,
2562 "RDMA ERROR: connecting to destination!");
2563 goto err_rdma_source_connect;
2564 }
2565
2566 if (return_path) {
2567 ret = qemu_get_cm_event_timeout(rdma, &cm_event, 5000, errp);
2568 } else {
2569 ret = rdma_get_cm_event(rdma->channel, &cm_event);
2570 if (ret < 0) {
2571 error_setg_errno(errp, errno,
2572 "RDMA ERROR: failed to get cm event");
2573 }
2574 }
2575 if (ret < 0) {
2576 goto err_rdma_source_connect;
2577 }
2578
2579 if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) {
2580 error_setg(errp, "RDMA ERROR: connecting to destination!");
2581 rdma_ack_cm_event(cm_event);
2582 goto err_rdma_source_connect;
2583 }
2584 rdma->connected = true;
2585
2586 memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap));
2587 network_to_caps(&cap);
2588
2589 /*
2590 * Verify that the *requested* capabilities are supported by the destination
2591 * and disable them otherwise.
2592 */
2593 if (rdma->pin_all && !(cap.flags & RDMA_CAPABILITY_PIN_ALL)) {
2594 warn_report("RDMA: Server cannot support pinning all memory. "
2595 "Will register memory dynamically.");
2596 rdma->pin_all = false;
2597 }
2598
2599 trace_qemu_rdma_connect_pin_all_outcome(rdma->pin_all);
2600
2601 rdma_ack_cm_event(cm_event);
2602
2603 rdma->control_ready_expected = 1;
2604 rdma->nb_sent = 0;
2605 return 0;
2606
2607 err_rdma_source_connect:
2608 qemu_rdma_cleanup(rdma);
2609 return -1;
2610 }
2611
qemu_rdma_dest_init(RDMAContext * rdma,Error ** errp)2612 static int qemu_rdma_dest_init(RDMAContext *rdma, Error **errp)
2613 {
2614 Error *err = NULL;
2615 int ret;
2616 struct rdma_cm_id *listen_id;
2617 char ip[40] = "unknown";
2618 struct rdma_addrinfo *res, *e;
2619 char port_str[16];
2620 int reuse = 1;
2621
2622 for (int i = 0; i < RDMA_WRID_MAX; i++) {
2623 rdma->wr_data[i].control_len = 0;
2624 rdma->wr_data[i].control_curr = NULL;
2625 }
2626
2627 if (!rdma->host || !rdma->host[0]) {
2628 error_setg(errp, "RDMA ERROR: RDMA host is not set!");
2629 rdma->errored = true;
2630 return -1;
2631 }
2632 /* create CM channel */
2633 rdma->channel = rdma_create_event_channel();
2634 if (!rdma->channel) {
2635 error_setg(errp, "RDMA ERROR: could not create rdma event channel");
2636 rdma->errored = true;
2637 return -1;
2638 }
2639
2640 /* create CM id */
2641 ret = rdma_create_id(rdma->channel, &listen_id, NULL, RDMA_PS_TCP);
2642 if (ret < 0) {
2643 error_setg(errp, "RDMA ERROR: could not create cm_id!");
2644 goto err_dest_init_create_listen_id;
2645 }
2646
2647 snprintf(port_str, 16, "%d", rdma->port);
2648 port_str[15] = '\0';
2649
2650 ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res);
2651 if (ret) {
2652 error_setg(errp, "RDMA ERROR: could not rdma_getaddrinfo address %s",
2653 rdma->host);
2654 goto err_dest_init_bind_addr;
2655 }
2656
2657 ret = rdma_set_option(listen_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR,
2658 &reuse, sizeof reuse);
2659 if (ret < 0) {
2660 error_setg(errp, "RDMA ERROR: Error: could not set REUSEADDR option");
2661 goto err_dest_init_bind_addr;
2662 }
2663
2664 /* Try all addresses, saving the first error in @err */
2665 for (e = res; e != NULL; e = e->ai_next) {
2666 Error **local_errp = err ? NULL : &err;
2667
2668 inet_ntop(e->ai_family,
2669 &((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip);
2670 trace_qemu_rdma_dest_init_trying(rdma->host, ip);
2671 ret = rdma_bind_addr(listen_id, e->ai_dst_addr);
2672 if (ret < 0) {
2673 continue;
2674 }
2675 if (e->ai_family == AF_INET6) {
2676 ret = qemu_rdma_broken_ipv6_kernel(listen_id->verbs,
2677 local_errp);
2678 if (ret < 0) {
2679 continue;
2680 }
2681 }
2682 error_free(err);
2683 break;
2684 }
2685
2686 rdma_freeaddrinfo(res);
2687 if (!e) {
2688 if (err) {
2689 error_propagate(errp, err);
2690 } else {
2691 error_setg(errp, "RDMA ERROR: Error: could not rdma_bind_addr!");
2692 }
2693 goto err_dest_init_bind_addr;
2694 }
2695
2696 rdma->listen_id = listen_id;
2697 qemu_rdma_dump_gid("dest_init", listen_id);
2698 return 0;
2699
2700 err_dest_init_bind_addr:
2701 rdma_destroy_id(listen_id);
2702 err_dest_init_create_listen_id:
2703 rdma_destroy_event_channel(rdma->channel);
2704 rdma->channel = NULL;
2705 rdma->errored = true;
2706 return -1;
2707
2708 }
2709
qemu_rdma_return_path_dest_init(RDMAContext * rdma_return_path,RDMAContext * rdma)2710 static void qemu_rdma_return_path_dest_init(RDMAContext *rdma_return_path,
2711 RDMAContext *rdma)
2712 {
2713 for (int i = 0; i < RDMA_WRID_MAX; i++) {
2714 rdma_return_path->wr_data[i].control_len = 0;
2715 rdma_return_path->wr_data[i].control_curr = NULL;
2716 }
2717
2718 /*the CM channel and CM id is shared*/
2719 rdma_return_path->channel = rdma->channel;
2720 rdma_return_path->listen_id = rdma->listen_id;
2721
2722 rdma->return_path = rdma_return_path;
2723 rdma_return_path->return_path = rdma;
2724 rdma_return_path->is_return_path = true;
2725 }
2726
qemu_rdma_data_init(InetSocketAddress * saddr,Error ** errp)2727 static RDMAContext *qemu_rdma_data_init(InetSocketAddress *saddr, Error **errp)
2728 {
2729 RDMAContext *rdma = NULL;
2730
2731 rdma = g_new0(RDMAContext, 1);
2732 rdma->current_index = -1;
2733 rdma->current_chunk = -1;
2734
2735 rdma->host = g_strdup(saddr->host);
2736 rdma->port = atoi(saddr->port);
2737 return rdma;
2738 }
2739
2740 /*
2741 * QEMUFile interface to the control channel.
2742 * SEND messages for control only.
2743 * VM's ram is handled with regular RDMA messages.
2744 */
qio_channel_rdma_writev(QIOChannel * ioc,const struct iovec * iov,size_t niov,int * fds,size_t nfds,int flags,Error ** errp)2745 static ssize_t qio_channel_rdma_writev(QIOChannel *ioc,
2746 const struct iovec *iov,
2747 size_t niov,
2748 int *fds,
2749 size_t nfds,
2750 int flags,
2751 Error **errp)
2752 {
2753 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
2754 RDMAContext *rdma;
2755 int ret;
2756 ssize_t done = 0;
2757 size_t len;
2758
2759 RCU_READ_LOCK_GUARD();
2760 rdma = qatomic_rcu_read(&rioc->rdmaout);
2761
2762 if (!rdma) {
2763 error_setg(errp, "RDMA control channel output is not set");
2764 return -1;
2765 }
2766
2767 if (rdma->errored) {
2768 error_setg(errp,
2769 "RDMA is in an error state waiting migration to abort!");
2770 return -1;
2771 }
2772
2773 /*
2774 * Push out any writes that
2775 * we're queued up for VM's ram.
2776 */
2777 ret = qemu_rdma_write_flush(rdma, errp);
2778 if (ret < 0) {
2779 rdma->errored = true;
2780 return -1;
2781 }
2782
2783 for (int i = 0; i < niov; i++) {
2784 size_t remaining = iov[i].iov_len;
2785 uint8_t * data = (void *)iov[i].iov_base;
2786 while (remaining) {
2787 RDMAControlHeader head = {};
2788
2789 len = MIN(remaining, RDMA_SEND_INCREMENT);
2790 remaining -= len;
2791
2792 head.len = len;
2793 head.type = RDMA_CONTROL_QEMU_FILE;
2794
2795 ret = qemu_rdma_exchange_send(rdma, &head,
2796 data, NULL, NULL, NULL, errp);
2797
2798 if (ret < 0) {
2799 rdma->errored = true;
2800 return -1;
2801 }
2802
2803 data += len;
2804 done += len;
2805 }
2806 }
2807
2808 return done;
2809 }
2810
qemu_rdma_fill(RDMAContext * rdma,uint8_t * buf,size_t size,int idx)2811 static size_t qemu_rdma_fill(RDMAContext *rdma, uint8_t *buf,
2812 size_t size, int idx)
2813 {
2814 size_t len = 0;
2815
2816 if (rdma->wr_data[idx].control_len) {
2817 trace_qemu_rdma_fill(rdma->wr_data[idx].control_len, size);
2818
2819 len = MIN(size, rdma->wr_data[idx].control_len);
2820 memcpy(buf, rdma->wr_data[idx].control_curr, len);
2821 rdma->wr_data[idx].control_curr += len;
2822 rdma->wr_data[idx].control_len -= len;
2823 }
2824
2825 return len;
2826 }
2827
2828 /*
2829 * QEMUFile interface to the control channel.
2830 * RDMA links don't use bytestreams, so we have to
2831 * return bytes to QEMUFile opportunistically.
2832 */
qio_channel_rdma_readv(QIOChannel * ioc,const struct iovec * iov,size_t niov,int ** fds,size_t * nfds,int flags,Error ** errp)2833 static ssize_t qio_channel_rdma_readv(QIOChannel *ioc,
2834 const struct iovec *iov,
2835 size_t niov,
2836 int **fds,
2837 size_t *nfds,
2838 int flags,
2839 Error **errp)
2840 {
2841 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
2842 RDMAContext *rdma;
2843 RDMAControlHeader head;
2844 int ret;
2845 ssize_t done = 0;
2846 size_t len;
2847
2848 RCU_READ_LOCK_GUARD();
2849 rdma = qatomic_rcu_read(&rioc->rdmain);
2850
2851 if (!rdma) {
2852 error_setg(errp, "RDMA control channel input is not set");
2853 return -1;
2854 }
2855
2856 if (rdma->errored) {
2857 error_setg(errp,
2858 "RDMA is in an error state waiting migration to abort!");
2859 return -1;
2860 }
2861
2862 for (int i = 0; i < niov; i++) {
2863 size_t want = iov[i].iov_len;
2864 uint8_t *data = (void *)iov[i].iov_base;
2865
2866 /*
2867 * First, we hold on to the last SEND message we
2868 * were given and dish out the bytes until we run
2869 * out of bytes.
2870 */
2871 len = qemu_rdma_fill(rdma, data, want, 0);
2872 done += len;
2873 want -= len;
2874 /* Got what we needed, so go to next iovec */
2875 if (want == 0) {
2876 continue;
2877 }
2878
2879 /* If we got any data so far, then don't wait
2880 * for more, just return what we have */
2881 if (done > 0) {
2882 break;
2883 }
2884
2885
2886 /* We've got nothing at all, so lets wait for
2887 * more to arrive
2888 */
2889 ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_QEMU_FILE,
2890 errp);
2891
2892 if (ret < 0) {
2893 rdma->errored = true;
2894 return -1;
2895 }
2896
2897 /*
2898 * SEND was received with new bytes, now try again.
2899 */
2900 len = qemu_rdma_fill(rdma, data, want, 0);
2901 done += len;
2902 want -= len;
2903
2904 /* Still didn't get enough, so lets just return */
2905 if (want) {
2906 if (done == 0) {
2907 return QIO_CHANNEL_ERR_BLOCK;
2908 } else {
2909 break;
2910 }
2911 }
2912 }
2913 return done;
2914 }
2915
2916 /*
2917 * Block until all the outstanding chunks have been delivered by the hardware.
2918 */
qemu_rdma_drain_cq(RDMAContext * rdma)2919 static int qemu_rdma_drain_cq(RDMAContext *rdma)
2920 {
2921 Error *err = NULL;
2922
2923 if (qemu_rdma_write_flush(rdma, &err) < 0) {
2924 error_report_err(err);
2925 return -1;
2926 }
2927
2928 while (rdma->nb_sent) {
2929 if (qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL) < 0) {
2930 error_report("rdma migration: complete polling error!");
2931 return -1;
2932 }
2933 }
2934
2935 qemu_rdma_unregister_waiting(rdma);
2936
2937 return 0;
2938 }
2939
2940
qio_channel_rdma_set_blocking(QIOChannel * ioc,bool blocking,Error ** errp)2941 static int qio_channel_rdma_set_blocking(QIOChannel *ioc,
2942 bool blocking,
2943 Error **errp)
2944 {
2945 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
2946 /* XXX we should make readv/writev actually honour this :-) */
2947 rioc->blocking = blocking;
2948 return 0;
2949 }
2950
2951
2952 typedef struct QIOChannelRDMASource QIOChannelRDMASource;
2953 struct QIOChannelRDMASource {
2954 GSource parent;
2955 QIOChannelRDMA *rioc;
2956 GIOCondition condition;
2957 };
2958
2959 static gboolean
qio_channel_rdma_source_prepare(GSource * source,gint * timeout)2960 qio_channel_rdma_source_prepare(GSource *source,
2961 gint *timeout)
2962 {
2963 QIOChannelRDMASource *rsource = (QIOChannelRDMASource *)source;
2964 RDMAContext *rdma;
2965 GIOCondition cond = 0;
2966 *timeout = -1;
2967
2968 RCU_READ_LOCK_GUARD();
2969 if (rsource->condition == G_IO_IN) {
2970 rdma = qatomic_rcu_read(&rsource->rioc->rdmain);
2971 } else {
2972 rdma = qatomic_rcu_read(&rsource->rioc->rdmaout);
2973 }
2974
2975 if (!rdma) {
2976 error_report("RDMAContext is NULL when prepare Gsource");
2977 return FALSE;
2978 }
2979
2980 if (rdma->wr_data[0].control_len) {
2981 cond |= G_IO_IN;
2982 }
2983 cond |= G_IO_OUT;
2984
2985 return cond & rsource->condition;
2986 }
2987
2988 static gboolean
qio_channel_rdma_source_check(GSource * source)2989 qio_channel_rdma_source_check(GSource *source)
2990 {
2991 QIOChannelRDMASource *rsource = (QIOChannelRDMASource *)source;
2992 RDMAContext *rdma;
2993 GIOCondition cond = 0;
2994
2995 RCU_READ_LOCK_GUARD();
2996 if (rsource->condition == G_IO_IN) {
2997 rdma = qatomic_rcu_read(&rsource->rioc->rdmain);
2998 } else {
2999 rdma = qatomic_rcu_read(&rsource->rioc->rdmaout);
3000 }
3001
3002 if (!rdma) {
3003 error_report("RDMAContext is NULL when check Gsource");
3004 return FALSE;
3005 }
3006
3007 if (rdma->wr_data[0].control_len) {
3008 cond |= G_IO_IN;
3009 }
3010 cond |= G_IO_OUT;
3011
3012 return cond & rsource->condition;
3013 }
3014
3015 static gboolean
qio_channel_rdma_source_dispatch(GSource * source,GSourceFunc callback,gpointer user_data)3016 qio_channel_rdma_source_dispatch(GSource *source,
3017 GSourceFunc callback,
3018 gpointer user_data)
3019 {
3020 QIOChannelFunc func = (QIOChannelFunc)callback;
3021 QIOChannelRDMASource *rsource = (QIOChannelRDMASource *)source;
3022 RDMAContext *rdma;
3023 GIOCondition cond = 0;
3024
3025 RCU_READ_LOCK_GUARD();
3026 if (rsource->condition == G_IO_IN) {
3027 rdma = qatomic_rcu_read(&rsource->rioc->rdmain);
3028 } else {
3029 rdma = qatomic_rcu_read(&rsource->rioc->rdmaout);
3030 }
3031
3032 if (!rdma) {
3033 error_report("RDMAContext is NULL when dispatch Gsource");
3034 return FALSE;
3035 }
3036
3037 if (rdma->wr_data[0].control_len) {
3038 cond |= G_IO_IN;
3039 }
3040 cond |= G_IO_OUT;
3041
3042 return (*func)(QIO_CHANNEL(rsource->rioc),
3043 (cond & rsource->condition),
3044 user_data);
3045 }
3046
3047 static void
qio_channel_rdma_source_finalize(GSource * source)3048 qio_channel_rdma_source_finalize(GSource *source)
3049 {
3050 QIOChannelRDMASource *ssource = (QIOChannelRDMASource *)source;
3051
3052 object_unref(OBJECT(ssource->rioc));
3053 }
3054
3055 static GSourceFuncs qio_channel_rdma_source_funcs = {
3056 qio_channel_rdma_source_prepare,
3057 qio_channel_rdma_source_check,
3058 qio_channel_rdma_source_dispatch,
3059 qio_channel_rdma_source_finalize
3060 };
3061
qio_channel_rdma_create_watch(QIOChannel * ioc,GIOCondition condition)3062 static GSource *qio_channel_rdma_create_watch(QIOChannel *ioc,
3063 GIOCondition condition)
3064 {
3065 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
3066 QIOChannelRDMASource *ssource;
3067 GSource *source;
3068
3069 source = g_source_new(&qio_channel_rdma_source_funcs,
3070 sizeof(QIOChannelRDMASource));
3071 ssource = (QIOChannelRDMASource *)source;
3072
3073 ssource->rioc = rioc;
3074 object_ref(OBJECT(rioc));
3075
3076 ssource->condition = condition;
3077
3078 return source;
3079 }
3080
qio_channel_rdma_set_aio_fd_handler(QIOChannel * ioc,AioContext * read_ctx,IOHandler * io_read,AioContext * write_ctx,IOHandler * io_write,void * opaque)3081 static void qio_channel_rdma_set_aio_fd_handler(QIOChannel *ioc,
3082 AioContext *read_ctx,
3083 IOHandler *io_read,
3084 AioContext *write_ctx,
3085 IOHandler *io_write,
3086 void *opaque)
3087 {
3088 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
3089 if (io_read) {
3090 aio_set_fd_handler(read_ctx, rioc->rdmain->recv_comp_channel->fd,
3091 io_read, io_write, NULL, NULL, opaque);
3092 aio_set_fd_handler(read_ctx, rioc->rdmain->send_comp_channel->fd,
3093 io_read, io_write, NULL, NULL, opaque);
3094 } else {
3095 aio_set_fd_handler(write_ctx, rioc->rdmaout->recv_comp_channel->fd,
3096 io_read, io_write, NULL, NULL, opaque);
3097 aio_set_fd_handler(write_ctx, rioc->rdmaout->send_comp_channel->fd,
3098 io_read, io_write, NULL, NULL, opaque);
3099 }
3100 }
3101
3102 struct rdma_close_rcu {
3103 struct rcu_head rcu;
3104 RDMAContext *rdmain;
3105 RDMAContext *rdmaout;
3106 };
3107
3108 /* callback from qio_channel_rdma_close via call_rcu */
qio_channel_rdma_close_rcu(struct rdma_close_rcu * rcu)3109 static void qio_channel_rdma_close_rcu(struct rdma_close_rcu *rcu)
3110 {
3111 if (rcu->rdmain) {
3112 qemu_rdma_cleanup(rcu->rdmain);
3113 }
3114
3115 if (rcu->rdmaout) {
3116 qemu_rdma_cleanup(rcu->rdmaout);
3117 }
3118
3119 g_free(rcu->rdmain);
3120 g_free(rcu->rdmaout);
3121 g_free(rcu);
3122 }
3123
qio_channel_rdma_close(QIOChannel * ioc,Error ** errp)3124 static int qio_channel_rdma_close(QIOChannel *ioc,
3125 Error **errp)
3126 {
3127 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
3128 RDMAContext *rdmain, *rdmaout;
3129 struct rdma_close_rcu *rcu = g_new(struct rdma_close_rcu, 1);
3130
3131 trace_qemu_rdma_close();
3132
3133 rdmain = rioc->rdmain;
3134 if (rdmain) {
3135 qatomic_rcu_set(&rioc->rdmain, NULL);
3136 }
3137
3138 rdmaout = rioc->rdmaout;
3139 if (rdmaout) {
3140 qatomic_rcu_set(&rioc->rdmaout, NULL);
3141 }
3142
3143 rcu->rdmain = rdmain;
3144 rcu->rdmaout = rdmaout;
3145 call_rcu(rcu, qio_channel_rdma_close_rcu, rcu);
3146
3147 return 0;
3148 }
3149
3150 static int
qio_channel_rdma_shutdown(QIOChannel * ioc,QIOChannelShutdown how,Error ** errp)3151 qio_channel_rdma_shutdown(QIOChannel *ioc,
3152 QIOChannelShutdown how,
3153 Error **errp)
3154 {
3155 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc);
3156 RDMAContext *rdmain, *rdmaout;
3157
3158 RCU_READ_LOCK_GUARD();
3159
3160 rdmain = qatomic_rcu_read(&rioc->rdmain);
3161 rdmaout = qatomic_rcu_read(&rioc->rdmain);
3162
3163 switch (how) {
3164 case QIO_CHANNEL_SHUTDOWN_READ:
3165 if (rdmain) {
3166 rdmain->errored = true;
3167 }
3168 break;
3169 case QIO_CHANNEL_SHUTDOWN_WRITE:
3170 if (rdmaout) {
3171 rdmaout->errored = true;
3172 }
3173 break;
3174 case QIO_CHANNEL_SHUTDOWN_BOTH:
3175 default:
3176 if (rdmain) {
3177 rdmain->errored = true;
3178 }
3179 if (rdmaout) {
3180 rdmaout->errored = true;
3181 }
3182 break;
3183 }
3184
3185 return 0;
3186 }
3187
3188 /*
3189 * Parameters:
3190 * @offset == 0 :
3191 * This means that 'block_offset' is a full virtual address that does not
3192 * belong to a RAMBlock of the virtual machine and instead
3193 * represents a private malloc'd memory area that the caller wishes to
3194 * transfer.
3195 *
3196 * @offset != 0 :
3197 * Offset is an offset to be added to block_offset and used
3198 * to also lookup the corresponding RAMBlock.
3199 *
3200 * @size : Number of bytes to transfer
3201 *
3202 * @pages_sent : User-specificed pointer to indicate how many pages were
3203 * sent. Usually, this will not be more than a few bytes of
3204 * the protocol because most transfers are sent asynchronously.
3205 */
qemu_rdma_save_page(QEMUFile * f,ram_addr_t block_offset,ram_addr_t offset,size_t size)3206 static int qemu_rdma_save_page(QEMUFile *f, ram_addr_t block_offset,
3207 ram_addr_t offset, size_t size)
3208 {
3209 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(qemu_file_get_ioc(f));
3210 Error *err = NULL;
3211 RDMAContext *rdma;
3212 int ret;
3213
3214 RCU_READ_LOCK_GUARD();
3215 rdma = qatomic_rcu_read(&rioc->rdmaout);
3216
3217 if (!rdma) {
3218 return -1;
3219 }
3220
3221 if (rdma_errored(rdma)) {
3222 return -1;
3223 }
3224
3225 qemu_fflush(f);
3226
3227 /*
3228 * Add this page to the current 'chunk'. If the chunk
3229 * is full, or the page doesn't belong to the current chunk,
3230 * an actual RDMA write will occur and a new chunk will be formed.
3231 */
3232 ret = qemu_rdma_write(rdma, block_offset, offset, size, &err);
3233 if (ret < 0) {
3234 error_report_err(err);
3235 goto err;
3236 }
3237
3238 /*
3239 * Drain the Completion Queue if possible, but do not block,
3240 * just poll.
3241 *
3242 * If nothing to poll, the end of the iteration will do this
3243 * again to make sure we don't overflow the request queue.
3244 */
3245 while (1) {
3246 uint64_t wr_id, wr_id_in;
3247 ret = qemu_rdma_poll(rdma, rdma->recv_cq, &wr_id_in, NULL);
3248
3249 if (ret < 0) {
3250 error_report("rdma migration: polling error");
3251 goto err;
3252 }
3253
3254 wr_id = wr_id_in & RDMA_WRID_TYPE_MASK;
3255
3256 if (wr_id == RDMA_WRID_NONE) {
3257 break;
3258 }
3259 }
3260
3261 while (1) {
3262 uint64_t wr_id, wr_id_in;
3263 ret = qemu_rdma_poll(rdma, rdma->send_cq, &wr_id_in, NULL);
3264
3265 if (ret < 0) {
3266 error_report("rdma migration: polling error");
3267 goto err;
3268 }
3269
3270 wr_id = wr_id_in & RDMA_WRID_TYPE_MASK;
3271
3272 if (wr_id == RDMA_WRID_NONE) {
3273 break;
3274 }
3275 }
3276
3277 return RAM_SAVE_CONTROL_DELAYED;
3278
3279 err:
3280 rdma->errored = true;
3281 return -1;
3282 }
3283
rdma_control_save_page(QEMUFile * f,ram_addr_t block_offset,ram_addr_t offset,size_t size)3284 int rdma_control_save_page(QEMUFile *f, ram_addr_t block_offset,
3285 ram_addr_t offset, size_t size)
3286 {
3287 if (!migrate_rdma() || migration_in_postcopy()) {
3288 return RAM_SAVE_CONTROL_NOT_SUPP;
3289 }
3290
3291 int ret = qemu_rdma_save_page(f, block_offset, offset, size);
3292
3293 if (ret != RAM_SAVE_CONTROL_DELAYED &&
3294 ret != RAM_SAVE_CONTROL_NOT_SUPP) {
3295 if (ret < 0) {
3296 qemu_file_set_error(f, ret);
3297 }
3298 }
3299 return ret;
3300 }
3301
3302 static void rdma_accept_incoming_migration(void *opaque);
3303
rdma_cm_poll_handler(void * opaque)3304 static void rdma_cm_poll_handler(void *opaque)
3305 {
3306 RDMAContext *rdma = opaque;
3307 struct rdma_cm_event *cm_event;
3308 MigrationIncomingState *mis = migration_incoming_get_current();
3309
3310 if (rdma_get_cm_event(rdma->channel, &cm_event) < 0) {
3311 error_report("get_cm_event failed %d", errno);
3312 return;
3313 }
3314
3315 if (cm_event->event == RDMA_CM_EVENT_DISCONNECTED ||
3316 cm_event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) {
3317 if (!rdma->errored &&
3318 migration_incoming_get_current()->state !=
3319 MIGRATION_STATUS_COMPLETED) {
3320 error_report("receive cm event, cm event is %d", cm_event->event);
3321 rdma->errored = true;
3322 if (rdma->return_path) {
3323 rdma->return_path->errored = true;
3324 }
3325 }
3326 rdma_ack_cm_event(cm_event);
3327 if (mis->loadvm_co) {
3328 qemu_coroutine_enter(mis->loadvm_co);
3329 }
3330 return;
3331 }
3332 rdma_ack_cm_event(cm_event);
3333 }
3334
qemu_rdma_accept(RDMAContext * rdma)3335 static int qemu_rdma_accept(RDMAContext *rdma)
3336 {
3337 Error *err = NULL;
3338 RDMACapabilities cap;
3339 struct rdma_conn_param conn_param = {
3340 .responder_resources = 2,
3341 .private_data = &cap,
3342 .private_data_len = sizeof(cap),
3343 };
3344 RDMAContext *rdma_return_path = NULL;
3345 g_autoptr(InetSocketAddress) isock = g_new0(InetSocketAddress, 1);
3346 struct rdma_cm_event *cm_event;
3347 struct ibv_context *verbs;
3348 int ret;
3349
3350 ret = rdma_get_cm_event(rdma->channel, &cm_event);
3351 if (ret < 0) {
3352 goto err_rdma_dest_wait;
3353 }
3354
3355 if (cm_event->event != RDMA_CM_EVENT_CONNECT_REQUEST) {
3356 rdma_ack_cm_event(cm_event);
3357 goto err_rdma_dest_wait;
3358 }
3359
3360 isock->host = g_strdup(rdma->host);
3361 isock->port = g_strdup_printf("%d", rdma->port);
3362
3363 /*
3364 * initialize the RDMAContext for return path for postcopy after first
3365 * connection request reached.
3366 */
3367 if ((migrate_postcopy() || migrate_return_path())
3368 && !rdma->is_return_path) {
3369 rdma_return_path = qemu_rdma_data_init(isock, NULL);
3370 if (rdma_return_path == NULL) {
3371 rdma_ack_cm_event(cm_event);
3372 goto err_rdma_dest_wait;
3373 }
3374
3375 qemu_rdma_return_path_dest_init(rdma_return_path, rdma);
3376 }
3377
3378 memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap));
3379
3380 network_to_caps(&cap);
3381
3382 if (cap.version < 1 || cap.version > RDMA_CONTROL_VERSION_CURRENT) {
3383 error_report("Unknown source RDMA version: %d, bailing...",
3384 cap.version);
3385 rdma_ack_cm_event(cm_event);
3386 goto err_rdma_dest_wait;
3387 }
3388
3389 /*
3390 * Respond with only the capabilities this version of QEMU knows about.
3391 */
3392 cap.flags &= known_capabilities;
3393
3394 /*
3395 * Enable the ones that we do know about.
3396 * Add other checks here as new ones are introduced.
3397 */
3398 if (cap.flags & RDMA_CAPABILITY_PIN_ALL) {
3399 rdma->pin_all = true;
3400 }
3401
3402 rdma->cm_id = cm_event->id;
3403 verbs = cm_event->id->verbs;
3404
3405 rdma_ack_cm_event(cm_event);
3406
3407 trace_qemu_rdma_accept_pin_state(rdma->pin_all);
3408
3409 caps_to_network(&cap);
3410
3411 trace_qemu_rdma_accept_pin_verbsc(verbs);
3412
3413 if (!rdma->verbs) {
3414 rdma->verbs = verbs;
3415 } else if (rdma->verbs != verbs) {
3416 error_report("ibv context not matching %p, %p!", rdma->verbs,
3417 verbs);
3418 goto err_rdma_dest_wait;
3419 }
3420
3421 qemu_rdma_dump_id("dest_init", verbs);
3422
3423 ret = qemu_rdma_alloc_pd_cq(rdma, &err);
3424 if (ret < 0) {
3425 error_report_err(err);
3426 goto err_rdma_dest_wait;
3427 }
3428
3429 ret = qemu_rdma_alloc_qp(rdma);
3430 if (ret < 0) {
3431 error_report("rdma migration: error allocating qp!");
3432 goto err_rdma_dest_wait;
3433 }
3434
3435 qemu_rdma_init_ram_blocks(rdma);
3436
3437 for (int i = 0; i < RDMA_WRID_MAX; i++) {
3438 ret = qemu_rdma_reg_control(rdma, i);
3439 if (ret < 0) {
3440 error_report("rdma: error registering %d control", i);
3441 goto err_rdma_dest_wait;
3442 }
3443 }
3444
3445 /* Accept the second connection request for return path */
3446 if ((migrate_postcopy() || migrate_return_path())
3447 && !rdma->is_return_path) {
3448 qemu_set_fd_handler(rdma->channel->fd, rdma_accept_incoming_migration,
3449 NULL,
3450 (void *)(intptr_t)rdma->return_path);
3451 } else {
3452 qemu_set_fd_handler(rdma->channel->fd, rdma_cm_poll_handler,
3453 NULL, rdma);
3454 }
3455
3456 ret = rdma_accept(rdma->cm_id, &conn_param);
3457 if (ret < 0) {
3458 error_report("rdma_accept failed");
3459 goto err_rdma_dest_wait;
3460 }
3461
3462 ret = rdma_get_cm_event(rdma->channel, &cm_event);
3463 if (ret < 0) {
3464 error_report("rdma_accept get_cm_event failed");
3465 goto err_rdma_dest_wait;
3466 }
3467
3468 if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) {
3469 error_report("rdma_accept not event established");
3470 rdma_ack_cm_event(cm_event);
3471 goto err_rdma_dest_wait;
3472 }
3473
3474 rdma_ack_cm_event(cm_event);
3475 rdma->connected = true;
3476
3477 ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY, &err);
3478 if (ret < 0) {
3479 error_report_err(err);
3480 goto err_rdma_dest_wait;
3481 }
3482
3483 qemu_rdma_dump_gid("dest_connect", rdma->cm_id);
3484
3485 return 0;
3486
3487 err_rdma_dest_wait:
3488 rdma->errored = true;
3489 qemu_rdma_cleanup(rdma);
3490 g_free(rdma_return_path);
3491 return -1;
3492 }
3493
dest_ram_sort_func(const void * a,const void * b)3494 static int dest_ram_sort_func(const void *a, const void *b)
3495 {
3496 unsigned int a_index = ((const RDMALocalBlock *)a)->src_index;
3497 unsigned int b_index = ((const RDMALocalBlock *)b)->src_index;
3498
3499 return (a_index < b_index) ? -1 : (a_index != b_index);
3500 }
3501
3502 /*
3503 * During each iteration of the migration, we listen for instructions
3504 * by the source VM to perform dynamic page registrations before they
3505 * can perform RDMA operations.
3506 *
3507 * We respond with the 'rkey'.
3508 *
3509 * Keep doing this until the source tells us to stop.
3510 */
rdma_registration_handle(QEMUFile * f)3511 int rdma_registration_handle(QEMUFile *f)
3512 {
3513 RDMAControlHeader reg_resp = { .len = sizeof(RDMARegisterResult),
3514 .type = RDMA_CONTROL_REGISTER_RESULT,
3515 .repeat = 0,
3516 };
3517 RDMAControlHeader unreg_resp = { .len = 0,
3518 .type = RDMA_CONTROL_UNREGISTER_FINISHED,
3519 .repeat = 0,
3520 };
3521 RDMAControlHeader blocks = { .type = RDMA_CONTROL_RAM_BLOCKS_RESULT,
3522 .repeat = 1 };
3523 QIOChannelRDMA *rioc;
3524 Error *err = NULL;
3525 RDMAContext *rdma;
3526 RDMALocalBlocks *local;
3527 RDMAControlHeader head;
3528 RDMARegister *reg, *registers;
3529 RDMACompress *comp;
3530 RDMARegisterResult *reg_result;
3531 static RDMARegisterResult results[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE];
3532 RDMALocalBlock *block;
3533 void *host_addr;
3534 int ret;
3535 int idx = 0;
3536
3537 if (!migrate_rdma()) {
3538 return 0;
3539 }
3540
3541 RCU_READ_LOCK_GUARD();
3542 rioc = QIO_CHANNEL_RDMA(qemu_file_get_ioc(f));
3543 rdma = qatomic_rcu_read(&rioc->rdmain);
3544
3545 if (!rdma) {
3546 return -1;
3547 }
3548
3549 if (rdma_errored(rdma)) {
3550 return -1;
3551 }
3552
3553 local = &rdma->local_ram_blocks;
3554 do {
3555 trace_rdma_registration_handle_wait();
3556
3557 ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_NONE, &err);
3558
3559 if (ret < 0) {
3560 error_report_err(err);
3561 break;
3562 }
3563
3564 if (head.repeat > RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE) {
3565 error_report("rdma: Too many requests in this message (%d)."
3566 "Bailing.", head.repeat);
3567 break;
3568 }
3569
3570 switch (head.type) {
3571 case RDMA_CONTROL_COMPRESS:
3572 comp = (RDMACompress *) rdma->wr_data[idx].control_curr;
3573 network_to_compress(comp);
3574
3575 trace_rdma_registration_handle_compress(comp->length,
3576 comp->block_idx,
3577 comp->offset);
3578 if (comp->block_idx >= rdma->local_ram_blocks.nb_blocks) {
3579 error_report("rdma: 'compress' bad block index %u (vs %d)",
3580 (unsigned int)comp->block_idx,
3581 rdma->local_ram_blocks.nb_blocks);
3582 goto err;
3583 }
3584 block = &(rdma->local_ram_blocks.block[comp->block_idx]);
3585
3586 host_addr = block->local_host_addr +
3587 (comp->offset - block->offset);
3588 if (comp->value) {
3589 error_report("rdma: Zero page with non-zero (%d) value",
3590 comp->value);
3591 goto err;
3592 }
3593 ram_handle_zero(host_addr, comp->length);
3594 break;
3595
3596 case RDMA_CONTROL_REGISTER_FINISHED:
3597 trace_rdma_registration_handle_finished();
3598 return 0;
3599
3600 case RDMA_CONTROL_RAM_BLOCKS_REQUEST:
3601 trace_rdma_registration_handle_ram_blocks();
3602
3603 /* Sort our local RAM Block list so it's the same as the source,
3604 * we can do this since we've filled in a src_index in the list
3605 * as we received the RAMBlock list earlier.
3606 */
3607 qsort(rdma->local_ram_blocks.block,
3608 rdma->local_ram_blocks.nb_blocks,
3609 sizeof(RDMALocalBlock), dest_ram_sort_func);
3610 for (int i = 0; i < local->nb_blocks; i++) {
3611 local->block[i].index = i;
3612 }
3613
3614 if (rdma->pin_all) {
3615 ret = qemu_rdma_reg_whole_ram_blocks(rdma, &err);
3616 if (ret < 0) {
3617 error_report_err(err);
3618 goto err;
3619 }
3620 }
3621
3622 /*
3623 * Dest uses this to prepare to transmit the RAMBlock descriptions
3624 * to the source VM after connection setup.
3625 * Both sides use the "remote" structure to communicate and update
3626 * their "local" descriptions with what was sent.
3627 */
3628 for (int i = 0; i < local->nb_blocks; i++) {
3629 rdma->dest_blocks[i].remote_host_addr =
3630 (uintptr_t)(local->block[i].local_host_addr);
3631
3632 if (rdma->pin_all) {
3633 rdma->dest_blocks[i].remote_rkey = local->block[i].mr->rkey;
3634 }
3635
3636 rdma->dest_blocks[i].offset = local->block[i].offset;
3637 rdma->dest_blocks[i].length = local->block[i].length;
3638
3639 dest_block_to_network(&rdma->dest_blocks[i]);
3640 trace_rdma_registration_handle_ram_blocks_loop(
3641 local->block[i].block_name,
3642 local->block[i].offset,
3643 local->block[i].length,
3644 local->block[i].local_host_addr,
3645 local->block[i].src_index);
3646 }
3647
3648 blocks.len = rdma->local_ram_blocks.nb_blocks
3649 * sizeof(RDMADestBlock);
3650
3651
3652 ret = qemu_rdma_post_send_control(rdma,
3653 (uint8_t *) rdma->dest_blocks, &blocks,
3654 &err);
3655
3656 if (ret < 0) {
3657 error_report_err(err);
3658 goto err;
3659 }
3660
3661 break;
3662 case RDMA_CONTROL_REGISTER_REQUEST:
3663 trace_rdma_registration_handle_register(head.repeat);
3664
3665 reg_resp.repeat = head.repeat;
3666 registers = (RDMARegister *) rdma->wr_data[idx].control_curr;
3667
3668 for (int count = 0; count < head.repeat; count++) {
3669 uint64_t chunk;
3670 uint8_t *chunk_start, *chunk_end;
3671
3672 reg = ®isters[count];
3673 network_to_register(reg);
3674
3675 reg_result = &results[count];
3676
3677 trace_rdma_registration_handle_register_loop(count,
3678 reg->current_index, reg->key.current_addr, reg->chunks);
3679
3680 if (reg->current_index >= rdma->local_ram_blocks.nb_blocks) {
3681 error_report("rdma: 'register' bad block index %u (vs %d)",
3682 (unsigned int)reg->current_index,
3683 rdma->local_ram_blocks.nb_blocks);
3684 goto err;
3685 }
3686 block = &(rdma->local_ram_blocks.block[reg->current_index]);
3687 if (block->is_ram_block) {
3688 if (block->offset > reg->key.current_addr) {
3689 error_report("rdma: bad register address for block %s"
3690 " offset: %" PRIx64 " current_addr: %" PRIx64,
3691 block->block_name, block->offset,
3692 reg->key.current_addr);
3693 goto err;
3694 }
3695 host_addr = (block->local_host_addr +
3696 (reg->key.current_addr - block->offset));
3697 chunk = ram_chunk_index(block->local_host_addr,
3698 (uint8_t *) host_addr);
3699 } else {
3700 chunk = reg->key.chunk;
3701 host_addr = block->local_host_addr +
3702 (reg->key.chunk * (1UL << RDMA_REG_CHUNK_SHIFT));
3703 /* Check for particularly bad chunk value */
3704 if (host_addr < (void *)block->local_host_addr) {
3705 error_report("rdma: bad chunk for block %s"
3706 " chunk: %" PRIx64,
3707 block->block_name, reg->key.chunk);
3708 goto err;
3709 }
3710 }
3711 chunk_start = ram_chunk_start(block, chunk);
3712 chunk_end = ram_chunk_end(block, chunk + reg->chunks);
3713 /* avoid "-Waddress-of-packed-member" warning */
3714 uint32_t tmp_rkey = 0;
3715 if (qemu_rdma_register_and_get_keys(rdma, block,
3716 (uintptr_t)host_addr, NULL, &tmp_rkey,
3717 chunk, chunk_start, chunk_end)) {
3718 error_report("cannot get rkey");
3719 goto err;
3720 }
3721 reg_result->rkey = tmp_rkey;
3722
3723 reg_result->host_addr = (uintptr_t)block->local_host_addr;
3724
3725 trace_rdma_registration_handle_register_rkey(reg_result->rkey);
3726
3727 result_to_network(reg_result);
3728 }
3729
3730 ret = qemu_rdma_post_send_control(rdma,
3731 (uint8_t *) results, ®_resp, &err);
3732
3733 if (ret < 0) {
3734 error_report_err(err);
3735 goto err;
3736 }
3737 break;
3738 case RDMA_CONTROL_UNREGISTER_REQUEST:
3739 trace_rdma_registration_handle_unregister(head.repeat);
3740 unreg_resp.repeat = head.repeat;
3741 registers = (RDMARegister *) rdma->wr_data[idx].control_curr;
3742
3743 for (int count = 0; count < head.repeat; count++) {
3744 reg = ®isters[count];
3745 network_to_register(reg);
3746
3747 trace_rdma_registration_handle_unregister_loop(count,
3748 reg->current_index, reg->key.chunk);
3749
3750 block = &(rdma->local_ram_blocks.block[reg->current_index]);
3751
3752 ret = ibv_dereg_mr(block->pmr[reg->key.chunk]);
3753 block->pmr[reg->key.chunk] = NULL;
3754
3755 if (ret != 0) {
3756 error_report("rdma unregistration chunk failed: %s",
3757 strerror(errno));
3758 goto err;
3759 }
3760
3761 rdma->total_registrations--;
3762
3763 trace_rdma_registration_handle_unregister_success(reg->key.chunk);
3764 }
3765
3766 ret = qemu_rdma_post_send_control(rdma, NULL, &unreg_resp, &err);
3767
3768 if (ret < 0) {
3769 error_report_err(err);
3770 goto err;
3771 }
3772 break;
3773 case RDMA_CONTROL_REGISTER_RESULT:
3774 error_report("Invalid RESULT message at dest.");
3775 goto err;
3776 default:
3777 error_report("Unknown control message %s", control_desc(head.type));
3778 goto err;
3779 }
3780 } while (1);
3781
3782 err:
3783 rdma->errored = true;
3784 return -1;
3785 }
3786
3787 /* Destination:
3788 * Called during the initial RAM load section which lists the
3789 * RAMBlocks by name. This lets us know the order of the RAMBlocks on
3790 * the source. We've already built our local RAMBlock list, but not
3791 * yet sent the list to the source.
3792 */
rdma_block_notification_handle(QEMUFile * f,const char * name)3793 int rdma_block_notification_handle(QEMUFile *f, const char *name)
3794 {
3795 int curr;
3796 int found = -1;
3797
3798 if (!migrate_rdma()) {
3799 return 0;
3800 }
3801
3802 RCU_READ_LOCK_GUARD();
3803 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(qemu_file_get_ioc(f));
3804 RDMAContext *rdma = qatomic_rcu_read(&rioc->rdmain);
3805
3806 if (!rdma) {
3807 return -1;
3808 }
3809
3810 /* Find the matching RAMBlock in our local list */
3811 for (curr = 0; curr < rdma->local_ram_blocks.nb_blocks; curr++) {
3812 if (!strcmp(rdma->local_ram_blocks.block[curr].block_name, name)) {
3813 found = curr;
3814 break;
3815 }
3816 }
3817
3818 if (found == -1) {
3819 error_report("RAMBlock '%s' not found on destination", name);
3820 return -1;
3821 }
3822
3823 rdma->local_ram_blocks.block[curr].src_index = rdma->next_src_index;
3824 trace_rdma_block_notification_handle(name, rdma->next_src_index);
3825 rdma->next_src_index++;
3826
3827 return 0;
3828 }
3829
rdma_registration_start(QEMUFile * f,uint64_t flags)3830 int rdma_registration_start(QEMUFile *f, uint64_t flags)
3831 {
3832 if (!migrate_rdma() || migration_in_postcopy()) {
3833 return 0;
3834 }
3835
3836 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(qemu_file_get_ioc(f));
3837 RCU_READ_LOCK_GUARD();
3838 RDMAContext *rdma = qatomic_rcu_read(&rioc->rdmaout);
3839 if (!rdma) {
3840 return -1;
3841 }
3842
3843 if (rdma_errored(rdma)) {
3844 return -1;
3845 }
3846
3847 trace_rdma_registration_start(flags);
3848 qemu_put_be64(f, RAM_SAVE_FLAG_HOOK);
3849 return qemu_fflush(f);
3850 }
3851
3852 /*
3853 * Inform dest that dynamic registrations are done for now.
3854 * First, flush writes, if any.
3855 */
rdma_registration_stop(QEMUFile * f,uint64_t flags)3856 int rdma_registration_stop(QEMUFile *f, uint64_t flags)
3857 {
3858 QIOChannelRDMA *rioc;
3859 Error *err = NULL;
3860 RDMAContext *rdma;
3861 RDMAControlHeader head = { .len = 0, .repeat = 1 };
3862 int ret;
3863
3864 if (!migrate_rdma() || migration_in_postcopy()) {
3865 return 0;
3866 }
3867
3868 RCU_READ_LOCK_GUARD();
3869 rioc = QIO_CHANNEL_RDMA(qemu_file_get_ioc(f));
3870 rdma = qatomic_rcu_read(&rioc->rdmaout);
3871 if (!rdma) {
3872 return -1;
3873 }
3874
3875 if (rdma_errored(rdma)) {
3876 return -1;
3877 }
3878
3879 qemu_fflush(f);
3880 ret = qemu_rdma_drain_cq(rdma);
3881
3882 if (ret < 0) {
3883 goto err;
3884 }
3885
3886 if (flags == RAM_CONTROL_SETUP) {
3887 RDMAControlHeader resp = {.type = RDMA_CONTROL_RAM_BLOCKS_RESULT };
3888 RDMALocalBlocks *local = &rdma->local_ram_blocks;
3889 int reg_result_idx, nb_dest_blocks;
3890
3891 head.type = RDMA_CONTROL_RAM_BLOCKS_REQUEST;
3892 trace_rdma_registration_stop_ram();
3893
3894 /*
3895 * Make sure that we parallelize the pinning on both sides.
3896 * For very large guests, doing this serially takes a really
3897 * long time, so we have to 'interleave' the pinning locally
3898 * with the control messages by performing the pinning on this
3899 * side before we receive the control response from the other
3900 * side that the pinning has completed.
3901 */
3902 ret = qemu_rdma_exchange_send(rdma, &head, NULL, &resp,
3903 ®_result_idx, rdma->pin_all ?
3904 qemu_rdma_reg_whole_ram_blocks : NULL,
3905 &err);
3906 if (ret < 0) {
3907 error_report_err(err);
3908 return -1;
3909 }
3910
3911 nb_dest_blocks = resp.len / sizeof(RDMADestBlock);
3912
3913 /*
3914 * The protocol uses two different sets of rkeys (mutually exclusive):
3915 * 1. One key to represent the virtual address of the entire ram block.
3916 * (dynamic chunk registration disabled - pin everything with one rkey.)
3917 * 2. One to represent individual chunks within a ram block.
3918 * (dynamic chunk registration enabled - pin individual chunks.)
3919 *
3920 * Once the capability is successfully negotiated, the destination transmits
3921 * the keys to use (or sends them later) including the virtual addresses
3922 * and then propagates the remote ram block descriptions to his local copy.
3923 */
3924
3925 if (local->nb_blocks != nb_dest_blocks) {
3926 error_report("ram blocks mismatch (Number of blocks %d vs %d)",
3927 local->nb_blocks, nb_dest_blocks);
3928 error_printf("Your QEMU command line parameters are probably "
3929 "not identical on both the source and destination.");
3930 rdma->errored = true;
3931 return -1;
3932 }
3933
3934 qemu_rdma_move_header(rdma, reg_result_idx, &resp);
3935 memcpy(rdma->dest_blocks,
3936 rdma->wr_data[reg_result_idx].control_curr, resp.len);
3937 for (int i = 0; i < nb_dest_blocks; i++) {
3938 network_to_dest_block(&rdma->dest_blocks[i]);
3939
3940 /* We require that the blocks are in the same order */
3941 if (rdma->dest_blocks[i].length != local->block[i].length) {
3942 error_report("Block %s/%d has a different length %" PRIu64
3943 "vs %" PRIu64,
3944 local->block[i].block_name, i,
3945 local->block[i].length,
3946 rdma->dest_blocks[i].length);
3947 rdma->errored = true;
3948 return -1;
3949 }
3950 local->block[i].remote_host_addr =
3951 rdma->dest_blocks[i].remote_host_addr;
3952 local->block[i].remote_rkey = rdma->dest_blocks[i].remote_rkey;
3953 }
3954 }
3955
3956 trace_rdma_registration_stop(flags);
3957
3958 head.type = RDMA_CONTROL_REGISTER_FINISHED;
3959 ret = qemu_rdma_exchange_send(rdma, &head, NULL, NULL, NULL, NULL, &err);
3960
3961 if (ret < 0) {
3962 error_report_err(err);
3963 goto err;
3964 }
3965
3966 return 0;
3967 err:
3968 rdma->errored = true;
3969 return -1;
3970 }
3971
qio_channel_rdma_finalize(Object * obj)3972 static void qio_channel_rdma_finalize(Object *obj)
3973 {
3974 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(obj);
3975 if (rioc->rdmain) {
3976 qemu_rdma_cleanup(rioc->rdmain);
3977 g_free(rioc->rdmain);
3978 rioc->rdmain = NULL;
3979 }
3980 if (rioc->rdmaout) {
3981 qemu_rdma_cleanup(rioc->rdmaout);
3982 g_free(rioc->rdmaout);
3983 rioc->rdmaout = NULL;
3984 }
3985 }
3986
qio_channel_rdma_class_init(ObjectClass * klass,void * class_data G_GNUC_UNUSED)3987 static void qio_channel_rdma_class_init(ObjectClass *klass,
3988 void *class_data G_GNUC_UNUSED)
3989 {
3990 QIOChannelClass *ioc_klass = QIO_CHANNEL_CLASS(klass);
3991
3992 ioc_klass->io_writev = qio_channel_rdma_writev;
3993 ioc_klass->io_readv = qio_channel_rdma_readv;
3994 ioc_klass->io_set_blocking = qio_channel_rdma_set_blocking;
3995 ioc_klass->io_close = qio_channel_rdma_close;
3996 ioc_klass->io_create_watch = qio_channel_rdma_create_watch;
3997 ioc_klass->io_set_aio_fd_handler = qio_channel_rdma_set_aio_fd_handler;
3998 ioc_klass->io_shutdown = qio_channel_rdma_shutdown;
3999 }
4000
4001 static const TypeInfo qio_channel_rdma_info = {
4002 .parent = TYPE_QIO_CHANNEL,
4003 .name = TYPE_QIO_CHANNEL_RDMA,
4004 .instance_size = sizeof(QIOChannelRDMA),
4005 .instance_finalize = qio_channel_rdma_finalize,
4006 .class_init = qio_channel_rdma_class_init,
4007 };
4008
qio_channel_rdma_register_types(void)4009 static void qio_channel_rdma_register_types(void)
4010 {
4011 type_register_static(&qio_channel_rdma_info);
4012 }
4013
4014 type_init(qio_channel_rdma_register_types);
4015
rdma_new_input(RDMAContext * rdma)4016 static QEMUFile *rdma_new_input(RDMAContext *rdma)
4017 {
4018 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA));
4019
4020 rioc->file = qemu_file_new_input(QIO_CHANNEL(rioc));
4021 rioc->rdmain = rdma;
4022 rioc->rdmaout = rdma->return_path;
4023
4024 return rioc->file;
4025 }
4026
rdma_new_output(RDMAContext * rdma)4027 static QEMUFile *rdma_new_output(RDMAContext *rdma)
4028 {
4029 QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA));
4030
4031 rioc->file = qemu_file_new_output(QIO_CHANNEL(rioc));
4032 rioc->rdmaout = rdma;
4033 rioc->rdmain = rdma->return_path;
4034
4035 return rioc->file;
4036 }
4037
rdma_accept_incoming_migration(void * opaque)4038 static void rdma_accept_incoming_migration(void *opaque)
4039 {
4040 RDMAContext *rdma = opaque;
4041 QEMUFile *f;
4042
4043 trace_qemu_rdma_accept_incoming_migration();
4044 if (qemu_rdma_accept(rdma) < 0) {
4045 error_report("RDMA ERROR: Migration initialization failed");
4046 return;
4047 }
4048
4049 trace_qemu_rdma_accept_incoming_migration_accepted();
4050
4051 if (rdma->is_return_path) {
4052 return;
4053 }
4054
4055 f = rdma_new_input(rdma);
4056 if (f == NULL) {
4057 error_report("RDMA ERROR: could not open RDMA for input");
4058 qemu_rdma_cleanup(rdma);
4059 return;
4060 }
4061
4062 rdma->migration_started_on_destination = 1;
4063 migration_fd_process_incoming(f);
4064 }
4065
rdma_start_incoming_migration(InetSocketAddress * host_port,Error ** errp)4066 void rdma_start_incoming_migration(InetSocketAddress *host_port,
4067 Error **errp)
4068 {
4069 MigrationState *s = migrate_get_current();
4070 int ret;
4071 RDMAContext *rdma;
4072
4073 trace_rdma_start_incoming_migration();
4074
4075 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4076 if (ram_block_discard_is_required()) {
4077 error_setg(errp, "RDMA: cannot disable RAM discard");
4078 return;
4079 }
4080
4081 rdma = qemu_rdma_data_init(host_port, errp);
4082 if (rdma == NULL) {
4083 goto err;
4084 }
4085
4086 ret = qemu_rdma_dest_init(rdma, errp);
4087 if (ret < 0) {
4088 goto err;
4089 }
4090
4091 trace_rdma_start_incoming_migration_after_dest_init();
4092
4093 ret = rdma_listen(rdma->listen_id, 5);
4094
4095 if (ret < 0) {
4096 error_setg(errp, "RDMA ERROR: listening on socket!");
4097 goto cleanup_rdma;
4098 }
4099
4100 trace_rdma_start_incoming_migration_after_rdma_listen();
4101 s->rdma_migration = true;
4102 qemu_set_fd_handler(rdma->channel->fd, rdma_accept_incoming_migration,
4103 NULL, (void *)(intptr_t)rdma);
4104 return;
4105
4106 cleanup_rdma:
4107 qemu_rdma_cleanup(rdma);
4108 err:
4109 if (rdma) {
4110 g_free(rdma->host);
4111 }
4112 g_free(rdma);
4113 }
4114
rdma_start_outgoing_migration(void * opaque,InetSocketAddress * host_port,Error ** errp)4115 void rdma_start_outgoing_migration(void *opaque,
4116 InetSocketAddress *host_port, Error **errp)
4117 {
4118 MigrationState *s = opaque;
4119 RDMAContext *rdma_return_path = NULL;
4120 RDMAContext *rdma;
4121 int ret;
4122
4123 /* Avoid ram_block_discard_disable(), cannot change during migration. */
4124 if (ram_block_discard_is_required()) {
4125 error_setg(errp, "RDMA: cannot disable RAM discard");
4126 return;
4127 }
4128
4129 rdma = qemu_rdma_data_init(host_port, errp);
4130 if (rdma == NULL) {
4131 goto err;
4132 }
4133
4134 ret = qemu_rdma_source_init(rdma, migrate_rdma_pin_all(), errp);
4135
4136 if (ret < 0) {
4137 goto err;
4138 }
4139
4140 trace_rdma_start_outgoing_migration_after_rdma_source_init();
4141 ret = qemu_rdma_connect(rdma, false, errp);
4142
4143 if (ret < 0) {
4144 goto err;
4145 }
4146
4147 /* RDMA postcopy need a separate queue pair for return path */
4148 if (migrate_postcopy() || migrate_return_path()) {
4149 rdma_return_path = qemu_rdma_data_init(host_port, errp);
4150
4151 if (rdma_return_path == NULL) {
4152 goto return_path_err;
4153 }
4154
4155 ret = qemu_rdma_source_init(rdma_return_path,
4156 migrate_rdma_pin_all(), errp);
4157
4158 if (ret < 0) {
4159 goto return_path_err;
4160 }
4161
4162 ret = qemu_rdma_connect(rdma_return_path, true, errp);
4163
4164 if (ret < 0) {
4165 goto return_path_err;
4166 }
4167
4168 rdma->return_path = rdma_return_path;
4169 rdma_return_path->return_path = rdma;
4170 rdma_return_path->is_return_path = true;
4171 }
4172
4173 trace_rdma_start_outgoing_migration_after_rdma_connect();
4174
4175 s->to_dst_file = rdma_new_output(rdma);
4176 s->rdma_migration = true;
4177 migrate_fd_connect(s, NULL);
4178 return;
4179 return_path_err:
4180 qemu_rdma_cleanup(rdma);
4181 err:
4182 g_free(rdma);
4183 g_free(rdma_return_path);
4184 }
4185