]>
Commit | Line | Data |
---|---|---|
2da776db MH |
1 | /* |
2 | * RDMA protocol and interfaces | |
3 | * | |
4 | * Copyright IBM, Corp. 2010-2013 | |
5 | * | |
6 | * Authors: | |
7 | * Michael R. Hines <mrhines@us.ibm.com> | |
8 | * Jiuxing Liu <jl@us.ibm.com> | |
9 | * | |
10 | * This work is licensed under the terms of the GNU GPL, version 2 or | |
11 | * later. See the COPYING file in the top-level directory. | |
12 | * | |
13 | */ | |
14 | #include "qemu-common.h" | |
15 | #include "migration/migration.h" | |
16 | #include "migration/qemu-file.h" | |
17 | #include "exec/cpu-common.h" | |
d49b6836 | 18 | #include "qemu/error-report.h" |
2da776db MH |
19 | #include "qemu/main-loop.h" |
20 | #include "qemu/sockets.h" | |
21 | #include "qemu/bitmap.h" | |
22 | #include "block/coroutine.h" | |
23 | #include <stdio.h> | |
24 | #include <sys/types.h> | |
25 | #include <sys/socket.h> | |
26 | #include <netdb.h> | |
27 | #include <arpa/inet.h> | |
28 | #include <string.h> | |
29 | #include <rdma/rdma_cma.h> | |
733252de | 30 | #include "trace.h" |
2da776db MH |
31 | |
32 | /* | |
33 | * Print and error on both the Monitor and the Log file. | |
34 | */ | |
35 | #define ERROR(errp, fmt, ...) \ | |
36 | do { \ | |
66988941 | 37 | fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \ |
2da776db MH |
38 | if (errp && (*(errp) == NULL)) { \ |
39 | error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \ | |
40 | } \ | |
41 | } while (0) | |
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 | #define CHECK_ERROR_STATE() \ | |
78 | do { \ | |
79 | if (rdma->error_state) { \ | |
80 | if (!rdma->error_reported) { \ | |
733252de DDAG |
81 | error_report("RDMA is in an error state waiting migration" \ |
82 | " to abort!"); \ | |
2da776db MH |
83 | rdma->error_reported = 1; \ |
84 | } \ | |
85 | return rdma->error_state; \ | |
86 | } \ | |
87 | } while (0); | |
88 | ||
89 | /* | |
90 | * A work request ID is 64-bits and we split up these bits | |
91 | * into 3 parts: | |
92 | * | |
93 | * bits 0-15 : type of control message, 2^16 | |
94 | * bits 16-29: ram block index, 2^14 | |
95 | * bits 30-63: ram block chunk number, 2^34 | |
96 | * | |
97 | * The last two bit ranges are only used for RDMA writes, | |
98 | * in order to track their completion and potentially | |
99 | * also track unregistration status of the message. | |
100 | */ | |
101 | #define RDMA_WRID_TYPE_SHIFT 0UL | |
102 | #define RDMA_WRID_BLOCK_SHIFT 16UL | |
103 | #define RDMA_WRID_CHUNK_SHIFT 30UL | |
104 | ||
105 | #define RDMA_WRID_TYPE_MASK \ | |
106 | ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL) | |
107 | ||
108 | #define RDMA_WRID_BLOCK_MASK \ | |
109 | (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL)) | |
110 | ||
111 | #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK) | |
112 | ||
113 | /* | |
114 | * RDMA migration protocol: | |
115 | * 1. RDMA Writes (data messages, i.e. RAM) | |
116 | * 2. IB Send/Recv (control channel messages) | |
117 | */ | |
118 | enum { | |
119 | RDMA_WRID_NONE = 0, | |
120 | RDMA_WRID_RDMA_WRITE = 1, | |
121 | RDMA_WRID_SEND_CONTROL = 2000, | |
122 | RDMA_WRID_RECV_CONTROL = 4000, | |
123 | }; | |
124 | ||
2ae31aea | 125 | static const char *wrid_desc[] = { |
2da776db MH |
126 | [RDMA_WRID_NONE] = "NONE", |
127 | [RDMA_WRID_RDMA_WRITE] = "WRITE RDMA", | |
128 | [RDMA_WRID_SEND_CONTROL] = "CONTROL SEND", | |
129 | [RDMA_WRID_RECV_CONTROL] = "CONTROL RECV", | |
130 | }; | |
131 | ||
132 | /* | |
133 | * Work request IDs for IB SEND messages only (not RDMA writes). | |
134 | * This is used by the migration protocol to transmit | |
135 | * control messages (such as device state and registration commands) | |
136 | * | |
137 | * We could use more WRs, but we have enough for now. | |
138 | */ | |
139 | enum { | |
140 | RDMA_WRID_READY = 0, | |
141 | RDMA_WRID_DATA, | |
142 | RDMA_WRID_CONTROL, | |
143 | RDMA_WRID_MAX, | |
144 | }; | |
145 | ||
146 | /* | |
147 | * SEND/RECV IB Control Messages. | |
148 | */ | |
149 | enum { | |
150 | RDMA_CONTROL_NONE = 0, | |
151 | RDMA_CONTROL_ERROR, | |
152 | RDMA_CONTROL_READY, /* ready to receive */ | |
153 | RDMA_CONTROL_QEMU_FILE, /* QEMUFile-transmitted bytes */ | |
154 | RDMA_CONTROL_RAM_BLOCKS_REQUEST, /* RAMBlock synchronization */ | |
155 | RDMA_CONTROL_RAM_BLOCKS_RESULT, /* RAMBlock synchronization */ | |
156 | RDMA_CONTROL_COMPRESS, /* page contains repeat values */ | |
157 | RDMA_CONTROL_REGISTER_REQUEST, /* dynamic page registration */ | |
158 | RDMA_CONTROL_REGISTER_RESULT, /* key to use after registration */ | |
159 | RDMA_CONTROL_REGISTER_FINISHED, /* current iteration finished */ | |
160 | RDMA_CONTROL_UNREGISTER_REQUEST, /* dynamic UN-registration */ | |
161 | RDMA_CONTROL_UNREGISTER_FINISHED, /* unpinning finished */ | |
162 | }; | |
163 | ||
2ae31aea | 164 | static const char *control_desc[] = { |
2da776db MH |
165 | [RDMA_CONTROL_NONE] = "NONE", |
166 | [RDMA_CONTROL_ERROR] = "ERROR", | |
167 | [RDMA_CONTROL_READY] = "READY", | |
168 | [RDMA_CONTROL_QEMU_FILE] = "QEMU FILE", | |
169 | [RDMA_CONTROL_RAM_BLOCKS_REQUEST] = "RAM BLOCKS REQUEST", | |
170 | [RDMA_CONTROL_RAM_BLOCKS_RESULT] = "RAM BLOCKS RESULT", | |
171 | [RDMA_CONTROL_COMPRESS] = "COMPRESS", | |
172 | [RDMA_CONTROL_REGISTER_REQUEST] = "REGISTER REQUEST", | |
173 | [RDMA_CONTROL_REGISTER_RESULT] = "REGISTER RESULT", | |
174 | [RDMA_CONTROL_REGISTER_FINISHED] = "REGISTER FINISHED", | |
175 | [RDMA_CONTROL_UNREGISTER_REQUEST] = "UNREGISTER REQUEST", | |
176 | [RDMA_CONTROL_UNREGISTER_FINISHED] = "UNREGISTER FINISHED", | |
177 | }; | |
178 | ||
179 | /* | |
180 | * Memory and MR structures used to represent an IB Send/Recv work request. | |
181 | * This is *not* used for RDMA writes, only IB Send/Recv. | |
182 | */ | |
183 | typedef struct { | |
184 | uint8_t control[RDMA_CONTROL_MAX_BUFFER]; /* actual buffer to register */ | |
185 | struct ibv_mr *control_mr; /* registration metadata */ | |
186 | size_t control_len; /* length of the message */ | |
187 | uint8_t *control_curr; /* start of unconsumed bytes */ | |
188 | } RDMAWorkRequestData; | |
189 | ||
190 | /* | |
191 | * Negotiate RDMA capabilities during connection-setup time. | |
192 | */ | |
193 | typedef struct { | |
194 | uint32_t version; | |
195 | uint32_t flags; | |
196 | } RDMACapabilities; | |
197 | ||
198 | static void caps_to_network(RDMACapabilities *cap) | |
199 | { | |
200 | cap->version = htonl(cap->version); | |
201 | cap->flags = htonl(cap->flags); | |
202 | } | |
203 | ||
204 | static void network_to_caps(RDMACapabilities *cap) | |
205 | { | |
206 | cap->version = ntohl(cap->version); | |
207 | cap->flags = ntohl(cap->flags); | |
208 | } | |
209 | ||
210 | /* | |
211 | * Representation of a RAMBlock from an RDMA perspective. | |
212 | * This is not transmitted, only local. | |
213 | * This and subsequent structures cannot be linked lists | |
214 | * because we're using a single IB message to transmit | |
215 | * the information. It's small anyway, so a list is overkill. | |
216 | */ | |
217 | typedef struct RDMALocalBlock { | |
4fb5364b DDAG |
218 | char *block_name; |
219 | uint8_t *local_host_addr; /* local virtual address */ | |
220 | uint64_t remote_host_addr; /* remote virtual address */ | |
221 | uint64_t offset; | |
222 | uint64_t length; | |
223 | struct ibv_mr **pmr; /* MRs for chunk-level registration */ | |
224 | struct ibv_mr *mr; /* MR for non-chunk-level registration */ | |
225 | uint32_t *remote_keys; /* rkeys for chunk-level registration */ | |
226 | uint32_t remote_rkey; /* rkeys for non-chunk-level registration */ | |
227 | int index; /* which block are we */ | |
228 | bool is_ram_block; | |
229 | int nb_chunks; | |
2da776db MH |
230 | unsigned long *transit_bitmap; |
231 | unsigned long *unregister_bitmap; | |
232 | } RDMALocalBlock; | |
233 | ||
234 | /* | |
235 | * Also represents a RAMblock, but only on the dest. | |
236 | * This gets transmitted by the dest during connection-time | |
237 | * to the source VM and then is used to populate the | |
238 | * corresponding RDMALocalBlock with | |
239 | * the information needed to perform the actual RDMA. | |
240 | */ | |
a97270ad | 241 | typedef struct QEMU_PACKED RDMADestBlock { |
2da776db MH |
242 | uint64_t remote_host_addr; |
243 | uint64_t offset; | |
244 | uint64_t length; | |
245 | uint32_t remote_rkey; | |
246 | uint32_t padding; | |
a97270ad | 247 | } RDMADestBlock; |
2da776db MH |
248 | |
249 | static uint64_t htonll(uint64_t v) | |
250 | { | |
251 | union { uint32_t lv[2]; uint64_t llv; } u; | |
252 | u.lv[0] = htonl(v >> 32); | |
253 | u.lv[1] = htonl(v & 0xFFFFFFFFULL); | |
254 | return u.llv; | |
255 | } | |
256 | ||
257 | static uint64_t ntohll(uint64_t v) { | |
258 | union { uint32_t lv[2]; uint64_t llv; } u; | |
259 | u.llv = v; | |
260 | return ((uint64_t)ntohl(u.lv[0]) << 32) | (uint64_t) ntohl(u.lv[1]); | |
261 | } | |
262 | ||
a97270ad | 263 | static void dest_block_to_network(RDMADestBlock *db) |
2da776db | 264 | { |
a97270ad DDAG |
265 | db->remote_host_addr = htonll(db->remote_host_addr); |
266 | db->offset = htonll(db->offset); | |
267 | db->length = htonll(db->length); | |
268 | db->remote_rkey = htonl(db->remote_rkey); | |
2da776db MH |
269 | } |
270 | ||
a97270ad | 271 | static void network_to_dest_block(RDMADestBlock *db) |
2da776db | 272 | { |
a97270ad DDAG |
273 | db->remote_host_addr = ntohll(db->remote_host_addr); |
274 | db->offset = ntohll(db->offset); | |
275 | db->length = ntohll(db->length); | |
276 | db->remote_rkey = ntohl(db->remote_rkey); | |
2da776db MH |
277 | } |
278 | ||
279 | /* | |
280 | * Virtual address of the above structures used for transmitting | |
281 | * the RAMBlock descriptions at connection-time. | |
282 | * This structure is *not* transmitted. | |
283 | */ | |
284 | typedef struct RDMALocalBlocks { | |
285 | int nb_blocks; | |
286 | bool init; /* main memory init complete */ | |
287 | RDMALocalBlock *block; | |
288 | } RDMALocalBlocks; | |
289 | ||
290 | /* | |
291 | * Main data structure for RDMA state. | |
292 | * While there is only one copy of this structure being allocated right now, | |
293 | * this is the place where one would start if you wanted to consider | |
294 | * having more than one RDMA connection open at the same time. | |
295 | */ | |
296 | typedef struct RDMAContext { | |
297 | char *host; | |
298 | int port; | |
299 | ||
1f22364b | 300 | RDMAWorkRequestData wr_data[RDMA_WRID_MAX]; |
2da776db MH |
301 | |
302 | /* | |
303 | * This is used by *_exchange_send() to figure out whether or not | |
304 | * the initial "READY" message has already been received or not. | |
305 | * This is because other functions may potentially poll() and detect | |
306 | * the READY message before send() does, in which case we need to | |
307 | * know if it completed. | |
308 | */ | |
309 | int control_ready_expected; | |
310 | ||
311 | /* number of outstanding writes */ | |
312 | int nb_sent; | |
313 | ||
314 | /* store info about current buffer so that we can | |
315 | merge it with future sends */ | |
316 | uint64_t current_addr; | |
317 | uint64_t current_length; | |
318 | /* index of ram block the current buffer belongs to */ | |
319 | int current_index; | |
320 | /* index of the chunk in the current ram block */ | |
321 | int current_chunk; | |
322 | ||
323 | bool pin_all; | |
324 | ||
325 | /* | |
326 | * infiniband-specific variables for opening the device | |
327 | * and maintaining connection state and so forth. | |
328 | * | |
329 | * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in | |
330 | * cm_id->verbs, cm_id->channel, and cm_id->qp. | |
331 | */ | |
332 | struct rdma_cm_id *cm_id; /* connection manager ID */ | |
333 | struct rdma_cm_id *listen_id; | |
5a91337c | 334 | bool connected; |
2da776db MH |
335 | |
336 | struct ibv_context *verbs; | |
337 | struct rdma_event_channel *channel; | |
338 | struct ibv_qp *qp; /* queue pair */ | |
339 | struct ibv_comp_channel *comp_channel; /* completion channel */ | |
340 | struct ibv_pd *pd; /* protection domain */ | |
341 | struct ibv_cq *cq; /* completion queue */ | |
342 | ||
343 | /* | |
344 | * If a previous write failed (perhaps because of a failed | |
345 | * memory registration, then do not attempt any future work | |
346 | * and remember the error state. | |
347 | */ | |
348 | int error_state; | |
349 | int error_reported; | |
350 | ||
351 | /* | |
352 | * Description of ram blocks used throughout the code. | |
353 | */ | |
354 | RDMALocalBlocks local_ram_blocks; | |
a97270ad | 355 | RDMADestBlock *dest_blocks; |
2da776db MH |
356 | |
357 | /* | |
358 | * Migration on *destination* started. | |
359 | * Then use coroutine yield function. | |
360 | * Source runs in a thread, so we don't care. | |
361 | */ | |
362 | int migration_started_on_destination; | |
363 | ||
364 | int total_registrations; | |
365 | int total_writes; | |
366 | ||
367 | int unregister_current, unregister_next; | |
368 | uint64_t unregistrations[RDMA_SIGNALED_SEND_MAX]; | |
369 | ||
370 | GHashTable *blockmap; | |
371 | } RDMAContext; | |
372 | ||
373 | /* | |
374 | * Interface to the rest of the migration call stack. | |
375 | */ | |
376 | typedef struct QEMUFileRDMA { | |
377 | RDMAContext *rdma; | |
378 | size_t len; | |
379 | void *file; | |
380 | } QEMUFileRDMA; | |
381 | ||
382 | /* | |
383 | * Main structure for IB Send/Recv control messages. | |
384 | * This gets prepended at the beginning of every Send/Recv. | |
385 | */ | |
386 | typedef struct QEMU_PACKED { | |
387 | uint32_t len; /* Total length of data portion */ | |
388 | uint32_t type; /* which control command to perform */ | |
389 | uint32_t repeat; /* number of commands in data portion of same type */ | |
390 | uint32_t padding; | |
391 | } RDMAControlHeader; | |
392 | ||
393 | static void control_to_network(RDMAControlHeader *control) | |
394 | { | |
395 | control->type = htonl(control->type); | |
396 | control->len = htonl(control->len); | |
397 | control->repeat = htonl(control->repeat); | |
398 | } | |
399 | ||
400 | static void network_to_control(RDMAControlHeader *control) | |
401 | { | |
402 | control->type = ntohl(control->type); | |
403 | control->len = ntohl(control->len); | |
404 | control->repeat = ntohl(control->repeat); | |
405 | } | |
406 | ||
407 | /* | |
408 | * Register a single Chunk. | |
409 | * Information sent by the source VM to inform the dest | |
410 | * to register an single chunk of memory before we can perform | |
411 | * the actual RDMA operation. | |
412 | */ | |
413 | typedef struct QEMU_PACKED { | |
414 | union QEMU_PACKED { | |
b12f7777 | 415 | uint64_t current_addr; /* offset into the ram_addr_t space */ |
2da776db MH |
416 | uint64_t chunk; /* chunk to lookup if unregistering */ |
417 | } key; | |
418 | uint32_t current_index; /* which ramblock the chunk belongs to */ | |
419 | uint32_t padding; | |
420 | uint64_t chunks; /* how many sequential chunks to register */ | |
421 | } RDMARegister; | |
422 | ||
b12f7777 | 423 | static void register_to_network(RDMAContext *rdma, RDMARegister *reg) |
2da776db | 424 | { |
b12f7777 DDAG |
425 | RDMALocalBlock *local_block; |
426 | local_block = &rdma->local_ram_blocks.block[reg->current_index]; | |
427 | ||
428 | if (local_block->is_ram_block) { | |
429 | /* | |
430 | * current_addr as passed in is an address in the local ram_addr_t | |
431 | * space, we need to translate this for the destination | |
432 | */ | |
433 | reg->key.current_addr -= local_block->offset; | |
434 | reg->key.current_addr += rdma->dest_blocks[reg->current_index].offset; | |
435 | } | |
2da776db MH |
436 | reg->key.current_addr = htonll(reg->key.current_addr); |
437 | reg->current_index = htonl(reg->current_index); | |
438 | reg->chunks = htonll(reg->chunks); | |
439 | } | |
440 | ||
441 | static void network_to_register(RDMARegister *reg) | |
442 | { | |
443 | reg->key.current_addr = ntohll(reg->key.current_addr); | |
444 | reg->current_index = ntohl(reg->current_index); | |
445 | reg->chunks = ntohll(reg->chunks); | |
446 | } | |
447 | ||
448 | typedef struct QEMU_PACKED { | |
449 | uint32_t value; /* if zero, we will madvise() */ | |
450 | uint32_t block_idx; /* which ram block index */ | |
b12f7777 | 451 | uint64_t offset; /* Address in remote ram_addr_t space */ |
2da776db MH |
452 | uint64_t length; /* length of the chunk */ |
453 | } RDMACompress; | |
454 | ||
b12f7777 | 455 | static void compress_to_network(RDMAContext *rdma, RDMACompress *comp) |
2da776db MH |
456 | { |
457 | comp->value = htonl(comp->value); | |
b12f7777 DDAG |
458 | /* |
459 | * comp->offset as passed in is an address in the local ram_addr_t | |
460 | * space, we need to translate this for the destination | |
461 | */ | |
462 | comp->offset -= rdma->local_ram_blocks.block[comp->block_idx].offset; | |
463 | comp->offset += rdma->dest_blocks[comp->block_idx].offset; | |
2da776db MH |
464 | comp->block_idx = htonl(comp->block_idx); |
465 | comp->offset = htonll(comp->offset); | |
466 | comp->length = htonll(comp->length); | |
467 | } | |
468 | ||
469 | static void network_to_compress(RDMACompress *comp) | |
470 | { | |
471 | comp->value = ntohl(comp->value); | |
472 | comp->block_idx = ntohl(comp->block_idx); | |
473 | comp->offset = ntohll(comp->offset); | |
474 | comp->length = ntohll(comp->length); | |
475 | } | |
476 | ||
477 | /* | |
478 | * The result of the dest's memory registration produces an "rkey" | |
479 | * which the source VM must reference in order to perform | |
480 | * the RDMA operation. | |
481 | */ | |
482 | typedef struct QEMU_PACKED { | |
483 | uint32_t rkey; | |
484 | uint32_t padding; | |
485 | uint64_t host_addr; | |
486 | } RDMARegisterResult; | |
487 | ||
488 | static void result_to_network(RDMARegisterResult *result) | |
489 | { | |
490 | result->rkey = htonl(result->rkey); | |
491 | result->host_addr = htonll(result->host_addr); | |
492 | }; | |
493 | ||
494 | static void network_to_result(RDMARegisterResult *result) | |
495 | { | |
496 | result->rkey = ntohl(result->rkey); | |
497 | result->host_addr = ntohll(result->host_addr); | |
498 | }; | |
499 | ||
500 | const char *print_wrid(int wrid); | |
501 | static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, | |
502 | uint8_t *data, RDMAControlHeader *resp, | |
503 | int *resp_idx, | |
504 | int (*callback)(RDMAContext *rdma)); | |
505 | ||
dd286ed7 IY |
506 | static inline uint64_t ram_chunk_index(const uint8_t *start, |
507 | const uint8_t *host) | |
2da776db MH |
508 | { |
509 | return ((uintptr_t) host - (uintptr_t) start) >> RDMA_REG_CHUNK_SHIFT; | |
510 | } | |
511 | ||
dd286ed7 | 512 | static inline uint8_t *ram_chunk_start(const RDMALocalBlock *rdma_ram_block, |
2da776db MH |
513 | uint64_t i) |
514 | { | |
fbce8c25 SW |
515 | return (uint8_t *)(uintptr_t)(rdma_ram_block->local_host_addr + |
516 | (i << RDMA_REG_CHUNK_SHIFT)); | |
2da776db MH |
517 | } |
518 | ||
dd286ed7 IY |
519 | static inline uint8_t *ram_chunk_end(const RDMALocalBlock *rdma_ram_block, |
520 | uint64_t i) | |
2da776db MH |
521 | { |
522 | uint8_t *result = ram_chunk_start(rdma_ram_block, i) + | |
523 | (1UL << RDMA_REG_CHUNK_SHIFT); | |
524 | ||
525 | if (result > (rdma_ram_block->local_host_addr + rdma_ram_block->length)) { | |
526 | result = rdma_ram_block->local_host_addr + rdma_ram_block->length; | |
527 | } | |
528 | ||
529 | return result; | |
530 | } | |
531 | ||
4fb5364b DDAG |
532 | static int rdma_add_block(RDMAContext *rdma, const char *block_name, |
533 | void *host_addr, | |
2da776db MH |
534 | ram_addr_t block_offset, uint64_t length) |
535 | { | |
536 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
537 | RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, | |
fbce8c25 | 538 | (void *)(uintptr_t)block_offset); |
2da776db MH |
539 | RDMALocalBlock *old = local->block; |
540 | ||
541 | assert(block == NULL); | |
542 | ||
543 | local->block = g_malloc0(sizeof(RDMALocalBlock) * (local->nb_blocks + 1)); | |
544 | ||
545 | if (local->nb_blocks) { | |
546 | int x; | |
547 | ||
548 | for (x = 0; x < local->nb_blocks; x++) { | |
fbce8c25 SW |
549 | g_hash_table_remove(rdma->blockmap, |
550 | (void *)(uintptr_t)old[x].offset); | |
551 | g_hash_table_insert(rdma->blockmap, | |
552 | (void *)(uintptr_t)old[x].offset, | |
553 | &local->block[x]); | |
2da776db MH |
554 | } |
555 | memcpy(local->block, old, sizeof(RDMALocalBlock) * local->nb_blocks); | |
556 | g_free(old); | |
557 | } | |
558 | ||
559 | block = &local->block[local->nb_blocks]; | |
560 | ||
4fb5364b | 561 | block->block_name = g_strdup(block_name); |
2da776db MH |
562 | block->local_host_addr = host_addr; |
563 | block->offset = block_offset; | |
564 | block->length = length; | |
565 | block->index = local->nb_blocks; | |
566 | block->nb_chunks = ram_chunk_index(host_addr, host_addr + length) + 1UL; | |
567 | block->transit_bitmap = bitmap_new(block->nb_chunks); | |
568 | bitmap_clear(block->transit_bitmap, 0, block->nb_chunks); | |
569 | block->unregister_bitmap = bitmap_new(block->nb_chunks); | |
570 | bitmap_clear(block->unregister_bitmap, 0, block->nb_chunks); | |
571 | block->remote_keys = g_malloc0(block->nb_chunks * sizeof(uint32_t)); | |
572 | ||
573 | block->is_ram_block = local->init ? false : true; | |
574 | ||
575 | g_hash_table_insert(rdma->blockmap, (void *) block_offset, block); | |
576 | ||
4fb5364b DDAG |
577 | trace_rdma_add_block(block_name, local->nb_blocks, |
578 | (uintptr_t) block->local_host_addr, | |
ba795761 | 579 | block->offset, block->length, |
fbce8c25 | 580 | (uintptr_t) (block->local_host_addr + block->length), |
ba795761 DDAG |
581 | BITS_TO_LONGS(block->nb_chunks) * |
582 | sizeof(unsigned long) * 8, | |
583 | block->nb_chunks); | |
2da776db MH |
584 | |
585 | local->nb_blocks++; | |
586 | ||
587 | return 0; | |
588 | } | |
589 | ||
590 | /* | |
591 | * Memory regions need to be registered with the device and queue pairs setup | |
592 | * in advanced before the migration starts. This tells us where the RAM blocks | |
593 | * are so that we can register them individually. | |
594 | */ | |
e3807054 | 595 | static int qemu_rdma_init_one_block(const char *block_name, void *host_addr, |
2da776db MH |
596 | ram_addr_t block_offset, ram_addr_t length, void *opaque) |
597 | { | |
4fb5364b | 598 | return rdma_add_block(opaque, block_name, host_addr, block_offset, length); |
2da776db MH |
599 | } |
600 | ||
601 | /* | |
602 | * Identify the RAMBlocks and their quantity. They will be references to | |
603 | * identify chunk boundaries inside each RAMBlock and also be referenced | |
604 | * during dynamic page registration. | |
605 | */ | |
606 | static int qemu_rdma_init_ram_blocks(RDMAContext *rdma) | |
607 | { | |
608 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
609 | ||
610 | assert(rdma->blockmap == NULL); | |
611 | rdma->blockmap = g_hash_table_new(g_direct_hash, g_direct_equal); | |
612 | memset(local, 0, sizeof *local); | |
613 | qemu_ram_foreach_block(qemu_rdma_init_one_block, rdma); | |
733252de | 614 | trace_qemu_rdma_init_ram_blocks(local->nb_blocks); |
a97270ad | 615 | rdma->dest_blocks = (RDMADestBlock *) g_malloc0(sizeof(RDMADestBlock) * |
2da776db MH |
616 | rdma->local_ram_blocks.nb_blocks); |
617 | local->init = true; | |
618 | return 0; | |
619 | } | |
620 | ||
03fcab38 DDAG |
621 | /* |
622 | * Note: If used outside of cleanup, the caller must ensure that the destination | |
623 | * block structures are also updated | |
624 | */ | |
625 | static int rdma_delete_block(RDMAContext *rdma, RDMALocalBlock *block) | |
2da776db MH |
626 | { |
627 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
2da776db MH |
628 | RDMALocalBlock *old = local->block; |
629 | int x; | |
630 | ||
03fcab38 DDAG |
631 | if (rdma->blockmap) { |
632 | g_hash_table_remove(rdma->blockmap, (void *)(uintptr_t)block->offset); | |
633 | } | |
2da776db MH |
634 | if (block->pmr) { |
635 | int j; | |
636 | ||
637 | for (j = 0; j < block->nb_chunks; j++) { | |
638 | if (!block->pmr[j]) { | |
639 | continue; | |
640 | } | |
641 | ibv_dereg_mr(block->pmr[j]); | |
642 | rdma->total_registrations--; | |
643 | } | |
644 | g_free(block->pmr); | |
645 | block->pmr = NULL; | |
646 | } | |
647 | ||
648 | if (block->mr) { | |
649 | ibv_dereg_mr(block->mr); | |
650 | rdma->total_registrations--; | |
651 | block->mr = NULL; | |
652 | } | |
653 | ||
654 | g_free(block->transit_bitmap); | |
655 | block->transit_bitmap = NULL; | |
656 | ||
657 | g_free(block->unregister_bitmap); | |
658 | block->unregister_bitmap = NULL; | |
659 | ||
660 | g_free(block->remote_keys); | |
661 | block->remote_keys = NULL; | |
662 | ||
4fb5364b DDAG |
663 | g_free(block->block_name); |
664 | block->block_name = NULL; | |
665 | ||
03fcab38 DDAG |
666 | if (rdma->blockmap) { |
667 | for (x = 0; x < local->nb_blocks; x++) { | |
668 | g_hash_table_remove(rdma->blockmap, | |
669 | (void *)(uintptr_t)old[x].offset); | |
670 | } | |
2da776db MH |
671 | } |
672 | ||
673 | if (local->nb_blocks > 1) { | |
674 | ||
675 | local->block = g_malloc0(sizeof(RDMALocalBlock) * | |
676 | (local->nb_blocks - 1)); | |
677 | ||
678 | if (block->index) { | |
679 | memcpy(local->block, old, sizeof(RDMALocalBlock) * block->index); | |
680 | } | |
681 | ||
682 | if (block->index < (local->nb_blocks - 1)) { | |
683 | memcpy(local->block + block->index, old + (block->index + 1), | |
684 | sizeof(RDMALocalBlock) * | |
685 | (local->nb_blocks - (block->index + 1))); | |
686 | } | |
687 | } else { | |
688 | assert(block == local->block); | |
689 | local->block = NULL; | |
690 | } | |
691 | ||
03fcab38 | 692 | trace_rdma_delete_block(block, (uintptr_t)block->local_host_addr, |
733252de | 693 | block->offset, block->length, |
fbce8c25 | 694 | (uintptr_t)(block->local_host_addr + block->length), |
733252de DDAG |
695 | BITS_TO_LONGS(block->nb_chunks) * |
696 | sizeof(unsigned long) * 8, block->nb_chunks); | |
2da776db MH |
697 | |
698 | g_free(old); | |
699 | ||
700 | local->nb_blocks--; | |
701 | ||
03fcab38 | 702 | if (local->nb_blocks && rdma->blockmap) { |
2da776db | 703 | for (x = 0; x < local->nb_blocks; x++) { |
fbce8c25 SW |
704 | g_hash_table_insert(rdma->blockmap, |
705 | (void *)(uintptr_t)local->block[x].offset, | |
706 | &local->block[x]); | |
2da776db MH |
707 | } |
708 | } | |
709 | ||
710 | return 0; | |
711 | } | |
712 | ||
713 | /* | |
714 | * Put in the log file which RDMA device was opened and the details | |
715 | * associated with that device. | |
716 | */ | |
717 | static void qemu_rdma_dump_id(const char *who, struct ibv_context *verbs) | |
718 | { | |
7fc5b13f MH |
719 | struct ibv_port_attr port; |
720 | ||
721 | if (ibv_query_port(verbs, 1, &port)) { | |
733252de | 722 | error_report("Failed to query port information"); |
7fc5b13f MH |
723 | return; |
724 | } | |
725 | ||
2da776db MH |
726 | printf("%s RDMA Device opened: kernel name %s " |
727 | "uverbs device name %s, " | |
7fc5b13f MH |
728 | "infiniband_verbs class device path %s, " |
729 | "infiniband class device path %s, " | |
730 | "transport: (%d) %s\n", | |
2da776db MH |
731 | who, |
732 | verbs->device->name, | |
733 | verbs->device->dev_name, | |
734 | verbs->device->dev_path, | |
7fc5b13f MH |
735 | verbs->device->ibdev_path, |
736 | port.link_layer, | |
737 | (port.link_layer == IBV_LINK_LAYER_INFINIBAND) ? "Infiniband" : | |
02942db7 | 738 | ((port.link_layer == IBV_LINK_LAYER_ETHERNET) |
7fc5b13f | 739 | ? "Ethernet" : "Unknown")); |
2da776db MH |
740 | } |
741 | ||
742 | /* | |
743 | * Put in the log file the RDMA gid addressing information, | |
744 | * useful for folks who have trouble understanding the | |
745 | * RDMA device hierarchy in the kernel. | |
746 | */ | |
747 | static void qemu_rdma_dump_gid(const char *who, struct rdma_cm_id *id) | |
748 | { | |
749 | char sgid[33]; | |
750 | char dgid[33]; | |
751 | inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.sgid, sgid, sizeof sgid); | |
752 | inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.dgid, dgid, sizeof dgid); | |
733252de | 753 | trace_qemu_rdma_dump_gid(who, sgid, dgid); |
2da776db MH |
754 | } |
755 | ||
7fc5b13f MH |
756 | /* |
757 | * As of now, IPv6 over RoCE / iWARP is not supported by linux. | |
758 | * We will try the next addrinfo struct, and fail if there are | |
759 | * no other valid addresses to bind against. | |
760 | * | |
761 | * If user is listening on '[::]', then we will not have a opened a device | |
762 | * yet and have no way of verifying if the device is RoCE or not. | |
763 | * | |
764 | * In this case, the source VM will throw an error for ALL types of | |
765 | * connections (both IPv4 and IPv6) if the destination machine does not have | |
766 | * a regular infiniband network available for use. | |
767 | * | |
4c293dc6 | 768 | * The only way to guarantee that an error is thrown for broken kernels is |
7fc5b13f MH |
769 | * for the management software to choose a *specific* interface at bind time |
770 | * and validate what time of hardware it is. | |
771 | * | |
772 | * Unfortunately, this puts the user in a fix: | |
02942db7 | 773 | * |
7fc5b13f MH |
774 | * If the source VM connects with an IPv4 address without knowing that the |
775 | * destination has bound to '[::]' the migration will unconditionally fail | |
776 | * unless the management software is explicitly listening on the the IPv4 | |
777 | * address while using a RoCE-based device. | |
778 | * | |
779 | * If the source VM connects with an IPv6 address, then we're OK because we can | |
780 | * throw an error on the source (and similarly on the destination). | |
02942db7 | 781 | * |
7fc5b13f MH |
782 | * But in mixed environments, this will be broken for a while until it is fixed |
783 | * inside linux. | |
784 | * | |
785 | * We do provide a *tiny* bit of help in this function: We can list all of the | |
786 | * devices in the system and check to see if all the devices are RoCE or | |
02942db7 | 787 | * Infiniband. |
7fc5b13f MH |
788 | * |
789 | * If we detect that we have a *pure* RoCE environment, then we can safely | |
4c293dc6 | 790 | * thrown an error even if the management software has specified '[::]' as the |
7fc5b13f MH |
791 | * bind address. |
792 | * | |
793 | * However, if there is are multiple hetergeneous devices, then we cannot make | |
794 | * this assumption and the user just has to be sure they know what they are | |
795 | * doing. | |
796 | * | |
797 | * Patches are being reviewed on linux-rdma. | |
798 | */ | |
799 | static int qemu_rdma_broken_ipv6_kernel(Error **errp, struct ibv_context *verbs) | |
800 | { | |
801 | struct ibv_port_attr port_attr; | |
802 | ||
803 | /* This bug only exists in linux, to our knowledge. */ | |
804 | #ifdef CONFIG_LINUX | |
805 | ||
02942db7 | 806 | /* |
7fc5b13f | 807 | * Verbs are only NULL if management has bound to '[::]'. |
02942db7 | 808 | * |
7fc5b13f MH |
809 | * Let's iterate through all the devices and see if there any pure IB |
810 | * devices (non-ethernet). | |
02942db7 | 811 | * |
7fc5b13f | 812 | * If not, then we can safely proceed with the migration. |
4c293dc6 | 813 | * Otherwise, there are no guarantees until the bug is fixed in linux. |
7fc5b13f MH |
814 | */ |
815 | if (!verbs) { | |
02942db7 | 816 | int num_devices, x; |
7fc5b13f MH |
817 | struct ibv_device ** dev_list = ibv_get_device_list(&num_devices); |
818 | bool roce_found = false; | |
819 | bool ib_found = false; | |
820 | ||
821 | for (x = 0; x < num_devices; x++) { | |
822 | verbs = ibv_open_device(dev_list[x]); | |
5b61d575 PR |
823 | if (!verbs) { |
824 | if (errno == EPERM) { | |
825 | continue; | |
826 | } else { | |
827 | return -EINVAL; | |
828 | } | |
829 | } | |
7fc5b13f MH |
830 | |
831 | if (ibv_query_port(verbs, 1, &port_attr)) { | |
832 | ibv_close_device(verbs); | |
833 | ERROR(errp, "Could not query initial IB port"); | |
834 | return -EINVAL; | |
835 | } | |
836 | ||
837 | if (port_attr.link_layer == IBV_LINK_LAYER_INFINIBAND) { | |
838 | ib_found = true; | |
839 | } else if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) { | |
840 | roce_found = true; | |
841 | } | |
842 | ||
843 | ibv_close_device(verbs); | |
844 | ||
845 | } | |
846 | ||
847 | if (roce_found) { | |
848 | if (ib_found) { | |
849 | fprintf(stderr, "WARN: migrations may fail:" | |
850 | " IPv6 over RoCE / iWARP in linux" | |
851 | " is broken. But since you appear to have a" | |
852 | " mixed RoCE / IB environment, be sure to only" | |
853 | " migrate over the IB fabric until the kernel " | |
854 | " fixes the bug.\n"); | |
855 | } else { | |
856 | ERROR(errp, "You only have RoCE / iWARP devices in your systems" | |
857 | " and your management software has specified '[::]'" | |
858 | ", but IPv6 over RoCE / iWARP is not supported in Linux."); | |
859 | return -ENONET; | |
860 | } | |
861 | } | |
862 | ||
863 | return 0; | |
864 | } | |
865 | ||
866 | /* | |
867 | * If we have a verbs context, that means that some other than '[::]' was | |
02942db7 SW |
868 | * used by the management software for binding. In which case we can |
869 | * actually warn the user about a potentially broken kernel. | |
7fc5b13f MH |
870 | */ |
871 | ||
872 | /* IB ports start with 1, not 0 */ | |
873 | if (ibv_query_port(verbs, 1, &port_attr)) { | |
874 | ERROR(errp, "Could not query initial IB port"); | |
875 | return -EINVAL; | |
876 | } | |
877 | ||
878 | if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) { | |
879 | ERROR(errp, "Linux kernel's RoCE / iWARP does not support IPv6 " | |
880 | "(but patches on linux-rdma in progress)"); | |
881 | return -ENONET; | |
882 | } | |
883 | ||
884 | #endif | |
885 | ||
886 | return 0; | |
887 | } | |
888 | ||
2da776db MH |
889 | /* |
890 | * Figure out which RDMA device corresponds to the requested IP hostname | |
891 | * Also create the initial connection manager identifiers for opening | |
892 | * the connection. | |
893 | */ | |
894 | static int qemu_rdma_resolve_host(RDMAContext *rdma, Error **errp) | |
895 | { | |
896 | int ret; | |
7fc5b13f | 897 | struct rdma_addrinfo *res; |
2da776db MH |
898 | char port_str[16]; |
899 | struct rdma_cm_event *cm_event; | |
900 | char ip[40] = "unknown"; | |
7fc5b13f | 901 | struct rdma_addrinfo *e; |
2da776db MH |
902 | |
903 | if (rdma->host == NULL || !strcmp(rdma->host, "")) { | |
66988941 | 904 | ERROR(errp, "RDMA hostname has not been set"); |
7fc5b13f | 905 | return -EINVAL; |
2da776db MH |
906 | } |
907 | ||
908 | /* create CM channel */ | |
909 | rdma->channel = rdma_create_event_channel(); | |
910 | if (!rdma->channel) { | |
66988941 | 911 | ERROR(errp, "could not create CM channel"); |
7fc5b13f | 912 | return -EINVAL; |
2da776db MH |
913 | } |
914 | ||
915 | /* create CM id */ | |
916 | ret = rdma_create_id(rdma->channel, &rdma->cm_id, NULL, RDMA_PS_TCP); | |
917 | if (ret) { | |
66988941 | 918 | ERROR(errp, "could not create channel id"); |
2da776db MH |
919 | goto err_resolve_create_id; |
920 | } | |
921 | ||
922 | snprintf(port_str, 16, "%d", rdma->port); | |
923 | port_str[15] = '\0'; | |
924 | ||
7fc5b13f | 925 | ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res); |
2da776db | 926 | if (ret < 0) { |
7fc5b13f | 927 | ERROR(errp, "could not rdma_getaddrinfo address %s", rdma->host); |
2da776db MH |
928 | goto err_resolve_get_addr; |
929 | } | |
930 | ||
6470215b MH |
931 | for (e = res; e != NULL; e = e->ai_next) { |
932 | inet_ntop(e->ai_family, | |
7fc5b13f | 933 | &((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip); |
733252de | 934 | trace_qemu_rdma_resolve_host_trying(rdma->host, ip); |
2da776db | 935 | |
7fc5b13f | 936 | ret = rdma_resolve_addr(rdma->cm_id, NULL, e->ai_dst_addr, |
6470215b MH |
937 | RDMA_RESOLVE_TIMEOUT_MS); |
938 | if (!ret) { | |
c89aa2f1 MH |
939 | if (e->ai_family == AF_INET6) { |
940 | ret = qemu_rdma_broken_ipv6_kernel(errp, rdma->cm_id->verbs); | |
941 | if (ret) { | |
942 | continue; | |
943 | } | |
7fc5b13f | 944 | } |
6470215b MH |
945 | goto route; |
946 | } | |
2da776db MH |
947 | } |
948 | ||
6470215b MH |
949 | ERROR(errp, "could not resolve address %s", rdma->host); |
950 | goto err_resolve_get_addr; | |
951 | ||
952 | route: | |
2da776db MH |
953 | qemu_rdma_dump_gid("source_resolve_addr", rdma->cm_id); |
954 | ||
955 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
956 | if (ret) { | |
66988941 | 957 | ERROR(errp, "could not perform event_addr_resolved"); |
2da776db MH |
958 | goto err_resolve_get_addr; |
959 | } | |
960 | ||
961 | if (cm_event->event != RDMA_CM_EVENT_ADDR_RESOLVED) { | |
66988941 | 962 | ERROR(errp, "result not equal to event_addr_resolved %s", |
2da776db MH |
963 | rdma_event_str(cm_event->event)); |
964 | perror("rdma_resolve_addr"); | |
2a934347 | 965 | rdma_ack_cm_event(cm_event); |
7fc5b13f | 966 | ret = -EINVAL; |
2da776db MH |
967 | goto err_resolve_get_addr; |
968 | } | |
969 | rdma_ack_cm_event(cm_event); | |
970 | ||
971 | /* resolve route */ | |
972 | ret = rdma_resolve_route(rdma->cm_id, RDMA_RESOLVE_TIMEOUT_MS); | |
973 | if (ret) { | |
66988941 | 974 | ERROR(errp, "could not resolve rdma route"); |
2da776db MH |
975 | goto err_resolve_get_addr; |
976 | } | |
977 | ||
978 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
979 | if (ret) { | |
66988941 | 980 | ERROR(errp, "could not perform event_route_resolved"); |
2da776db MH |
981 | goto err_resolve_get_addr; |
982 | } | |
983 | if (cm_event->event != RDMA_CM_EVENT_ROUTE_RESOLVED) { | |
66988941 | 984 | ERROR(errp, "result not equal to event_route_resolved: %s", |
2da776db MH |
985 | rdma_event_str(cm_event->event)); |
986 | rdma_ack_cm_event(cm_event); | |
7fc5b13f | 987 | ret = -EINVAL; |
2da776db MH |
988 | goto err_resolve_get_addr; |
989 | } | |
990 | rdma_ack_cm_event(cm_event); | |
991 | rdma->verbs = rdma->cm_id->verbs; | |
992 | qemu_rdma_dump_id("source_resolve_host", rdma->cm_id->verbs); | |
993 | qemu_rdma_dump_gid("source_resolve_host", rdma->cm_id); | |
994 | return 0; | |
995 | ||
996 | err_resolve_get_addr: | |
997 | rdma_destroy_id(rdma->cm_id); | |
998 | rdma->cm_id = NULL; | |
999 | err_resolve_create_id: | |
1000 | rdma_destroy_event_channel(rdma->channel); | |
1001 | rdma->channel = NULL; | |
7fc5b13f | 1002 | return ret; |
2da776db MH |
1003 | } |
1004 | ||
1005 | /* | |
1006 | * Create protection domain and completion queues | |
1007 | */ | |
1008 | static int qemu_rdma_alloc_pd_cq(RDMAContext *rdma) | |
1009 | { | |
1010 | /* allocate pd */ | |
1011 | rdma->pd = ibv_alloc_pd(rdma->verbs); | |
1012 | if (!rdma->pd) { | |
733252de | 1013 | error_report("failed to allocate protection domain"); |
2da776db MH |
1014 | return -1; |
1015 | } | |
1016 | ||
1017 | /* create completion channel */ | |
1018 | rdma->comp_channel = ibv_create_comp_channel(rdma->verbs); | |
1019 | if (!rdma->comp_channel) { | |
733252de | 1020 | error_report("failed to allocate completion channel"); |
2da776db MH |
1021 | goto err_alloc_pd_cq; |
1022 | } | |
1023 | ||
1024 | /* | |
1025 | * Completion queue can be filled by both read and write work requests, | |
1026 | * so must reflect the sum of both possible queue sizes. | |
1027 | */ | |
1028 | rdma->cq = ibv_create_cq(rdma->verbs, (RDMA_SIGNALED_SEND_MAX * 3), | |
1029 | NULL, rdma->comp_channel, 0); | |
1030 | if (!rdma->cq) { | |
733252de | 1031 | error_report("failed to allocate completion queue"); |
2da776db MH |
1032 | goto err_alloc_pd_cq; |
1033 | } | |
1034 | ||
1035 | return 0; | |
1036 | ||
1037 | err_alloc_pd_cq: | |
1038 | if (rdma->pd) { | |
1039 | ibv_dealloc_pd(rdma->pd); | |
1040 | } | |
1041 | if (rdma->comp_channel) { | |
1042 | ibv_destroy_comp_channel(rdma->comp_channel); | |
1043 | } | |
1044 | rdma->pd = NULL; | |
1045 | rdma->comp_channel = NULL; | |
1046 | return -1; | |
1047 | ||
1048 | } | |
1049 | ||
1050 | /* | |
1051 | * Create queue pairs. | |
1052 | */ | |
1053 | static int qemu_rdma_alloc_qp(RDMAContext *rdma) | |
1054 | { | |
1055 | struct ibv_qp_init_attr attr = { 0 }; | |
1056 | int ret; | |
1057 | ||
1058 | attr.cap.max_send_wr = RDMA_SIGNALED_SEND_MAX; | |
1059 | attr.cap.max_recv_wr = 3; | |
1060 | attr.cap.max_send_sge = 1; | |
1061 | attr.cap.max_recv_sge = 1; | |
1062 | attr.send_cq = rdma->cq; | |
1063 | attr.recv_cq = rdma->cq; | |
1064 | attr.qp_type = IBV_QPT_RC; | |
1065 | ||
1066 | ret = rdma_create_qp(rdma->cm_id, rdma->pd, &attr); | |
1067 | if (ret) { | |
1068 | return -1; | |
1069 | } | |
1070 | ||
1071 | rdma->qp = rdma->cm_id->qp; | |
1072 | return 0; | |
1073 | } | |
1074 | ||
1075 | static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) | |
1076 | { | |
1077 | int i; | |
1078 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
1079 | ||
1080 | for (i = 0; i < local->nb_blocks; i++) { | |
1081 | local->block[i].mr = | |
1082 | ibv_reg_mr(rdma->pd, | |
1083 | local->block[i].local_host_addr, | |
1084 | local->block[i].length, | |
1085 | IBV_ACCESS_LOCAL_WRITE | | |
1086 | IBV_ACCESS_REMOTE_WRITE | |
1087 | ); | |
1088 | if (!local->block[i].mr) { | |
1089 | perror("Failed to register local dest ram block!\n"); | |
1090 | break; | |
1091 | } | |
1092 | rdma->total_registrations++; | |
1093 | } | |
1094 | ||
1095 | if (i >= local->nb_blocks) { | |
1096 | return 0; | |
1097 | } | |
1098 | ||
1099 | for (i--; i >= 0; i--) { | |
1100 | ibv_dereg_mr(local->block[i].mr); | |
1101 | rdma->total_registrations--; | |
1102 | } | |
1103 | ||
1104 | return -1; | |
1105 | ||
1106 | } | |
1107 | ||
1108 | /* | |
1109 | * Find the ram block that corresponds to the page requested to be | |
1110 | * transmitted by QEMU. | |
1111 | * | |
1112 | * Once the block is found, also identify which 'chunk' within that | |
1113 | * block that the page belongs to. | |
1114 | * | |
1115 | * This search cannot fail or the migration will fail. | |
1116 | */ | |
1117 | static int qemu_rdma_search_ram_block(RDMAContext *rdma, | |
fbce8c25 | 1118 | uintptr_t block_offset, |
2da776db MH |
1119 | uint64_t offset, |
1120 | uint64_t length, | |
1121 | uint64_t *block_index, | |
1122 | uint64_t *chunk_index) | |
1123 | { | |
1124 | uint64_t current_addr = block_offset + offset; | |
1125 | RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, | |
1126 | (void *) block_offset); | |
1127 | assert(block); | |
1128 | assert(current_addr >= block->offset); | |
1129 | assert((current_addr + length) <= (block->offset + block->length)); | |
1130 | ||
1131 | *block_index = block->index; | |
1132 | *chunk_index = ram_chunk_index(block->local_host_addr, | |
1133 | block->local_host_addr + (current_addr - block->offset)); | |
1134 | ||
1135 | return 0; | |
1136 | } | |
1137 | ||
1138 | /* | |
1139 | * Register a chunk with IB. If the chunk was already registered | |
1140 | * previously, then skip. | |
1141 | * | |
1142 | * Also return the keys associated with the registration needed | |
1143 | * to perform the actual RDMA operation. | |
1144 | */ | |
1145 | static int qemu_rdma_register_and_get_keys(RDMAContext *rdma, | |
3ac040c0 | 1146 | RDMALocalBlock *block, uintptr_t host_addr, |
2da776db MH |
1147 | uint32_t *lkey, uint32_t *rkey, int chunk, |
1148 | uint8_t *chunk_start, uint8_t *chunk_end) | |
1149 | { | |
1150 | if (block->mr) { | |
1151 | if (lkey) { | |
1152 | *lkey = block->mr->lkey; | |
1153 | } | |
1154 | if (rkey) { | |
1155 | *rkey = block->mr->rkey; | |
1156 | } | |
1157 | return 0; | |
1158 | } | |
1159 | ||
1160 | /* allocate memory to store chunk MRs */ | |
1161 | if (!block->pmr) { | |
1162 | block->pmr = g_malloc0(block->nb_chunks * sizeof(struct ibv_mr *)); | |
2da776db MH |
1163 | } |
1164 | ||
1165 | /* | |
1166 | * If 'rkey', then we're the destination, so grant access to the source. | |
1167 | * | |
1168 | * If 'lkey', then we're the source VM, so grant access only to ourselves. | |
1169 | */ | |
1170 | if (!block->pmr[chunk]) { | |
1171 | uint64_t len = chunk_end - chunk_start; | |
1172 | ||
733252de | 1173 | trace_qemu_rdma_register_and_get_keys(len, chunk_start); |
2da776db MH |
1174 | |
1175 | block->pmr[chunk] = ibv_reg_mr(rdma->pd, | |
1176 | chunk_start, len, | |
1177 | (rkey ? (IBV_ACCESS_LOCAL_WRITE | | |
1178 | IBV_ACCESS_REMOTE_WRITE) : 0)); | |
1179 | ||
1180 | if (!block->pmr[chunk]) { | |
1181 | perror("Failed to register chunk!"); | |
1182 | fprintf(stderr, "Chunk details: block: %d chunk index %d" | |
3ac040c0 SW |
1183 | " start %" PRIuPTR " end %" PRIuPTR |
1184 | " host %" PRIuPTR | |
1185 | " local %" PRIuPTR " registrations: %d\n", | |
1186 | block->index, chunk, (uintptr_t)chunk_start, | |
1187 | (uintptr_t)chunk_end, host_addr, | |
1188 | (uintptr_t)block->local_host_addr, | |
2da776db MH |
1189 | rdma->total_registrations); |
1190 | return -1; | |
1191 | } | |
1192 | rdma->total_registrations++; | |
1193 | } | |
1194 | ||
1195 | if (lkey) { | |
1196 | *lkey = block->pmr[chunk]->lkey; | |
1197 | } | |
1198 | if (rkey) { | |
1199 | *rkey = block->pmr[chunk]->rkey; | |
1200 | } | |
1201 | return 0; | |
1202 | } | |
1203 | ||
1204 | /* | |
1205 | * Register (at connection time) the memory used for control | |
1206 | * channel messages. | |
1207 | */ | |
1208 | static int qemu_rdma_reg_control(RDMAContext *rdma, int idx) | |
1209 | { | |
1210 | rdma->wr_data[idx].control_mr = ibv_reg_mr(rdma->pd, | |
1211 | rdma->wr_data[idx].control, RDMA_CONTROL_MAX_BUFFER, | |
1212 | IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); | |
1213 | if (rdma->wr_data[idx].control_mr) { | |
1214 | rdma->total_registrations++; | |
1215 | return 0; | |
1216 | } | |
733252de | 1217 | error_report("qemu_rdma_reg_control failed"); |
2da776db MH |
1218 | return -1; |
1219 | } | |
1220 | ||
1221 | const char *print_wrid(int wrid) | |
1222 | { | |
1223 | if (wrid >= RDMA_WRID_RECV_CONTROL) { | |
1224 | return wrid_desc[RDMA_WRID_RECV_CONTROL]; | |
1225 | } | |
1226 | return wrid_desc[wrid]; | |
1227 | } | |
1228 | ||
1229 | /* | |
1230 | * RDMA requires memory registration (mlock/pinning), but this is not good for | |
1231 | * overcommitment. | |
1232 | * | |
1233 | * In preparation for the future where LRU information or workload-specific | |
1234 | * writable writable working set memory access behavior is available to QEMU | |
1235 | * it would be nice to have in place the ability to UN-register/UN-pin | |
1236 | * particular memory regions from the RDMA hardware when it is determine that | |
1237 | * those regions of memory will likely not be accessed again in the near future. | |
1238 | * | |
1239 | * While we do not yet have such information right now, the following | |
1240 | * compile-time option allows us to perform a non-optimized version of this | |
1241 | * behavior. | |
1242 | * | |
1243 | * By uncommenting this option, you will cause *all* RDMA transfers to be | |
1244 | * unregistered immediately after the transfer completes on both sides of the | |
1245 | * connection. This has no effect in 'rdma-pin-all' mode, only regular mode. | |
1246 | * | |
1247 | * This will have a terrible impact on migration performance, so until future | |
1248 | * workload information or LRU information is available, do not attempt to use | |
1249 | * this feature except for basic testing. | |
1250 | */ | |
1251 | //#define RDMA_UNREGISTRATION_EXAMPLE | |
1252 | ||
1253 | /* | |
1254 | * Perform a non-optimized memory unregistration after every transfer | |
24ec68ef | 1255 | * for demonstration purposes, only if pin-all is not requested. |
2da776db MH |
1256 | * |
1257 | * Potential optimizations: | |
1258 | * 1. Start a new thread to run this function continuously | |
1259 | - for bit clearing | |
1260 | - and for receipt of unregister messages | |
1261 | * 2. Use an LRU. | |
1262 | * 3. Use workload hints. | |
1263 | */ | |
1264 | static int qemu_rdma_unregister_waiting(RDMAContext *rdma) | |
1265 | { | |
1266 | while (rdma->unregistrations[rdma->unregister_current]) { | |
1267 | int ret; | |
1268 | uint64_t wr_id = rdma->unregistrations[rdma->unregister_current]; | |
1269 | uint64_t chunk = | |
1270 | (wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; | |
1271 | uint64_t index = | |
1272 | (wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; | |
1273 | RDMALocalBlock *block = | |
1274 | &(rdma->local_ram_blocks.block[index]); | |
1275 | RDMARegister reg = { .current_index = index }; | |
1276 | RDMAControlHeader resp = { .type = RDMA_CONTROL_UNREGISTER_FINISHED, | |
1277 | }; | |
1278 | RDMAControlHeader head = { .len = sizeof(RDMARegister), | |
1279 | .type = RDMA_CONTROL_UNREGISTER_REQUEST, | |
1280 | .repeat = 1, | |
1281 | }; | |
1282 | ||
733252de DDAG |
1283 | trace_qemu_rdma_unregister_waiting_proc(chunk, |
1284 | rdma->unregister_current); | |
2da776db MH |
1285 | |
1286 | rdma->unregistrations[rdma->unregister_current] = 0; | |
1287 | rdma->unregister_current++; | |
1288 | ||
1289 | if (rdma->unregister_current == RDMA_SIGNALED_SEND_MAX) { | |
1290 | rdma->unregister_current = 0; | |
1291 | } | |
1292 | ||
1293 | ||
1294 | /* | |
1295 | * Unregistration is speculative (because migration is single-threaded | |
1296 | * and we cannot break the protocol's inifinband message ordering). | |
1297 | * Thus, if the memory is currently being used for transmission, | |
1298 | * then abort the attempt to unregister and try again | |
1299 | * later the next time a completion is received for this memory. | |
1300 | */ | |
1301 | clear_bit(chunk, block->unregister_bitmap); | |
1302 | ||
1303 | if (test_bit(chunk, block->transit_bitmap)) { | |
733252de | 1304 | trace_qemu_rdma_unregister_waiting_inflight(chunk); |
2da776db MH |
1305 | continue; |
1306 | } | |
1307 | ||
733252de | 1308 | trace_qemu_rdma_unregister_waiting_send(chunk); |
2da776db MH |
1309 | |
1310 | ret = ibv_dereg_mr(block->pmr[chunk]); | |
1311 | block->pmr[chunk] = NULL; | |
1312 | block->remote_keys[chunk] = 0; | |
1313 | ||
1314 | if (ret != 0) { | |
1315 | perror("unregistration chunk failed"); | |
1316 | return -ret; | |
1317 | } | |
1318 | rdma->total_registrations--; | |
1319 | ||
1320 | reg.key.chunk = chunk; | |
b12f7777 | 1321 | register_to_network(rdma, ®); |
2da776db MH |
1322 | ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
1323 | &resp, NULL, NULL); | |
1324 | if (ret < 0) { | |
1325 | return ret; | |
1326 | } | |
1327 | ||
733252de | 1328 | trace_qemu_rdma_unregister_waiting_complete(chunk); |
2da776db MH |
1329 | } |
1330 | ||
1331 | return 0; | |
1332 | } | |
1333 | ||
1334 | static uint64_t qemu_rdma_make_wrid(uint64_t wr_id, uint64_t index, | |
1335 | uint64_t chunk) | |
1336 | { | |
1337 | uint64_t result = wr_id & RDMA_WRID_TYPE_MASK; | |
1338 | ||
1339 | result |= (index << RDMA_WRID_BLOCK_SHIFT); | |
1340 | result |= (chunk << RDMA_WRID_CHUNK_SHIFT); | |
1341 | ||
1342 | return result; | |
1343 | } | |
1344 | ||
1345 | /* | |
1346 | * Set bit for unregistration in the next iteration. | |
1347 | * We cannot transmit right here, but will unpin later. | |
1348 | */ | |
1349 | static void qemu_rdma_signal_unregister(RDMAContext *rdma, uint64_t index, | |
1350 | uint64_t chunk, uint64_t wr_id) | |
1351 | { | |
1352 | if (rdma->unregistrations[rdma->unregister_next] != 0) { | |
733252de | 1353 | error_report("rdma migration: queue is full"); |
2da776db MH |
1354 | } else { |
1355 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); | |
1356 | ||
1357 | if (!test_and_set_bit(chunk, block->unregister_bitmap)) { | |
733252de DDAG |
1358 | trace_qemu_rdma_signal_unregister_append(chunk, |
1359 | rdma->unregister_next); | |
2da776db MH |
1360 | |
1361 | rdma->unregistrations[rdma->unregister_next++] = | |
1362 | qemu_rdma_make_wrid(wr_id, index, chunk); | |
1363 | ||
1364 | if (rdma->unregister_next == RDMA_SIGNALED_SEND_MAX) { | |
1365 | rdma->unregister_next = 0; | |
1366 | } | |
1367 | } else { | |
733252de | 1368 | trace_qemu_rdma_signal_unregister_already(chunk); |
2da776db MH |
1369 | } |
1370 | } | |
1371 | } | |
1372 | ||
1373 | /* | |
1374 | * Consult the connection manager to see a work request | |
1375 | * (of any kind) has completed. | |
1376 | * Return the work request ID that completed. | |
1377 | */ | |
88571882 IY |
1378 | static uint64_t qemu_rdma_poll(RDMAContext *rdma, uint64_t *wr_id_out, |
1379 | uint32_t *byte_len) | |
2da776db MH |
1380 | { |
1381 | int ret; | |
1382 | struct ibv_wc wc; | |
1383 | uint64_t wr_id; | |
1384 | ||
1385 | ret = ibv_poll_cq(rdma->cq, 1, &wc); | |
1386 | ||
1387 | if (!ret) { | |
1388 | *wr_id_out = RDMA_WRID_NONE; | |
1389 | return 0; | |
1390 | } | |
1391 | ||
1392 | if (ret < 0) { | |
733252de | 1393 | error_report("ibv_poll_cq return %d", ret); |
2da776db MH |
1394 | return ret; |
1395 | } | |
1396 | ||
1397 | wr_id = wc.wr_id & RDMA_WRID_TYPE_MASK; | |
1398 | ||
1399 | if (wc.status != IBV_WC_SUCCESS) { | |
1400 | fprintf(stderr, "ibv_poll_cq wc.status=%d %s!\n", | |
1401 | wc.status, ibv_wc_status_str(wc.status)); | |
1402 | fprintf(stderr, "ibv_poll_cq wrid=%s!\n", wrid_desc[wr_id]); | |
1403 | ||
1404 | return -1; | |
1405 | } | |
1406 | ||
1407 | if (rdma->control_ready_expected && | |
1408 | (wr_id >= RDMA_WRID_RECV_CONTROL)) { | |
733252de | 1409 | trace_qemu_rdma_poll_recv(wrid_desc[RDMA_WRID_RECV_CONTROL], |
2da776db MH |
1410 | wr_id - RDMA_WRID_RECV_CONTROL, wr_id, rdma->nb_sent); |
1411 | rdma->control_ready_expected = 0; | |
1412 | } | |
1413 | ||
1414 | if (wr_id == RDMA_WRID_RDMA_WRITE) { | |
1415 | uint64_t chunk = | |
1416 | (wc.wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; | |
1417 | uint64_t index = | |
1418 | (wc.wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; | |
1419 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); | |
1420 | ||
733252de | 1421 | trace_qemu_rdma_poll_write(print_wrid(wr_id), wr_id, rdma->nb_sent, |
fbce8c25 SW |
1422 | index, chunk, block->local_host_addr, |
1423 | (void *)(uintptr_t)block->remote_host_addr); | |
2da776db MH |
1424 | |
1425 | clear_bit(chunk, block->transit_bitmap); | |
1426 | ||
1427 | if (rdma->nb_sent > 0) { | |
1428 | rdma->nb_sent--; | |
1429 | } | |
1430 | ||
1431 | if (!rdma->pin_all) { | |
1432 | /* | |
1433 | * FYI: If one wanted to signal a specific chunk to be unregistered | |
1434 | * using LRU or workload-specific information, this is the function | |
1435 | * you would call to do so. That chunk would then get asynchronously | |
1436 | * unregistered later. | |
1437 | */ | |
1438 | #ifdef RDMA_UNREGISTRATION_EXAMPLE | |
1439 | qemu_rdma_signal_unregister(rdma, index, chunk, wc.wr_id); | |
1440 | #endif | |
1441 | } | |
1442 | } else { | |
733252de | 1443 | trace_qemu_rdma_poll_other(print_wrid(wr_id), wr_id, rdma->nb_sent); |
2da776db MH |
1444 | } |
1445 | ||
1446 | *wr_id_out = wc.wr_id; | |
88571882 IY |
1447 | if (byte_len) { |
1448 | *byte_len = wc.byte_len; | |
1449 | } | |
2da776db MH |
1450 | |
1451 | return 0; | |
1452 | } | |
1453 | ||
1454 | /* | |
1455 | * Block until the next work request has completed. | |
1456 | * | |
1457 | * First poll to see if a work request has already completed, | |
1458 | * otherwise block. | |
1459 | * | |
1460 | * If we encounter completed work requests for IDs other than | |
1461 | * the one we're interested in, then that's generally an error. | |
1462 | * | |
1463 | * The only exception is actual RDMA Write completions. These | |
1464 | * completions only need to be recorded, but do not actually | |
1465 | * need further processing. | |
1466 | */ | |
88571882 IY |
1467 | static int qemu_rdma_block_for_wrid(RDMAContext *rdma, int wrid_requested, |
1468 | uint32_t *byte_len) | |
2da776db MH |
1469 | { |
1470 | int num_cq_events = 0, ret = 0; | |
1471 | struct ibv_cq *cq; | |
1472 | void *cq_ctx; | |
1473 | uint64_t wr_id = RDMA_WRID_NONE, wr_id_in; | |
1474 | ||
1475 | if (ibv_req_notify_cq(rdma->cq, 0)) { | |
1476 | return -1; | |
1477 | } | |
1478 | /* poll cq first */ | |
1479 | while (wr_id != wrid_requested) { | |
88571882 | 1480 | ret = qemu_rdma_poll(rdma, &wr_id_in, byte_len); |
2da776db MH |
1481 | if (ret < 0) { |
1482 | return ret; | |
1483 | } | |
1484 | ||
1485 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; | |
1486 | ||
1487 | if (wr_id == RDMA_WRID_NONE) { | |
1488 | break; | |
1489 | } | |
1490 | if (wr_id != wrid_requested) { | |
733252de DDAG |
1491 | trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested), |
1492 | wrid_requested, print_wrid(wr_id), wr_id); | |
2da776db MH |
1493 | } |
1494 | } | |
1495 | ||
1496 | if (wr_id == wrid_requested) { | |
1497 | return 0; | |
1498 | } | |
1499 | ||
1500 | while (1) { | |
1501 | /* | |
1502 | * Coroutine doesn't start until process_incoming_migration() | |
1503 | * so don't yield unless we know we're running inside of a coroutine. | |
1504 | */ | |
1505 | if (rdma->migration_started_on_destination) { | |
1506 | yield_until_fd_readable(rdma->comp_channel->fd); | |
1507 | } | |
1508 | ||
1509 | if (ibv_get_cq_event(rdma->comp_channel, &cq, &cq_ctx)) { | |
1510 | perror("ibv_get_cq_event"); | |
1511 | goto err_block_for_wrid; | |
1512 | } | |
1513 | ||
1514 | num_cq_events++; | |
1515 | ||
1516 | if (ibv_req_notify_cq(cq, 0)) { | |
1517 | goto err_block_for_wrid; | |
1518 | } | |
1519 | ||
1520 | while (wr_id != wrid_requested) { | |
88571882 | 1521 | ret = qemu_rdma_poll(rdma, &wr_id_in, byte_len); |
2da776db MH |
1522 | if (ret < 0) { |
1523 | goto err_block_for_wrid; | |
1524 | } | |
1525 | ||
1526 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; | |
1527 | ||
1528 | if (wr_id == RDMA_WRID_NONE) { | |
1529 | break; | |
1530 | } | |
1531 | if (wr_id != wrid_requested) { | |
733252de DDAG |
1532 | trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested), |
1533 | wrid_requested, print_wrid(wr_id), wr_id); | |
2da776db MH |
1534 | } |
1535 | } | |
1536 | ||
1537 | if (wr_id == wrid_requested) { | |
1538 | goto success_block_for_wrid; | |
1539 | } | |
1540 | } | |
1541 | ||
1542 | success_block_for_wrid: | |
1543 | if (num_cq_events) { | |
1544 | ibv_ack_cq_events(cq, num_cq_events); | |
1545 | } | |
1546 | return 0; | |
1547 | ||
1548 | err_block_for_wrid: | |
1549 | if (num_cq_events) { | |
1550 | ibv_ack_cq_events(cq, num_cq_events); | |
1551 | } | |
1552 | return ret; | |
1553 | } | |
1554 | ||
1555 | /* | |
1556 | * Post a SEND message work request for the control channel | |
1557 | * containing some data and block until the post completes. | |
1558 | */ | |
1559 | static int qemu_rdma_post_send_control(RDMAContext *rdma, uint8_t *buf, | |
1560 | RDMAControlHeader *head) | |
1561 | { | |
1562 | int ret = 0; | |
1f22364b | 1563 | RDMAWorkRequestData *wr = &rdma->wr_data[RDMA_WRID_CONTROL]; |
2da776db MH |
1564 | struct ibv_send_wr *bad_wr; |
1565 | struct ibv_sge sge = { | |
fbce8c25 | 1566 | .addr = (uintptr_t)(wr->control), |
2da776db MH |
1567 | .length = head->len + sizeof(RDMAControlHeader), |
1568 | .lkey = wr->control_mr->lkey, | |
1569 | }; | |
1570 | struct ibv_send_wr send_wr = { | |
1571 | .wr_id = RDMA_WRID_SEND_CONTROL, | |
1572 | .opcode = IBV_WR_SEND, | |
1573 | .send_flags = IBV_SEND_SIGNALED, | |
1574 | .sg_list = &sge, | |
1575 | .num_sge = 1, | |
1576 | }; | |
1577 | ||
733252de | 1578 | trace_qemu_rdma_post_send_control(control_desc[head->type]); |
2da776db MH |
1579 | |
1580 | /* | |
1581 | * We don't actually need to do a memcpy() in here if we used | |
1582 | * the "sge" properly, but since we're only sending control messages | |
1583 | * (not RAM in a performance-critical path), then its OK for now. | |
1584 | * | |
1585 | * The copy makes the RDMAControlHeader simpler to manipulate | |
1586 | * for the time being. | |
1587 | */ | |
6f1484ed | 1588 | assert(head->len <= RDMA_CONTROL_MAX_BUFFER - sizeof(*head)); |
2da776db MH |
1589 | memcpy(wr->control, head, sizeof(RDMAControlHeader)); |
1590 | control_to_network((void *) wr->control); | |
1591 | ||
1592 | if (buf) { | |
1593 | memcpy(wr->control + sizeof(RDMAControlHeader), buf, head->len); | |
1594 | } | |
1595 | ||
1596 | ||
e325b49a | 1597 | ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr); |
2da776db | 1598 | |
e325b49a | 1599 | if (ret > 0) { |
733252de | 1600 | error_report("Failed to use post IB SEND for control"); |
e325b49a | 1601 | return -ret; |
2da776db MH |
1602 | } |
1603 | ||
88571882 | 1604 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_SEND_CONTROL, NULL); |
2da776db | 1605 | if (ret < 0) { |
733252de | 1606 | error_report("rdma migration: send polling control error"); |
2da776db MH |
1607 | } |
1608 | ||
1609 | return ret; | |
1610 | } | |
1611 | ||
1612 | /* | |
1613 | * Post a RECV work request in anticipation of some future receipt | |
1614 | * of data on the control channel. | |
1615 | */ | |
1616 | static int qemu_rdma_post_recv_control(RDMAContext *rdma, int idx) | |
1617 | { | |
1618 | struct ibv_recv_wr *bad_wr; | |
1619 | struct ibv_sge sge = { | |
fbce8c25 | 1620 | .addr = (uintptr_t)(rdma->wr_data[idx].control), |
2da776db MH |
1621 | .length = RDMA_CONTROL_MAX_BUFFER, |
1622 | .lkey = rdma->wr_data[idx].control_mr->lkey, | |
1623 | }; | |
1624 | ||
1625 | struct ibv_recv_wr recv_wr = { | |
1626 | .wr_id = RDMA_WRID_RECV_CONTROL + idx, | |
1627 | .sg_list = &sge, | |
1628 | .num_sge = 1, | |
1629 | }; | |
1630 | ||
1631 | ||
1632 | if (ibv_post_recv(rdma->qp, &recv_wr, &bad_wr)) { | |
1633 | return -1; | |
1634 | } | |
1635 | ||
1636 | return 0; | |
1637 | } | |
1638 | ||
1639 | /* | |
1640 | * Block and wait for a RECV control channel message to arrive. | |
1641 | */ | |
1642 | static int qemu_rdma_exchange_get_response(RDMAContext *rdma, | |
1643 | RDMAControlHeader *head, int expecting, int idx) | |
1644 | { | |
88571882 IY |
1645 | uint32_t byte_len; |
1646 | int ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RECV_CONTROL + idx, | |
1647 | &byte_len); | |
2da776db MH |
1648 | |
1649 | if (ret < 0) { | |
733252de | 1650 | error_report("rdma migration: recv polling control error!"); |
2da776db MH |
1651 | return ret; |
1652 | } | |
1653 | ||
1654 | network_to_control((void *) rdma->wr_data[idx].control); | |
1655 | memcpy(head, rdma->wr_data[idx].control, sizeof(RDMAControlHeader)); | |
1656 | ||
733252de | 1657 | trace_qemu_rdma_exchange_get_response_start(control_desc[expecting]); |
2da776db MH |
1658 | |
1659 | if (expecting == RDMA_CONTROL_NONE) { | |
733252de DDAG |
1660 | trace_qemu_rdma_exchange_get_response_none(control_desc[head->type], |
1661 | head->type); | |
2da776db | 1662 | } else if (head->type != expecting || head->type == RDMA_CONTROL_ERROR) { |
733252de DDAG |
1663 | error_report("Was expecting a %s (%d) control message" |
1664 | ", but got: %s (%d), length: %d", | |
2da776db MH |
1665 | control_desc[expecting], expecting, |
1666 | control_desc[head->type], head->type, head->len); | |
1667 | return -EIO; | |
1668 | } | |
6f1484ed | 1669 | if (head->len > RDMA_CONTROL_MAX_BUFFER - sizeof(*head)) { |
81b07353 | 1670 | error_report("too long length: %d", head->len); |
6f1484ed IY |
1671 | return -EINVAL; |
1672 | } | |
88571882 | 1673 | if (sizeof(*head) + head->len != byte_len) { |
733252de | 1674 | error_report("Malformed length: %d byte_len %d", head->len, byte_len); |
88571882 IY |
1675 | return -EINVAL; |
1676 | } | |
2da776db MH |
1677 | |
1678 | return 0; | |
1679 | } | |
1680 | ||
1681 | /* | |
1682 | * When a RECV work request has completed, the work request's | |
1683 | * buffer is pointed at the header. | |
1684 | * | |
1685 | * This will advance the pointer to the data portion | |
1686 | * of the control message of the work request's buffer that | |
1687 | * was populated after the work request finished. | |
1688 | */ | |
1689 | static void qemu_rdma_move_header(RDMAContext *rdma, int idx, | |
1690 | RDMAControlHeader *head) | |
1691 | { | |
1692 | rdma->wr_data[idx].control_len = head->len; | |
1693 | rdma->wr_data[idx].control_curr = | |
1694 | rdma->wr_data[idx].control + sizeof(RDMAControlHeader); | |
1695 | } | |
1696 | ||
1697 | /* | |
1698 | * This is an 'atomic' high-level operation to deliver a single, unified | |
1699 | * control-channel message. | |
1700 | * | |
1701 | * Additionally, if the user is expecting some kind of reply to this message, | |
1702 | * they can request a 'resp' response message be filled in by posting an | |
1703 | * additional work request on behalf of the user and waiting for an additional | |
1704 | * completion. | |
1705 | * | |
1706 | * The extra (optional) response is used during registration to us from having | |
1707 | * to perform an *additional* exchange of message just to provide a response by | |
1708 | * instead piggy-backing on the acknowledgement. | |
1709 | */ | |
1710 | static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, | |
1711 | uint8_t *data, RDMAControlHeader *resp, | |
1712 | int *resp_idx, | |
1713 | int (*callback)(RDMAContext *rdma)) | |
1714 | { | |
1715 | int ret = 0; | |
1716 | ||
1717 | /* | |
1718 | * Wait until the dest is ready before attempting to deliver the message | |
1719 | * by waiting for a READY message. | |
1720 | */ | |
1721 | if (rdma->control_ready_expected) { | |
1722 | RDMAControlHeader resp; | |
1723 | ret = qemu_rdma_exchange_get_response(rdma, | |
1724 | &resp, RDMA_CONTROL_READY, RDMA_WRID_READY); | |
1725 | if (ret < 0) { | |
1726 | return ret; | |
1727 | } | |
1728 | } | |
1729 | ||
1730 | /* | |
1731 | * If the user is expecting a response, post a WR in anticipation of it. | |
1732 | */ | |
1733 | if (resp) { | |
1734 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_DATA); | |
1735 | if (ret) { | |
733252de | 1736 | error_report("rdma migration: error posting" |
2da776db MH |
1737 | " extra control recv for anticipated result!"); |
1738 | return ret; | |
1739 | } | |
1740 | } | |
1741 | ||
1742 | /* | |
1743 | * Post a WR to replace the one we just consumed for the READY message. | |
1744 | */ | |
1745 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); | |
1746 | if (ret) { | |
733252de | 1747 | error_report("rdma migration: error posting first control recv!"); |
2da776db MH |
1748 | return ret; |
1749 | } | |
1750 | ||
1751 | /* | |
1752 | * Deliver the control message that was requested. | |
1753 | */ | |
1754 | ret = qemu_rdma_post_send_control(rdma, data, head); | |
1755 | ||
1756 | if (ret < 0) { | |
733252de | 1757 | error_report("Failed to send control buffer!"); |
2da776db MH |
1758 | return ret; |
1759 | } | |
1760 | ||
1761 | /* | |
1762 | * If we're expecting a response, block and wait for it. | |
1763 | */ | |
1764 | if (resp) { | |
1765 | if (callback) { | |
733252de | 1766 | trace_qemu_rdma_exchange_send_issue_callback(); |
2da776db MH |
1767 | ret = callback(rdma); |
1768 | if (ret < 0) { | |
1769 | return ret; | |
1770 | } | |
1771 | } | |
1772 | ||
733252de | 1773 | trace_qemu_rdma_exchange_send_waiting(control_desc[resp->type]); |
2da776db MH |
1774 | ret = qemu_rdma_exchange_get_response(rdma, resp, |
1775 | resp->type, RDMA_WRID_DATA); | |
1776 | ||
1777 | if (ret < 0) { | |
1778 | return ret; | |
1779 | } | |
1780 | ||
1781 | qemu_rdma_move_header(rdma, RDMA_WRID_DATA, resp); | |
1782 | if (resp_idx) { | |
1783 | *resp_idx = RDMA_WRID_DATA; | |
1784 | } | |
733252de | 1785 | trace_qemu_rdma_exchange_send_received(control_desc[resp->type]); |
2da776db MH |
1786 | } |
1787 | ||
1788 | rdma->control_ready_expected = 1; | |
1789 | ||
1790 | return 0; | |
1791 | } | |
1792 | ||
1793 | /* | |
1794 | * This is an 'atomic' high-level operation to receive a single, unified | |
1795 | * control-channel message. | |
1796 | */ | |
1797 | static int qemu_rdma_exchange_recv(RDMAContext *rdma, RDMAControlHeader *head, | |
1798 | int expecting) | |
1799 | { | |
1800 | RDMAControlHeader ready = { | |
1801 | .len = 0, | |
1802 | .type = RDMA_CONTROL_READY, | |
1803 | .repeat = 1, | |
1804 | }; | |
1805 | int ret; | |
1806 | ||
1807 | /* | |
1808 | * Inform the source that we're ready to receive a message. | |
1809 | */ | |
1810 | ret = qemu_rdma_post_send_control(rdma, NULL, &ready); | |
1811 | ||
1812 | if (ret < 0) { | |
733252de | 1813 | error_report("Failed to send control buffer!"); |
2da776db MH |
1814 | return ret; |
1815 | } | |
1816 | ||
1817 | /* | |
1818 | * Block and wait for the message. | |
1819 | */ | |
1820 | ret = qemu_rdma_exchange_get_response(rdma, head, | |
1821 | expecting, RDMA_WRID_READY); | |
1822 | ||
1823 | if (ret < 0) { | |
1824 | return ret; | |
1825 | } | |
1826 | ||
1827 | qemu_rdma_move_header(rdma, RDMA_WRID_READY, head); | |
1828 | ||
1829 | /* | |
1830 | * Post a new RECV work request to replace the one we just consumed. | |
1831 | */ | |
1832 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); | |
1833 | if (ret) { | |
733252de | 1834 | error_report("rdma migration: error posting second control recv!"); |
2da776db MH |
1835 | return ret; |
1836 | } | |
1837 | ||
1838 | return 0; | |
1839 | } | |
1840 | ||
1841 | /* | |
1842 | * Write an actual chunk of memory using RDMA. | |
1843 | * | |
1844 | * If we're using dynamic registration on the dest-side, we have to | |
1845 | * send a registration command first. | |
1846 | */ | |
1847 | static int qemu_rdma_write_one(QEMUFile *f, RDMAContext *rdma, | |
1848 | int current_index, uint64_t current_addr, | |
1849 | uint64_t length) | |
1850 | { | |
1851 | struct ibv_sge sge; | |
1852 | struct ibv_send_wr send_wr = { 0 }; | |
1853 | struct ibv_send_wr *bad_wr; | |
1854 | int reg_result_idx, ret, count = 0; | |
1855 | uint64_t chunk, chunks; | |
1856 | uint8_t *chunk_start, *chunk_end; | |
1857 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[current_index]); | |
1858 | RDMARegister reg; | |
1859 | RDMARegisterResult *reg_result; | |
1860 | RDMAControlHeader resp = { .type = RDMA_CONTROL_REGISTER_RESULT }; | |
1861 | RDMAControlHeader head = { .len = sizeof(RDMARegister), | |
1862 | .type = RDMA_CONTROL_REGISTER_REQUEST, | |
1863 | .repeat = 1, | |
1864 | }; | |
1865 | ||
1866 | retry: | |
fbce8c25 | 1867 | sge.addr = (uintptr_t)(block->local_host_addr + |
2da776db MH |
1868 | (current_addr - block->offset)); |
1869 | sge.length = length; | |
1870 | ||
fbce8c25 SW |
1871 | chunk = ram_chunk_index(block->local_host_addr, |
1872 | (uint8_t *)(uintptr_t)sge.addr); | |
2da776db MH |
1873 | chunk_start = ram_chunk_start(block, chunk); |
1874 | ||
1875 | if (block->is_ram_block) { | |
1876 | chunks = length / (1UL << RDMA_REG_CHUNK_SHIFT); | |
1877 | ||
1878 | if (chunks && ((length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { | |
1879 | chunks--; | |
1880 | } | |
1881 | } else { | |
1882 | chunks = block->length / (1UL << RDMA_REG_CHUNK_SHIFT); | |
1883 | ||
1884 | if (chunks && ((block->length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { | |
1885 | chunks--; | |
1886 | } | |
1887 | } | |
1888 | ||
733252de DDAG |
1889 | trace_qemu_rdma_write_one_top(chunks + 1, |
1890 | (chunks + 1) * | |
1891 | (1UL << RDMA_REG_CHUNK_SHIFT) / 1024 / 1024); | |
2da776db MH |
1892 | |
1893 | chunk_end = ram_chunk_end(block, chunk + chunks); | |
1894 | ||
1895 | if (!rdma->pin_all) { | |
1896 | #ifdef RDMA_UNREGISTRATION_EXAMPLE | |
1897 | qemu_rdma_unregister_waiting(rdma); | |
1898 | #endif | |
1899 | } | |
1900 | ||
1901 | while (test_bit(chunk, block->transit_bitmap)) { | |
1902 | (void)count; | |
733252de | 1903 | trace_qemu_rdma_write_one_block(count++, current_index, chunk, |
2da776db MH |
1904 | sge.addr, length, rdma->nb_sent, block->nb_chunks); |
1905 | ||
88571882 | 1906 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL); |
2da776db MH |
1907 | |
1908 | if (ret < 0) { | |
733252de | 1909 | error_report("Failed to Wait for previous write to complete " |
2da776db | 1910 | "block %d chunk %" PRIu64 |
733252de | 1911 | " current %" PRIu64 " len %" PRIu64 " %d", |
2da776db MH |
1912 | current_index, chunk, sge.addr, length, rdma->nb_sent); |
1913 | return ret; | |
1914 | } | |
1915 | } | |
1916 | ||
1917 | if (!rdma->pin_all || !block->is_ram_block) { | |
1918 | if (!block->remote_keys[chunk]) { | |
1919 | /* | |
1920 | * This chunk has not yet been registered, so first check to see | |
1921 | * if the entire chunk is zero. If so, tell the other size to | |
1922 | * memset() + madvise() the entire chunk without RDMA. | |
1923 | */ | |
1924 | ||
fbce8c25 SW |
1925 | if (can_use_buffer_find_nonzero_offset((void *)(uintptr_t)sge.addr, |
1926 | length) | |
1927 | && buffer_find_nonzero_offset((void *)(uintptr_t)sge.addr, | |
2da776db MH |
1928 | length) == length) { |
1929 | RDMACompress comp = { | |
1930 | .offset = current_addr, | |
1931 | .value = 0, | |
1932 | .block_idx = current_index, | |
1933 | .length = length, | |
1934 | }; | |
1935 | ||
1936 | head.len = sizeof(comp); | |
1937 | head.type = RDMA_CONTROL_COMPRESS; | |
1938 | ||
733252de DDAG |
1939 | trace_qemu_rdma_write_one_zero(chunk, sge.length, |
1940 | current_index, current_addr); | |
2da776db | 1941 | |
b12f7777 | 1942 | compress_to_network(rdma, &comp); |
2da776db MH |
1943 | ret = qemu_rdma_exchange_send(rdma, &head, |
1944 | (uint8_t *) &comp, NULL, NULL, NULL); | |
1945 | ||
1946 | if (ret < 0) { | |
1947 | return -EIO; | |
1948 | } | |
1949 | ||
1950 | acct_update_position(f, sge.length, true); | |
1951 | ||
1952 | return 1; | |
1953 | } | |
1954 | ||
1955 | /* | |
1956 | * Otherwise, tell other side to register. | |
1957 | */ | |
1958 | reg.current_index = current_index; | |
1959 | if (block->is_ram_block) { | |
1960 | reg.key.current_addr = current_addr; | |
1961 | } else { | |
1962 | reg.key.chunk = chunk; | |
1963 | } | |
1964 | reg.chunks = chunks; | |
1965 | ||
733252de DDAG |
1966 | trace_qemu_rdma_write_one_sendreg(chunk, sge.length, current_index, |
1967 | current_addr); | |
2da776db | 1968 | |
b12f7777 | 1969 | register_to_network(rdma, ®); |
2da776db MH |
1970 | ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
1971 | &resp, ®_result_idx, NULL); | |
1972 | if (ret < 0) { | |
1973 | return ret; | |
1974 | } | |
1975 | ||
1976 | /* try to overlap this single registration with the one we sent. */ | |
3ac040c0 | 1977 | if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr, |
2da776db MH |
1978 | &sge.lkey, NULL, chunk, |
1979 | chunk_start, chunk_end)) { | |
733252de | 1980 | error_report("cannot get lkey"); |
2da776db MH |
1981 | return -EINVAL; |
1982 | } | |
1983 | ||
1984 | reg_result = (RDMARegisterResult *) | |
1985 | rdma->wr_data[reg_result_idx].control_curr; | |
1986 | ||
1987 | network_to_result(reg_result); | |
1988 | ||
733252de DDAG |
1989 | trace_qemu_rdma_write_one_recvregres(block->remote_keys[chunk], |
1990 | reg_result->rkey, chunk); | |
2da776db MH |
1991 | |
1992 | block->remote_keys[chunk] = reg_result->rkey; | |
1993 | block->remote_host_addr = reg_result->host_addr; | |
1994 | } else { | |
1995 | /* already registered before */ | |
3ac040c0 | 1996 | if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr, |
2da776db MH |
1997 | &sge.lkey, NULL, chunk, |
1998 | chunk_start, chunk_end)) { | |
733252de | 1999 | error_report("cannot get lkey!"); |
2da776db MH |
2000 | return -EINVAL; |
2001 | } | |
2002 | } | |
2003 | ||
2004 | send_wr.wr.rdma.rkey = block->remote_keys[chunk]; | |
2005 | } else { | |
2006 | send_wr.wr.rdma.rkey = block->remote_rkey; | |
2007 | ||
3ac040c0 | 2008 | if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr, |
2da776db MH |
2009 | &sge.lkey, NULL, chunk, |
2010 | chunk_start, chunk_end)) { | |
733252de | 2011 | error_report("cannot get lkey!"); |
2da776db MH |
2012 | return -EINVAL; |
2013 | } | |
2014 | } | |
2015 | ||
2016 | /* | |
2017 | * Encode the ram block index and chunk within this wrid. | |
2018 | * We will use this information at the time of completion | |
2019 | * to figure out which bitmap to check against and then which | |
2020 | * chunk in the bitmap to look for. | |
2021 | */ | |
2022 | send_wr.wr_id = qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE, | |
2023 | current_index, chunk); | |
2024 | ||
2025 | send_wr.opcode = IBV_WR_RDMA_WRITE; | |
2026 | send_wr.send_flags = IBV_SEND_SIGNALED; | |
2027 | send_wr.sg_list = &sge; | |
2028 | send_wr.num_sge = 1; | |
2029 | send_wr.wr.rdma.remote_addr = block->remote_host_addr + | |
2030 | (current_addr - block->offset); | |
2031 | ||
733252de DDAG |
2032 | trace_qemu_rdma_write_one_post(chunk, sge.addr, send_wr.wr.rdma.remote_addr, |
2033 | sge.length); | |
2da776db MH |
2034 | |
2035 | /* | |
2036 | * ibv_post_send() does not return negative error numbers, | |
2037 | * per the specification they are positive - no idea why. | |
2038 | */ | |
2039 | ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr); | |
2040 | ||
2041 | if (ret == ENOMEM) { | |
733252de | 2042 | trace_qemu_rdma_write_one_queue_full(); |
88571882 | 2043 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL); |
2da776db | 2044 | if (ret < 0) { |
733252de DDAG |
2045 | error_report("rdma migration: failed to make " |
2046 | "room in full send queue! %d", ret); | |
2da776db MH |
2047 | return ret; |
2048 | } | |
2049 | ||
2050 | goto retry; | |
2051 | ||
2052 | } else if (ret > 0) { | |
2053 | perror("rdma migration: post rdma write failed"); | |
2054 | return -ret; | |
2055 | } | |
2056 | ||
2057 | set_bit(chunk, block->transit_bitmap); | |
2058 | acct_update_position(f, sge.length, false); | |
2059 | rdma->total_writes++; | |
2060 | ||
2061 | return 0; | |
2062 | } | |
2063 | ||
2064 | /* | |
2065 | * Push out any unwritten RDMA operations. | |
2066 | * | |
2067 | * We support sending out multiple chunks at the same time. | |
2068 | * Not all of them need to get signaled in the completion queue. | |
2069 | */ | |
2070 | static int qemu_rdma_write_flush(QEMUFile *f, RDMAContext *rdma) | |
2071 | { | |
2072 | int ret; | |
2073 | ||
2074 | if (!rdma->current_length) { | |
2075 | return 0; | |
2076 | } | |
2077 | ||
2078 | ret = qemu_rdma_write_one(f, rdma, | |
2079 | rdma->current_index, rdma->current_addr, rdma->current_length); | |
2080 | ||
2081 | if (ret < 0) { | |
2082 | return ret; | |
2083 | } | |
2084 | ||
2085 | if (ret == 0) { | |
2086 | rdma->nb_sent++; | |
733252de | 2087 | trace_qemu_rdma_write_flush(rdma->nb_sent); |
2da776db MH |
2088 | } |
2089 | ||
2090 | rdma->current_length = 0; | |
2091 | rdma->current_addr = 0; | |
2092 | ||
2093 | return 0; | |
2094 | } | |
2095 | ||
2096 | static inline int qemu_rdma_buffer_mergable(RDMAContext *rdma, | |
2097 | uint64_t offset, uint64_t len) | |
2098 | { | |
44b59494 IY |
2099 | RDMALocalBlock *block; |
2100 | uint8_t *host_addr; | |
2101 | uint8_t *chunk_end; | |
2102 | ||
2103 | if (rdma->current_index < 0) { | |
2104 | return 0; | |
2105 | } | |
2106 | ||
2107 | if (rdma->current_chunk < 0) { | |
2108 | return 0; | |
2109 | } | |
2110 | ||
2111 | block = &(rdma->local_ram_blocks.block[rdma->current_index]); | |
2112 | host_addr = block->local_host_addr + (offset - block->offset); | |
2113 | chunk_end = ram_chunk_end(block, rdma->current_chunk); | |
2da776db MH |
2114 | |
2115 | if (rdma->current_length == 0) { | |
2116 | return 0; | |
2117 | } | |
2118 | ||
2119 | /* | |
2120 | * Only merge into chunk sequentially. | |
2121 | */ | |
2122 | if (offset != (rdma->current_addr + rdma->current_length)) { | |
2123 | return 0; | |
2124 | } | |
2125 | ||
2da776db MH |
2126 | if (offset < block->offset) { |
2127 | return 0; | |
2128 | } | |
2129 | ||
2130 | if ((offset + len) > (block->offset + block->length)) { | |
2131 | return 0; | |
2132 | } | |
2133 | ||
2da776db MH |
2134 | if ((host_addr + len) > chunk_end) { |
2135 | return 0; | |
2136 | } | |
2137 | ||
2138 | return 1; | |
2139 | } | |
2140 | ||
2141 | /* | |
2142 | * We're not actually writing here, but doing three things: | |
2143 | * | |
2144 | * 1. Identify the chunk the buffer belongs to. | |
2145 | * 2. If the chunk is full or the buffer doesn't belong to the current | |
2146 | * chunk, then start a new chunk and flush() the old chunk. | |
2147 | * 3. To keep the hardware busy, we also group chunks into batches | |
2148 | * and only require that a batch gets acknowledged in the completion | |
2149 | * qeueue instead of each individual chunk. | |
2150 | */ | |
2151 | static int qemu_rdma_write(QEMUFile *f, RDMAContext *rdma, | |
2152 | uint64_t block_offset, uint64_t offset, | |
2153 | uint64_t len) | |
2154 | { | |
2155 | uint64_t current_addr = block_offset + offset; | |
2156 | uint64_t index = rdma->current_index; | |
2157 | uint64_t chunk = rdma->current_chunk; | |
2158 | int ret; | |
2159 | ||
2160 | /* If we cannot merge it, we flush the current buffer first. */ | |
2161 | if (!qemu_rdma_buffer_mergable(rdma, current_addr, len)) { | |
2162 | ret = qemu_rdma_write_flush(f, rdma); | |
2163 | if (ret) { | |
2164 | return ret; | |
2165 | } | |
2166 | rdma->current_length = 0; | |
2167 | rdma->current_addr = current_addr; | |
2168 | ||
2169 | ret = qemu_rdma_search_ram_block(rdma, block_offset, | |
2170 | offset, len, &index, &chunk); | |
2171 | if (ret) { | |
733252de | 2172 | error_report("ram block search failed"); |
2da776db MH |
2173 | return ret; |
2174 | } | |
2175 | rdma->current_index = index; | |
2176 | rdma->current_chunk = chunk; | |
2177 | } | |
2178 | ||
2179 | /* merge it */ | |
2180 | rdma->current_length += len; | |
2181 | ||
2182 | /* flush it if buffer is too large */ | |
2183 | if (rdma->current_length >= RDMA_MERGE_MAX) { | |
2184 | return qemu_rdma_write_flush(f, rdma); | |
2185 | } | |
2186 | ||
2187 | return 0; | |
2188 | } | |
2189 | ||
2190 | static void qemu_rdma_cleanup(RDMAContext *rdma) | |
2191 | { | |
2192 | struct rdma_cm_event *cm_event; | |
2193 | int ret, idx; | |
2194 | ||
5a91337c | 2195 | if (rdma->cm_id && rdma->connected) { |
2da776db MH |
2196 | if (rdma->error_state) { |
2197 | RDMAControlHeader head = { .len = 0, | |
2198 | .type = RDMA_CONTROL_ERROR, | |
2199 | .repeat = 1, | |
2200 | }; | |
733252de | 2201 | error_report("Early error. Sending error."); |
2da776db MH |
2202 | qemu_rdma_post_send_control(rdma, NULL, &head); |
2203 | } | |
2204 | ||
2205 | ret = rdma_disconnect(rdma->cm_id); | |
2206 | if (!ret) { | |
733252de | 2207 | trace_qemu_rdma_cleanup_waiting_for_disconnect(); |
2da776db MH |
2208 | ret = rdma_get_cm_event(rdma->channel, &cm_event); |
2209 | if (!ret) { | |
2210 | rdma_ack_cm_event(cm_event); | |
2211 | } | |
2212 | } | |
733252de | 2213 | trace_qemu_rdma_cleanup_disconnect(); |
5a91337c | 2214 | rdma->connected = false; |
2da776db MH |
2215 | } |
2216 | ||
a97270ad DDAG |
2217 | g_free(rdma->dest_blocks); |
2218 | rdma->dest_blocks = NULL; | |
2da776db | 2219 | |
1f22364b | 2220 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2221 | if (rdma->wr_data[idx].control_mr) { |
2222 | rdma->total_registrations--; | |
2223 | ibv_dereg_mr(rdma->wr_data[idx].control_mr); | |
2224 | } | |
2225 | rdma->wr_data[idx].control_mr = NULL; | |
2226 | } | |
2227 | ||
2228 | if (rdma->local_ram_blocks.block) { | |
2229 | while (rdma->local_ram_blocks.nb_blocks) { | |
03fcab38 | 2230 | rdma_delete_block(rdma, &rdma->local_ram_blocks.block[0]); |
2da776db MH |
2231 | } |
2232 | } | |
2233 | ||
80b262e1 PR |
2234 | if (rdma->qp) { |
2235 | rdma_destroy_qp(rdma->cm_id); | |
2236 | rdma->qp = NULL; | |
2237 | } | |
2da776db MH |
2238 | if (rdma->cq) { |
2239 | ibv_destroy_cq(rdma->cq); | |
2240 | rdma->cq = NULL; | |
2241 | } | |
2242 | if (rdma->comp_channel) { | |
2243 | ibv_destroy_comp_channel(rdma->comp_channel); | |
2244 | rdma->comp_channel = NULL; | |
2245 | } | |
2246 | if (rdma->pd) { | |
2247 | ibv_dealloc_pd(rdma->pd); | |
2248 | rdma->pd = NULL; | |
2249 | } | |
2da776db MH |
2250 | if (rdma->cm_id) { |
2251 | rdma_destroy_id(rdma->cm_id); | |
2252 | rdma->cm_id = NULL; | |
2253 | } | |
80b262e1 PR |
2254 | if (rdma->listen_id) { |
2255 | rdma_destroy_id(rdma->listen_id); | |
2256 | rdma->listen_id = NULL; | |
2257 | } | |
2da776db MH |
2258 | if (rdma->channel) { |
2259 | rdma_destroy_event_channel(rdma->channel); | |
2260 | rdma->channel = NULL; | |
2261 | } | |
e1d0fb37 IY |
2262 | g_free(rdma->host); |
2263 | rdma->host = NULL; | |
2da776db MH |
2264 | } |
2265 | ||
2266 | ||
2267 | static int qemu_rdma_source_init(RDMAContext *rdma, Error **errp, bool pin_all) | |
2268 | { | |
2269 | int ret, idx; | |
2270 | Error *local_err = NULL, **temp = &local_err; | |
2271 | ||
2272 | /* | |
2273 | * Will be validated against destination's actual capabilities | |
2274 | * after the connect() completes. | |
2275 | */ | |
2276 | rdma->pin_all = pin_all; | |
2277 | ||
2278 | ret = qemu_rdma_resolve_host(rdma, temp); | |
2279 | if (ret) { | |
2280 | goto err_rdma_source_init; | |
2281 | } | |
2282 | ||
2283 | ret = qemu_rdma_alloc_pd_cq(rdma); | |
2284 | if (ret) { | |
2285 | ERROR(temp, "rdma migration: error allocating pd and cq! Your mlock()" | |
2286 | " limits may be too low. Please check $ ulimit -a # and " | |
66988941 | 2287 | "search for 'ulimit -l' in the output"); |
2da776db MH |
2288 | goto err_rdma_source_init; |
2289 | } | |
2290 | ||
2291 | ret = qemu_rdma_alloc_qp(rdma); | |
2292 | if (ret) { | |
66988941 | 2293 | ERROR(temp, "rdma migration: error allocating qp!"); |
2da776db MH |
2294 | goto err_rdma_source_init; |
2295 | } | |
2296 | ||
2297 | ret = qemu_rdma_init_ram_blocks(rdma); | |
2298 | if (ret) { | |
66988941 | 2299 | ERROR(temp, "rdma migration: error initializing ram blocks!"); |
2da776db MH |
2300 | goto err_rdma_source_init; |
2301 | } | |
2302 | ||
1f22364b | 2303 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2304 | ret = qemu_rdma_reg_control(rdma, idx); |
2305 | if (ret) { | |
66988941 | 2306 | ERROR(temp, "rdma migration: error registering %d control!", |
2da776db MH |
2307 | idx); |
2308 | goto err_rdma_source_init; | |
2309 | } | |
2310 | } | |
2311 | ||
2312 | return 0; | |
2313 | ||
2314 | err_rdma_source_init: | |
2315 | error_propagate(errp, local_err); | |
2316 | qemu_rdma_cleanup(rdma); | |
2317 | return -1; | |
2318 | } | |
2319 | ||
2320 | static int qemu_rdma_connect(RDMAContext *rdma, Error **errp) | |
2321 | { | |
2322 | RDMACapabilities cap = { | |
2323 | .version = RDMA_CONTROL_VERSION_CURRENT, | |
2324 | .flags = 0, | |
2325 | }; | |
2326 | struct rdma_conn_param conn_param = { .initiator_depth = 2, | |
2327 | .retry_count = 5, | |
2328 | .private_data = &cap, | |
2329 | .private_data_len = sizeof(cap), | |
2330 | }; | |
2331 | struct rdma_cm_event *cm_event; | |
2332 | int ret; | |
2333 | ||
2334 | /* | |
2335 | * Only negotiate the capability with destination if the user | |
2336 | * on the source first requested the capability. | |
2337 | */ | |
2338 | if (rdma->pin_all) { | |
733252de | 2339 | trace_qemu_rdma_connect_pin_all_requested(); |
2da776db MH |
2340 | cap.flags |= RDMA_CAPABILITY_PIN_ALL; |
2341 | } | |
2342 | ||
2343 | caps_to_network(&cap); | |
2344 | ||
2345 | ret = rdma_connect(rdma->cm_id, &conn_param); | |
2346 | if (ret) { | |
2347 | perror("rdma_connect"); | |
66988941 | 2348 | ERROR(errp, "connecting to destination!"); |
2da776db MH |
2349 | goto err_rdma_source_connect; |
2350 | } | |
2351 | ||
2352 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
2353 | if (ret) { | |
2354 | perror("rdma_get_cm_event after rdma_connect"); | |
66988941 | 2355 | ERROR(errp, "connecting to destination!"); |
2da776db | 2356 | rdma_ack_cm_event(cm_event); |
2da776db MH |
2357 | goto err_rdma_source_connect; |
2358 | } | |
2359 | ||
2360 | if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) { | |
2361 | perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect"); | |
66988941 | 2362 | ERROR(errp, "connecting to destination!"); |
2da776db | 2363 | rdma_ack_cm_event(cm_event); |
2da776db MH |
2364 | goto err_rdma_source_connect; |
2365 | } | |
5a91337c | 2366 | rdma->connected = true; |
2da776db MH |
2367 | |
2368 | memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap)); | |
2369 | network_to_caps(&cap); | |
2370 | ||
2371 | /* | |
2372 | * Verify that the *requested* capabilities are supported by the destination | |
2373 | * and disable them otherwise. | |
2374 | */ | |
2375 | if (rdma->pin_all && !(cap.flags & RDMA_CAPABILITY_PIN_ALL)) { | |
2376 | ERROR(errp, "Server cannot support pinning all memory. " | |
66988941 | 2377 | "Will register memory dynamically."); |
2da776db MH |
2378 | rdma->pin_all = false; |
2379 | } | |
2380 | ||
733252de | 2381 | trace_qemu_rdma_connect_pin_all_outcome(rdma->pin_all); |
2da776db MH |
2382 | |
2383 | rdma_ack_cm_event(cm_event); | |
2384 | ||
87772639 | 2385 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
2da776db | 2386 | if (ret) { |
66988941 | 2387 | ERROR(errp, "posting second control recv!"); |
2da776db MH |
2388 | goto err_rdma_source_connect; |
2389 | } | |
2390 | ||
2391 | rdma->control_ready_expected = 1; | |
2392 | rdma->nb_sent = 0; | |
2393 | return 0; | |
2394 | ||
2395 | err_rdma_source_connect: | |
2396 | qemu_rdma_cleanup(rdma); | |
2397 | return -1; | |
2398 | } | |
2399 | ||
2400 | static int qemu_rdma_dest_init(RDMAContext *rdma, Error **errp) | |
2401 | { | |
1dbd2fd9 | 2402 | int ret, idx; |
2da776db MH |
2403 | struct rdma_cm_id *listen_id; |
2404 | char ip[40] = "unknown"; | |
1dbd2fd9 | 2405 | struct rdma_addrinfo *res, *e; |
b58c8552 | 2406 | char port_str[16]; |
2da776db | 2407 | |
1f22364b | 2408 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2409 | rdma->wr_data[idx].control_len = 0; |
2410 | rdma->wr_data[idx].control_curr = NULL; | |
2411 | } | |
2412 | ||
1dbd2fd9 | 2413 | if (!rdma->host || !rdma->host[0]) { |
66988941 | 2414 | ERROR(errp, "RDMA host is not set!"); |
2da776db MH |
2415 | rdma->error_state = -EINVAL; |
2416 | return -1; | |
2417 | } | |
2418 | /* create CM channel */ | |
2419 | rdma->channel = rdma_create_event_channel(); | |
2420 | if (!rdma->channel) { | |
66988941 | 2421 | ERROR(errp, "could not create rdma event channel"); |
2da776db MH |
2422 | rdma->error_state = -EINVAL; |
2423 | return -1; | |
2424 | } | |
2425 | ||
2426 | /* create CM id */ | |
2427 | ret = rdma_create_id(rdma->channel, &listen_id, NULL, RDMA_PS_TCP); | |
2428 | if (ret) { | |
66988941 | 2429 | ERROR(errp, "could not create cm_id!"); |
2da776db MH |
2430 | goto err_dest_init_create_listen_id; |
2431 | } | |
2432 | ||
b58c8552 MH |
2433 | snprintf(port_str, 16, "%d", rdma->port); |
2434 | port_str[15] = '\0'; | |
2da776db | 2435 | |
1dbd2fd9 MT |
2436 | ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res); |
2437 | if (ret < 0) { | |
2438 | ERROR(errp, "could not rdma_getaddrinfo address %s", rdma->host); | |
2439 | goto err_dest_init_bind_addr; | |
2440 | } | |
6470215b | 2441 | |
1dbd2fd9 MT |
2442 | for (e = res; e != NULL; e = e->ai_next) { |
2443 | inet_ntop(e->ai_family, | |
2444 | &((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip); | |
2445 | trace_qemu_rdma_dest_init_trying(rdma->host, ip); | |
2446 | ret = rdma_bind_addr(listen_id, e->ai_dst_addr); | |
2447 | if (ret) { | |
2448 | continue; | |
2da776db | 2449 | } |
1dbd2fd9 MT |
2450 | if (e->ai_family == AF_INET6) { |
2451 | ret = qemu_rdma_broken_ipv6_kernel(errp, listen_id->verbs); | |
2452 | if (ret) { | |
2453 | continue; | |
6470215b MH |
2454 | } |
2455 | } | |
1dbd2fd9 MT |
2456 | break; |
2457 | } | |
b58c8552 | 2458 | |
1dbd2fd9 | 2459 | if (!e) { |
6470215b MH |
2460 | ERROR(errp, "Error: could not rdma_bind_addr!"); |
2461 | goto err_dest_init_bind_addr; | |
2da776db | 2462 | } |
2da776db MH |
2463 | |
2464 | rdma->listen_id = listen_id; | |
2465 | qemu_rdma_dump_gid("dest_init", listen_id); | |
2466 | return 0; | |
2467 | ||
2468 | err_dest_init_bind_addr: | |
2469 | rdma_destroy_id(listen_id); | |
2470 | err_dest_init_create_listen_id: | |
2471 | rdma_destroy_event_channel(rdma->channel); | |
2472 | rdma->channel = NULL; | |
2473 | rdma->error_state = ret; | |
2474 | return ret; | |
2475 | ||
2476 | } | |
2477 | ||
2478 | static void *qemu_rdma_data_init(const char *host_port, Error **errp) | |
2479 | { | |
2480 | RDMAContext *rdma = NULL; | |
2481 | InetSocketAddress *addr; | |
2482 | ||
2483 | if (host_port) { | |
2484 | rdma = g_malloc0(sizeof(RDMAContext)); | |
2da776db MH |
2485 | rdma->current_index = -1; |
2486 | rdma->current_chunk = -1; | |
2487 | ||
2488 | addr = inet_parse(host_port, NULL); | |
2489 | if (addr != NULL) { | |
2490 | rdma->port = atoi(addr->port); | |
2491 | rdma->host = g_strdup(addr->host); | |
2492 | } else { | |
2493 | ERROR(errp, "bad RDMA migration address '%s'", host_port); | |
2494 | g_free(rdma); | |
e325b49a | 2495 | rdma = NULL; |
2da776db | 2496 | } |
e325b49a MH |
2497 | |
2498 | qapi_free_InetSocketAddress(addr); | |
2da776db MH |
2499 | } |
2500 | ||
2501 | return rdma; | |
2502 | } | |
2503 | ||
2504 | /* | |
2505 | * QEMUFile interface to the control channel. | |
2506 | * SEND messages for control only. | |
971ae6ef | 2507 | * VM's ram is handled with regular RDMA messages. |
2da776db MH |
2508 | */ |
2509 | static int qemu_rdma_put_buffer(void *opaque, const uint8_t *buf, | |
2510 | int64_t pos, int size) | |
2511 | { | |
2512 | QEMUFileRDMA *r = opaque; | |
2513 | QEMUFile *f = r->file; | |
2514 | RDMAContext *rdma = r->rdma; | |
2515 | size_t remaining = size; | |
2516 | uint8_t * data = (void *) buf; | |
2517 | int ret; | |
2518 | ||
2519 | CHECK_ERROR_STATE(); | |
2520 | ||
2521 | /* | |
2522 | * Push out any writes that | |
971ae6ef | 2523 | * we're queued up for VM's ram. |
2da776db MH |
2524 | */ |
2525 | ret = qemu_rdma_write_flush(f, rdma); | |
2526 | if (ret < 0) { | |
2527 | rdma->error_state = ret; | |
2528 | return ret; | |
2529 | } | |
2530 | ||
2531 | while (remaining) { | |
2532 | RDMAControlHeader head; | |
2533 | ||
2534 | r->len = MIN(remaining, RDMA_SEND_INCREMENT); | |
2535 | remaining -= r->len; | |
2536 | ||
2537 | head.len = r->len; | |
2538 | head.type = RDMA_CONTROL_QEMU_FILE; | |
2539 | ||
2540 | ret = qemu_rdma_exchange_send(rdma, &head, data, NULL, NULL, NULL); | |
2541 | ||
2542 | if (ret < 0) { | |
2543 | rdma->error_state = ret; | |
2544 | return ret; | |
2545 | } | |
2546 | ||
2547 | data += r->len; | |
2548 | } | |
2549 | ||
2550 | return size; | |
2551 | } | |
2552 | ||
2553 | static size_t qemu_rdma_fill(RDMAContext *rdma, uint8_t *buf, | |
2554 | int size, int idx) | |
2555 | { | |
2556 | size_t len = 0; | |
2557 | ||
2558 | if (rdma->wr_data[idx].control_len) { | |
733252de | 2559 | trace_qemu_rdma_fill(rdma->wr_data[idx].control_len, size); |
2da776db MH |
2560 | |
2561 | len = MIN(size, rdma->wr_data[idx].control_len); | |
2562 | memcpy(buf, rdma->wr_data[idx].control_curr, len); | |
2563 | rdma->wr_data[idx].control_curr += len; | |
2564 | rdma->wr_data[idx].control_len -= len; | |
2565 | } | |
2566 | ||
2567 | return len; | |
2568 | } | |
2569 | ||
2570 | /* | |
2571 | * QEMUFile interface to the control channel. | |
2572 | * RDMA links don't use bytestreams, so we have to | |
2573 | * return bytes to QEMUFile opportunistically. | |
2574 | */ | |
2575 | static int qemu_rdma_get_buffer(void *opaque, uint8_t *buf, | |
2576 | int64_t pos, int size) | |
2577 | { | |
2578 | QEMUFileRDMA *r = opaque; | |
2579 | RDMAContext *rdma = r->rdma; | |
2580 | RDMAControlHeader head; | |
2581 | int ret = 0; | |
2582 | ||
2583 | CHECK_ERROR_STATE(); | |
2584 | ||
2585 | /* | |
2586 | * First, we hold on to the last SEND message we | |
2587 | * were given and dish out the bytes until we run | |
2588 | * out of bytes. | |
2589 | */ | |
2590 | r->len = qemu_rdma_fill(r->rdma, buf, size, 0); | |
2591 | if (r->len) { | |
2592 | return r->len; | |
2593 | } | |
2594 | ||
2595 | /* | |
2596 | * Once we run out, we block and wait for another | |
2597 | * SEND message to arrive. | |
2598 | */ | |
2599 | ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_QEMU_FILE); | |
2600 | ||
2601 | if (ret < 0) { | |
2602 | rdma->error_state = ret; | |
2603 | return ret; | |
2604 | } | |
2605 | ||
2606 | /* | |
2607 | * SEND was received with new bytes, now try again. | |
2608 | */ | |
2609 | return qemu_rdma_fill(r->rdma, buf, size, 0); | |
2610 | } | |
2611 | ||
2612 | /* | |
2613 | * Block until all the outstanding chunks have been delivered by the hardware. | |
2614 | */ | |
2615 | static int qemu_rdma_drain_cq(QEMUFile *f, RDMAContext *rdma) | |
2616 | { | |
2617 | int ret; | |
2618 | ||
2619 | if (qemu_rdma_write_flush(f, rdma) < 0) { | |
2620 | return -EIO; | |
2621 | } | |
2622 | ||
2623 | while (rdma->nb_sent) { | |
88571882 | 2624 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL); |
2da776db | 2625 | if (ret < 0) { |
733252de | 2626 | error_report("rdma migration: complete polling error!"); |
2da776db MH |
2627 | return -EIO; |
2628 | } | |
2629 | } | |
2630 | ||
2631 | qemu_rdma_unregister_waiting(rdma); | |
2632 | ||
2633 | return 0; | |
2634 | } | |
2635 | ||
2636 | static int qemu_rdma_close(void *opaque) | |
2637 | { | |
733252de | 2638 | trace_qemu_rdma_close(); |
2da776db MH |
2639 | QEMUFileRDMA *r = opaque; |
2640 | if (r->rdma) { | |
2641 | qemu_rdma_cleanup(r->rdma); | |
2642 | g_free(r->rdma); | |
2643 | } | |
2644 | g_free(r); | |
2645 | return 0; | |
2646 | } | |
2647 | ||
2648 | /* | |
2649 | * Parameters: | |
2650 | * @offset == 0 : | |
2651 | * This means that 'block_offset' is a full virtual address that does not | |
2652 | * belong to a RAMBlock of the virtual machine and instead | |
2653 | * represents a private malloc'd memory area that the caller wishes to | |
2654 | * transfer. | |
2655 | * | |
2656 | * @offset != 0 : | |
2657 | * Offset is an offset to be added to block_offset and used | |
2658 | * to also lookup the corresponding RAMBlock. | |
2659 | * | |
2660 | * @size > 0 : | |
2661 | * Initiate an transfer this size. | |
2662 | * | |
2663 | * @size == 0 : | |
2664 | * A 'hint' or 'advice' that means that we wish to speculatively | |
2665 | * and asynchronously unregister this memory. In this case, there is no | |
52f35022 | 2666 | * guarantee that the unregister will actually happen, for example, |
2da776db MH |
2667 | * if the memory is being actively transmitted. Additionally, the memory |
2668 | * may be re-registered at any future time if a write within the same | |
2669 | * chunk was requested again, even if you attempted to unregister it | |
2670 | * here. | |
2671 | * | |
2672 | * @size < 0 : TODO, not yet supported | |
2673 | * Unregister the memory NOW. This means that the caller does not | |
2674 | * expect there to be any future RDMA transfers and we just want to clean | |
2675 | * things up. This is used in case the upper layer owns the memory and | |
2676 | * cannot wait for qemu_fclose() to occur. | |
2677 | * | |
2678 | * @bytes_sent : User-specificed pointer to indicate how many bytes were | |
2679 | * sent. Usually, this will not be more than a few bytes of | |
2680 | * the protocol because most transfers are sent asynchronously. | |
2681 | */ | |
2682 | static size_t qemu_rdma_save_page(QEMUFile *f, void *opaque, | |
2683 | ram_addr_t block_offset, ram_addr_t offset, | |
6e1dea46 | 2684 | size_t size, uint64_t *bytes_sent) |
2da776db MH |
2685 | { |
2686 | QEMUFileRDMA *rfile = opaque; | |
2687 | RDMAContext *rdma = rfile->rdma; | |
2688 | int ret; | |
2689 | ||
2690 | CHECK_ERROR_STATE(); | |
2691 | ||
2692 | qemu_fflush(f); | |
2693 | ||
2694 | if (size > 0) { | |
2695 | /* | |
2696 | * Add this page to the current 'chunk'. If the chunk | |
2697 | * is full, or the page doen't belong to the current chunk, | |
2698 | * an actual RDMA write will occur and a new chunk will be formed. | |
2699 | */ | |
2700 | ret = qemu_rdma_write(f, rdma, block_offset, offset, size); | |
2701 | if (ret < 0) { | |
733252de | 2702 | error_report("rdma migration: write error! %d", ret); |
2da776db MH |
2703 | goto err; |
2704 | } | |
2705 | ||
2706 | /* | |
2707 | * We always return 1 bytes because the RDMA | |
2708 | * protocol is completely asynchronous. We do not yet know | |
2709 | * whether an identified chunk is zero or not because we're | |
2710 | * waiting for other pages to potentially be merged with | |
2711 | * the current chunk. So, we have to call qemu_update_position() | |
2712 | * later on when the actual write occurs. | |
2713 | */ | |
2714 | if (bytes_sent) { | |
2715 | *bytes_sent = 1; | |
2716 | } | |
2717 | } else { | |
2718 | uint64_t index, chunk; | |
2719 | ||
2720 | /* TODO: Change QEMUFileOps prototype to be signed: size_t => long | |
2721 | if (size < 0) { | |
2722 | ret = qemu_rdma_drain_cq(f, rdma); | |
2723 | if (ret < 0) { | |
2724 | fprintf(stderr, "rdma: failed to synchronously drain" | |
2725 | " completion queue before unregistration.\n"); | |
2726 | goto err; | |
2727 | } | |
2728 | } | |
2729 | */ | |
2730 | ||
2731 | ret = qemu_rdma_search_ram_block(rdma, block_offset, | |
2732 | offset, size, &index, &chunk); | |
2733 | ||
2734 | if (ret) { | |
733252de | 2735 | error_report("ram block search failed"); |
2da776db MH |
2736 | goto err; |
2737 | } | |
2738 | ||
2739 | qemu_rdma_signal_unregister(rdma, index, chunk, 0); | |
2740 | ||
2741 | /* | |
52f35022 | 2742 | * TODO: Synchronous, guaranteed unregistration (should not occur during |
2da776db MH |
2743 | * fast-path). Otherwise, unregisters will process on the next call to |
2744 | * qemu_rdma_drain_cq() | |
2745 | if (size < 0) { | |
2746 | qemu_rdma_unregister_waiting(rdma); | |
2747 | } | |
2748 | */ | |
2749 | } | |
2750 | ||
2751 | /* | |
2752 | * Drain the Completion Queue if possible, but do not block, | |
2753 | * just poll. | |
2754 | * | |
2755 | * If nothing to poll, the end of the iteration will do this | |
2756 | * again to make sure we don't overflow the request queue. | |
2757 | */ | |
2758 | while (1) { | |
2759 | uint64_t wr_id, wr_id_in; | |
88571882 | 2760 | int ret = qemu_rdma_poll(rdma, &wr_id_in, NULL); |
2da776db | 2761 | if (ret < 0) { |
733252de | 2762 | error_report("rdma migration: polling error! %d", ret); |
2da776db MH |
2763 | goto err; |
2764 | } | |
2765 | ||
2766 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; | |
2767 | ||
2768 | if (wr_id == RDMA_WRID_NONE) { | |
2769 | break; | |
2770 | } | |
2771 | } | |
2772 | ||
2773 | return RAM_SAVE_CONTROL_DELAYED; | |
2774 | err: | |
2775 | rdma->error_state = ret; | |
2776 | return ret; | |
2777 | } | |
2778 | ||
2779 | static int qemu_rdma_accept(RDMAContext *rdma) | |
2780 | { | |
2781 | RDMACapabilities cap; | |
2782 | struct rdma_conn_param conn_param = { | |
2783 | .responder_resources = 2, | |
2784 | .private_data = &cap, | |
2785 | .private_data_len = sizeof(cap), | |
2786 | }; | |
2787 | struct rdma_cm_event *cm_event; | |
2788 | struct ibv_context *verbs; | |
2789 | int ret = -EINVAL; | |
2790 | int idx; | |
2791 | ||
2792 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
2793 | if (ret) { | |
2794 | goto err_rdma_dest_wait; | |
2795 | } | |
2796 | ||
2797 | if (cm_event->event != RDMA_CM_EVENT_CONNECT_REQUEST) { | |
2798 | rdma_ack_cm_event(cm_event); | |
2799 | goto err_rdma_dest_wait; | |
2800 | } | |
2801 | ||
2802 | memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap)); | |
2803 | ||
2804 | network_to_caps(&cap); | |
2805 | ||
2806 | if (cap.version < 1 || cap.version > RDMA_CONTROL_VERSION_CURRENT) { | |
733252de | 2807 | error_report("Unknown source RDMA version: %d, bailing...", |
2da776db MH |
2808 | cap.version); |
2809 | rdma_ack_cm_event(cm_event); | |
2810 | goto err_rdma_dest_wait; | |
2811 | } | |
2812 | ||
2813 | /* | |
2814 | * Respond with only the capabilities this version of QEMU knows about. | |
2815 | */ | |
2816 | cap.flags &= known_capabilities; | |
2817 | ||
2818 | /* | |
2819 | * Enable the ones that we do know about. | |
2820 | * Add other checks here as new ones are introduced. | |
2821 | */ | |
2822 | if (cap.flags & RDMA_CAPABILITY_PIN_ALL) { | |
2823 | rdma->pin_all = true; | |
2824 | } | |
2825 | ||
2826 | rdma->cm_id = cm_event->id; | |
2827 | verbs = cm_event->id->verbs; | |
2828 | ||
2829 | rdma_ack_cm_event(cm_event); | |
2830 | ||
733252de | 2831 | trace_qemu_rdma_accept_pin_state(rdma->pin_all); |
2da776db MH |
2832 | |
2833 | caps_to_network(&cap); | |
2834 | ||
733252de | 2835 | trace_qemu_rdma_accept_pin_verbsc(verbs); |
2da776db MH |
2836 | |
2837 | if (!rdma->verbs) { | |
2838 | rdma->verbs = verbs; | |
2839 | } else if (rdma->verbs != verbs) { | |
733252de DDAG |
2840 | error_report("ibv context not matching %p, %p!", rdma->verbs, |
2841 | verbs); | |
2da776db MH |
2842 | goto err_rdma_dest_wait; |
2843 | } | |
2844 | ||
2845 | qemu_rdma_dump_id("dest_init", verbs); | |
2846 | ||
2847 | ret = qemu_rdma_alloc_pd_cq(rdma); | |
2848 | if (ret) { | |
733252de | 2849 | error_report("rdma migration: error allocating pd and cq!"); |
2da776db MH |
2850 | goto err_rdma_dest_wait; |
2851 | } | |
2852 | ||
2853 | ret = qemu_rdma_alloc_qp(rdma); | |
2854 | if (ret) { | |
733252de | 2855 | error_report("rdma migration: error allocating qp!"); |
2da776db MH |
2856 | goto err_rdma_dest_wait; |
2857 | } | |
2858 | ||
2859 | ret = qemu_rdma_init_ram_blocks(rdma); | |
2860 | if (ret) { | |
733252de | 2861 | error_report("rdma migration: error initializing ram blocks!"); |
2da776db MH |
2862 | goto err_rdma_dest_wait; |
2863 | } | |
2864 | ||
1f22364b | 2865 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2866 | ret = qemu_rdma_reg_control(rdma, idx); |
2867 | if (ret) { | |
733252de | 2868 | error_report("rdma: error registering %d control", idx); |
2da776db MH |
2869 | goto err_rdma_dest_wait; |
2870 | } | |
2871 | } | |
2872 | ||
82e1cc4b | 2873 | qemu_set_fd_handler(rdma->channel->fd, NULL, NULL, NULL); |
2da776db MH |
2874 | |
2875 | ret = rdma_accept(rdma->cm_id, &conn_param); | |
2876 | if (ret) { | |
733252de | 2877 | error_report("rdma_accept returns %d", ret); |
2da776db MH |
2878 | goto err_rdma_dest_wait; |
2879 | } | |
2880 | ||
2881 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
2882 | if (ret) { | |
733252de | 2883 | error_report("rdma_accept get_cm_event failed %d", ret); |
2da776db MH |
2884 | goto err_rdma_dest_wait; |
2885 | } | |
2886 | ||
2887 | if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) { | |
733252de | 2888 | error_report("rdma_accept not event established"); |
2da776db MH |
2889 | rdma_ack_cm_event(cm_event); |
2890 | goto err_rdma_dest_wait; | |
2891 | } | |
2892 | ||
2893 | rdma_ack_cm_event(cm_event); | |
5a91337c | 2894 | rdma->connected = true; |
2da776db | 2895 | |
87772639 | 2896 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
2da776db | 2897 | if (ret) { |
733252de | 2898 | error_report("rdma migration: error posting second control recv"); |
2da776db MH |
2899 | goto err_rdma_dest_wait; |
2900 | } | |
2901 | ||
2902 | qemu_rdma_dump_gid("dest_connect", rdma->cm_id); | |
2903 | ||
2904 | return 0; | |
2905 | ||
2906 | err_rdma_dest_wait: | |
2907 | rdma->error_state = ret; | |
2908 | qemu_rdma_cleanup(rdma); | |
2909 | return ret; | |
2910 | } | |
2911 | ||
2912 | /* | |
2913 | * During each iteration of the migration, we listen for instructions | |
2914 | * by the source VM to perform dynamic page registrations before they | |
2915 | * can perform RDMA operations. | |
2916 | * | |
2917 | * We respond with the 'rkey'. | |
2918 | * | |
2919 | * Keep doing this until the source tells us to stop. | |
2920 | */ | |
632e3a5c | 2921 | static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque) |
2da776db MH |
2922 | { |
2923 | RDMAControlHeader reg_resp = { .len = sizeof(RDMARegisterResult), | |
2924 | .type = RDMA_CONTROL_REGISTER_RESULT, | |
2925 | .repeat = 0, | |
2926 | }; | |
2927 | RDMAControlHeader unreg_resp = { .len = 0, | |
2928 | .type = RDMA_CONTROL_UNREGISTER_FINISHED, | |
2929 | .repeat = 0, | |
2930 | }; | |
2931 | RDMAControlHeader blocks = { .type = RDMA_CONTROL_RAM_BLOCKS_RESULT, | |
2932 | .repeat = 1 }; | |
2933 | QEMUFileRDMA *rfile = opaque; | |
2934 | RDMAContext *rdma = rfile->rdma; | |
2935 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
2936 | RDMAControlHeader head; | |
2937 | RDMARegister *reg, *registers; | |
2938 | RDMACompress *comp; | |
2939 | RDMARegisterResult *reg_result; | |
2940 | static RDMARegisterResult results[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE]; | |
2941 | RDMALocalBlock *block; | |
2942 | void *host_addr; | |
2943 | int ret = 0; | |
2944 | int idx = 0; | |
2945 | int count = 0; | |
2946 | int i = 0; | |
2947 | ||
2948 | CHECK_ERROR_STATE(); | |
2949 | ||
2950 | do { | |
632e3a5c | 2951 | trace_qemu_rdma_registration_handle_wait(); |
2da776db MH |
2952 | |
2953 | ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_NONE); | |
2954 | ||
2955 | if (ret < 0) { | |
2956 | break; | |
2957 | } | |
2958 | ||
2959 | if (head.repeat > RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE) { | |
733252de DDAG |
2960 | error_report("rdma: Too many requests in this message (%d)." |
2961 | "Bailing.", head.repeat); | |
2da776db MH |
2962 | ret = -EIO; |
2963 | break; | |
2964 | } | |
2965 | ||
2966 | switch (head.type) { | |
2967 | case RDMA_CONTROL_COMPRESS: | |
2968 | comp = (RDMACompress *) rdma->wr_data[idx].control_curr; | |
2969 | network_to_compress(comp); | |
2970 | ||
733252de DDAG |
2971 | trace_qemu_rdma_registration_handle_compress(comp->length, |
2972 | comp->block_idx, | |
2973 | comp->offset); | |
2da776db MH |
2974 | block = &(rdma->local_ram_blocks.block[comp->block_idx]); |
2975 | ||
2976 | host_addr = block->local_host_addr + | |
2977 | (comp->offset - block->offset); | |
2978 | ||
2979 | ram_handle_compressed(host_addr, comp->value, comp->length); | |
2980 | break; | |
2981 | ||
2982 | case RDMA_CONTROL_REGISTER_FINISHED: | |
733252de | 2983 | trace_qemu_rdma_registration_handle_finished(); |
2da776db MH |
2984 | goto out; |
2985 | ||
2986 | case RDMA_CONTROL_RAM_BLOCKS_REQUEST: | |
733252de | 2987 | trace_qemu_rdma_registration_handle_ram_blocks(); |
2da776db MH |
2988 | |
2989 | if (rdma->pin_all) { | |
2990 | ret = qemu_rdma_reg_whole_ram_blocks(rdma); | |
2991 | if (ret) { | |
733252de DDAG |
2992 | error_report("rdma migration: error dest " |
2993 | "registering ram blocks"); | |
2da776db MH |
2994 | goto out; |
2995 | } | |
2996 | } | |
2997 | ||
2998 | /* | |
2999 | * Dest uses this to prepare to transmit the RAMBlock descriptions | |
3000 | * to the source VM after connection setup. | |
3001 | * Both sides use the "remote" structure to communicate and update | |
3002 | * their "local" descriptions with what was sent. | |
3003 | */ | |
3004 | for (i = 0; i < local->nb_blocks; i++) { | |
a97270ad | 3005 | rdma->dest_blocks[i].remote_host_addr = |
fbce8c25 | 3006 | (uintptr_t)(local->block[i].local_host_addr); |
2da776db MH |
3007 | |
3008 | if (rdma->pin_all) { | |
a97270ad | 3009 | rdma->dest_blocks[i].remote_rkey = local->block[i].mr->rkey; |
2da776db MH |
3010 | } |
3011 | ||
a97270ad DDAG |
3012 | rdma->dest_blocks[i].offset = local->block[i].offset; |
3013 | rdma->dest_blocks[i].length = local->block[i].length; | |
2da776db | 3014 | |
a97270ad | 3015 | dest_block_to_network(&rdma->dest_blocks[i]); |
2da776db MH |
3016 | } |
3017 | ||
3018 | blocks.len = rdma->local_ram_blocks.nb_blocks | |
a97270ad | 3019 | * sizeof(RDMADestBlock); |
2da776db MH |
3020 | |
3021 | ||
3022 | ret = qemu_rdma_post_send_control(rdma, | |
a97270ad | 3023 | (uint8_t *) rdma->dest_blocks, &blocks); |
2da776db MH |
3024 | |
3025 | if (ret < 0) { | |
733252de | 3026 | error_report("rdma migration: error sending remote info"); |
2da776db MH |
3027 | goto out; |
3028 | } | |
3029 | ||
3030 | break; | |
3031 | case RDMA_CONTROL_REGISTER_REQUEST: | |
733252de | 3032 | trace_qemu_rdma_registration_handle_register(head.repeat); |
2da776db MH |
3033 | |
3034 | reg_resp.repeat = head.repeat; | |
3035 | registers = (RDMARegister *) rdma->wr_data[idx].control_curr; | |
3036 | ||
3037 | for (count = 0; count < head.repeat; count++) { | |
3038 | uint64_t chunk; | |
3039 | uint8_t *chunk_start, *chunk_end; | |
3040 | ||
3041 | reg = ®isters[count]; | |
3042 | network_to_register(reg); | |
3043 | ||
3044 | reg_result = &results[count]; | |
3045 | ||
733252de | 3046 | trace_qemu_rdma_registration_handle_register_loop(count, |
2da776db MH |
3047 | reg->current_index, reg->key.current_addr, reg->chunks); |
3048 | ||
3049 | block = &(rdma->local_ram_blocks.block[reg->current_index]); | |
3050 | if (block->is_ram_block) { | |
3051 | host_addr = (block->local_host_addr + | |
3052 | (reg->key.current_addr - block->offset)); | |
3053 | chunk = ram_chunk_index(block->local_host_addr, | |
3054 | (uint8_t *) host_addr); | |
3055 | } else { | |
3056 | chunk = reg->key.chunk; | |
3057 | host_addr = block->local_host_addr + | |
3058 | (reg->key.chunk * (1UL << RDMA_REG_CHUNK_SHIFT)); | |
3059 | } | |
3060 | chunk_start = ram_chunk_start(block, chunk); | |
3061 | chunk_end = ram_chunk_end(block, chunk + reg->chunks); | |
3062 | if (qemu_rdma_register_and_get_keys(rdma, block, | |
3ac040c0 | 3063 | (uintptr_t)host_addr, NULL, ®_result->rkey, |
2da776db | 3064 | chunk, chunk_start, chunk_end)) { |
733252de | 3065 | error_report("cannot get rkey"); |
2da776db MH |
3066 | ret = -EINVAL; |
3067 | goto out; | |
3068 | } | |
3069 | ||
fbce8c25 | 3070 | reg_result->host_addr = (uintptr_t)block->local_host_addr; |
2da776db | 3071 | |
733252de DDAG |
3072 | trace_qemu_rdma_registration_handle_register_rkey( |
3073 | reg_result->rkey); | |
2da776db MH |
3074 | |
3075 | result_to_network(reg_result); | |
3076 | } | |
3077 | ||
3078 | ret = qemu_rdma_post_send_control(rdma, | |
3079 | (uint8_t *) results, ®_resp); | |
3080 | ||
3081 | if (ret < 0) { | |
733252de | 3082 | error_report("Failed to send control buffer"); |
2da776db MH |
3083 | goto out; |
3084 | } | |
3085 | break; | |
3086 | case RDMA_CONTROL_UNREGISTER_REQUEST: | |
733252de | 3087 | trace_qemu_rdma_registration_handle_unregister(head.repeat); |
2da776db MH |
3088 | unreg_resp.repeat = head.repeat; |
3089 | registers = (RDMARegister *) rdma->wr_data[idx].control_curr; | |
3090 | ||
3091 | for (count = 0; count < head.repeat; count++) { | |
3092 | reg = ®isters[count]; | |
3093 | network_to_register(reg); | |
3094 | ||
733252de DDAG |
3095 | trace_qemu_rdma_registration_handle_unregister_loop(count, |
3096 | reg->current_index, reg->key.chunk); | |
2da776db MH |
3097 | |
3098 | block = &(rdma->local_ram_blocks.block[reg->current_index]); | |
3099 | ||
3100 | ret = ibv_dereg_mr(block->pmr[reg->key.chunk]); | |
3101 | block->pmr[reg->key.chunk] = NULL; | |
3102 | ||
3103 | if (ret != 0) { | |
3104 | perror("rdma unregistration chunk failed"); | |
3105 | ret = -ret; | |
3106 | goto out; | |
3107 | } | |
3108 | ||
3109 | rdma->total_registrations--; | |
3110 | ||
733252de DDAG |
3111 | trace_qemu_rdma_registration_handle_unregister_success( |
3112 | reg->key.chunk); | |
2da776db MH |
3113 | } |
3114 | ||
3115 | ret = qemu_rdma_post_send_control(rdma, NULL, &unreg_resp); | |
3116 | ||
3117 | if (ret < 0) { | |
733252de | 3118 | error_report("Failed to send control buffer"); |
2da776db MH |
3119 | goto out; |
3120 | } | |
3121 | break; | |
3122 | case RDMA_CONTROL_REGISTER_RESULT: | |
733252de | 3123 | error_report("Invalid RESULT message at dest."); |
2da776db MH |
3124 | ret = -EIO; |
3125 | goto out; | |
3126 | default: | |
733252de | 3127 | error_report("Unknown control message %s", control_desc[head.type]); |
2da776db MH |
3128 | ret = -EIO; |
3129 | goto out; | |
3130 | } | |
3131 | } while (1); | |
3132 | out: | |
3133 | if (ret < 0) { | |
3134 | rdma->error_state = ret; | |
3135 | } | |
3136 | return ret; | |
3137 | } | |
3138 | ||
632e3a5c DDAG |
3139 | static int rdma_load_hook(QEMUFile *f, void *opaque, uint64_t flags, void *data) |
3140 | { | |
3141 | switch (flags) { | |
3142 | case RAM_CONTROL_BLOCK_REG: | |
3143 | /* TODO A later patch */ | |
3144 | return 0; | |
3145 | break; | |
3146 | ||
3147 | case RAM_CONTROL_HOOK: | |
3148 | return qemu_rdma_registration_handle(f, opaque); | |
3149 | ||
3150 | default: | |
3151 | /* Shouldn't be called with any other values */ | |
3152 | abort(); | |
3153 | } | |
3154 | } | |
3155 | ||
2da776db | 3156 | static int qemu_rdma_registration_start(QEMUFile *f, void *opaque, |
632e3a5c | 3157 | uint64_t flags, void *data) |
2da776db MH |
3158 | { |
3159 | QEMUFileRDMA *rfile = opaque; | |
3160 | RDMAContext *rdma = rfile->rdma; | |
3161 | ||
3162 | CHECK_ERROR_STATE(); | |
3163 | ||
733252de | 3164 | trace_qemu_rdma_registration_start(flags); |
2da776db MH |
3165 | qemu_put_be64(f, RAM_SAVE_FLAG_HOOK); |
3166 | qemu_fflush(f); | |
3167 | ||
3168 | return 0; | |
3169 | } | |
3170 | ||
3171 | /* | |
3172 | * Inform dest that dynamic registrations are done for now. | |
3173 | * First, flush writes, if any. | |
3174 | */ | |
3175 | static int qemu_rdma_registration_stop(QEMUFile *f, void *opaque, | |
632e3a5c | 3176 | uint64_t flags, void *data) |
2da776db MH |
3177 | { |
3178 | Error *local_err = NULL, **errp = &local_err; | |
3179 | QEMUFileRDMA *rfile = opaque; | |
3180 | RDMAContext *rdma = rfile->rdma; | |
3181 | RDMAControlHeader head = { .len = 0, .repeat = 1 }; | |
3182 | int ret = 0; | |
3183 | ||
3184 | CHECK_ERROR_STATE(); | |
3185 | ||
3186 | qemu_fflush(f); | |
3187 | ret = qemu_rdma_drain_cq(f, rdma); | |
3188 | ||
3189 | if (ret < 0) { | |
3190 | goto err; | |
3191 | } | |
3192 | ||
3193 | if (flags == RAM_CONTROL_SETUP) { | |
3194 | RDMAControlHeader resp = {.type = RDMA_CONTROL_RAM_BLOCKS_RESULT }; | |
3195 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
a97270ad | 3196 | int reg_result_idx, i, j, nb_dest_blocks; |
2da776db MH |
3197 | |
3198 | head.type = RDMA_CONTROL_RAM_BLOCKS_REQUEST; | |
733252de | 3199 | trace_qemu_rdma_registration_stop_ram(); |
2da776db MH |
3200 | |
3201 | /* | |
3202 | * Make sure that we parallelize the pinning on both sides. | |
3203 | * For very large guests, doing this serially takes a really | |
3204 | * long time, so we have to 'interleave' the pinning locally | |
3205 | * with the control messages by performing the pinning on this | |
3206 | * side before we receive the control response from the other | |
3207 | * side that the pinning has completed. | |
3208 | */ | |
3209 | ret = qemu_rdma_exchange_send(rdma, &head, NULL, &resp, | |
3210 | ®_result_idx, rdma->pin_all ? | |
3211 | qemu_rdma_reg_whole_ram_blocks : NULL); | |
3212 | if (ret < 0) { | |
66988941 | 3213 | ERROR(errp, "receiving remote info!"); |
2da776db MH |
3214 | return ret; |
3215 | } | |
3216 | ||
a97270ad | 3217 | nb_dest_blocks = resp.len / sizeof(RDMADestBlock); |
2da776db MH |
3218 | |
3219 | /* | |
3220 | * The protocol uses two different sets of rkeys (mutually exclusive): | |
3221 | * 1. One key to represent the virtual address of the entire ram block. | |
3222 | * (dynamic chunk registration disabled - pin everything with one rkey.) | |
3223 | * 2. One to represent individual chunks within a ram block. | |
3224 | * (dynamic chunk registration enabled - pin individual chunks.) | |
3225 | * | |
3226 | * Once the capability is successfully negotiated, the destination transmits | |
3227 | * the keys to use (or sends them later) including the virtual addresses | |
3228 | * and then propagates the remote ram block descriptions to his local copy. | |
3229 | */ | |
3230 | ||
a97270ad | 3231 | if (local->nb_blocks != nb_dest_blocks) { |
2da776db MH |
3232 | ERROR(errp, "ram blocks mismatch #1! " |
3233 | "Your QEMU command line parameters are probably " | |
66988941 | 3234 | "not identical on both the source and destination."); |
2da776db MH |
3235 | return -EINVAL; |
3236 | } | |
3237 | ||
885e8f98 | 3238 | qemu_rdma_move_header(rdma, reg_result_idx, &resp); |
a97270ad | 3239 | memcpy(rdma->dest_blocks, |
885e8f98 | 3240 | rdma->wr_data[reg_result_idx].control_curr, resp.len); |
a97270ad DDAG |
3241 | for (i = 0; i < nb_dest_blocks; i++) { |
3242 | network_to_dest_block(&rdma->dest_blocks[i]); | |
2da776db MH |
3243 | |
3244 | /* search local ram blocks */ | |
3245 | for (j = 0; j < local->nb_blocks; j++) { | |
a97270ad | 3246 | if (rdma->dest_blocks[i].offset != local->block[j].offset) { |
2da776db MH |
3247 | continue; |
3248 | } | |
3249 | ||
a97270ad | 3250 | if (rdma->dest_blocks[i].length != local->block[j].length) { |
2da776db MH |
3251 | ERROR(errp, "ram blocks mismatch #2! " |
3252 | "Your QEMU command line parameters are probably " | |
66988941 | 3253 | "not identical on both the source and destination."); |
2da776db MH |
3254 | return -EINVAL; |
3255 | } | |
3256 | local->block[j].remote_host_addr = | |
a97270ad DDAG |
3257 | rdma->dest_blocks[i].remote_host_addr; |
3258 | local->block[j].remote_rkey = rdma->dest_blocks[i].remote_rkey; | |
2da776db MH |
3259 | break; |
3260 | } | |
3261 | ||
3262 | if (j >= local->nb_blocks) { | |
3263 | ERROR(errp, "ram blocks mismatch #3! " | |
3264 | "Your QEMU command line parameters are probably " | |
66988941 | 3265 | "not identical on both the source and destination."); |
2da776db MH |
3266 | return -EINVAL; |
3267 | } | |
3268 | } | |
3269 | } | |
3270 | ||
733252de | 3271 | trace_qemu_rdma_registration_stop(flags); |
2da776db MH |
3272 | |
3273 | head.type = RDMA_CONTROL_REGISTER_FINISHED; | |
3274 | ret = qemu_rdma_exchange_send(rdma, &head, NULL, NULL, NULL, NULL); | |
3275 | ||
3276 | if (ret < 0) { | |
3277 | goto err; | |
3278 | } | |
3279 | ||
3280 | return 0; | |
3281 | err: | |
3282 | rdma->error_state = ret; | |
3283 | return ret; | |
3284 | } | |
3285 | ||
3286 | static int qemu_rdma_get_fd(void *opaque) | |
3287 | { | |
3288 | QEMUFileRDMA *rfile = opaque; | |
3289 | RDMAContext *rdma = rfile->rdma; | |
3290 | ||
3291 | return rdma->comp_channel->fd; | |
3292 | } | |
3293 | ||
2ae31aea | 3294 | static const QEMUFileOps rdma_read_ops = { |
2da776db MH |
3295 | .get_buffer = qemu_rdma_get_buffer, |
3296 | .get_fd = qemu_rdma_get_fd, | |
3297 | .close = qemu_rdma_close, | |
632e3a5c | 3298 | .hook_ram_load = rdma_load_hook, |
2da776db MH |
3299 | }; |
3300 | ||
2ae31aea | 3301 | static const QEMUFileOps rdma_write_ops = { |
2da776db MH |
3302 | .put_buffer = qemu_rdma_put_buffer, |
3303 | .close = qemu_rdma_close, | |
3304 | .before_ram_iterate = qemu_rdma_registration_start, | |
3305 | .after_ram_iterate = qemu_rdma_registration_stop, | |
3306 | .save_page = qemu_rdma_save_page, | |
3307 | }; | |
3308 | ||
3309 | static void *qemu_fopen_rdma(RDMAContext *rdma, const char *mode) | |
3310 | { | |
728470be | 3311 | QEMUFileRDMA *r; |
2da776db MH |
3312 | |
3313 | if (qemu_file_mode_is_not_valid(mode)) { | |
3314 | return NULL; | |
3315 | } | |
3316 | ||
728470be | 3317 | r = g_malloc0(sizeof(QEMUFileRDMA)); |
2da776db MH |
3318 | r->rdma = rdma; |
3319 | ||
3320 | if (mode[0] == 'w') { | |
3321 | r->file = qemu_fopen_ops(r, &rdma_write_ops); | |
3322 | } else { | |
3323 | r->file = qemu_fopen_ops(r, &rdma_read_ops); | |
3324 | } | |
3325 | ||
3326 | return r->file; | |
3327 | } | |
3328 | ||
3329 | static void rdma_accept_incoming_migration(void *opaque) | |
3330 | { | |
3331 | RDMAContext *rdma = opaque; | |
3332 | int ret; | |
3333 | QEMUFile *f; | |
3334 | Error *local_err = NULL, **errp = &local_err; | |
3335 | ||
24ec68ef | 3336 | trace_qemu_rdma_accept_incoming_migration(); |
2da776db MH |
3337 | ret = qemu_rdma_accept(rdma); |
3338 | ||
3339 | if (ret) { | |
66988941 | 3340 | ERROR(errp, "RDMA Migration initialization failed!"); |
2da776db MH |
3341 | return; |
3342 | } | |
3343 | ||
24ec68ef | 3344 | trace_qemu_rdma_accept_incoming_migration_accepted(); |
2da776db MH |
3345 | |
3346 | f = qemu_fopen_rdma(rdma, "rb"); | |
3347 | if (f == NULL) { | |
66988941 | 3348 | ERROR(errp, "could not qemu_fopen_rdma!"); |
2da776db MH |
3349 | qemu_rdma_cleanup(rdma); |
3350 | return; | |
3351 | } | |
3352 | ||
3353 | rdma->migration_started_on_destination = 1; | |
3354 | process_incoming_migration(f); | |
3355 | } | |
3356 | ||
3357 | void rdma_start_incoming_migration(const char *host_port, Error **errp) | |
3358 | { | |
3359 | int ret; | |
3360 | RDMAContext *rdma; | |
3361 | Error *local_err = NULL; | |
3362 | ||
733252de | 3363 | trace_rdma_start_incoming_migration(); |
2da776db MH |
3364 | rdma = qemu_rdma_data_init(host_port, &local_err); |
3365 | ||
3366 | if (rdma == NULL) { | |
3367 | goto err; | |
3368 | } | |
3369 | ||
3370 | ret = qemu_rdma_dest_init(rdma, &local_err); | |
3371 | ||
3372 | if (ret) { | |
3373 | goto err; | |
3374 | } | |
3375 | ||
733252de | 3376 | trace_rdma_start_incoming_migration_after_dest_init(); |
2da776db MH |
3377 | |
3378 | ret = rdma_listen(rdma->listen_id, 5); | |
3379 | ||
3380 | if (ret) { | |
66988941 | 3381 | ERROR(errp, "listening on socket!"); |
2da776db MH |
3382 | goto err; |
3383 | } | |
3384 | ||
733252de | 3385 | trace_rdma_start_incoming_migration_after_rdma_listen(); |
2da776db | 3386 | |
82e1cc4b FZ |
3387 | qemu_set_fd_handler(rdma->channel->fd, rdma_accept_incoming_migration, |
3388 | NULL, (void *)(intptr_t)rdma); | |
2da776db MH |
3389 | return; |
3390 | err: | |
3391 | error_propagate(errp, local_err); | |
3392 | g_free(rdma); | |
3393 | } | |
3394 | ||
3395 | void rdma_start_outgoing_migration(void *opaque, | |
3396 | const char *host_port, Error **errp) | |
3397 | { | |
3398 | MigrationState *s = opaque; | |
3399 | Error *local_err = NULL, **temp = &local_err; | |
3400 | RDMAContext *rdma = qemu_rdma_data_init(host_port, &local_err); | |
3401 | int ret = 0; | |
3402 | ||
3403 | if (rdma == NULL) { | |
66988941 | 3404 | ERROR(temp, "Failed to initialize RDMA data structures! %d", ret); |
2da776db MH |
3405 | goto err; |
3406 | } | |
3407 | ||
3408 | ret = qemu_rdma_source_init(rdma, &local_err, | |
41310c68 | 3409 | s->enabled_capabilities[MIGRATION_CAPABILITY_RDMA_PIN_ALL]); |
2da776db MH |
3410 | |
3411 | if (ret) { | |
3412 | goto err; | |
3413 | } | |
3414 | ||
733252de | 3415 | trace_rdma_start_outgoing_migration_after_rdma_source_init(); |
2da776db MH |
3416 | ret = qemu_rdma_connect(rdma, &local_err); |
3417 | ||
3418 | if (ret) { | |
3419 | goto err; | |
3420 | } | |
3421 | ||
733252de | 3422 | trace_rdma_start_outgoing_migration_after_rdma_connect(); |
2da776db MH |
3423 | |
3424 | s->file = qemu_fopen_rdma(rdma, "wb"); | |
3425 | migrate_fd_connect(s); | |
3426 | return; | |
3427 | err: | |
3428 | error_propagate(errp, local_err); | |
3429 | g_free(rdma); | |
3430 | migrate_fd_error(s); | |
3431 | } |