]>
Commit | Line | Data |
---|---|---|
2da776db MH |
1 | /* |
2 | * RDMA protocol and interfaces | |
3 | * | |
4 | * Copyright IBM, Corp. 2010-2013 | |
6ddd2d76 | 5 | * Copyright Red Hat, Inc. 2015-2016 |
2da776db MH |
6 | * |
7 | * Authors: | |
8 | * Michael R. Hines <mrhines@us.ibm.com> | |
9 | * Jiuxing Liu <jl@us.ibm.com> | |
6ddd2d76 | 10 | * Daniel P. Berrange <berrange@redhat.com> |
2da776db MH |
11 | * |
12 | * This work is licensed under the terms of the GNU GPL, version 2 or | |
13 | * later. See the COPYING file in the top-level directory. | |
14 | * | |
15 | */ | |
1393a485 | 16 | #include "qemu/osdep.h" |
da34e65c | 17 | #include "qapi/error.h" |
2da776db | 18 | #include "qemu-common.h" |
f348b6d1 | 19 | #include "qemu/cutils.h" |
e1a3ecee | 20 | #include "rdma.h" |
6666c96a | 21 | #include "migration.h" |
08a0aee1 | 22 | #include "qemu-file.h" |
7b1e1a22 | 23 | #include "ram.h" |
40014d81 | 24 | #include "qemu-file-channel.h" |
d49b6836 | 25 | #include "qemu/error-report.h" |
2da776db MH |
26 | #include "qemu/main-loop.h" |
27 | #include "qemu/sockets.h" | |
28 | #include "qemu/bitmap.h" | |
10817bf0 | 29 | #include "qemu/coroutine.h" |
2da776db MH |
30 | #include <sys/socket.h> |
31 | #include <netdb.h> | |
32 | #include <arpa/inet.h> | |
2da776db | 33 | #include <rdma/rdma_cma.h> |
733252de | 34 | #include "trace.h" |
2da776db MH |
35 | |
36 | /* | |
37 | * Print and error on both the Monitor and the Log file. | |
38 | */ | |
39 | #define ERROR(errp, fmt, ...) \ | |
40 | do { \ | |
66988941 | 41 | fprintf(stderr, "RDMA ERROR: " fmt "\n", ## __VA_ARGS__); \ |
2da776db MH |
42 | if (errp && (*(errp) == NULL)) { \ |
43 | error_setg(errp, "RDMA ERROR: " fmt, ## __VA_ARGS__); \ | |
44 | } \ | |
45 | } while (0) | |
46 | ||
47 | #define RDMA_RESOLVE_TIMEOUT_MS 10000 | |
48 | ||
49 | /* Do not merge data if larger than this. */ | |
50 | #define RDMA_MERGE_MAX (2 * 1024 * 1024) | |
51 | #define RDMA_SIGNALED_SEND_MAX (RDMA_MERGE_MAX / 4096) | |
52 | ||
53 | #define RDMA_REG_CHUNK_SHIFT 20 /* 1 MB */ | |
54 | ||
55 | /* | |
56 | * This is only for non-live state being migrated. | |
57 | * Instead of RDMA_WRITE messages, we use RDMA_SEND | |
58 | * messages for that state, which requires a different | |
59 | * delivery design than main memory. | |
60 | */ | |
61 | #define RDMA_SEND_INCREMENT 32768 | |
62 | ||
63 | /* | |
64 | * Maximum size infiniband SEND message | |
65 | */ | |
66 | #define RDMA_CONTROL_MAX_BUFFER (512 * 1024) | |
67 | #define RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE 4096 | |
68 | ||
69 | #define RDMA_CONTROL_VERSION_CURRENT 1 | |
70 | /* | |
71 | * Capabilities for negotiation. | |
72 | */ | |
73 | #define RDMA_CAPABILITY_PIN_ALL 0x01 | |
74 | ||
75 | /* | |
76 | * Add the other flags above to this list of known capabilities | |
77 | * as they are introduced. | |
78 | */ | |
79 | static uint32_t known_capabilities = RDMA_CAPABILITY_PIN_ALL; | |
80 | ||
81 | #define CHECK_ERROR_STATE() \ | |
82 | do { \ | |
83 | if (rdma->error_state) { \ | |
84 | if (!rdma->error_reported) { \ | |
733252de DDAG |
85 | error_report("RDMA is in an error state waiting migration" \ |
86 | " to abort!"); \ | |
2da776db MH |
87 | rdma->error_reported = 1; \ |
88 | } \ | |
74637e6f | 89 | rcu_read_unlock(); \ |
2da776db MH |
90 | return rdma->error_state; \ |
91 | } \ | |
2562755e | 92 | } while (0) |
2da776db MH |
93 | |
94 | /* | |
95 | * A work request ID is 64-bits and we split up these bits | |
96 | * into 3 parts: | |
97 | * | |
98 | * bits 0-15 : type of control message, 2^16 | |
99 | * bits 16-29: ram block index, 2^14 | |
100 | * bits 30-63: ram block chunk number, 2^34 | |
101 | * | |
102 | * The last two bit ranges are only used for RDMA writes, | |
103 | * in order to track their completion and potentially | |
104 | * also track unregistration status of the message. | |
105 | */ | |
106 | #define RDMA_WRID_TYPE_SHIFT 0UL | |
107 | #define RDMA_WRID_BLOCK_SHIFT 16UL | |
108 | #define RDMA_WRID_CHUNK_SHIFT 30UL | |
109 | ||
110 | #define RDMA_WRID_TYPE_MASK \ | |
111 | ((1UL << RDMA_WRID_BLOCK_SHIFT) - 1UL) | |
112 | ||
113 | #define RDMA_WRID_BLOCK_MASK \ | |
114 | (~RDMA_WRID_TYPE_MASK & ((1UL << RDMA_WRID_CHUNK_SHIFT) - 1UL)) | |
115 | ||
116 | #define RDMA_WRID_CHUNK_MASK (~RDMA_WRID_BLOCK_MASK & ~RDMA_WRID_TYPE_MASK) | |
117 | ||
118 | /* | |
119 | * RDMA migration protocol: | |
120 | * 1. RDMA Writes (data messages, i.e. RAM) | |
121 | * 2. IB Send/Recv (control channel messages) | |
122 | */ | |
123 | enum { | |
124 | RDMA_WRID_NONE = 0, | |
125 | RDMA_WRID_RDMA_WRITE = 1, | |
126 | RDMA_WRID_SEND_CONTROL = 2000, | |
127 | RDMA_WRID_RECV_CONTROL = 4000, | |
128 | }; | |
129 | ||
2ae31aea | 130 | static const char *wrid_desc[] = { |
2da776db MH |
131 | [RDMA_WRID_NONE] = "NONE", |
132 | [RDMA_WRID_RDMA_WRITE] = "WRITE RDMA", | |
133 | [RDMA_WRID_SEND_CONTROL] = "CONTROL SEND", | |
134 | [RDMA_WRID_RECV_CONTROL] = "CONTROL RECV", | |
135 | }; | |
136 | ||
137 | /* | |
138 | * Work request IDs for IB SEND messages only (not RDMA writes). | |
139 | * This is used by the migration protocol to transmit | |
140 | * control messages (such as device state and registration commands) | |
141 | * | |
142 | * We could use more WRs, but we have enough for now. | |
143 | */ | |
144 | enum { | |
145 | RDMA_WRID_READY = 0, | |
146 | RDMA_WRID_DATA, | |
147 | RDMA_WRID_CONTROL, | |
148 | RDMA_WRID_MAX, | |
149 | }; | |
150 | ||
151 | /* | |
152 | * SEND/RECV IB Control Messages. | |
153 | */ | |
154 | enum { | |
155 | RDMA_CONTROL_NONE = 0, | |
156 | RDMA_CONTROL_ERROR, | |
157 | RDMA_CONTROL_READY, /* ready to receive */ | |
158 | RDMA_CONTROL_QEMU_FILE, /* QEMUFile-transmitted bytes */ | |
159 | RDMA_CONTROL_RAM_BLOCKS_REQUEST, /* RAMBlock synchronization */ | |
160 | RDMA_CONTROL_RAM_BLOCKS_RESULT, /* RAMBlock synchronization */ | |
161 | RDMA_CONTROL_COMPRESS, /* page contains repeat values */ | |
162 | RDMA_CONTROL_REGISTER_REQUEST, /* dynamic page registration */ | |
163 | RDMA_CONTROL_REGISTER_RESULT, /* key to use after registration */ | |
164 | RDMA_CONTROL_REGISTER_FINISHED, /* current iteration finished */ | |
165 | RDMA_CONTROL_UNREGISTER_REQUEST, /* dynamic UN-registration */ | |
166 | RDMA_CONTROL_UNREGISTER_FINISHED, /* unpinning finished */ | |
167 | }; | |
168 | ||
2da776db MH |
169 | |
170 | /* | |
171 | * Memory and MR structures used to represent an IB Send/Recv work request. | |
172 | * This is *not* used for RDMA writes, only IB Send/Recv. | |
173 | */ | |
174 | typedef struct { | |
175 | uint8_t control[RDMA_CONTROL_MAX_BUFFER]; /* actual buffer to register */ | |
176 | struct ibv_mr *control_mr; /* registration metadata */ | |
177 | size_t control_len; /* length of the message */ | |
178 | uint8_t *control_curr; /* start of unconsumed bytes */ | |
179 | } RDMAWorkRequestData; | |
180 | ||
181 | /* | |
182 | * Negotiate RDMA capabilities during connection-setup time. | |
183 | */ | |
184 | typedef struct { | |
185 | uint32_t version; | |
186 | uint32_t flags; | |
187 | } RDMACapabilities; | |
188 | ||
189 | static void caps_to_network(RDMACapabilities *cap) | |
190 | { | |
191 | cap->version = htonl(cap->version); | |
192 | cap->flags = htonl(cap->flags); | |
193 | } | |
194 | ||
195 | static void network_to_caps(RDMACapabilities *cap) | |
196 | { | |
197 | cap->version = ntohl(cap->version); | |
198 | cap->flags = ntohl(cap->flags); | |
199 | } | |
200 | ||
201 | /* | |
202 | * Representation of a RAMBlock from an RDMA perspective. | |
203 | * This is not transmitted, only local. | |
204 | * This and subsequent structures cannot be linked lists | |
205 | * because we're using a single IB message to transmit | |
206 | * the information. It's small anyway, so a list is overkill. | |
207 | */ | |
208 | typedef struct RDMALocalBlock { | |
4fb5364b DDAG |
209 | char *block_name; |
210 | uint8_t *local_host_addr; /* local virtual address */ | |
211 | uint64_t remote_host_addr; /* remote virtual address */ | |
212 | uint64_t offset; | |
213 | uint64_t length; | |
214 | struct ibv_mr **pmr; /* MRs for chunk-level registration */ | |
215 | struct ibv_mr *mr; /* MR for non-chunk-level registration */ | |
216 | uint32_t *remote_keys; /* rkeys for chunk-level registration */ | |
217 | uint32_t remote_rkey; /* rkeys for non-chunk-level registration */ | |
218 | int index; /* which block are we */ | |
e4d63320 | 219 | unsigned int src_index; /* (Only used on dest) */ |
4fb5364b DDAG |
220 | bool is_ram_block; |
221 | int nb_chunks; | |
2da776db MH |
222 | unsigned long *transit_bitmap; |
223 | unsigned long *unregister_bitmap; | |
224 | } RDMALocalBlock; | |
225 | ||
226 | /* | |
227 | * Also represents a RAMblock, but only on the dest. | |
228 | * This gets transmitted by the dest during connection-time | |
229 | * to the source VM and then is used to populate the | |
230 | * corresponding RDMALocalBlock with | |
231 | * the information needed to perform the actual RDMA. | |
232 | */ | |
a97270ad | 233 | typedef struct QEMU_PACKED RDMADestBlock { |
2da776db MH |
234 | uint64_t remote_host_addr; |
235 | uint64_t offset; | |
236 | uint64_t length; | |
237 | uint32_t remote_rkey; | |
238 | uint32_t padding; | |
a97270ad | 239 | } RDMADestBlock; |
2da776db | 240 | |
482a33c5 DDAG |
241 | static const char *control_desc(unsigned int rdma_control) |
242 | { | |
243 | static const char *strs[] = { | |
244 | [RDMA_CONTROL_NONE] = "NONE", | |
245 | [RDMA_CONTROL_ERROR] = "ERROR", | |
246 | [RDMA_CONTROL_READY] = "READY", | |
247 | [RDMA_CONTROL_QEMU_FILE] = "QEMU FILE", | |
248 | [RDMA_CONTROL_RAM_BLOCKS_REQUEST] = "RAM BLOCKS REQUEST", | |
249 | [RDMA_CONTROL_RAM_BLOCKS_RESULT] = "RAM BLOCKS RESULT", | |
250 | [RDMA_CONTROL_COMPRESS] = "COMPRESS", | |
251 | [RDMA_CONTROL_REGISTER_REQUEST] = "REGISTER REQUEST", | |
252 | [RDMA_CONTROL_REGISTER_RESULT] = "REGISTER RESULT", | |
253 | [RDMA_CONTROL_REGISTER_FINISHED] = "REGISTER FINISHED", | |
254 | [RDMA_CONTROL_UNREGISTER_REQUEST] = "UNREGISTER REQUEST", | |
255 | [RDMA_CONTROL_UNREGISTER_FINISHED] = "UNREGISTER FINISHED", | |
256 | }; | |
257 | ||
258 | if (rdma_control > RDMA_CONTROL_UNREGISTER_FINISHED) { | |
259 | return "??BAD CONTROL VALUE??"; | |
260 | } | |
261 | ||
262 | return strs[rdma_control]; | |
263 | } | |
264 | ||
2da776db MH |
265 | static uint64_t htonll(uint64_t v) |
266 | { | |
267 | union { uint32_t lv[2]; uint64_t llv; } u; | |
268 | u.lv[0] = htonl(v >> 32); | |
269 | u.lv[1] = htonl(v & 0xFFFFFFFFULL); | |
270 | return u.llv; | |
271 | } | |
272 | ||
273 | static uint64_t ntohll(uint64_t v) { | |
274 | union { uint32_t lv[2]; uint64_t llv; } u; | |
275 | u.llv = v; | |
276 | return ((uint64_t)ntohl(u.lv[0]) << 32) | (uint64_t) ntohl(u.lv[1]); | |
277 | } | |
278 | ||
a97270ad | 279 | static void dest_block_to_network(RDMADestBlock *db) |
2da776db | 280 | { |
a97270ad DDAG |
281 | db->remote_host_addr = htonll(db->remote_host_addr); |
282 | db->offset = htonll(db->offset); | |
283 | db->length = htonll(db->length); | |
284 | db->remote_rkey = htonl(db->remote_rkey); | |
2da776db MH |
285 | } |
286 | ||
a97270ad | 287 | static void network_to_dest_block(RDMADestBlock *db) |
2da776db | 288 | { |
a97270ad DDAG |
289 | db->remote_host_addr = ntohll(db->remote_host_addr); |
290 | db->offset = ntohll(db->offset); | |
291 | db->length = ntohll(db->length); | |
292 | db->remote_rkey = ntohl(db->remote_rkey); | |
2da776db MH |
293 | } |
294 | ||
295 | /* | |
296 | * Virtual address of the above structures used for transmitting | |
297 | * the RAMBlock descriptions at connection-time. | |
298 | * This structure is *not* transmitted. | |
299 | */ | |
300 | typedef struct RDMALocalBlocks { | |
301 | int nb_blocks; | |
302 | bool init; /* main memory init complete */ | |
303 | RDMALocalBlock *block; | |
304 | } RDMALocalBlocks; | |
305 | ||
306 | /* | |
307 | * Main data structure for RDMA state. | |
308 | * While there is only one copy of this structure being allocated right now, | |
309 | * this is the place where one would start if you wanted to consider | |
310 | * having more than one RDMA connection open at the same time. | |
311 | */ | |
312 | typedef struct RDMAContext { | |
313 | char *host; | |
314 | int port; | |
315 | ||
1f22364b | 316 | RDMAWorkRequestData wr_data[RDMA_WRID_MAX]; |
2da776db MH |
317 | |
318 | /* | |
319 | * This is used by *_exchange_send() to figure out whether or not | |
320 | * the initial "READY" message has already been received or not. | |
321 | * This is because other functions may potentially poll() and detect | |
322 | * the READY message before send() does, in which case we need to | |
323 | * know if it completed. | |
324 | */ | |
325 | int control_ready_expected; | |
326 | ||
327 | /* number of outstanding writes */ | |
328 | int nb_sent; | |
329 | ||
330 | /* store info about current buffer so that we can | |
331 | merge it with future sends */ | |
332 | uint64_t current_addr; | |
333 | uint64_t current_length; | |
334 | /* index of ram block the current buffer belongs to */ | |
335 | int current_index; | |
336 | /* index of the chunk in the current ram block */ | |
337 | int current_chunk; | |
338 | ||
339 | bool pin_all; | |
340 | ||
341 | /* | |
342 | * infiniband-specific variables for opening the device | |
343 | * and maintaining connection state and so forth. | |
344 | * | |
345 | * cm_id also has ibv_context, rdma_event_channel, and ibv_qp in | |
346 | * cm_id->verbs, cm_id->channel, and cm_id->qp. | |
347 | */ | |
348 | struct rdma_cm_id *cm_id; /* connection manager ID */ | |
349 | struct rdma_cm_id *listen_id; | |
5a91337c | 350 | bool connected; |
2da776db MH |
351 | |
352 | struct ibv_context *verbs; | |
353 | struct rdma_event_channel *channel; | |
354 | struct ibv_qp *qp; /* queue pair */ | |
355 | struct ibv_comp_channel *comp_channel; /* completion channel */ | |
356 | struct ibv_pd *pd; /* protection domain */ | |
357 | struct ibv_cq *cq; /* completion queue */ | |
358 | ||
359 | /* | |
360 | * If a previous write failed (perhaps because of a failed | |
361 | * memory registration, then do not attempt any future work | |
362 | * and remember the error state. | |
363 | */ | |
364 | int error_state; | |
365 | int error_reported; | |
cd5ea070 | 366 | int received_error; |
2da776db MH |
367 | |
368 | /* | |
369 | * Description of ram blocks used throughout the code. | |
370 | */ | |
371 | RDMALocalBlocks local_ram_blocks; | |
a97270ad | 372 | RDMADestBlock *dest_blocks; |
2da776db | 373 | |
e4d63320 DDAG |
374 | /* Index of the next RAMBlock received during block registration */ |
375 | unsigned int next_src_index; | |
376 | ||
2da776db MH |
377 | /* |
378 | * Migration on *destination* started. | |
379 | * Then use coroutine yield function. | |
380 | * Source runs in a thread, so we don't care. | |
381 | */ | |
382 | int migration_started_on_destination; | |
383 | ||
384 | int total_registrations; | |
385 | int total_writes; | |
386 | ||
387 | int unregister_current, unregister_next; | |
388 | uint64_t unregistrations[RDMA_SIGNALED_SEND_MAX]; | |
389 | ||
390 | GHashTable *blockmap; | |
55cc1b59 LC |
391 | |
392 | /* the RDMAContext for return path */ | |
393 | struct RDMAContext *return_path; | |
394 | bool is_return_path; | |
2da776db MH |
395 | } RDMAContext; |
396 | ||
6ddd2d76 DB |
397 | #define TYPE_QIO_CHANNEL_RDMA "qio-channel-rdma" |
398 | #define QIO_CHANNEL_RDMA(obj) \ | |
399 | OBJECT_CHECK(QIOChannelRDMA, (obj), TYPE_QIO_CHANNEL_RDMA) | |
400 | ||
401 | typedef struct QIOChannelRDMA QIOChannelRDMA; | |
402 | ||
403 | ||
404 | struct QIOChannelRDMA { | |
405 | QIOChannel parent; | |
74637e6f LC |
406 | RDMAContext *rdmain; |
407 | RDMAContext *rdmaout; | |
6ddd2d76 | 408 | QEMUFile *file; |
6ddd2d76 DB |
409 | bool blocking; /* XXX we don't actually honour this yet */ |
410 | }; | |
2da776db MH |
411 | |
412 | /* | |
413 | * Main structure for IB Send/Recv control messages. | |
414 | * This gets prepended at the beginning of every Send/Recv. | |
415 | */ | |
416 | typedef struct QEMU_PACKED { | |
417 | uint32_t len; /* Total length of data portion */ | |
418 | uint32_t type; /* which control command to perform */ | |
419 | uint32_t repeat; /* number of commands in data portion of same type */ | |
420 | uint32_t padding; | |
421 | } RDMAControlHeader; | |
422 | ||
423 | static void control_to_network(RDMAControlHeader *control) | |
424 | { | |
425 | control->type = htonl(control->type); | |
426 | control->len = htonl(control->len); | |
427 | control->repeat = htonl(control->repeat); | |
428 | } | |
429 | ||
430 | static void network_to_control(RDMAControlHeader *control) | |
431 | { | |
432 | control->type = ntohl(control->type); | |
433 | control->len = ntohl(control->len); | |
434 | control->repeat = ntohl(control->repeat); | |
435 | } | |
436 | ||
437 | /* | |
438 | * Register a single Chunk. | |
439 | * Information sent by the source VM to inform the dest | |
440 | * to register an single chunk of memory before we can perform | |
441 | * the actual RDMA operation. | |
442 | */ | |
443 | typedef struct QEMU_PACKED { | |
444 | union QEMU_PACKED { | |
b12f7777 | 445 | uint64_t current_addr; /* offset into the ram_addr_t space */ |
2da776db MH |
446 | uint64_t chunk; /* chunk to lookup if unregistering */ |
447 | } key; | |
448 | uint32_t current_index; /* which ramblock the chunk belongs to */ | |
449 | uint32_t padding; | |
450 | uint64_t chunks; /* how many sequential chunks to register */ | |
451 | } RDMARegister; | |
452 | ||
b12f7777 | 453 | static void register_to_network(RDMAContext *rdma, RDMARegister *reg) |
2da776db | 454 | { |
b12f7777 DDAG |
455 | RDMALocalBlock *local_block; |
456 | local_block = &rdma->local_ram_blocks.block[reg->current_index]; | |
457 | ||
458 | if (local_block->is_ram_block) { | |
459 | /* | |
460 | * current_addr as passed in is an address in the local ram_addr_t | |
461 | * space, we need to translate this for the destination | |
462 | */ | |
463 | reg->key.current_addr -= local_block->offset; | |
464 | reg->key.current_addr += rdma->dest_blocks[reg->current_index].offset; | |
465 | } | |
2da776db MH |
466 | reg->key.current_addr = htonll(reg->key.current_addr); |
467 | reg->current_index = htonl(reg->current_index); | |
468 | reg->chunks = htonll(reg->chunks); | |
469 | } | |
470 | ||
471 | static void network_to_register(RDMARegister *reg) | |
472 | { | |
473 | reg->key.current_addr = ntohll(reg->key.current_addr); | |
474 | reg->current_index = ntohl(reg->current_index); | |
475 | reg->chunks = ntohll(reg->chunks); | |
476 | } | |
477 | ||
478 | typedef struct QEMU_PACKED { | |
479 | uint32_t value; /* if zero, we will madvise() */ | |
480 | uint32_t block_idx; /* which ram block index */ | |
b12f7777 | 481 | uint64_t offset; /* Address in remote ram_addr_t space */ |
2da776db MH |
482 | uint64_t length; /* length of the chunk */ |
483 | } RDMACompress; | |
484 | ||
b12f7777 | 485 | static void compress_to_network(RDMAContext *rdma, RDMACompress *comp) |
2da776db MH |
486 | { |
487 | comp->value = htonl(comp->value); | |
b12f7777 DDAG |
488 | /* |
489 | * comp->offset as passed in is an address in the local ram_addr_t | |
490 | * space, we need to translate this for the destination | |
491 | */ | |
492 | comp->offset -= rdma->local_ram_blocks.block[comp->block_idx].offset; | |
493 | comp->offset += rdma->dest_blocks[comp->block_idx].offset; | |
2da776db MH |
494 | comp->block_idx = htonl(comp->block_idx); |
495 | comp->offset = htonll(comp->offset); | |
496 | comp->length = htonll(comp->length); | |
497 | } | |
498 | ||
499 | static void network_to_compress(RDMACompress *comp) | |
500 | { | |
501 | comp->value = ntohl(comp->value); | |
502 | comp->block_idx = ntohl(comp->block_idx); | |
503 | comp->offset = ntohll(comp->offset); | |
504 | comp->length = ntohll(comp->length); | |
505 | } | |
506 | ||
507 | /* | |
508 | * The result of the dest's memory registration produces an "rkey" | |
509 | * which the source VM must reference in order to perform | |
510 | * the RDMA operation. | |
511 | */ | |
512 | typedef struct QEMU_PACKED { | |
513 | uint32_t rkey; | |
514 | uint32_t padding; | |
515 | uint64_t host_addr; | |
516 | } RDMARegisterResult; | |
517 | ||
518 | static void result_to_network(RDMARegisterResult *result) | |
519 | { | |
520 | result->rkey = htonl(result->rkey); | |
521 | result->host_addr = htonll(result->host_addr); | |
522 | }; | |
523 | ||
524 | static void network_to_result(RDMARegisterResult *result) | |
525 | { | |
526 | result->rkey = ntohl(result->rkey); | |
527 | result->host_addr = ntohll(result->host_addr); | |
528 | }; | |
529 | ||
530 | const char *print_wrid(int wrid); | |
531 | static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, | |
532 | uint8_t *data, RDMAControlHeader *resp, | |
533 | int *resp_idx, | |
534 | int (*callback)(RDMAContext *rdma)); | |
535 | ||
dd286ed7 IY |
536 | static inline uint64_t ram_chunk_index(const uint8_t *start, |
537 | const uint8_t *host) | |
2da776db MH |
538 | { |
539 | return ((uintptr_t) host - (uintptr_t) start) >> RDMA_REG_CHUNK_SHIFT; | |
540 | } | |
541 | ||
dd286ed7 | 542 | static inline uint8_t *ram_chunk_start(const RDMALocalBlock *rdma_ram_block, |
2da776db MH |
543 | uint64_t i) |
544 | { | |
fbce8c25 SW |
545 | return (uint8_t *)(uintptr_t)(rdma_ram_block->local_host_addr + |
546 | (i << RDMA_REG_CHUNK_SHIFT)); | |
2da776db MH |
547 | } |
548 | ||
dd286ed7 IY |
549 | static inline uint8_t *ram_chunk_end(const RDMALocalBlock *rdma_ram_block, |
550 | uint64_t i) | |
2da776db MH |
551 | { |
552 | uint8_t *result = ram_chunk_start(rdma_ram_block, i) + | |
553 | (1UL << RDMA_REG_CHUNK_SHIFT); | |
554 | ||
555 | if (result > (rdma_ram_block->local_host_addr + rdma_ram_block->length)) { | |
556 | result = rdma_ram_block->local_host_addr + rdma_ram_block->length; | |
557 | } | |
558 | ||
559 | return result; | |
560 | } | |
561 | ||
4fb5364b DDAG |
562 | static int rdma_add_block(RDMAContext *rdma, const char *block_name, |
563 | void *host_addr, | |
2da776db MH |
564 | ram_addr_t block_offset, uint64_t length) |
565 | { | |
566 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
760ff4be | 567 | RDMALocalBlock *block; |
2da776db MH |
568 | RDMALocalBlock *old = local->block; |
569 | ||
97f3ad35 | 570 | local->block = g_new0(RDMALocalBlock, local->nb_blocks + 1); |
2da776db MH |
571 | |
572 | if (local->nb_blocks) { | |
573 | int x; | |
574 | ||
760ff4be DDAG |
575 | if (rdma->blockmap) { |
576 | for (x = 0; x < local->nb_blocks; x++) { | |
577 | g_hash_table_remove(rdma->blockmap, | |
578 | (void *)(uintptr_t)old[x].offset); | |
579 | g_hash_table_insert(rdma->blockmap, | |
580 | (void *)(uintptr_t)old[x].offset, | |
581 | &local->block[x]); | |
582 | } | |
2da776db MH |
583 | } |
584 | memcpy(local->block, old, sizeof(RDMALocalBlock) * local->nb_blocks); | |
585 | g_free(old); | |
586 | } | |
587 | ||
588 | block = &local->block[local->nb_blocks]; | |
589 | ||
4fb5364b | 590 | block->block_name = g_strdup(block_name); |
2da776db MH |
591 | block->local_host_addr = host_addr; |
592 | block->offset = block_offset; | |
593 | block->length = length; | |
594 | block->index = local->nb_blocks; | |
e4d63320 | 595 | block->src_index = ~0U; /* Filled in by the receipt of the block list */ |
2da776db MH |
596 | block->nb_chunks = ram_chunk_index(host_addr, host_addr + length) + 1UL; |
597 | block->transit_bitmap = bitmap_new(block->nb_chunks); | |
598 | bitmap_clear(block->transit_bitmap, 0, block->nb_chunks); | |
599 | block->unregister_bitmap = bitmap_new(block->nb_chunks); | |
600 | bitmap_clear(block->unregister_bitmap, 0, block->nb_chunks); | |
97f3ad35 | 601 | block->remote_keys = g_new0(uint32_t, block->nb_chunks); |
2da776db MH |
602 | |
603 | block->is_ram_block = local->init ? false : true; | |
604 | ||
760ff4be | 605 | if (rdma->blockmap) { |
80e60c6e | 606 | g_hash_table_insert(rdma->blockmap, (void *)(uintptr_t)block_offset, block); |
760ff4be | 607 | } |
2da776db | 608 | |
4fb5364b DDAG |
609 | trace_rdma_add_block(block_name, local->nb_blocks, |
610 | (uintptr_t) block->local_host_addr, | |
ba795761 | 611 | block->offset, block->length, |
fbce8c25 | 612 | (uintptr_t) (block->local_host_addr + block->length), |
ba795761 DDAG |
613 | BITS_TO_LONGS(block->nb_chunks) * |
614 | sizeof(unsigned long) * 8, | |
615 | block->nb_chunks); | |
2da776db MH |
616 | |
617 | local->nb_blocks++; | |
618 | ||
619 | return 0; | |
620 | } | |
621 | ||
622 | /* | |
623 | * Memory regions need to be registered with the device and queue pairs setup | |
624 | * in advanced before the migration starts. This tells us where the RAM blocks | |
625 | * are so that we can register them individually. | |
626 | */ | |
e3807054 | 627 | static int qemu_rdma_init_one_block(const char *block_name, void *host_addr, |
2da776db MH |
628 | ram_addr_t block_offset, ram_addr_t length, void *opaque) |
629 | { | |
4fb5364b | 630 | return rdma_add_block(opaque, block_name, host_addr, block_offset, length); |
2da776db MH |
631 | } |
632 | ||
633 | /* | |
634 | * Identify the RAMBlocks and their quantity. They will be references to | |
635 | * identify chunk boundaries inside each RAMBlock and also be referenced | |
636 | * during dynamic page registration. | |
637 | */ | |
638 | static int qemu_rdma_init_ram_blocks(RDMAContext *rdma) | |
639 | { | |
640 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
641 | ||
642 | assert(rdma->blockmap == NULL); | |
2da776db | 643 | memset(local, 0, sizeof *local); |
ff0769a4 | 644 | qemu_ram_foreach_migratable_block(qemu_rdma_init_one_block, rdma); |
733252de | 645 | trace_qemu_rdma_init_ram_blocks(local->nb_blocks); |
97f3ad35 MA |
646 | rdma->dest_blocks = g_new0(RDMADestBlock, |
647 | rdma->local_ram_blocks.nb_blocks); | |
2da776db MH |
648 | local->init = true; |
649 | return 0; | |
650 | } | |
651 | ||
03fcab38 DDAG |
652 | /* |
653 | * Note: If used outside of cleanup, the caller must ensure that the destination | |
654 | * block structures are also updated | |
655 | */ | |
656 | static int rdma_delete_block(RDMAContext *rdma, RDMALocalBlock *block) | |
2da776db MH |
657 | { |
658 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
2da776db MH |
659 | RDMALocalBlock *old = local->block; |
660 | int x; | |
661 | ||
03fcab38 DDAG |
662 | if (rdma->blockmap) { |
663 | g_hash_table_remove(rdma->blockmap, (void *)(uintptr_t)block->offset); | |
664 | } | |
2da776db MH |
665 | if (block->pmr) { |
666 | int j; | |
667 | ||
668 | for (j = 0; j < block->nb_chunks; j++) { | |
669 | if (!block->pmr[j]) { | |
670 | continue; | |
671 | } | |
672 | ibv_dereg_mr(block->pmr[j]); | |
673 | rdma->total_registrations--; | |
674 | } | |
675 | g_free(block->pmr); | |
676 | block->pmr = NULL; | |
677 | } | |
678 | ||
679 | if (block->mr) { | |
680 | ibv_dereg_mr(block->mr); | |
681 | rdma->total_registrations--; | |
682 | block->mr = NULL; | |
683 | } | |
684 | ||
685 | g_free(block->transit_bitmap); | |
686 | block->transit_bitmap = NULL; | |
687 | ||
688 | g_free(block->unregister_bitmap); | |
689 | block->unregister_bitmap = NULL; | |
690 | ||
691 | g_free(block->remote_keys); | |
692 | block->remote_keys = NULL; | |
693 | ||
4fb5364b DDAG |
694 | g_free(block->block_name); |
695 | block->block_name = NULL; | |
696 | ||
03fcab38 DDAG |
697 | if (rdma->blockmap) { |
698 | for (x = 0; x < local->nb_blocks; x++) { | |
699 | g_hash_table_remove(rdma->blockmap, | |
700 | (void *)(uintptr_t)old[x].offset); | |
701 | } | |
2da776db MH |
702 | } |
703 | ||
704 | if (local->nb_blocks > 1) { | |
705 | ||
97f3ad35 | 706 | local->block = g_new0(RDMALocalBlock, local->nb_blocks - 1); |
2da776db MH |
707 | |
708 | if (block->index) { | |
709 | memcpy(local->block, old, sizeof(RDMALocalBlock) * block->index); | |
710 | } | |
711 | ||
712 | if (block->index < (local->nb_blocks - 1)) { | |
713 | memcpy(local->block + block->index, old + (block->index + 1), | |
714 | sizeof(RDMALocalBlock) * | |
715 | (local->nb_blocks - (block->index + 1))); | |
71cd7306 LC |
716 | for (x = block->index; x < local->nb_blocks - 1; x++) { |
717 | local->block[x].index--; | |
718 | } | |
2da776db MH |
719 | } |
720 | } else { | |
721 | assert(block == local->block); | |
722 | local->block = NULL; | |
723 | } | |
724 | ||
03fcab38 | 725 | trace_rdma_delete_block(block, (uintptr_t)block->local_host_addr, |
733252de | 726 | block->offset, block->length, |
fbce8c25 | 727 | (uintptr_t)(block->local_host_addr + block->length), |
733252de DDAG |
728 | BITS_TO_LONGS(block->nb_chunks) * |
729 | sizeof(unsigned long) * 8, block->nb_chunks); | |
2da776db MH |
730 | |
731 | g_free(old); | |
732 | ||
733 | local->nb_blocks--; | |
734 | ||
03fcab38 | 735 | if (local->nb_blocks && rdma->blockmap) { |
2da776db | 736 | for (x = 0; x < local->nb_blocks; x++) { |
fbce8c25 SW |
737 | g_hash_table_insert(rdma->blockmap, |
738 | (void *)(uintptr_t)local->block[x].offset, | |
739 | &local->block[x]); | |
2da776db MH |
740 | } |
741 | } | |
742 | ||
743 | return 0; | |
744 | } | |
745 | ||
746 | /* | |
747 | * Put in the log file which RDMA device was opened and the details | |
748 | * associated with that device. | |
749 | */ | |
750 | static void qemu_rdma_dump_id(const char *who, struct ibv_context *verbs) | |
751 | { | |
7fc5b13f MH |
752 | struct ibv_port_attr port; |
753 | ||
754 | if (ibv_query_port(verbs, 1, &port)) { | |
733252de | 755 | error_report("Failed to query port information"); |
7fc5b13f MH |
756 | return; |
757 | } | |
758 | ||
2da776db MH |
759 | printf("%s RDMA Device opened: kernel name %s " |
760 | "uverbs device name %s, " | |
7fc5b13f MH |
761 | "infiniband_verbs class device path %s, " |
762 | "infiniband class device path %s, " | |
763 | "transport: (%d) %s\n", | |
2da776db MH |
764 | who, |
765 | verbs->device->name, | |
766 | verbs->device->dev_name, | |
767 | verbs->device->dev_path, | |
7fc5b13f MH |
768 | verbs->device->ibdev_path, |
769 | port.link_layer, | |
770 | (port.link_layer == IBV_LINK_LAYER_INFINIBAND) ? "Infiniband" : | |
02942db7 | 771 | ((port.link_layer == IBV_LINK_LAYER_ETHERNET) |
7fc5b13f | 772 | ? "Ethernet" : "Unknown")); |
2da776db MH |
773 | } |
774 | ||
775 | /* | |
776 | * Put in the log file the RDMA gid addressing information, | |
777 | * useful for folks who have trouble understanding the | |
778 | * RDMA device hierarchy in the kernel. | |
779 | */ | |
780 | static void qemu_rdma_dump_gid(const char *who, struct rdma_cm_id *id) | |
781 | { | |
782 | char sgid[33]; | |
783 | char dgid[33]; | |
784 | inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.sgid, sgid, sizeof sgid); | |
785 | inet_ntop(AF_INET6, &id->route.addr.addr.ibaddr.dgid, dgid, sizeof dgid); | |
733252de | 786 | trace_qemu_rdma_dump_gid(who, sgid, dgid); |
2da776db MH |
787 | } |
788 | ||
7fc5b13f MH |
789 | /* |
790 | * As of now, IPv6 over RoCE / iWARP is not supported by linux. | |
791 | * We will try the next addrinfo struct, and fail if there are | |
792 | * no other valid addresses to bind against. | |
793 | * | |
794 | * If user is listening on '[::]', then we will not have a opened a device | |
795 | * yet and have no way of verifying if the device is RoCE or not. | |
796 | * | |
797 | * In this case, the source VM will throw an error for ALL types of | |
798 | * connections (both IPv4 and IPv6) if the destination machine does not have | |
799 | * a regular infiniband network available for use. | |
800 | * | |
4c293dc6 | 801 | * The only way to guarantee that an error is thrown for broken kernels is |
7fc5b13f MH |
802 | * for the management software to choose a *specific* interface at bind time |
803 | * and validate what time of hardware it is. | |
804 | * | |
805 | * Unfortunately, this puts the user in a fix: | |
02942db7 | 806 | * |
7fc5b13f MH |
807 | * If the source VM connects with an IPv4 address without knowing that the |
808 | * destination has bound to '[::]' the migration will unconditionally fail | |
b6af0975 | 809 | * unless the management software is explicitly listening on the IPv4 |
7fc5b13f MH |
810 | * address while using a RoCE-based device. |
811 | * | |
812 | * If the source VM connects with an IPv6 address, then we're OK because we can | |
813 | * throw an error on the source (and similarly on the destination). | |
02942db7 | 814 | * |
7fc5b13f MH |
815 | * But in mixed environments, this will be broken for a while until it is fixed |
816 | * inside linux. | |
817 | * | |
818 | * We do provide a *tiny* bit of help in this function: We can list all of the | |
819 | * devices in the system and check to see if all the devices are RoCE or | |
02942db7 | 820 | * Infiniband. |
7fc5b13f MH |
821 | * |
822 | * If we detect that we have a *pure* RoCE environment, then we can safely | |
4c293dc6 | 823 | * thrown an error even if the management software has specified '[::]' as the |
7fc5b13f MH |
824 | * bind address. |
825 | * | |
826 | * However, if there is are multiple hetergeneous devices, then we cannot make | |
827 | * this assumption and the user just has to be sure they know what they are | |
828 | * doing. | |
829 | * | |
830 | * Patches are being reviewed on linux-rdma. | |
831 | */ | |
bbfb89e3 | 832 | static int qemu_rdma_broken_ipv6_kernel(struct ibv_context *verbs, Error **errp) |
7fc5b13f MH |
833 | { |
834 | struct ibv_port_attr port_attr; | |
835 | ||
836 | /* This bug only exists in linux, to our knowledge. */ | |
837 | #ifdef CONFIG_LINUX | |
838 | ||
02942db7 | 839 | /* |
7fc5b13f | 840 | * Verbs are only NULL if management has bound to '[::]'. |
02942db7 | 841 | * |
7fc5b13f MH |
842 | * Let's iterate through all the devices and see if there any pure IB |
843 | * devices (non-ethernet). | |
02942db7 | 844 | * |
7fc5b13f | 845 | * If not, then we can safely proceed with the migration. |
4c293dc6 | 846 | * Otherwise, there are no guarantees until the bug is fixed in linux. |
7fc5b13f MH |
847 | */ |
848 | if (!verbs) { | |
02942db7 | 849 | int num_devices, x; |
7fc5b13f MH |
850 | struct ibv_device ** dev_list = ibv_get_device_list(&num_devices); |
851 | bool roce_found = false; | |
852 | bool ib_found = false; | |
853 | ||
854 | for (x = 0; x < num_devices; x++) { | |
855 | verbs = ibv_open_device(dev_list[x]); | |
5b61d575 PR |
856 | if (!verbs) { |
857 | if (errno == EPERM) { | |
858 | continue; | |
859 | } else { | |
860 | return -EINVAL; | |
861 | } | |
862 | } | |
7fc5b13f MH |
863 | |
864 | if (ibv_query_port(verbs, 1, &port_attr)) { | |
865 | ibv_close_device(verbs); | |
866 | ERROR(errp, "Could not query initial IB port"); | |
867 | return -EINVAL; | |
868 | } | |
869 | ||
870 | if (port_attr.link_layer == IBV_LINK_LAYER_INFINIBAND) { | |
871 | ib_found = true; | |
872 | } else if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) { | |
873 | roce_found = true; | |
874 | } | |
875 | ||
876 | ibv_close_device(verbs); | |
877 | ||
878 | } | |
879 | ||
880 | if (roce_found) { | |
881 | if (ib_found) { | |
882 | fprintf(stderr, "WARN: migrations may fail:" | |
883 | " IPv6 over RoCE / iWARP in linux" | |
884 | " is broken. But since you appear to have a" | |
885 | " mixed RoCE / IB environment, be sure to only" | |
886 | " migrate over the IB fabric until the kernel " | |
887 | " fixes the bug.\n"); | |
888 | } else { | |
889 | ERROR(errp, "You only have RoCE / iWARP devices in your systems" | |
890 | " and your management software has specified '[::]'" | |
891 | ", but IPv6 over RoCE / iWARP is not supported in Linux."); | |
892 | return -ENONET; | |
893 | } | |
894 | } | |
895 | ||
896 | return 0; | |
897 | } | |
898 | ||
899 | /* | |
900 | * If we have a verbs context, that means that some other than '[::]' was | |
02942db7 SW |
901 | * used by the management software for binding. In which case we can |
902 | * actually warn the user about a potentially broken kernel. | |
7fc5b13f MH |
903 | */ |
904 | ||
905 | /* IB ports start with 1, not 0 */ | |
906 | if (ibv_query_port(verbs, 1, &port_attr)) { | |
907 | ERROR(errp, "Could not query initial IB port"); | |
908 | return -EINVAL; | |
909 | } | |
910 | ||
911 | if (port_attr.link_layer == IBV_LINK_LAYER_ETHERNET) { | |
912 | ERROR(errp, "Linux kernel's RoCE / iWARP does not support IPv6 " | |
913 | "(but patches on linux-rdma in progress)"); | |
914 | return -ENONET; | |
915 | } | |
916 | ||
917 | #endif | |
918 | ||
919 | return 0; | |
920 | } | |
921 | ||
2da776db MH |
922 | /* |
923 | * Figure out which RDMA device corresponds to the requested IP hostname | |
924 | * Also create the initial connection manager identifiers for opening | |
925 | * the connection. | |
926 | */ | |
927 | static int qemu_rdma_resolve_host(RDMAContext *rdma, Error **errp) | |
928 | { | |
929 | int ret; | |
7fc5b13f | 930 | struct rdma_addrinfo *res; |
2da776db MH |
931 | char port_str[16]; |
932 | struct rdma_cm_event *cm_event; | |
933 | char ip[40] = "unknown"; | |
7fc5b13f | 934 | struct rdma_addrinfo *e; |
2da776db MH |
935 | |
936 | if (rdma->host == NULL || !strcmp(rdma->host, "")) { | |
66988941 | 937 | ERROR(errp, "RDMA hostname has not been set"); |
7fc5b13f | 938 | return -EINVAL; |
2da776db MH |
939 | } |
940 | ||
941 | /* create CM channel */ | |
942 | rdma->channel = rdma_create_event_channel(); | |
943 | if (!rdma->channel) { | |
66988941 | 944 | ERROR(errp, "could not create CM channel"); |
7fc5b13f | 945 | return -EINVAL; |
2da776db MH |
946 | } |
947 | ||
948 | /* create CM id */ | |
949 | ret = rdma_create_id(rdma->channel, &rdma->cm_id, NULL, RDMA_PS_TCP); | |
950 | if (ret) { | |
66988941 | 951 | ERROR(errp, "could not create channel id"); |
2da776db MH |
952 | goto err_resolve_create_id; |
953 | } | |
954 | ||
955 | snprintf(port_str, 16, "%d", rdma->port); | |
956 | port_str[15] = '\0'; | |
957 | ||
7fc5b13f | 958 | ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res); |
2da776db | 959 | if (ret < 0) { |
7fc5b13f | 960 | ERROR(errp, "could not rdma_getaddrinfo address %s", rdma->host); |
2da776db MH |
961 | goto err_resolve_get_addr; |
962 | } | |
963 | ||
6470215b MH |
964 | for (e = res; e != NULL; e = e->ai_next) { |
965 | inet_ntop(e->ai_family, | |
7fc5b13f | 966 | &((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip); |
733252de | 967 | trace_qemu_rdma_resolve_host_trying(rdma->host, ip); |
2da776db | 968 | |
7fc5b13f | 969 | ret = rdma_resolve_addr(rdma->cm_id, NULL, e->ai_dst_addr, |
6470215b MH |
970 | RDMA_RESOLVE_TIMEOUT_MS); |
971 | if (!ret) { | |
c89aa2f1 | 972 | if (e->ai_family == AF_INET6) { |
bbfb89e3 | 973 | ret = qemu_rdma_broken_ipv6_kernel(rdma->cm_id->verbs, errp); |
c89aa2f1 MH |
974 | if (ret) { |
975 | continue; | |
976 | } | |
7fc5b13f | 977 | } |
6470215b MH |
978 | goto route; |
979 | } | |
2da776db MH |
980 | } |
981 | ||
6470215b MH |
982 | ERROR(errp, "could not resolve address %s", rdma->host); |
983 | goto err_resolve_get_addr; | |
984 | ||
985 | route: | |
2da776db MH |
986 | qemu_rdma_dump_gid("source_resolve_addr", rdma->cm_id); |
987 | ||
988 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
989 | if (ret) { | |
66988941 | 990 | ERROR(errp, "could not perform event_addr_resolved"); |
2da776db MH |
991 | goto err_resolve_get_addr; |
992 | } | |
993 | ||
994 | if (cm_event->event != RDMA_CM_EVENT_ADDR_RESOLVED) { | |
66988941 | 995 | ERROR(errp, "result not equal to event_addr_resolved %s", |
2da776db MH |
996 | rdma_event_str(cm_event->event)); |
997 | perror("rdma_resolve_addr"); | |
2a934347 | 998 | rdma_ack_cm_event(cm_event); |
7fc5b13f | 999 | ret = -EINVAL; |
2da776db MH |
1000 | goto err_resolve_get_addr; |
1001 | } | |
1002 | rdma_ack_cm_event(cm_event); | |
1003 | ||
1004 | /* resolve route */ | |
1005 | ret = rdma_resolve_route(rdma->cm_id, RDMA_RESOLVE_TIMEOUT_MS); | |
1006 | if (ret) { | |
66988941 | 1007 | ERROR(errp, "could not resolve rdma route"); |
2da776db MH |
1008 | goto err_resolve_get_addr; |
1009 | } | |
1010 | ||
1011 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
1012 | if (ret) { | |
66988941 | 1013 | ERROR(errp, "could not perform event_route_resolved"); |
2da776db MH |
1014 | goto err_resolve_get_addr; |
1015 | } | |
1016 | if (cm_event->event != RDMA_CM_EVENT_ROUTE_RESOLVED) { | |
66988941 | 1017 | ERROR(errp, "result not equal to event_route_resolved: %s", |
2da776db MH |
1018 | rdma_event_str(cm_event->event)); |
1019 | rdma_ack_cm_event(cm_event); | |
7fc5b13f | 1020 | ret = -EINVAL; |
2da776db MH |
1021 | goto err_resolve_get_addr; |
1022 | } | |
1023 | rdma_ack_cm_event(cm_event); | |
1024 | rdma->verbs = rdma->cm_id->verbs; | |
1025 | qemu_rdma_dump_id("source_resolve_host", rdma->cm_id->verbs); | |
1026 | qemu_rdma_dump_gid("source_resolve_host", rdma->cm_id); | |
1027 | return 0; | |
1028 | ||
1029 | err_resolve_get_addr: | |
1030 | rdma_destroy_id(rdma->cm_id); | |
1031 | rdma->cm_id = NULL; | |
1032 | err_resolve_create_id: | |
1033 | rdma_destroy_event_channel(rdma->channel); | |
1034 | rdma->channel = NULL; | |
7fc5b13f | 1035 | return ret; |
2da776db MH |
1036 | } |
1037 | ||
1038 | /* | |
1039 | * Create protection domain and completion queues | |
1040 | */ | |
1041 | static int qemu_rdma_alloc_pd_cq(RDMAContext *rdma) | |
1042 | { | |
1043 | /* allocate pd */ | |
1044 | rdma->pd = ibv_alloc_pd(rdma->verbs); | |
1045 | if (!rdma->pd) { | |
733252de | 1046 | error_report("failed to allocate protection domain"); |
2da776db MH |
1047 | return -1; |
1048 | } | |
1049 | ||
1050 | /* create completion channel */ | |
1051 | rdma->comp_channel = ibv_create_comp_channel(rdma->verbs); | |
1052 | if (!rdma->comp_channel) { | |
733252de | 1053 | error_report("failed to allocate completion channel"); |
2da776db MH |
1054 | goto err_alloc_pd_cq; |
1055 | } | |
1056 | ||
1057 | /* | |
1058 | * Completion queue can be filled by both read and write work requests, | |
1059 | * so must reflect the sum of both possible queue sizes. | |
1060 | */ | |
1061 | rdma->cq = ibv_create_cq(rdma->verbs, (RDMA_SIGNALED_SEND_MAX * 3), | |
1062 | NULL, rdma->comp_channel, 0); | |
1063 | if (!rdma->cq) { | |
733252de | 1064 | error_report("failed to allocate completion queue"); |
2da776db MH |
1065 | goto err_alloc_pd_cq; |
1066 | } | |
1067 | ||
1068 | return 0; | |
1069 | ||
1070 | err_alloc_pd_cq: | |
1071 | if (rdma->pd) { | |
1072 | ibv_dealloc_pd(rdma->pd); | |
1073 | } | |
1074 | if (rdma->comp_channel) { | |
1075 | ibv_destroy_comp_channel(rdma->comp_channel); | |
1076 | } | |
1077 | rdma->pd = NULL; | |
1078 | rdma->comp_channel = NULL; | |
1079 | return -1; | |
1080 | ||
1081 | } | |
1082 | ||
1083 | /* | |
1084 | * Create queue pairs. | |
1085 | */ | |
1086 | static int qemu_rdma_alloc_qp(RDMAContext *rdma) | |
1087 | { | |
1088 | struct ibv_qp_init_attr attr = { 0 }; | |
1089 | int ret; | |
1090 | ||
1091 | attr.cap.max_send_wr = RDMA_SIGNALED_SEND_MAX; | |
1092 | attr.cap.max_recv_wr = 3; | |
1093 | attr.cap.max_send_sge = 1; | |
1094 | attr.cap.max_recv_sge = 1; | |
1095 | attr.send_cq = rdma->cq; | |
1096 | attr.recv_cq = rdma->cq; | |
1097 | attr.qp_type = IBV_QPT_RC; | |
1098 | ||
1099 | ret = rdma_create_qp(rdma->cm_id, rdma->pd, &attr); | |
1100 | if (ret) { | |
1101 | return -1; | |
1102 | } | |
1103 | ||
1104 | rdma->qp = rdma->cm_id->qp; | |
1105 | return 0; | |
1106 | } | |
1107 | ||
1108 | static int qemu_rdma_reg_whole_ram_blocks(RDMAContext *rdma) | |
1109 | { | |
1110 | int i; | |
1111 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
1112 | ||
1113 | for (i = 0; i < local->nb_blocks; i++) { | |
1114 | local->block[i].mr = | |
1115 | ibv_reg_mr(rdma->pd, | |
1116 | local->block[i].local_host_addr, | |
1117 | local->block[i].length, | |
1118 | IBV_ACCESS_LOCAL_WRITE | | |
1119 | IBV_ACCESS_REMOTE_WRITE | |
1120 | ); | |
1121 | if (!local->block[i].mr) { | |
1122 | perror("Failed to register local dest ram block!\n"); | |
1123 | break; | |
1124 | } | |
1125 | rdma->total_registrations++; | |
1126 | } | |
1127 | ||
1128 | if (i >= local->nb_blocks) { | |
1129 | return 0; | |
1130 | } | |
1131 | ||
1132 | for (i--; i >= 0; i--) { | |
1133 | ibv_dereg_mr(local->block[i].mr); | |
1134 | rdma->total_registrations--; | |
1135 | } | |
1136 | ||
1137 | return -1; | |
1138 | ||
1139 | } | |
1140 | ||
1141 | /* | |
1142 | * Find the ram block that corresponds to the page requested to be | |
1143 | * transmitted by QEMU. | |
1144 | * | |
1145 | * Once the block is found, also identify which 'chunk' within that | |
1146 | * block that the page belongs to. | |
1147 | * | |
1148 | * This search cannot fail or the migration will fail. | |
1149 | */ | |
1150 | static int qemu_rdma_search_ram_block(RDMAContext *rdma, | |
fbce8c25 | 1151 | uintptr_t block_offset, |
2da776db MH |
1152 | uint64_t offset, |
1153 | uint64_t length, | |
1154 | uint64_t *block_index, | |
1155 | uint64_t *chunk_index) | |
1156 | { | |
1157 | uint64_t current_addr = block_offset + offset; | |
1158 | RDMALocalBlock *block = g_hash_table_lookup(rdma->blockmap, | |
1159 | (void *) block_offset); | |
1160 | assert(block); | |
1161 | assert(current_addr >= block->offset); | |
1162 | assert((current_addr + length) <= (block->offset + block->length)); | |
1163 | ||
1164 | *block_index = block->index; | |
1165 | *chunk_index = ram_chunk_index(block->local_host_addr, | |
1166 | block->local_host_addr + (current_addr - block->offset)); | |
1167 | ||
1168 | return 0; | |
1169 | } | |
1170 | ||
1171 | /* | |
1172 | * Register a chunk with IB. If the chunk was already registered | |
1173 | * previously, then skip. | |
1174 | * | |
1175 | * Also return the keys associated with the registration needed | |
1176 | * to perform the actual RDMA operation. | |
1177 | */ | |
1178 | static int qemu_rdma_register_and_get_keys(RDMAContext *rdma, | |
3ac040c0 | 1179 | RDMALocalBlock *block, uintptr_t host_addr, |
2da776db MH |
1180 | uint32_t *lkey, uint32_t *rkey, int chunk, |
1181 | uint8_t *chunk_start, uint8_t *chunk_end) | |
1182 | { | |
1183 | if (block->mr) { | |
1184 | if (lkey) { | |
1185 | *lkey = block->mr->lkey; | |
1186 | } | |
1187 | if (rkey) { | |
1188 | *rkey = block->mr->rkey; | |
1189 | } | |
1190 | return 0; | |
1191 | } | |
1192 | ||
1193 | /* allocate memory to store chunk MRs */ | |
1194 | if (!block->pmr) { | |
97f3ad35 | 1195 | block->pmr = g_new0(struct ibv_mr *, block->nb_chunks); |
2da776db MH |
1196 | } |
1197 | ||
1198 | /* | |
1199 | * If 'rkey', then we're the destination, so grant access to the source. | |
1200 | * | |
1201 | * If 'lkey', then we're the source VM, so grant access only to ourselves. | |
1202 | */ | |
1203 | if (!block->pmr[chunk]) { | |
1204 | uint64_t len = chunk_end - chunk_start; | |
1205 | ||
733252de | 1206 | trace_qemu_rdma_register_and_get_keys(len, chunk_start); |
2da776db MH |
1207 | |
1208 | block->pmr[chunk] = ibv_reg_mr(rdma->pd, | |
1209 | chunk_start, len, | |
1210 | (rkey ? (IBV_ACCESS_LOCAL_WRITE | | |
1211 | IBV_ACCESS_REMOTE_WRITE) : 0)); | |
1212 | ||
1213 | if (!block->pmr[chunk]) { | |
1214 | perror("Failed to register chunk!"); | |
1215 | fprintf(stderr, "Chunk details: block: %d chunk index %d" | |
3ac040c0 SW |
1216 | " start %" PRIuPTR " end %" PRIuPTR |
1217 | " host %" PRIuPTR | |
1218 | " local %" PRIuPTR " registrations: %d\n", | |
1219 | block->index, chunk, (uintptr_t)chunk_start, | |
1220 | (uintptr_t)chunk_end, host_addr, | |
1221 | (uintptr_t)block->local_host_addr, | |
2da776db MH |
1222 | rdma->total_registrations); |
1223 | return -1; | |
1224 | } | |
1225 | rdma->total_registrations++; | |
1226 | } | |
1227 | ||
1228 | if (lkey) { | |
1229 | *lkey = block->pmr[chunk]->lkey; | |
1230 | } | |
1231 | if (rkey) { | |
1232 | *rkey = block->pmr[chunk]->rkey; | |
1233 | } | |
1234 | return 0; | |
1235 | } | |
1236 | ||
1237 | /* | |
1238 | * Register (at connection time) the memory used for control | |
1239 | * channel messages. | |
1240 | */ | |
1241 | static int qemu_rdma_reg_control(RDMAContext *rdma, int idx) | |
1242 | { | |
1243 | rdma->wr_data[idx].control_mr = ibv_reg_mr(rdma->pd, | |
1244 | rdma->wr_data[idx].control, RDMA_CONTROL_MAX_BUFFER, | |
1245 | IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE); | |
1246 | if (rdma->wr_data[idx].control_mr) { | |
1247 | rdma->total_registrations++; | |
1248 | return 0; | |
1249 | } | |
733252de | 1250 | error_report("qemu_rdma_reg_control failed"); |
2da776db MH |
1251 | return -1; |
1252 | } | |
1253 | ||
1254 | const char *print_wrid(int wrid) | |
1255 | { | |
1256 | if (wrid >= RDMA_WRID_RECV_CONTROL) { | |
1257 | return wrid_desc[RDMA_WRID_RECV_CONTROL]; | |
1258 | } | |
1259 | return wrid_desc[wrid]; | |
1260 | } | |
1261 | ||
1262 | /* | |
1263 | * RDMA requires memory registration (mlock/pinning), but this is not good for | |
1264 | * overcommitment. | |
1265 | * | |
1266 | * In preparation for the future where LRU information or workload-specific | |
1267 | * writable writable working set memory access behavior is available to QEMU | |
1268 | * it would be nice to have in place the ability to UN-register/UN-pin | |
1269 | * particular memory regions from the RDMA hardware when it is determine that | |
1270 | * those regions of memory will likely not be accessed again in the near future. | |
1271 | * | |
1272 | * While we do not yet have such information right now, the following | |
1273 | * compile-time option allows us to perform a non-optimized version of this | |
1274 | * behavior. | |
1275 | * | |
1276 | * By uncommenting this option, you will cause *all* RDMA transfers to be | |
1277 | * unregistered immediately after the transfer completes on both sides of the | |
1278 | * connection. This has no effect in 'rdma-pin-all' mode, only regular mode. | |
1279 | * | |
1280 | * This will have a terrible impact on migration performance, so until future | |
1281 | * workload information or LRU information is available, do not attempt to use | |
1282 | * this feature except for basic testing. | |
1283 | */ | |
1284 | //#define RDMA_UNREGISTRATION_EXAMPLE | |
1285 | ||
1286 | /* | |
1287 | * Perform a non-optimized memory unregistration after every transfer | |
24ec68ef | 1288 | * for demonstration purposes, only if pin-all is not requested. |
2da776db MH |
1289 | * |
1290 | * Potential optimizations: | |
1291 | * 1. Start a new thread to run this function continuously | |
1292 | - for bit clearing | |
1293 | - and for receipt of unregister messages | |
1294 | * 2. Use an LRU. | |
1295 | * 3. Use workload hints. | |
1296 | */ | |
1297 | static int qemu_rdma_unregister_waiting(RDMAContext *rdma) | |
1298 | { | |
1299 | while (rdma->unregistrations[rdma->unregister_current]) { | |
1300 | int ret; | |
1301 | uint64_t wr_id = rdma->unregistrations[rdma->unregister_current]; | |
1302 | uint64_t chunk = | |
1303 | (wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; | |
1304 | uint64_t index = | |
1305 | (wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; | |
1306 | RDMALocalBlock *block = | |
1307 | &(rdma->local_ram_blocks.block[index]); | |
1308 | RDMARegister reg = { .current_index = index }; | |
1309 | RDMAControlHeader resp = { .type = RDMA_CONTROL_UNREGISTER_FINISHED, | |
1310 | }; | |
1311 | RDMAControlHeader head = { .len = sizeof(RDMARegister), | |
1312 | .type = RDMA_CONTROL_UNREGISTER_REQUEST, | |
1313 | .repeat = 1, | |
1314 | }; | |
1315 | ||
733252de DDAG |
1316 | trace_qemu_rdma_unregister_waiting_proc(chunk, |
1317 | rdma->unregister_current); | |
2da776db MH |
1318 | |
1319 | rdma->unregistrations[rdma->unregister_current] = 0; | |
1320 | rdma->unregister_current++; | |
1321 | ||
1322 | if (rdma->unregister_current == RDMA_SIGNALED_SEND_MAX) { | |
1323 | rdma->unregister_current = 0; | |
1324 | } | |
1325 | ||
1326 | ||
1327 | /* | |
1328 | * Unregistration is speculative (because migration is single-threaded | |
1329 | * and we cannot break the protocol's inifinband message ordering). | |
1330 | * Thus, if the memory is currently being used for transmission, | |
1331 | * then abort the attempt to unregister and try again | |
1332 | * later the next time a completion is received for this memory. | |
1333 | */ | |
1334 | clear_bit(chunk, block->unregister_bitmap); | |
1335 | ||
1336 | if (test_bit(chunk, block->transit_bitmap)) { | |
733252de | 1337 | trace_qemu_rdma_unregister_waiting_inflight(chunk); |
2da776db MH |
1338 | continue; |
1339 | } | |
1340 | ||
733252de | 1341 | trace_qemu_rdma_unregister_waiting_send(chunk); |
2da776db MH |
1342 | |
1343 | ret = ibv_dereg_mr(block->pmr[chunk]); | |
1344 | block->pmr[chunk] = NULL; | |
1345 | block->remote_keys[chunk] = 0; | |
1346 | ||
1347 | if (ret != 0) { | |
1348 | perror("unregistration chunk failed"); | |
1349 | return -ret; | |
1350 | } | |
1351 | rdma->total_registrations--; | |
1352 | ||
1353 | reg.key.chunk = chunk; | |
b12f7777 | 1354 | register_to_network(rdma, ®); |
2da776db MH |
1355 | ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
1356 | &resp, NULL, NULL); | |
1357 | if (ret < 0) { | |
1358 | return ret; | |
1359 | } | |
1360 | ||
733252de | 1361 | trace_qemu_rdma_unregister_waiting_complete(chunk); |
2da776db MH |
1362 | } |
1363 | ||
1364 | return 0; | |
1365 | } | |
1366 | ||
1367 | static uint64_t qemu_rdma_make_wrid(uint64_t wr_id, uint64_t index, | |
1368 | uint64_t chunk) | |
1369 | { | |
1370 | uint64_t result = wr_id & RDMA_WRID_TYPE_MASK; | |
1371 | ||
1372 | result |= (index << RDMA_WRID_BLOCK_SHIFT); | |
1373 | result |= (chunk << RDMA_WRID_CHUNK_SHIFT); | |
1374 | ||
1375 | return result; | |
1376 | } | |
1377 | ||
1378 | /* | |
1379 | * Set bit for unregistration in the next iteration. | |
1380 | * We cannot transmit right here, but will unpin later. | |
1381 | */ | |
1382 | static void qemu_rdma_signal_unregister(RDMAContext *rdma, uint64_t index, | |
1383 | uint64_t chunk, uint64_t wr_id) | |
1384 | { | |
1385 | if (rdma->unregistrations[rdma->unregister_next] != 0) { | |
733252de | 1386 | error_report("rdma migration: queue is full"); |
2da776db MH |
1387 | } else { |
1388 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); | |
1389 | ||
1390 | if (!test_and_set_bit(chunk, block->unregister_bitmap)) { | |
733252de DDAG |
1391 | trace_qemu_rdma_signal_unregister_append(chunk, |
1392 | rdma->unregister_next); | |
2da776db MH |
1393 | |
1394 | rdma->unregistrations[rdma->unregister_next++] = | |
1395 | qemu_rdma_make_wrid(wr_id, index, chunk); | |
1396 | ||
1397 | if (rdma->unregister_next == RDMA_SIGNALED_SEND_MAX) { | |
1398 | rdma->unregister_next = 0; | |
1399 | } | |
1400 | } else { | |
733252de | 1401 | trace_qemu_rdma_signal_unregister_already(chunk); |
2da776db MH |
1402 | } |
1403 | } | |
1404 | } | |
1405 | ||
1406 | /* | |
1407 | * Consult the connection manager to see a work request | |
1408 | * (of any kind) has completed. | |
1409 | * Return the work request ID that completed. | |
1410 | */ | |
88571882 IY |
1411 | static uint64_t qemu_rdma_poll(RDMAContext *rdma, uint64_t *wr_id_out, |
1412 | uint32_t *byte_len) | |
2da776db MH |
1413 | { |
1414 | int ret; | |
1415 | struct ibv_wc wc; | |
1416 | uint64_t wr_id; | |
1417 | ||
1418 | ret = ibv_poll_cq(rdma->cq, 1, &wc); | |
1419 | ||
1420 | if (!ret) { | |
1421 | *wr_id_out = RDMA_WRID_NONE; | |
1422 | return 0; | |
1423 | } | |
1424 | ||
1425 | if (ret < 0) { | |
733252de | 1426 | error_report("ibv_poll_cq return %d", ret); |
2da776db MH |
1427 | return ret; |
1428 | } | |
1429 | ||
1430 | wr_id = wc.wr_id & RDMA_WRID_TYPE_MASK; | |
1431 | ||
1432 | if (wc.status != IBV_WC_SUCCESS) { | |
1433 | fprintf(stderr, "ibv_poll_cq wc.status=%d %s!\n", | |
1434 | wc.status, ibv_wc_status_str(wc.status)); | |
1435 | fprintf(stderr, "ibv_poll_cq wrid=%s!\n", wrid_desc[wr_id]); | |
1436 | ||
1437 | return -1; | |
1438 | } | |
1439 | ||
1440 | if (rdma->control_ready_expected && | |
1441 | (wr_id >= RDMA_WRID_RECV_CONTROL)) { | |
733252de | 1442 | trace_qemu_rdma_poll_recv(wrid_desc[RDMA_WRID_RECV_CONTROL], |
2da776db MH |
1443 | wr_id - RDMA_WRID_RECV_CONTROL, wr_id, rdma->nb_sent); |
1444 | rdma->control_ready_expected = 0; | |
1445 | } | |
1446 | ||
1447 | if (wr_id == RDMA_WRID_RDMA_WRITE) { | |
1448 | uint64_t chunk = | |
1449 | (wc.wr_id & RDMA_WRID_CHUNK_MASK) >> RDMA_WRID_CHUNK_SHIFT; | |
1450 | uint64_t index = | |
1451 | (wc.wr_id & RDMA_WRID_BLOCK_MASK) >> RDMA_WRID_BLOCK_SHIFT; | |
1452 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[index]); | |
1453 | ||
733252de | 1454 | trace_qemu_rdma_poll_write(print_wrid(wr_id), wr_id, rdma->nb_sent, |
fbce8c25 SW |
1455 | index, chunk, block->local_host_addr, |
1456 | (void *)(uintptr_t)block->remote_host_addr); | |
2da776db MH |
1457 | |
1458 | clear_bit(chunk, block->transit_bitmap); | |
1459 | ||
1460 | if (rdma->nb_sent > 0) { | |
1461 | rdma->nb_sent--; | |
1462 | } | |
1463 | ||
1464 | if (!rdma->pin_all) { | |
1465 | /* | |
1466 | * FYI: If one wanted to signal a specific chunk to be unregistered | |
1467 | * using LRU or workload-specific information, this is the function | |
1468 | * you would call to do so. That chunk would then get asynchronously | |
1469 | * unregistered later. | |
1470 | */ | |
1471 | #ifdef RDMA_UNREGISTRATION_EXAMPLE | |
1472 | qemu_rdma_signal_unregister(rdma, index, chunk, wc.wr_id); | |
1473 | #endif | |
1474 | } | |
1475 | } else { | |
733252de | 1476 | trace_qemu_rdma_poll_other(print_wrid(wr_id), wr_id, rdma->nb_sent); |
2da776db MH |
1477 | } |
1478 | ||
1479 | *wr_id_out = wc.wr_id; | |
88571882 IY |
1480 | if (byte_len) { |
1481 | *byte_len = wc.byte_len; | |
1482 | } | |
2da776db MH |
1483 | |
1484 | return 0; | |
1485 | } | |
1486 | ||
9c98cfbe DDAG |
1487 | /* Wait for activity on the completion channel. |
1488 | * Returns 0 on success, none-0 on error. | |
1489 | */ | |
1490 | static int qemu_rdma_wait_comp_channel(RDMAContext *rdma) | |
1491 | { | |
1492 | /* | |
1493 | * Coroutine doesn't start until migration_fd_process_incoming() | |
1494 | * so don't yield unless we know we're running inside of a coroutine. | |
1495 | */ | |
f5627c2a LC |
1496 | if (rdma->migration_started_on_destination && |
1497 | migration_incoming_get_current()->state == MIGRATION_STATUS_ACTIVE) { | |
9c98cfbe DDAG |
1498 | yield_until_fd_readable(rdma->comp_channel->fd); |
1499 | } else { | |
1500 | /* This is the source side, we're in a separate thread | |
1501 | * or destination prior to migration_fd_process_incoming() | |
f5627c2a | 1502 | * after postcopy, the destination also in a seprate thread. |
9c98cfbe DDAG |
1503 | * we can't yield; so we have to poll the fd. |
1504 | * But we need to be able to handle 'cancel' or an error | |
1505 | * without hanging forever. | |
1506 | */ | |
1507 | while (!rdma->error_state && !rdma->received_error) { | |
1508 | GPollFD pfds[1]; | |
1509 | pfds[0].fd = rdma->comp_channel->fd; | |
1510 | pfds[0].events = G_IO_IN | G_IO_HUP | G_IO_ERR; | |
1511 | /* 0.1s timeout, should be fine for a 'cancel' */ | |
1512 | switch (qemu_poll_ns(pfds, 1, 100 * 1000 * 1000)) { | |
1513 | case 1: /* fd active */ | |
1514 | return 0; | |
1515 | ||
1516 | case 0: /* Timeout, go around again */ | |
1517 | break; | |
1518 | ||
1519 | default: /* Error of some type - | |
1520 | * I don't trust errno from qemu_poll_ns | |
1521 | */ | |
1522 | error_report("%s: poll failed", __func__); | |
1523 | return -EPIPE; | |
1524 | } | |
1525 | ||
1526 | if (migrate_get_current()->state == MIGRATION_STATUS_CANCELLING) { | |
1527 | /* Bail out and let the cancellation happen */ | |
1528 | return -EPIPE; | |
1529 | } | |
1530 | } | |
1531 | } | |
1532 | ||
1533 | if (rdma->received_error) { | |
1534 | return -EPIPE; | |
1535 | } | |
1536 | return rdma->error_state; | |
1537 | } | |
1538 | ||
2da776db MH |
1539 | /* |
1540 | * Block until the next work request has completed. | |
1541 | * | |
1542 | * First poll to see if a work request has already completed, | |
1543 | * otherwise block. | |
1544 | * | |
1545 | * If we encounter completed work requests for IDs other than | |
1546 | * the one we're interested in, then that's generally an error. | |
1547 | * | |
1548 | * The only exception is actual RDMA Write completions. These | |
1549 | * completions only need to be recorded, but do not actually | |
1550 | * need further processing. | |
1551 | */ | |
88571882 IY |
1552 | static int qemu_rdma_block_for_wrid(RDMAContext *rdma, int wrid_requested, |
1553 | uint32_t *byte_len) | |
2da776db MH |
1554 | { |
1555 | int num_cq_events = 0, ret = 0; | |
1556 | struct ibv_cq *cq; | |
1557 | void *cq_ctx; | |
1558 | uint64_t wr_id = RDMA_WRID_NONE, wr_id_in; | |
1559 | ||
1560 | if (ibv_req_notify_cq(rdma->cq, 0)) { | |
1561 | return -1; | |
1562 | } | |
1563 | /* poll cq first */ | |
1564 | while (wr_id != wrid_requested) { | |
88571882 | 1565 | ret = qemu_rdma_poll(rdma, &wr_id_in, byte_len); |
2da776db MH |
1566 | if (ret < 0) { |
1567 | return ret; | |
1568 | } | |
1569 | ||
1570 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; | |
1571 | ||
1572 | if (wr_id == RDMA_WRID_NONE) { | |
1573 | break; | |
1574 | } | |
1575 | if (wr_id != wrid_requested) { | |
733252de DDAG |
1576 | trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested), |
1577 | wrid_requested, print_wrid(wr_id), wr_id); | |
2da776db MH |
1578 | } |
1579 | } | |
1580 | ||
1581 | if (wr_id == wrid_requested) { | |
1582 | return 0; | |
1583 | } | |
1584 | ||
1585 | while (1) { | |
9c98cfbe DDAG |
1586 | ret = qemu_rdma_wait_comp_channel(rdma); |
1587 | if (ret) { | |
1588 | goto err_block_for_wrid; | |
2da776db MH |
1589 | } |
1590 | ||
0b3c15f0 DDAG |
1591 | ret = ibv_get_cq_event(rdma->comp_channel, &cq, &cq_ctx); |
1592 | if (ret) { | |
2da776db MH |
1593 | perror("ibv_get_cq_event"); |
1594 | goto err_block_for_wrid; | |
1595 | } | |
1596 | ||
1597 | num_cq_events++; | |
1598 | ||
0b3c15f0 DDAG |
1599 | ret = -ibv_req_notify_cq(cq, 0); |
1600 | if (ret) { | |
2da776db MH |
1601 | goto err_block_for_wrid; |
1602 | } | |
1603 | ||
1604 | while (wr_id != wrid_requested) { | |
88571882 | 1605 | ret = qemu_rdma_poll(rdma, &wr_id_in, byte_len); |
2da776db MH |
1606 | if (ret < 0) { |
1607 | goto err_block_for_wrid; | |
1608 | } | |
1609 | ||
1610 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; | |
1611 | ||
1612 | if (wr_id == RDMA_WRID_NONE) { | |
1613 | break; | |
1614 | } | |
1615 | if (wr_id != wrid_requested) { | |
733252de DDAG |
1616 | trace_qemu_rdma_block_for_wrid_miss(print_wrid(wrid_requested), |
1617 | wrid_requested, print_wrid(wr_id), wr_id); | |
2da776db MH |
1618 | } |
1619 | } | |
1620 | ||
1621 | if (wr_id == wrid_requested) { | |
1622 | goto success_block_for_wrid; | |
1623 | } | |
1624 | } | |
1625 | ||
1626 | success_block_for_wrid: | |
1627 | if (num_cq_events) { | |
1628 | ibv_ack_cq_events(cq, num_cq_events); | |
1629 | } | |
1630 | return 0; | |
1631 | ||
1632 | err_block_for_wrid: | |
1633 | if (num_cq_events) { | |
1634 | ibv_ack_cq_events(cq, num_cq_events); | |
1635 | } | |
0b3c15f0 DDAG |
1636 | |
1637 | rdma->error_state = ret; | |
2da776db MH |
1638 | return ret; |
1639 | } | |
1640 | ||
1641 | /* | |
1642 | * Post a SEND message work request for the control channel | |
1643 | * containing some data and block until the post completes. | |
1644 | */ | |
1645 | static int qemu_rdma_post_send_control(RDMAContext *rdma, uint8_t *buf, | |
1646 | RDMAControlHeader *head) | |
1647 | { | |
1648 | int ret = 0; | |
1f22364b | 1649 | RDMAWorkRequestData *wr = &rdma->wr_data[RDMA_WRID_CONTROL]; |
2da776db MH |
1650 | struct ibv_send_wr *bad_wr; |
1651 | struct ibv_sge sge = { | |
fbce8c25 | 1652 | .addr = (uintptr_t)(wr->control), |
2da776db MH |
1653 | .length = head->len + sizeof(RDMAControlHeader), |
1654 | .lkey = wr->control_mr->lkey, | |
1655 | }; | |
1656 | struct ibv_send_wr send_wr = { | |
1657 | .wr_id = RDMA_WRID_SEND_CONTROL, | |
1658 | .opcode = IBV_WR_SEND, | |
1659 | .send_flags = IBV_SEND_SIGNALED, | |
1660 | .sg_list = &sge, | |
1661 | .num_sge = 1, | |
1662 | }; | |
1663 | ||
482a33c5 | 1664 | trace_qemu_rdma_post_send_control(control_desc(head->type)); |
2da776db MH |
1665 | |
1666 | /* | |
1667 | * We don't actually need to do a memcpy() in here if we used | |
1668 | * the "sge" properly, but since we're only sending control messages | |
1669 | * (not RAM in a performance-critical path), then its OK for now. | |
1670 | * | |
1671 | * The copy makes the RDMAControlHeader simpler to manipulate | |
1672 | * for the time being. | |
1673 | */ | |
6f1484ed | 1674 | assert(head->len <= RDMA_CONTROL_MAX_BUFFER - sizeof(*head)); |
2da776db MH |
1675 | memcpy(wr->control, head, sizeof(RDMAControlHeader)); |
1676 | control_to_network((void *) wr->control); | |
1677 | ||
1678 | if (buf) { | |
1679 | memcpy(wr->control + sizeof(RDMAControlHeader), buf, head->len); | |
1680 | } | |
1681 | ||
1682 | ||
e325b49a | 1683 | ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr); |
2da776db | 1684 | |
e325b49a | 1685 | if (ret > 0) { |
733252de | 1686 | error_report("Failed to use post IB SEND for control"); |
e325b49a | 1687 | return -ret; |
2da776db MH |
1688 | } |
1689 | ||
88571882 | 1690 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_SEND_CONTROL, NULL); |
2da776db | 1691 | if (ret < 0) { |
733252de | 1692 | error_report("rdma migration: send polling control error"); |
2da776db MH |
1693 | } |
1694 | ||
1695 | return ret; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Post a RECV work request in anticipation of some future receipt | |
1700 | * of data on the control channel. | |
1701 | */ | |
1702 | static int qemu_rdma_post_recv_control(RDMAContext *rdma, int idx) | |
1703 | { | |
1704 | struct ibv_recv_wr *bad_wr; | |
1705 | struct ibv_sge sge = { | |
fbce8c25 | 1706 | .addr = (uintptr_t)(rdma->wr_data[idx].control), |
2da776db MH |
1707 | .length = RDMA_CONTROL_MAX_BUFFER, |
1708 | .lkey = rdma->wr_data[idx].control_mr->lkey, | |
1709 | }; | |
1710 | ||
1711 | struct ibv_recv_wr recv_wr = { | |
1712 | .wr_id = RDMA_WRID_RECV_CONTROL + idx, | |
1713 | .sg_list = &sge, | |
1714 | .num_sge = 1, | |
1715 | }; | |
1716 | ||
1717 | ||
1718 | if (ibv_post_recv(rdma->qp, &recv_wr, &bad_wr)) { | |
1719 | return -1; | |
1720 | } | |
1721 | ||
1722 | return 0; | |
1723 | } | |
1724 | ||
1725 | /* | |
1726 | * Block and wait for a RECV control channel message to arrive. | |
1727 | */ | |
1728 | static int qemu_rdma_exchange_get_response(RDMAContext *rdma, | |
1729 | RDMAControlHeader *head, int expecting, int idx) | |
1730 | { | |
88571882 IY |
1731 | uint32_t byte_len; |
1732 | int ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RECV_CONTROL + idx, | |
1733 | &byte_len); | |
2da776db MH |
1734 | |
1735 | if (ret < 0) { | |
733252de | 1736 | error_report("rdma migration: recv polling control error!"); |
2da776db MH |
1737 | return ret; |
1738 | } | |
1739 | ||
1740 | network_to_control((void *) rdma->wr_data[idx].control); | |
1741 | memcpy(head, rdma->wr_data[idx].control, sizeof(RDMAControlHeader)); | |
1742 | ||
482a33c5 | 1743 | trace_qemu_rdma_exchange_get_response_start(control_desc(expecting)); |
2da776db MH |
1744 | |
1745 | if (expecting == RDMA_CONTROL_NONE) { | |
482a33c5 | 1746 | trace_qemu_rdma_exchange_get_response_none(control_desc(head->type), |
733252de | 1747 | head->type); |
2da776db | 1748 | } else if (head->type != expecting || head->type == RDMA_CONTROL_ERROR) { |
733252de DDAG |
1749 | error_report("Was expecting a %s (%d) control message" |
1750 | ", but got: %s (%d), length: %d", | |
482a33c5 DDAG |
1751 | control_desc(expecting), expecting, |
1752 | control_desc(head->type), head->type, head->len); | |
cd5ea070 DDAG |
1753 | if (head->type == RDMA_CONTROL_ERROR) { |
1754 | rdma->received_error = true; | |
1755 | } | |
2da776db MH |
1756 | return -EIO; |
1757 | } | |
6f1484ed | 1758 | if (head->len > RDMA_CONTROL_MAX_BUFFER - sizeof(*head)) { |
81b07353 | 1759 | error_report("too long length: %d", head->len); |
6f1484ed IY |
1760 | return -EINVAL; |
1761 | } | |
88571882 | 1762 | if (sizeof(*head) + head->len != byte_len) { |
733252de | 1763 | error_report("Malformed length: %d byte_len %d", head->len, byte_len); |
88571882 IY |
1764 | return -EINVAL; |
1765 | } | |
2da776db MH |
1766 | |
1767 | return 0; | |
1768 | } | |
1769 | ||
1770 | /* | |
1771 | * When a RECV work request has completed, the work request's | |
1772 | * buffer is pointed at the header. | |
1773 | * | |
1774 | * This will advance the pointer to the data portion | |
1775 | * of the control message of the work request's buffer that | |
1776 | * was populated after the work request finished. | |
1777 | */ | |
1778 | static void qemu_rdma_move_header(RDMAContext *rdma, int idx, | |
1779 | RDMAControlHeader *head) | |
1780 | { | |
1781 | rdma->wr_data[idx].control_len = head->len; | |
1782 | rdma->wr_data[idx].control_curr = | |
1783 | rdma->wr_data[idx].control + sizeof(RDMAControlHeader); | |
1784 | } | |
1785 | ||
1786 | /* | |
1787 | * This is an 'atomic' high-level operation to deliver a single, unified | |
1788 | * control-channel message. | |
1789 | * | |
1790 | * Additionally, if the user is expecting some kind of reply to this message, | |
1791 | * they can request a 'resp' response message be filled in by posting an | |
1792 | * additional work request on behalf of the user and waiting for an additional | |
1793 | * completion. | |
1794 | * | |
1795 | * The extra (optional) response is used during registration to us from having | |
1796 | * to perform an *additional* exchange of message just to provide a response by | |
1797 | * instead piggy-backing on the acknowledgement. | |
1798 | */ | |
1799 | static int qemu_rdma_exchange_send(RDMAContext *rdma, RDMAControlHeader *head, | |
1800 | uint8_t *data, RDMAControlHeader *resp, | |
1801 | int *resp_idx, | |
1802 | int (*callback)(RDMAContext *rdma)) | |
1803 | { | |
1804 | int ret = 0; | |
1805 | ||
1806 | /* | |
1807 | * Wait until the dest is ready before attempting to deliver the message | |
1808 | * by waiting for a READY message. | |
1809 | */ | |
1810 | if (rdma->control_ready_expected) { | |
1811 | RDMAControlHeader resp; | |
1812 | ret = qemu_rdma_exchange_get_response(rdma, | |
1813 | &resp, RDMA_CONTROL_READY, RDMA_WRID_READY); | |
1814 | if (ret < 0) { | |
1815 | return ret; | |
1816 | } | |
1817 | } | |
1818 | ||
1819 | /* | |
1820 | * If the user is expecting a response, post a WR in anticipation of it. | |
1821 | */ | |
1822 | if (resp) { | |
1823 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_DATA); | |
1824 | if (ret) { | |
733252de | 1825 | error_report("rdma migration: error posting" |
2da776db MH |
1826 | " extra control recv for anticipated result!"); |
1827 | return ret; | |
1828 | } | |
1829 | } | |
1830 | ||
1831 | /* | |
1832 | * Post a WR to replace the one we just consumed for the READY message. | |
1833 | */ | |
1834 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); | |
1835 | if (ret) { | |
733252de | 1836 | error_report("rdma migration: error posting first control recv!"); |
2da776db MH |
1837 | return ret; |
1838 | } | |
1839 | ||
1840 | /* | |
1841 | * Deliver the control message that was requested. | |
1842 | */ | |
1843 | ret = qemu_rdma_post_send_control(rdma, data, head); | |
1844 | ||
1845 | if (ret < 0) { | |
733252de | 1846 | error_report("Failed to send control buffer!"); |
2da776db MH |
1847 | return ret; |
1848 | } | |
1849 | ||
1850 | /* | |
1851 | * If we're expecting a response, block and wait for it. | |
1852 | */ | |
1853 | if (resp) { | |
1854 | if (callback) { | |
733252de | 1855 | trace_qemu_rdma_exchange_send_issue_callback(); |
2da776db MH |
1856 | ret = callback(rdma); |
1857 | if (ret < 0) { | |
1858 | return ret; | |
1859 | } | |
1860 | } | |
1861 | ||
482a33c5 | 1862 | trace_qemu_rdma_exchange_send_waiting(control_desc(resp->type)); |
2da776db MH |
1863 | ret = qemu_rdma_exchange_get_response(rdma, resp, |
1864 | resp->type, RDMA_WRID_DATA); | |
1865 | ||
1866 | if (ret < 0) { | |
1867 | return ret; | |
1868 | } | |
1869 | ||
1870 | qemu_rdma_move_header(rdma, RDMA_WRID_DATA, resp); | |
1871 | if (resp_idx) { | |
1872 | *resp_idx = RDMA_WRID_DATA; | |
1873 | } | |
482a33c5 | 1874 | trace_qemu_rdma_exchange_send_received(control_desc(resp->type)); |
2da776db MH |
1875 | } |
1876 | ||
1877 | rdma->control_ready_expected = 1; | |
1878 | ||
1879 | return 0; | |
1880 | } | |
1881 | ||
1882 | /* | |
1883 | * This is an 'atomic' high-level operation to receive a single, unified | |
1884 | * control-channel message. | |
1885 | */ | |
1886 | static int qemu_rdma_exchange_recv(RDMAContext *rdma, RDMAControlHeader *head, | |
1887 | int expecting) | |
1888 | { | |
1889 | RDMAControlHeader ready = { | |
1890 | .len = 0, | |
1891 | .type = RDMA_CONTROL_READY, | |
1892 | .repeat = 1, | |
1893 | }; | |
1894 | int ret; | |
1895 | ||
1896 | /* | |
1897 | * Inform the source that we're ready to receive a message. | |
1898 | */ | |
1899 | ret = qemu_rdma_post_send_control(rdma, NULL, &ready); | |
1900 | ||
1901 | if (ret < 0) { | |
733252de | 1902 | error_report("Failed to send control buffer!"); |
2da776db MH |
1903 | return ret; |
1904 | } | |
1905 | ||
1906 | /* | |
1907 | * Block and wait for the message. | |
1908 | */ | |
1909 | ret = qemu_rdma_exchange_get_response(rdma, head, | |
1910 | expecting, RDMA_WRID_READY); | |
1911 | ||
1912 | if (ret < 0) { | |
1913 | return ret; | |
1914 | } | |
1915 | ||
1916 | qemu_rdma_move_header(rdma, RDMA_WRID_READY, head); | |
1917 | ||
1918 | /* | |
1919 | * Post a new RECV work request to replace the one we just consumed. | |
1920 | */ | |
1921 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); | |
1922 | if (ret) { | |
733252de | 1923 | error_report("rdma migration: error posting second control recv!"); |
2da776db MH |
1924 | return ret; |
1925 | } | |
1926 | ||
1927 | return 0; | |
1928 | } | |
1929 | ||
1930 | /* | |
1931 | * Write an actual chunk of memory using RDMA. | |
1932 | * | |
1933 | * If we're using dynamic registration on the dest-side, we have to | |
1934 | * send a registration command first. | |
1935 | */ | |
1936 | static int qemu_rdma_write_one(QEMUFile *f, RDMAContext *rdma, | |
1937 | int current_index, uint64_t current_addr, | |
1938 | uint64_t length) | |
1939 | { | |
1940 | struct ibv_sge sge; | |
1941 | struct ibv_send_wr send_wr = { 0 }; | |
1942 | struct ibv_send_wr *bad_wr; | |
1943 | int reg_result_idx, ret, count = 0; | |
1944 | uint64_t chunk, chunks; | |
1945 | uint8_t *chunk_start, *chunk_end; | |
1946 | RDMALocalBlock *block = &(rdma->local_ram_blocks.block[current_index]); | |
1947 | RDMARegister reg; | |
1948 | RDMARegisterResult *reg_result; | |
1949 | RDMAControlHeader resp = { .type = RDMA_CONTROL_REGISTER_RESULT }; | |
1950 | RDMAControlHeader head = { .len = sizeof(RDMARegister), | |
1951 | .type = RDMA_CONTROL_REGISTER_REQUEST, | |
1952 | .repeat = 1, | |
1953 | }; | |
1954 | ||
1955 | retry: | |
fbce8c25 | 1956 | sge.addr = (uintptr_t)(block->local_host_addr + |
2da776db MH |
1957 | (current_addr - block->offset)); |
1958 | sge.length = length; | |
1959 | ||
fbce8c25 SW |
1960 | chunk = ram_chunk_index(block->local_host_addr, |
1961 | (uint8_t *)(uintptr_t)sge.addr); | |
2da776db MH |
1962 | chunk_start = ram_chunk_start(block, chunk); |
1963 | ||
1964 | if (block->is_ram_block) { | |
1965 | chunks = length / (1UL << RDMA_REG_CHUNK_SHIFT); | |
1966 | ||
1967 | if (chunks && ((length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { | |
1968 | chunks--; | |
1969 | } | |
1970 | } else { | |
1971 | chunks = block->length / (1UL << RDMA_REG_CHUNK_SHIFT); | |
1972 | ||
1973 | if (chunks && ((block->length % (1UL << RDMA_REG_CHUNK_SHIFT)) == 0)) { | |
1974 | chunks--; | |
1975 | } | |
1976 | } | |
1977 | ||
733252de DDAG |
1978 | trace_qemu_rdma_write_one_top(chunks + 1, |
1979 | (chunks + 1) * | |
1980 | (1UL << RDMA_REG_CHUNK_SHIFT) / 1024 / 1024); | |
2da776db MH |
1981 | |
1982 | chunk_end = ram_chunk_end(block, chunk + chunks); | |
1983 | ||
1984 | if (!rdma->pin_all) { | |
1985 | #ifdef RDMA_UNREGISTRATION_EXAMPLE | |
1986 | qemu_rdma_unregister_waiting(rdma); | |
1987 | #endif | |
1988 | } | |
1989 | ||
1990 | while (test_bit(chunk, block->transit_bitmap)) { | |
1991 | (void)count; | |
733252de | 1992 | trace_qemu_rdma_write_one_block(count++, current_index, chunk, |
2da776db MH |
1993 | sge.addr, length, rdma->nb_sent, block->nb_chunks); |
1994 | ||
88571882 | 1995 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL); |
2da776db MH |
1996 | |
1997 | if (ret < 0) { | |
733252de | 1998 | error_report("Failed to Wait for previous write to complete " |
2da776db | 1999 | "block %d chunk %" PRIu64 |
733252de | 2000 | " current %" PRIu64 " len %" PRIu64 " %d", |
2da776db MH |
2001 | current_index, chunk, sge.addr, length, rdma->nb_sent); |
2002 | return ret; | |
2003 | } | |
2004 | } | |
2005 | ||
2006 | if (!rdma->pin_all || !block->is_ram_block) { | |
2007 | if (!block->remote_keys[chunk]) { | |
2008 | /* | |
2009 | * This chunk has not yet been registered, so first check to see | |
2010 | * if the entire chunk is zero. If so, tell the other size to | |
2011 | * memset() + madvise() the entire chunk without RDMA. | |
2012 | */ | |
2013 | ||
a1febc49 | 2014 | if (buffer_is_zero((void *)(uintptr_t)sge.addr, length)) { |
2da776db MH |
2015 | RDMACompress comp = { |
2016 | .offset = current_addr, | |
2017 | .value = 0, | |
2018 | .block_idx = current_index, | |
2019 | .length = length, | |
2020 | }; | |
2021 | ||
2022 | head.len = sizeof(comp); | |
2023 | head.type = RDMA_CONTROL_COMPRESS; | |
2024 | ||
733252de DDAG |
2025 | trace_qemu_rdma_write_one_zero(chunk, sge.length, |
2026 | current_index, current_addr); | |
2da776db | 2027 | |
b12f7777 | 2028 | compress_to_network(rdma, &comp); |
2da776db MH |
2029 | ret = qemu_rdma_exchange_send(rdma, &head, |
2030 | (uint8_t *) &comp, NULL, NULL, NULL); | |
2031 | ||
2032 | if (ret < 0) { | |
2033 | return -EIO; | |
2034 | } | |
2035 | ||
2036 | acct_update_position(f, sge.length, true); | |
2037 | ||
2038 | return 1; | |
2039 | } | |
2040 | ||
2041 | /* | |
2042 | * Otherwise, tell other side to register. | |
2043 | */ | |
2044 | reg.current_index = current_index; | |
2045 | if (block->is_ram_block) { | |
2046 | reg.key.current_addr = current_addr; | |
2047 | } else { | |
2048 | reg.key.chunk = chunk; | |
2049 | } | |
2050 | reg.chunks = chunks; | |
2051 | ||
733252de DDAG |
2052 | trace_qemu_rdma_write_one_sendreg(chunk, sge.length, current_index, |
2053 | current_addr); | |
2da776db | 2054 | |
b12f7777 | 2055 | register_to_network(rdma, ®); |
2da776db MH |
2056 | ret = qemu_rdma_exchange_send(rdma, &head, (uint8_t *) ®, |
2057 | &resp, ®_result_idx, NULL); | |
2058 | if (ret < 0) { | |
2059 | return ret; | |
2060 | } | |
2061 | ||
2062 | /* try to overlap this single registration with the one we sent. */ | |
3ac040c0 | 2063 | if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr, |
2da776db MH |
2064 | &sge.lkey, NULL, chunk, |
2065 | chunk_start, chunk_end)) { | |
733252de | 2066 | error_report("cannot get lkey"); |
2da776db MH |
2067 | return -EINVAL; |
2068 | } | |
2069 | ||
2070 | reg_result = (RDMARegisterResult *) | |
2071 | rdma->wr_data[reg_result_idx].control_curr; | |
2072 | ||
2073 | network_to_result(reg_result); | |
2074 | ||
733252de DDAG |
2075 | trace_qemu_rdma_write_one_recvregres(block->remote_keys[chunk], |
2076 | reg_result->rkey, chunk); | |
2da776db MH |
2077 | |
2078 | block->remote_keys[chunk] = reg_result->rkey; | |
2079 | block->remote_host_addr = reg_result->host_addr; | |
2080 | } else { | |
2081 | /* already registered before */ | |
3ac040c0 | 2082 | if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr, |
2da776db MH |
2083 | &sge.lkey, NULL, chunk, |
2084 | chunk_start, chunk_end)) { | |
733252de | 2085 | error_report("cannot get lkey!"); |
2da776db MH |
2086 | return -EINVAL; |
2087 | } | |
2088 | } | |
2089 | ||
2090 | send_wr.wr.rdma.rkey = block->remote_keys[chunk]; | |
2091 | } else { | |
2092 | send_wr.wr.rdma.rkey = block->remote_rkey; | |
2093 | ||
3ac040c0 | 2094 | if (qemu_rdma_register_and_get_keys(rdma, block, sge.addr, |
2da776db MH |
2095 | &sge.lkey, NULL, chunk, |
2096 | chunk_start, chunk_end)) { | |
733252de | 2097 | error_report("cannot get lkey!"); |
2da776db MH |
2098 | return -EINVAL; |
2099 | } | |
2100 | } | |
2101 | ||
2102 | /* | |
2103 | * Encode the ram block index and chunk within this wrid. | |
2104 | * We will use this information at the time of completion | |
2105 | * to figure out which bitmap to check against and then which | |
2106 | * chunk in the bitmap to look for. | |
2107 | */ | |
2108 | send_wr.wr_id = qemu_rdma_make_wrid(RDMA_WRID_RDMA_WRITE, | |
2109 | current_index, chunk); | |
2110 | ||
2111 | send_wr.opcode = IBV_WR_RDMA_WRITE; | |
2112 | send_wr.send_flags = IBV_SEND_SIGNALED; | |
2113 | send_wr.sg_list = &sge; | |
2114 | send_wr.num_sge = 1; | |
2115 | send_wr.wr.rdma.remote_addr = block->remote_host_addr + | |
2116 | (current_addr - block->offset); | |
2117 | ||
733252de DDAG |
2118 | trace_qemu_rdma_write_one_post(chunk, sge.addr, send_wr.wr.rdma.remote_addr, |
2119 | sge.length); | |
2da776db MH |
2120 | |
2121 | /* | |
2122 | * ibv_post_send() does not return negative error numbers, | |
2123 | * per the specification they are positive - no idea why. | |
2124 | */ | |
2125 | ret = ibv_post_send(rdma->qp, &send_wr, &bad_wr); | |
2126 | ||
2127 | if (ret == ENOMEM) { | |
733252de | 2128 | trace_qemu_rdma_write_one_queue_full(); |
88571882 | 2129 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL); |
2da776db | 2130 | if (ret < 0) { |
733252de DDAG |
2131 | error_report("rdma migration: failed to make " |
2132 | "room in full send queue! %d", ret); | |
2da776db MH |
2133 | return ret; |
2134 | } | |
2135 | ||
2136 | goto retry; | |
2137 | ||
2138 | } else if (ret > 0) { | |
2139 | perror("rdma migration: post rdma write failed"); | |
2140 | return -ret; | |
2141 | } | |
2142 | ||
2143 | set_bit(chunk, block->transit_bitmap); | |
2144 | acct_update_position(f, sge.length, false); | |
2145 | rdma->total_writes++; | |
2146 | ||
2147 | return 0; | |
2148 | } | |
2149 | ||
2150 | /* | |
2151 | * Push out any unwritten RDMA operations. | |
2152 | * | |
2153 | * We support sending out multiple chunks at the same time. | |
2154 | * Not all of them need to get signaled in the completion queue. | |
2155 | */ | |
2156 | static int qemu_rdma_write_flush(QEMUFile *f, RDMAContext *rdma) | |
2157 | { | |
2158 | int ret; | |
2159 | ||
2160 | if (!rdma->current_length) { | |
2161 | return 0; | |
2162 | } | |
2163 | ||
2164 | ret = qemu_rdma_write_one(f, rdma, | |
2165 | rdma->current_index, rdma->current_addr, rdma->current_length); | |
2166 | ||
2167 | if (ret < 0) { | |
2168 | return ret; | |
2169 | } | |
2170 | ||
2171 | if (ret == 0) { | |
2172 | rdma->nb_sent++; | |
733252de | 2173 | trace_qemu_rdma_write_flush(rdma->nb_sent); |
2da776db MH |
2174 | } |
2175 | ||
2176 | rdma->current_length = 0; | |
2177 | rdma->current_addr = 0; | |
2178 | ||
2179 | return 0; | |
2180 | } | |
2181 | ||
2182 | static inline int qemu_rdma_buffer_mergable(RDMAContext *rdma, | |
2183 | uint64_t offset, uint64_t len) | |
2184 | { | |
44b59494 IY |
2185 | RDMALocalBlock *block; |
2186 | uint8_t *host_addr; | |
2187 | uint8_t *chunk_end; | |
2188 | ||
2189 | if (rdma->current_index < 0) { | |
2190 | return 0; | |
2191 | } | |
2192 | ||
2193 | if (rdma->current_chunk < 0) { | |
2194 | return 0; | |
2195 | } | |
2196 | ||
2197 | block = &(rdma->local_ram_blocks.block[rdma->current_index]); | |
2198 | host_addr = block->local_host_addr + (offset - block->offset); | |
2199 | chunk_end = ram_chunk_end(block, rdma->current_chunk); | |
2da776db MH |
2200 | |
2201 | if (rdma->current_length == 0) { | |
2202 | return 0; | |
2203 | } | |
2204 | ||
2205 | /* | |
2206 | * Only merge into chunk sequentially. | |
2207 | */ | |
2208 | if (offset != (rdma->current_addr + rdma->current_length)) { | |
2209 | return 0; | |
2210 | } | |
2211 | ||
2da776db MH |
2212 | if (offset < block->offset) { |
2213 | return 0; | |
2214 | } | |
2215 | ||
2216 | if ((offset + len) > (block->offset + block->length)) { | |
2217 | return 0; | |
2218 | } | |
2219 | ||
2da776db MH |
2220 | if ((host_addr + len) > chunk_end) { |
2221 | return 0; | |
2222 | } | |
2223 | ||
2224 | return 1; | |
2225 | } | |
2226 | ||
2227 | /* | |
2228 | * We're not actually writing here, but doing three things: | |
2229 | * | |
2230 | * 1. Identify the chunk the buffer belongs to. | |
2231 | * 2. If the chunk is full or the buffer doesn't belong to the current | |
2232 | * chunk, then start a new chunk and flush() the old chunk. | |
2233 | * 3. To keep the hardware busy, we also group chunks into batches | |
2234 | * and only require that a batch gets acknowledged in the completion | |
2235 | * qeueue instead of each individual chunk. | |
2236 | */ | |
2237 | static int qemu_rdma_write(QEMUFile *f, RDMAContext *rdma, | |
2238 | uint64_t block_offset, uint64_t offset, | |
2239 | uint64_t len) | |
2240 | { | |
2241 | uint64_t current_addr = block_offset + offset; | |
2242 | uint64_t index = rdma->current_index; | |
2243 | uint64_t chunk = rdma->current_chunk; | |
2244 | int ret; | |
2245 | ||
2246 | /* If we cannot merge it, we flush the current buffer first. */ | |
2247 | if (!qemu_rdma_buffer_mergable(rdma, current_addr, len)) { | |
2248 | ret = qemu_rdma_write_flush(f, rdma); | |
2249 | if (ret) { | |
2250 | return ret; | |
2251 | } | |
2252 | rdma->current_length = 0; | |
2253 | rdma->current_addr = current_addr; | |
2254 | ||
2255 | ret = qemu_rdma_search_ram_block(rdma, block_offset, | |
2256 | offset, len, &index, &chunk); | |
2257 | if (ret) { | |
733252de | 2258 | error_report("ram block search failed"); |
2da776db MH |
2259 | return ret; |
2260 | } | |
2261 | rdma->current_index = index; | |
2262 | rdma->current_chunk = chunk; | |
2263 | } | |
2264 | ||
2265 | /* merge it */ | |
2266 | rdma->current_length += len; | |
2267 | ||
2268 | /* flush it if buffer is too large */ | |
2269 | if (rdma->current_length >= RDMA_MERGE_MAX) { | |
2270 | return qemu_rdma_write_flush(f, rdma); | |
2271 | } | |
2272 | ||
2273 | return 0; | |
2274 | } | |
2275 | ||
2276 | static void qemu_rdma_cleanup(RDMAContext *rdma) | |
2277 | { | |
c5e76115 | 2278 | int idx; |
2da776db | 2279 | |
5a91337c | 2280 | if (rdma->cm_id && rdma->connected) { |
32bce196 DDAG |
2281 | if ((rdma->error_state || |
2282 | migrate_get_current()->state == MIGRATION_STATUS_CANCELLING) && | |
2283 | !rdma->received_error) { | |
2da776db MH |
2284 | RDMAControlHeader head = { .len = 0, |
2285 | .type = RDMA_CONTROL_ERROR, | |
2286 | .repeat = 1, | |
2287 | }; | |
733252de | 2288 | error_report("Early error. Sending error."); |
2da776db MH |
2289 | qemu_rdma_post_send_control(rdma, NULL, &head); |
2290 | } | |
2291 | ||
c5e76115 | 2292 | rdma_disconnect(rdma->cm_id); |
733252de | 2293 | trace_qemu_rdma_cleanup_disconnect(); |
5a91337c | 2294 | rdma->connected = false; |
2da776db MH |
2295 | } |
2296 | ||
a97270ad DDAG |
2297 | g_free(rdma->dest_blocks); |
2298 | rdma->dest_blocks = NULL; | |
2da776db | 2299 | |
1f22364b | 2300 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2301 | if (rdma->wr_data[idx].control_mr) { |
2302 | rdma->total_registrations--; | |
2303 | ibv_dereg_mr(rdma->wr_data[idx].control_mr); | |
2304 | } | |
2305 | rdma->wr_data[idx].control_mr = NULL; | |
2306 | } | |
2307 | ||
2308 | if (rdma->local_ram_blocks.block) { | |
2309 | while (rdma->local_ram_blocks.nb_blocks) { | |
03fcab38 | 2310 | rdma_delete_block(rdma, &rdma->local_ram_blocks.block[0]); |
2da776db MH |
2311 | } |
2312 | } | |
2313 | ||
80b262e1 PR |
2314 | if (rdma->qp) { |
2315 | rdma_destroy_qp(rdma->cm_id); | |
2316 | rdma->qp = NULL; | |
2317 | } | |
2da776db MH |
2318 | if (rdma->cq) { |
2319 | ibv_destroy_cq(rdma->cq); | |
2320 | rdma->cq = NULL; | |
2321 | } | |
2322 | if (rdma->comp_channel) { | |
2323 | ibv_destroy_comp_channel(rdma->comp_channel); | |
2324 | rdma->comp_channel = NULL; | |
2325 | } | |
2326 | if (rdma->pd) { | |
2327 | ibv_dealloc_pd(rdma->pd); | |
2328 | rdma->pd = NULL; | |
2329 | } | |
2da776db MH |
2330 | if (rdma->cm_id) { |
2331 | rdma_destroy_id(rdma->cm_id); | |
2332 | rdma->cm_id = NULL; | |
2333 | } | |
55cc1b59 LC |
2334 | |
2335 | /* the destination side, listen_id and channel is shared */ | |
80b262e1 | 2336 | if (rdma->listen_id) { |
55cc1b59 LC |
2337 | if (!rdma->is_return_path) { |
2338 | rdma_destroy_id(rdma->listen_id); | |
2339 | } | |
80b262e1 | 2340 | rdma->listen_id = NULL; |
55cc1b59 LC |
2341 | |
2342 | if (rdma->channel) { | |
2343 | if (!rdma->is_return_path) { | |
2344 | rdma_destroy_event_channel(rdma->channel); | |
2345 | } | |
2346 | rdma->channel = NULL; | |
2347 | } | |
80b262e1 | 2348 | } |
55cc1b59 | 2349 | |
2da776db MH |
2350 | if (rdma->channel) { |
2351 | rdma_destroy_event_channel(rdma->channel); | |
2352 | rdma->channel = NULL; | |
2353 | } | |
e1d0fb37 IY |
2354 | g_free(rdma->host); |
2355 | rdma->host = NULL; | |
2da776db MH |
2356 | } |
2357 | ||
2358 | ||
bbfb89e3 | 2359 | static int qemu_rdma_source_init(RDMAContext *rdma, bool pin_all, Error **errp) |
2da776db MH |
2360 | { |
2361 | int ret, idx; | |
2362 | Error *local_err = NULL, **temp = &local_err; | |
2363 | ||
2364 | /* | |
2365 | * Will be validated against destination's actual capabilities | |
2366 | * after the connect() completes. | |
2367 | */ | |
2368 | rdma->pin_all = pin_all; | |
2369 | ||
2370 | ret = qemu_rdma_resolve_host(rdma, temp); | |
2371 | if (ret) { | |
2372 | goto err_rdma_source_init; | |
2373 | } | |
2374 | ||
2375 | ret = qemu_rdma_alloc_pd_cq(rdma); | |
2376 | if (ret) { | |
2377 | ERROR(temp, "rdma migration: error allocating pd and cq! Your mlock()" | |
2378 | " limits may be too low. Please check $ ulimit -a # and " | |
66988941 | 2379 | "search for 'ulimit -l' in the output"); |
2da776db MH |
2380 | goto err_rdma_source_init; |
2381 | } | |
2382 | ||
2383 | ret = qemu_rdma_alloc_qp(rdma); | |
2384 | if (ret) { | |
66988941 | 2385 | ERROR(temp, "rdma migration: error allocating qp!"); |
2da776db MH |
2386 | goto err_rdma_source_init; |
2387 | } | |
2388 | ||
2389 | ret = qemu_rdma_init_ram_blocks(rdma); | |
2390 | if (ret) { | |
66988941 | 2391 | ERROR(temp, "rdma migration: error initializing ram blocks!"); |
2da776db MH |
2392 | goto err_rdma_source_init; |
2393 | } | |
2394 | ||
760ff4be DDAG |
2395 | /* Build the hash that maps from offset to RAMBlock */ |
2396 | rdma->blockmap = g_hash_table_new(g_direct_hash, g_direct_equal); | |
2397 | for (idx = 0; idx < rdma->local_ram_blocks.nb_blocks; idx++) { | |
2398 | g_hash_table_insert(rdma->blockmap, | |
2399 | (void *)(uintptr_t)rdma->local_ram_blocks.block[idx].offset, | |
2400 | &rdma->local_ram_blocks.block[idx]); | |
2401 | } | |
2402 | ||
1f22364b | 2403 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2404 | ret = qemu_rdma_reg_control(rdma, idx); |
2405 | if (ret) { | |
66988941 | 2406 | ERROR(temp, "rdma migration: error registering %d control!", |
2da776db MH |
2407 | idx); |
2408 | goto err_rdma_source_init; | |
2409 | } | |
2410 | } | |
2411 | ||
2412 | return 0; | |
2413 | ||
2414 | err_rdma_source_init: | |
2415 | error_propagate(errp, local_err); | |
2416 | qemu_rdma_cleanup(rdma); | |
2417 | return -1; | |
2418 | } | |
2419 | ||
2420 | static int qemu_rdma_connect(RDMAContext *rdma, Error **errp) | |
2421 | { | |
2422 | RDMACapabilities cap = { | |
2423 | .version = RDMA_CONTROL_VERSION_CURRENT, | |
2424 | .flags = 0, | |
2425 | }; | |
2426 | struct rdma_conn_param conn_param = { .initiator_depth = 2, | |
2427 | .retry_count = 5, | |
2428 | .private_data = &cap, | |
2429 | .private_data_len = sizeof(cap), | |
2430 | }; | |
2431 | struct rdma_cm_event *cm_event; | |
2432 | int ret; | |
2433 | ||
2434 | /* | |
2435 | * Only negotiate the capability with destination if the user | |
2436 | * on the source first requested the capability. | |
2437 | */ | |
2438 | if (rdma->pin_all) { | |
733252de | 2439 | trace_qemu_rdma_connect_pin_all_requested(); |
2da776db MH |
2440 | cap.flags |= RDMA_CAPABILITY_PIN_ALL; |
2441 | } | |
2442 | ||
2443 | caps_to_network(&cap); | |
2444 | ||
9cf2bab2 DDAG |
2445 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
2446 | if (ret) { | |
2447 | ERROR(errp, "posting second control recv"); | |
2448 | goto err_rdma_source_connect; | |
2449 | } | |
2450 | ||
2da776db MH |
2451 | ret = rdma_connect(rdma->cm_id, &conn_param); |
2452 | if (ret) { | |
2453 | perror("rdma_connect"); | |
66988941 | 2454 | ERROR(errp, "connecting to destination!"); |
2da776db MH |
2455 | goto err_rdma_source_connect; |
2456 | } | |
2457 | ||
2458 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
2459 | if (ret) { | |
2460 | perror("rdma_get_cm_event after rdma_connect"); | |
66988941 | 2461 | ERROR(errp, "connecting to destination!"); |
2da776db | 2462 | rdma_ack_cm_event(cm_event); |
2da776db MH |
2463 | goto err_rdma_source_connect; |
2464 | } | |
2465 | ||
2466 | if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) { | |
2467 | perror("rdma_get_cm_event != EVENT_ESTABLISHED after rdma_connect"); | |
66988941 | 2468 | ERROR(errp, "connecting to destination!"); |
2da776db | 2469 | rdma_ack_cm_event(cm_event); |
2da776db MH |
2470 | goto err_rdma_source_connect; |
2471 | } | |
5a91337c | 2472 | rdma->connected = true; |
2da776db MH |
2473 | |
2474 | memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap)); | |
2475 | network_to_caps(&cap); | |
2476 | ||
2477 | /* | |
2478 | * Verify that the *requested* capabilities are supported by the destination | |
2479 | * and disable them otherwise. | |
2480 | */ | |
2481 | if (rdma->pin_all && !(cap.flags & RDMA_CAPABILITY_PIN_ALL)) { | |
2482 | ERROR(errp, "Server cannot support pinning all memory. " | |
66988941 | 2483 | "Will register memory dynamically."); |
2da776db MH |
2484 | rdma->pin_all = false; |
2485 | } | |
2486 | ||
733252de | 2487 | trace_qemu_rdma_connect_pin_all_outcome(rdma->pin_all); |
2da776db MH |
2488 | |
2489 | rdma_ack_cm_event(cm_event); | |
2490 | ||
2da776db MH |
2491 | rdma->control_ready_expected = 1; |
2492 | rdma->nb_sent = 0; | |
2493 | return 0; | |
2494 | ||
2495 | err_rdma_source_connect: | |
2496 | qemu_rdma_cleanup(rdma); | |
2497 | return -1; | |
2498 | } | |
2499 | ||
2500 | static int qemu_rdma_dest_init(RDMAContext *rdma, Error **errp) | |
2501 | { | |
1dbd2fd9 | 2502 | int ret, idx; |
2da776db MH |
2503 | struct rdma_cm_id *listen_id; |
2504 | char ip[40] = "unknown"; | |
1dbd2fd9 | 2505 | struct rdma_addrinfo *res, *e; |
b58c8552 | 2506 | char port_str[16]; |
2da776db | 2507 | |
1f22364b | 2508 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
2509 | rdma->wr_data[idx].control_len = 0; |
2510 | rdma->wr_data[idx].control_curr = NULL; | |
2511 | } | |
2512 | ||
1dbd2fd9 | 2513 | if (!rdma->host || !rdma->host[0]) { |
66988941 | 2514 | ERROR(errp, "RDMA host is not set!"); |
2da776db MH |
2515 | rdma->error_state = -EINVAL; |
2516 | return -1; | |
2517 | } | |
2518 | /* create CM channel */ | |
2519 | rdma->channel = rdma_create_event_channel(); | |
2520 | if (!rdma->channel) { | |
66988941 | 2521 | ERROR(errp, "could not create rdma event channel"); |
2da776db MH |
2522 | rdma->error_state = -EINVAL; |
2523 | return -1; | |
2524 | } | |
2525 | ||
2526 | /* create CM id */ | |
2527 | ret = rdma_create_id(rdma->channel, &listen_id, NULL, RDMA_PS_TCP); | |
2528 | if (ret) { | |
66988941 | 2529 | ERROR(errp, "could not create cm_id!"); |
2da776db MH |
2530 | goto err_dest_init_create_listen_id; |
2531 | } | |
2532 | ||
b58c8552 MH |
2533 | snprintf(port_str, 16, "%d", rdma->port); |
2534 | port_str[15] = '\0'; | |
2da776db | 2535 | |
1dbd2fd9 MT |
2536 | ret = rdma_getaddrinfo(rdma->host, port_str, NULL, &res); |
2537 | if (ret < 0) { | |
2538 | ERROR(errp, "could not rdma_getaddrinfo address %s", rdma->host); | |
2539 | goto err_dest_init_bind_addr; | |
2540 | } | |
6470215b | 2541 | |
1dbd2fd9 MT |
2542 | for (e = res; e != NULL; e = e->ai_next) { |
2543 | inet_ntop(e->ai_family, | |
2544 | &((struct sockaddr_in *) e->ai_dst_addr)->sin_addr, ip, sizeof ip); | |
2545 | trace_qemu_rdma_dest_init_trying(rdma->host, ip); | |
2546 | ret = rdma_bind_addr(listen_id, e->ai_dst_addr); | |
2547 | if (ret) { | |
2548 | continue; | |
2da776db | 2549 | } |
1dbd2fd9 | 2550 | if (e->ai_family == AF_INET6) { |
bbfb89e3 | 2551 | ret = qemu_rdma_broken_ipv6_kernel(listen_id->verbs, errp); |
1dbd2fd9 MT |
2552 | if (ret) { |
2553 | continue; | |
6470215b MH |
2554 | } |
2555 | } | |
1dbd2fd9 MT |
2556 | break; |
2557 | } | |
b58c8552 | 2558 | |
1dbd2fd9 | 2559 | if (!e) { |
6470215b MH |
2560 | ERROR(errp, "Error: could not rdma_bind_addr!"); |
2561 | goto err_dest_init_bind_addr; | |
2da776db | 2562 | } |
2da776db MH |
2563 | |
2564 | rdma->listen_id = listen_id; | |
2565 | qemu_rdma_dump_gid("dest_init", listen_id); | |
2566 | return 0; | |
2567 | ||
2568 | err_dest_init_bind_addr: | |
2569 | rdma_destroy_id(listen_id); | |
2570 | err_dest_init_create_listen_id: | |
2571 | rdma_destroy_event_channel(rdma->channel); | |
2572 | rdma->channel = NULL; | |
2573 | rdma->error_state = ret; | |
2574 | return ret; | |
2575 | ||
2576 | } | |
2577 | ||
55cc1b59 LC |
2578 | static void qemu_rdma_return_path_dest_init(RDMAContext *rdma_return_path, |
2579 | RDMAContext *rdma) | |
2580 | { | |
2581 | int idx; | |
2582 | ||
2583 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { | |
2584 | rdma_return_path->wr_data[idx].control_len = 0; | |
2585 | rdma_return_path->wr_data[idx].control_curr = NULL; | |
2586 | } | |
2587 | ||
2588 | /*the CM channel and CM id is shared*/ | |
2589 | rdma_return_path->channel = rdma->channel; | |
2590 | rdma_return_path->listen_id = rdma->listen_id; | |
2591 | ||
2592 | rdma->return_path = rdma_return_path; | |
2593 | rdma_return_path->return_path = rdma; | |
2594 | rdma_return_path->is_return_path = true; | |
2595 | } | |
2596 | ||
2da776db MH |
2597 | static void *qemu_rdma_data_init(const char *host_port, Error **errp) |
2598 | { | |
2599 | RDMAContext *rdma = NULL; | |
2600 | InetSocketAddress *addr; | |
2601 | ||
2602 | if (host_port) { | |
97f3ad35 | 2603 | rdma = g_new0(RDMAContext, 1); |
2da776db MH |
2604 | rdma->current_index = -1; |
2605 | rdma->current_chunk = -1; | |
2606 | ||
0785bd7a MA |
2607 | addr = g_new(InetSocketAddress, 1); |
2608 | if (!inet_parse(addr, host_port, NULL)) { | |
2da776db MH |
2609 | rdma->port = atoi(addr->port); |
2610 | rdma->host = g_strdup(addr->host); | |
2611 | } else { | |
2612 | ERROR(errp, "bad RDMA migration address '%s'", host_port); | |
2613 | g_free(rdma); | |
e325b49a | 2614 | rdma = NULL; |
2da776db | 2615 | } |
e325b49a MH |
2616 | |
2617 | qapi_free_InetSocketAddress(addr); | |
2da776db MH |
2618 | } |
2619 | ||
2620 | return rdma; | |
2621 | } | |
2622 | ||
2623 | /* | |
2624 | * QEMUFile interface to the control channel. | |
2625 | * SEND messages for control only. | |
971ae6ef | 2626 | * VM's ram is handled with regular RDMA messages. |
2da776db | 2627 | */ |
6ddd2d76 DB |
2628 | static ssize_t qio_channel_rdma_writev(QIOChannel *ioc, |
2629 | const struct iovec *iov, | |
2630 | size_t niov, | |
2631 | int *fds, | |
2632 | size_t nfds, | |
2633 | Error **errp) | |
2634 | { | |
2635 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc); | |
2636 | QEMUFile *f = rioc->file; | |
74637e6f | 2637 | RDMAContext *rdma; |
2da776db | 2638 | int ret; |
6ddd2d76 DB |
2639 | ssize_t done = 0; |
2640 | size_t i; | |
f38f6d41 | 2641 | size_t len = 0; |
2da776db | 2642 | |
74637e6f LC |
2643 | rcu_read_lock(); |
2644 | rdma = atomic_rcu_read(&rioc->rdmaout); | |
2645 | ||
2646 | if (!rdma) { | |
2647 | rcu_read_unlock(); | |
2648 | return -EIO; | |
2649 | } | |
2650 | ||
2da776db MH |
2651 | CHECK_ERROR_STATE(); |
2652 | ||
2653 | /* | |
2654 | * Push out any writes that | |
971ae6ef | 2655 | * we're queued up for VM's ram. |
2da776db MH |
2656 | */ |
2657 | ret = qemu_rdma_write_flush(f, rdma); | |
2658 | if (ret < 0) { | |
2659 | rdma->error_state = ret; | |
74637e6f | 2660 | rcu_read_unlock(); |
2da776db MH |
2661 | return ret; |
2662 | } | |
2663 | ||
6ddd2d76 DB |
2664 | for (i = 0; i < niov; i++) { |
2665 | size_t remaining = iov[i].iov_len; | |
2666 | uint8_t * data = (void *)iov[i].iov_base; | |
2667 | while (remaining) { | |
2668 | RDMAControlHeader head; | |
2da776db | 2669 | |
f38f6d41 LC |
2670 | len = MIN(remaining, RDMA_SEND_INCREMENT); |
2671 | remaining -= len; | |
2da776db | 2672 | |
f38f6d41 | 2673 | head.len = len; |
6ddd2d76 | 2674 | head.type = RDMA_CONTROL_QEMU_FILE; |
2da776db | 2675 | |
6ddd2d76 | 2676 | ret = qemu_rdma_exchange_send(rdma, &head, data, NULL, NULL, NULL); |
2da776db | 2677 | |
6ddd2d76 DB |
2678 | if (ret < 0) { |
2679 | rdma->error_state = ret; | |
74637e6f | 2680 | rcu_read_unlock(); |
6ddd2d76 DB |
2681 | return ret; |
2682 | } | |
2da776db | 2683 | |
f38f6d41 LC |
2684 | data += len; |
2685 | done += len; | |
6ddd2d76 | 2686 | } |
2da776db MH |
2687 | } |
2688 | ||
74637e6f | 2689 | rcu_read_unlock(); |
6ddd2d76 | 2690 | return done; |
2da776db MH |
2691 | } |
2692 | ||
2693 | static size_t qemu_rdma_fill(RDMAContext *rdma, uint8_t *buf, | |
a202a4c0 | 2694 | size_t size, int idx) |
2da776db MH |
2695 | { |
2696 | size_t len = 0; | |
2697 | ||
2698 | if (rdma->wr_data[idx].control_len) { | |
733252de | 2699 | trace_qemu_rdma_fill(rdma->wr_data[idx].control_len, size); |
2da776db MH |
2700 | |
2701 | len = MIN(size, rdma->wr_data[idx].control_len); | |
2702 | memcpy(buf, rdma->wr_data[idx].control_curr, len); | |
2703 | rdma->wr_data[idx].control_curr += len; | |
2704 | rdma->wr_data[idx].control_len -= len; | |
2705 | } | |
2706 | ||
2707 | return len; | |
2708 | } | |
2709 | ||
2710 | /* | |
2711 | * QEMUFile interface to the control channel. | |
2712 | * RDMA links don't use bytestreams, so we have to | |
2713 | * return bytes to QEMUFile opportunistically. | |
2714 | */ | |
6ddd2d76 DB |
2715 | static ssize_t qio_channel_rdma_readv(QIOChannel *ioc, |
2716 | const struct iovec *iov, | |
2717 | size_t niov, | |
2718 | int **fds, | |
2719 | size_t *nfds, | |
2720 | Error **errp) | |
2721 | { | |
2722 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc); | |
74637e6f | 2723 | RDMAContext *rdma; |
2da776db MH |
2724 | RDMAControlHeader head; |
2725 | int ret = 0; | |
6ddd2d76 DB |
2726 | ssize_t i; |
2727 | size_t done = 0; | |
2da776db | 2728 | |
74637e6f LC |
2729 | rcu_read_lock(); |
2730 | rdma = atomic_rcu_read(&rioc->rdmain); | |
2731 | ||
2732 | if (!rdma) { | |
2733 | rcu_read_unlock(); | |
2734 | return -EIO; | |
2735 | } | |
2736 | ||
2da776db MH |
2737 | CHECK_ERROR_STATE(); |
2738 | ||
6ddd2d76 DB |
2739 | for (i = 0; i < niov; i++) { |
2740 | size_t want = iov[i].iov_len; | |
2741 | uint8_t *data = (void *)iov[i].iov_base; | |
2da776db | 2742 | |
6ddd2d76 DB |
2743 | /* |
2744 | * First, we hold on to the last SEND message we | |
2745 | * were given and dish out the bytes until we run | |
2746 | * out of bytes. | |
2747 | */ | |
74637e6f | 2748 | ret = qemu_rdma_fill(rdma, data, want, 0); |
6ddd2d76 DB |
2749 | done += ret; |
2750 | want -= ret; | |
2751 | /* Got what we needed, so go to next iovec */ | |
2752 | if (want == 0) { | |
2753 | continue; | |
2754 | } | |
2da776db | 2755 | |
6ddd2d76 DB |
2756 | /* If we got any data so far, then don't wait |
2757 | * for more, just return what we have */ | |
2758 | if (done > 0) { | |
2759 | break; | |
2760 | } | |
2da776db | 2761 | |
6ddd2d76 DB |
2762 | |
2763 | /* We've got nothing at all, so lets wait for | |
2764 | * more to arrive | |
2765 | */ | |
2766 | ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_QEMU_FILE); | |
2767 | ||
2768 | if (ret < 0) { | |
2769 | rdma->error_state = ret; | |
74637e6f | 2770 | rcu_read_unlock(); |
6ddd2d76 DB |
2771 | return ret; |
2772 | } | |
2773 | ||
2774 | /* | |
2775 | * SEND was received with new bytes, now try again. | |
2776 | */ | |
74637e6f | 2777 | ret = qemu_rdma_fill(rdma, data, want, 0); |
6ddd2d76 DB |
2778 | done += ret; |
2779 | want -= ret; | |
2780 | ||
2781 | /* Still didn't get enough, so lets just return */ | |
2782 | if (want) { | |
2783 | if (done == 0) { | |
74637e6f | 2784 | rcu_read_unlock(); |
6ddd2d76 DB |
2785 | return QIO_CHANNEL_ERR_BLOCK; |
2786 | } else { | |
2787 | break; | |
2788 | } | |
2789 | } | |
2790 | } | |
74637e6f | 2791 | rcu_read_unlock(); |
f38f6d41 | 2792 | return done; |
2da776db MH |
2793 | } |
2794 | ||
2795 | /* | |
2796 | * Block until all the outstanding chunks have been delivered by the hardware. | |
2797 | */ | |
2798 | static int qemu_rdma_drain_cq(QEMUFile *f, RDMAContext *rdma) | |
2799 | { | |
2800 | int ret; | |
2801 | ||
2802 | if (qemu_rdma_write_flush(f, rdma) < 0) { | |
2803 | return -EIO; | |
2804 | } | |
2805 | ||
2806 | while (rdma->nb_sent) { | |
88571882 | 2807 | ret = qemu_rdma_block_for_wrid(rdma, RDMA_WRID_RDMA_WRITE, NULL); |
2da776db | 2808 | if (ret < 0) { |
733252de | 2809 | error_report("rdma migration: complete polling error!"); |
2da776db MH |
2810 | return -EIO; |
2811 | } | |
2812 | } | |
2813 | ||
2814 | qemu_rdma_unregister_waiting(rdma); | |
2815 | ||
2816 | return 0; | |
2817 | } | |
2818 | ||
6ddd2d76 DB |
2819 | |
2820 | static int qio_channel_rdma_set_blocking(QIOChannel *ioc, | |
2821 | bool blocking, | |
2822 | Error **errp) | |
2823 | { | |
2824 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc); | |
2825 | /* XXX we should make readv/writev actually honour this :-) */ | |
2826 | rioc->blocking = blocking; | |
2827 | return 0; | |
2828 | } | |
2829 | ||
2830 | ||
2831 | typedef struct QIOChannelRDMASource QIOChannelRDMASource; | |
2832 | struct QIOChannelRDMASource { | |
2833 | GSource parent; | |
2834 | QIOChannelRDMA *rioc; | |
2835 | GIOCondition condition; | |
2836 | }; | |
2837 | ||
2838 | static gboolean | |
2839 | qio_channel_rdma_source_prepare(GSource *source, | |
2840 | gint *timeout) | |
2841 | { | |
2842 | QIOChannelRDMASource *rsource = (QIOChannelRDMASource *)source; | |
74637e6f | 2843 | RDMAContext *rdma; |
6ddd2d76 DB |
2844 | GIOCondition cond = 0; |
2845 | *timeout = -1; | |
2846 | ||
74637e6f LC |
2847 | rcu_read_lock(); |
2848 | if (rsource->condition == G_IO_IN) { | |
2849 | rdma = atomic_rcu_read(&rsource->rioc->rdmain); | |
2850 | } else { | |
2851 | rdma = atomic_rcu_read(&rsource->rioc->rdmaout); | |
2852 | } | |
2853 | ||
2854 | if (!rdma) { | |
2855 | error_report("RDMAContext is NULL when prepare Gsource"); | |
2856 | rcu_read_unlock(); | |
2857 | return FALSE; | |
2858 | } | |
2859 | ||
6ddd2d76 DB |
2860 | if (rdma->wr_data[0].control_len) { |
2861 | cond |= G_IO_IN; | |
2862 | } | |
2863 | cond |= G_IO_OUT; | |
2864 | ||
74637e6f | 2865 | rcu_read_unlock(); |
6ddd2d76 DB |
2866 | return cond & rsource->condition; |
2867 | } | |
2868 | ||
2869 | static gboolean | |
2870 | qio_channel_rdma_source_check(GSource *source) | |
2871 | { | |
2872 | QIOChannelRDMASource *rsource = (QIOChannelRDMASource *)source; | |
74637e6f | 2873 | RDMAContext *rdma; |
6ddd2d76 DB |
2874 | GIOCondition cond = 0; |
2875 | ||
74637e6f LC |
2876 | rcu_read_lock(); |
2877 | if (rsource->condition == G_IO_IN) { | |
2878 | rdma = atomic_rcu_read(&rsource->rioc->rdmain); | |
2879 | } else { | |
2880 | rdma = atomic_rcu_read(&rsource->rioc->rdmaout); | |
2881 | } | |
2882 | ||
2883 | if (!rdma) { | |
2884 | error_report("RDMAContext is NULL when check Gsource"); | |
2885 | rcu_read_unlock(); | |
2886 | return FALSE; | |
2887 | } | |
2888 | ||
6ddd2d76 DB |
2889 | if (rdma->wr_data[0].control_len) { |
2890 | cond |= G_IO_IN; | |
2891 | } | |
2892 | cond |= G_IO_OUT; | |
2893 | ||
74637e6f | 2894 | rcu_read_unlock(); |
6ddd2d76 DB |
2895 | return cond & rsource->condition; |
2896 | } | |
2897 | ||
2898 | static gboolean | |
2899 | qio_channel_rdma_source_dispatch(GSource *source, | |
2900 | GSourceFunc callback, | |
2901 | gpointer user_data) | |
2902 | { | |
2903 | QIOChannelFunc func = (QIOChannelFunc)callback; | |
2904 | QIOChannelRDMASource *rsource = (QIOChannelRDMASource *)source; | |
74637e6f | 2905 | RDMAContext *rdma; |
6ddd2d76 DB |
2906 | GIOCondition cond = 0; |
2907 | ||
74637e6f LC |
2908 | rcu_read_lock(); |
2909 | if (rsource->condition == G_IO_IN) { | |
2910 | rdma = atomic_rcu_read(&rsource->rioc->rdmain); | |
2911 | } else { | |
2912 | rdma = atomic_rcu_read(&rsource->rioc->rdmaout); | |
2913 | } | |
2914 | ||
2915 | if (!rdma) { | |
2916 | error_report("RDMAContext is NULL when dispatch Gsource"); | |
2917 | rcu_read_unlock(); | |
2918 | return FALSE; | |
2919 | } | |
2920 | ||
6ddd2d76 DB |
2921 | if (rdma->wr_data[0].control_len) { |
2922 | cond |= G_IO_IN; | |
2923 | } | |
2924 | cond |= G_IO_OUT; | |
2925 | ||
74637e6f | 2926 | rcu_read_unlock(); |
6ddd2d76 DB |
2927 | return (*func)(QIO_CHANNEL(rsource->rioc), |
2928 | (cond & rsource->condition), | |
2929 | user_data); | |
2930 | } | |
2931 | ||
2932 | static void | |
2933 | qio_channel_rdma_source_finalize(GSource *source) | |
2934 | { | |
2935 | QIOChannelRDMASource *ssource = (QIOChannelRDMASource *)source; | |
2936 | ||
2937 | object_unref(OBJECT(ssource->rioc)); | |
2938 | } | |
2939 | ||
2940 | GSourceFuncs qio_channel_rdma_source_funcs = { | |
2941 | qio_channel_rdma_source_prepare, | |
2942 | qio_channel_rdma_source_check, | |
2943 | qio_channel_rdma_source_dispatch, | |
2944 | qio_channel_rdma_source_finalize | |
2945 | }; | |
2946 | ||
2947 | static GSource *qio_channel_rdma_create_watch(QIOChannel *ioc, | |
2948 | GIOCondition condition) | |
2da776db | 2949 | { |
6ddd2d76 DB |
2950 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc); |
2951 | QIOChannelRDMASource *ssource; | |
2952 | GSource *source; | |
2953 | ||
2954 | source = g_source_new(&qio_channel_rdma_source_funcs, | |
2955 | sizeof(QIOChannelRDMASource)); | |
2956 | ssource = (QIOChannelRDMASource *)source; | |
2957 | ||
2958 | ssource->rioc = rioc; | |
2959 | object_ref(OBJECT(rioc)); | |
2960 | ||
2961 | ssource->condition = condition; | |
2962 | ||
2963 | return source; | |
2964 | } | |
2965 | ||
4d9f675b LC |
2966 | static void qio_channel_rdma_set_aio_fd_handler(QIOChannel *ioc, |
2967 | AioContext *ctx, | |
2968 | IOHandler *io_read, | |
2969 | IOHandler *io_write, | |
2970 | void *opaque) | |
2971 | { | |
2972 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc); | |
2973 | if (io_read) { | |
2974 | aio_set_fd_handler(ctx, rioc->rdmain->comp_channel->fd, | |
2975 | false, io_read, io_write, NULL, opaque); | |
2976 | } else { | |
2977 | aio_set_fd_handler(ctx, rioc->rdmaout->comp_channel->fd, | |
2978 | false, io_read, io_write, NULL, opaque); | |
2979 | } | |
2980 | } | |
6ddd2d76 DB |
2981 | |
2982 | static int qio_channel_rdma_close(QIOChannel *ioc, | |
2983 | Error **errp) | |
2984 | { | |
2985 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(ioc); | |
74637e6f | 2986 | RDMAContext *rdmain, *rdmaout; |
733252de | 2987 | trace_qemu_rdma_close(); |
74637e6f LC |
2988 | |
2989 | rdmain = rioc->rdmain; | |
2990 | if (rdmain) { | |
2991 | atomic_rcu_set(&rioc->rdmain, NULL); | |
2992 | } | |
2993 | ||
2994 | rdmaout = rioc->rdmaout; | |
2995 | if (rdmaout) { | |
2996 | atomic_rcu_set(&rioc->rdmaout, NULL); | |
2da776db | 2997 | } |
74637e6f LC |
2998 | |
2999 | synchronize_rcu(); | |
3000 | ||
3001 | if (rdmain) { | |
3002 | qemu_rdma_cleanup(rdmain); | |
3003 | } | |
3004 | ||
3005 | if (rdmaout) { | |
3006 | qemu_rdma_cleanup(rdmaout); | |
3007 | } | |
3008 | ||
3009 | g_free(rdmain); | |
3010 | g_free(rdmaout); | |
3011 | ||
2da776db MH |
3012 | return 0; |
3013 | } | |
3014 | ||
3015 | /* | |
3016 | * Parameters: | |
3017 | * @offset == 0 : | |
3018 | * This means that 'block_offset' is a full virtual address that does not | |
3019 | * belong to a RAMBlock of the virtual machine and instead | |
3020 | * represents a private malloc'd memory area that the caller wishes to | |
3021 | * transfer. | |
3022 | * | |
3023 | * @offset != 0 : | |
3024 | * Offset is an offset to be added to block_offset and used | |
3025 | * to also lookup the corresponding RAMBlock. | |
3026 | * | |
3027 | * @size > 0 : | |
3028 | * Initiate an transfer this size. | |
3029 | * | |
3030 | * @size == 0 : | |
3031 | * A 'hint' or 'advice' that means that we wish to speculatively | |
3032 | * and asynchronously unregister this memory. In this case, there is no | |
52f35022 | 3033 | * guarantee that the unregister will actually happen, for example, |
2da776db MH |
3034 | * if the memory is being actively transmitted. Additionally, the memory |
3035 | * may be re-registered at any future time if a write within the same | |
3036 | * chunk was requested again, even if you attempted to unregister it | |
3037 | * here. | |
3038 | * | |
3039 | * @size < 0 : TODO, not yet supported | |
3040 | * Unregister the memory NOW. This means that the caller does not | |
3041 | * expect there to be any future RDMA transfers and we just want to clean | |
3042 | * things up. This is used in case the upper layer owns the memory and | |
3043 | * cannot wait for qemu_fclose() to occur. | |
3044 | * | |
3045 | * @bytes_sent : User-specificed pointer to indicate how many bytes were | |
3046 | * sent. Usually, this will not be more than a few bytes of | |
3047 | * the protocol because most transfers are sent asynchronously. | |
3048 | */ | |
3049 | static size_t qemu_rdma_save_page(QEMUFile *f, void *opaque, | |
3050 | ram_addr_t block_offset, ram_addr_t offset, | |
6e1dea46 | 3051 | size_t size, uint64_t *bytes_sent) |
2da776db | 3052 | { |
6ddd2d76 | 3053 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(opaque); |
74637e6f | 3054 | RDMAContext *rdma; |
2da776db MH |
3055 | int ret; |
3056 | ||
74637e6f LC |
3057 | rcu_read_lock(); |
3058 | rdma = atomic_rcu_read(&rioc->rdmaout); | |
3059 | ||
3060 | if (!rdma) { | |
3061 | rcu_read_unlock(); | |
3062 | return -EIO; | |
3063 | } | |
3064 | ||
2da776db MH |
3065 | CHECK_ERROR_STATE(); |
3066 | ||
ccb7e1b5 | 3067 | if (migrate_get_current()->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { |
74637e6f | 3068 | rcu_read_unlock(); |
ccb7e1b5 LC |
3069 | return RAM_SAVE_CONTROL_NOT_SUPP; |
3070 | } | |
3071 | ||
2da776db MH |
3072 | qemu_fflush(f); |
3073 | ||
3074 | if (size > 0) { | |
3075 | /* | |
3076 | * Add this page to the current 'chunk'. If the chunk | |
3077 | * is full, or the page doen't belong to the current chunk, | |
3078 | * an actual RDMA write will occur and a new chunk will be formed. | |
3079 | */ | |
3080 | ret = qemu_rdma_write(f, rdma, block_offset, offset, size); | |
3081 | if (ret < 0) { | |
733252de | 3082 | error_report("rdma migration: write error! %d", ret); |
2da776db MH |
3083 | goto err; |
3084 | } | |
3085 | ||
3086 | /* | |
3087 | * We always return 1 bytes because the RDMA | |
3088 | * protocol is completely asynchronous. We do not yet know | |
3089 | * whether an identified chunk is zero or not because we're | |
3090 | * waiting for other pages to potentially be merged with | |
3091 | * the current chunk. So, we have to call qemu_update_position() | |
3092 | * later on when the actual write occurs. | |
3093 | */ | |
3094 | if (bytes_sent) { | |
3095 | *bytes_sent = 1; | |
3096 | } | |
3097 | } else { | |
3098 | uint64_t index, chunk; | |
3099 | ||
3100 | /* TODO: Change QEMUFileOps prototype to be signed: size_t => long | |
3101 | if (size < 0) { | |
3102 | ret = qemu_rdma_drain_cq(f, rdma); | |
3103 | if (ret < 0) { | |
3104 | fprintf(stderr, "rdma: failed to synchronously drain" | |
3105 | " completion queue before unregistration.\n"); | |
3106 | goto err; | |
3107 | } | |
3108 | } | |
3109 | */ | |
3110 | ||
3111 | ret = qemu_rdma_search_ram_block(rdma, block_offset, | |
3112 | offset, size, &index, &chunk); | |
3113 | ||
3114 | if (ret) { | |
733252de | 3115 | error_report("ram block search failed"); |
2da776db MH |
3116 | goto err; |
3117 | } | |
3118 | ||
3119 | qemu_rdma_signal_unregister(rdma, index, chunk, 0); | |
3120 | ||
3121 | /* | |
52f35022 | 3122 | * TODO: Synchronous, guaranteed unregistration (should not occur during |
2da776db MH |
3123 | * fast-path). Otherwise, unregisters will process on the next call to |
3124 | * qemu_rdma_drain_cq() | |
3125 | if (size < 0) { | |
3126 | qemu_rdma_unregister_waiting(rdma); | |
3127 | } | |
3128 | */ | |
3129 | } | |
3130 | ||
3131 | /* | |
3132 | * Drain the Completion Queue if possible, but do not block, | |
3133 | * just poll. | |
3134 | * | |
3135 | * If nothing to poll, the end of the iteration will do this | |
3136 | * again to make sure we don't overflow the request queue. | |
3137 | */ | |
3138 | while (1) { | |
3139 | uint64_t wr_id, wr_id_in; | |
88571882 | 3140 | int ret = qemu_rdma_poll(rdma, &wr_id_in, NULL); |
2da776db | 3141 | if (ret < 0) { |
733252de | 3142 | error_report("rdma migration: polling error! %d", ret); |
2da776db MH |
3143 | goto err; |
3144 | } | |
3145 | ||
3146 | wr_id = wr_id_in & RDMA_WRID_TYPE_MASK; | |
3147 | ||
3148 | if (wr_id == RDMA_WRID_NONE) { | |
3149 | break; | |
3150 | } | |
3151 | } | |
3152 | ||
74637e6f | 3153 | rcu_read_unlock(); |
2da776db MH |
3154 | return RAM_SAVE_CONTROL_DELAYED; |
3155 | err: | |
3156 | rdma->error_state = ret; | |
74637e6f | 3157 | rcu_read_unlock(); |
2da776db MH |
3158 | return ret; |
3159 | } | |
3160 | ||
55cc1b59 LC |
3161 | static void rdma_accept_incoming_migration(void *opaque); |
3162 | ||
2da776db MH |
3163 | static int qemu_rdma_accept(RDMAContext *rdma) |
3164 | { | |
3165 | RDMACapabilities cap; | |
3166 | struct rdma_conn_param conn_param = { | |
3167 | .responder_resources = 2, | |
3168 | .private_data = &cap, | |
3169 | .private_data_len = sizeof(cap), | |
3170 | }; | |
3171 | struct rdma_cm_event *cm_event; | |
3172 | struct ibv_context *verbs; | |
3173 | int ret = -EINVAL; | |
3174 | int idx; | |
3175 | ||
3176 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
3177 | if (ret) { | |
3178 | goto err_rdma_dest_wait; | |
3179 | } | |
3180 | ||
3181 | if (cm_event->event != RDMA_CM_EVENT_CONNECT_REQUEST) { | |
3182 | rdma_ack_cm_event(cm_event); | |
3183 | goto err_rdma_dest_wait; | |
3184 | } | |
3185 | ||
3186 | memcpy(&cap, cm_event->param.conn.private_data, sizeof(cap)); | |
3187 | ||
3188 | network_to_caps(&cap); | |
3189 | ||
3190 | if (cap.version < 1 || cap.version > RDMA_CONTROL_VERSION_CURRENT) { | |
733252de | 3191 | error_report("Unknown source RDMA version: %d, bailing...", |
2da776db MH |
3192 | cap.version); |
3193 | rdma_ack_cm_event(cm_event); | |
3194 | goto err_rdma_dest_wait; | |
3195 | } | |
3196 | ||
3197 | /* | |
3198 | * Respond with only the capabilities this version of QEMU knows about. | |
3199 | */ | |
3200 | cap.flags &= known_capabilities; | |
3201 | ||
3202 | /* | |
3203 | * Enable the ones that we do know about. | |
3204 | * Add other checks here as new ones are introduced. | |
3205 | */ | |
3206 | if (cap.flags & RDMA_CAPABILITY_PIN_ALL) { | |
3207 | rdma->pin_all = true; | |
3208 | } | |
3209 | ||
3210 | rdma->cm_id = cm_event->id; | |
3211 | verbs = cm_event->id->verbs; | |
3212 | ||
3213 | rdma_ack_cm_event(cm_event); | |
3214 | ||
733252de | 3215 | trace_qemu_rdma_accept_pin_state(rdma->pin_all); |
2da776db MH |
3216 | |
3217 | caps_to_network(&cap); | |
3218 | ||
733252de | 3219 | trace_qemu_rdma_accept_pin_verbsc(verbs); |
2da776db MH |
3220 | |
3221 | if (!rdma->verbs) { | |
3222 | rdma->verbs = verbs; | |
3223 | } else if (rdma->verbs != verbs) { | |
733252de DDAG |
3224 | error_report("ibv context not matching %p, %p!", rdma->verbs, |
3225 | verbs); | |
2da776db MH |
3226 | goto err_rdma_dest_wait; |
3227 | } | |
3228 | ||
3229 | qemu_rdma_dump_id("dest_init", verbs); | |
3230 | ||
3231 | ret = qemu_rdma_alloc_pd_cq(rdma); | |
3232 | if (ret) { | |
733252de | 3233 | error_report("rdma migration: error allocating pd and cq!"); |
2da776db MH |
3234 | goto err_rdma_dest_wait; |
3235 | } | |
3236 | ||
3237 | ret = qemu_rdma_alloc_qp(rdma); | |
3238 | if (ret) { | |
733252de | 3239 | error_report("rdma migration: error allocating qp!"); |
2da776db MH |
3240 | goto err_rdma_dest_wait; |
3241 | } | |
3242 | ||
3243 | ret = qemu_rdma_init_ram_blocks(rdma); | |
3244 | if (ret) { | |
733252de | 3245 | error_report("rdma migration: error initializing ram blocks!"); |
2da776db MH |
3246 | goto err_rdma_dest_wait; |
3247 | } | |
3248 | ||
1f22364b | 3249 | for (idx = 0; idx < RDMA_WRID_MAX; idx++) { |
2da776db MH |
3250 | ret = qemu_rdma_reg_control(rdma, idx); |
3251 | if (ret) { | |
733252de | 3252 | error_report("rdma: error registering %d control", idx); |
2da776db MH |
3253 | goto err_rdma_dest_wait; |
3254 | } | |
3255 | } | |
3256 | ||
55cc1b59 LC |
3257 | /* Accept the second connection request for return path */ |
3258 | if (migrate_postcopy() && !rdma->is_return_path) { | |
3259 | qemu_set_fd_handler(rdma->channel->fd, rdma_accept_incoming_migration, | |
3260 | NULL, | |
3261 | (void *)(intptr_t)rdma->return_path); | |
3262 | } else { | |
3263 | qemu_set_fd_handler(rdma->channel->fd, NULL, NULL, NULL); | |
3264 | } | |
2da776db MH |
3265 | |
3266 | ret = rdma_accept(rdma->cm_id, &conn_param); | |
3267 | if (ret) { | |
733252de | 3268 | error_report("rdma_accept returns %d", ret); |
2da776db MH |
3269 | goto err_rdma_dest_wait; |
3270 | } | |
3271 | ||
3272 | ret = rdma_get_cm_event(rdma->channel, &cm_event); | |
3273 | if (ret) { | |
733252de | 3274 | error_report("rdma_accept get_cm_event failed %d", ret); |
2da776db MH |
3275 | goto err_rdma_dest_wait; |
3276 | } | |
3277 | ||
3278 | if (cm_event->event != RDMA_CM_EVENT_ESTABLISHED) { | |
733252de | 3279 | error_report("rdma_accept not event established"); |
2da776db MH |
3280 | rdma_ack_cm_event(cm_event); |
3281 | goto err_rdma_dest_wait; | |
3282 | } | |
3283 | ||
3284 | rdma_ack_cm_event(cm_event); | |
5a91337c | 3285 | rdma->connected = true; |
2da776db | 3286 | |
87772639 | 3287 | ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY); |
2da776db | 3288 | if (ret) { |
733252de | 3289 | error_report("rdma migration: error posting second control recv"); |
2da776db MH |
3290 | goto err_rdma_dest_wait; |
3291 | } | |
3292 | ||
3293 | qemu_rdma_dump_gid("dest_connect", rdma->cm_id); | |
3294 | ||
3295 | return 0; | |
3296 | ||
3297 | err_rdma_dest_wait: | |
3298 | rdma->error_state = ret; | |
3299 | qemu_rdma_cleanup(rdma); | |
3300 | return ret; | |
3301 | } | |
3302 | ||
e4d63320 DDAG |
3303 | static int dest_ram_sort_func(const void *a, const void *b) |
3304 | { | |
3305 | unsigned int a_index = ((const RDMALocalBlock *)a)->src_index; | |
3306 | unsigned int b_index = ((const RDMALocalBlock *)b)->src_index; | |
3307 | ||
3308 | return (a_index < b_index) ? -1 : (a_index != b_index); | |
3309 | } | |
3310 | ||
2da776db MH |
3311 | /* |
3312 | * During each iteration of the migration, we listen for instructions | |
3313 | * by the source VM to perform dynamic page registrations before they | |
3314 | * can perform RDMA operations. | |
3315 | * | |
3316 | * We respond with the 'rkey'. | |
3317 | * | |
3318 | * Keep doing this until the source tells us to stop. | |
3319 | */ | |
632e3a5c | 3320 | static int qemu_rdma_registration_handle(QEMUFile *f, void *opaque) |
2da776db MH |
3321 | { |
3322 | RDMAControlHeader reg_resp = { .len = sizeof(RDMARegisterResult), | |
3323 | .type = RDMA_CONTROL_REGISTER_RESULT, | |
3324 | .repeat = 0, | |
3325 | }; | |
3326 | RDMAControlHeader unreg_resp = { .len = 0, | |
3327 | .type = RDMA_CONTROL_UNREGISTER_FINISHED, | |
3328 | .repeat = 0, | |
3329 | }; | |
3330 | RDMAControlHeader blocks = { .type = RDMA_CONTROL_RAM_BLOCKS_RESULT, | |
3331 | .repeat = 1 }; | |
6ddd2d76 | 3332 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(opaque); |
74637e6f LC |
3333 | RDMAContext *rdma; |
3334 | RDMALocalBlocks *local; | |
2da776db MH |
3335 | RDMAControlHeader head; |
3336 | RDMARegister *reg, *registers; | |
3337 | RDMACompress *comp; | |
3338 | RDMARegisterResult *reg_result; | |
3339 | static RDMARegisterResult results[RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE]; | |
3340 | RDMALocalBlock *block; | |
3341 | void *host_addr; | |
3342 | int ret = 0; | |
3343 | int idx = 0; | |
3344 | int count = 0; | |
3345 | int i = 0; | |
3346 | ||
74637e6f LC |
3347 | rcu_read_lock(); |
3348 | rdma = atomic_rcu_read(&rioc->rdmain); | |
3349 | ||
3350 | if (!rdma) { | |
3351 | rcu_read_unlock(); | |
3352 | return -EIO; | |
3353 | } | |
3354 | ||
2da776db MH |
3355 | CHECK_ERROR_STATE(); |
3356 | ||
74637e6f | 3357 | local = &rdma->local_ram_blocks; |
2da776db | 3358 | do { |
632e3a5c | 3359 | trace_qemu_rdma_registration_handle_wait(); |
2da776db MH |
3360 | |
3361 | ret = qemu_rdma_exchange_recv(rdma, &head, RDMA_CONTROL_NONE); | |
3362 | ||
3363 | if (ret < 0) { | |
3364 | break; | |
3365 | } | |
3366 | ||
3367 | if (head.repeat > RDMA_CONTROL_MAX_COMMANDS_PER_MESSAGE) { | |
733252de DDAG |
3368 | error_report("rdma: Too many requests in this message (%d)." |
3369 | "Bailing.", head.repeat); | |
2da776db MH |
3370 | ret = -EIO; |
3371 | break; | |
3372 | } | |
3373 | ||
3374 | switch (head.type) { | |
3375 | case RDMA_CONTROL_COMPRESS: | |
3376 | comp = (RDMACompress *) rdma->wr_data[idx].control_curr; | |
3377 | network_to_compress(comp); | |
3378 | ||
733252de DDAG |
3379 | trace_qemu_rdma_registration_handle_compress(comp->length, |
3380 | comp->block_idx, | |
3381 | comp->offset); | |
afcddefd DDAG |
3382 | if (comp->block_idx >= rdma->local_ram_blocks.nb_blocks) { |
3383 | error_report("rdma: 'compress' bad block index %u (vs %d)", | |
3384 | (unsigned int)comp->block_idx, | |
3385 | rdma->local_ram_blocks.nb_blocks); | |
3386 | ret = -EIO; | |
24b41d66 | 3387 | goto out; |
afcddefd | 3388 | } |
2da776db MH |
3389 | block = &(rdma->local_ram_blocks.block[comp->block_idx]); |
3390 | ||
3391 | host_addr = block->local_host_addr + | |
3392 | (comp->offset - block->offset); | |
3393 | ||
3394 | ram_handle_compressed(host_addr, comp->value, comp->length); | |
3395 | break; | |
3396 | ||
3397 | case RDMA_CONTROL_REGISTER_FINISHED: | |
733252de | 3398 | trace_qemu_rdma_registration_handle_finished(); |
2da776db MH |
3399 | goto out; |
3400 | ||
3401 | case RDMA_CONTROL_RAM_BLOCKS_REQUEST: | |
733252de | 3402 | trace_qemu_rdma_registration_handle_ram_blocks(); |
2da776db | 3403 | |
e4d63320 DDAG |
3404 | /* Sort our local RAM Block list so it's the same as the source, |
3405 | * we can do this since we've filled in a src_index in the list | |
3406 | * as we received the RAMBlock list earlier. | |
3407 | */ | |
3408 | qsort(rdma->local_ram_blocks.block, | |
3409 | rdma->local_ram_blocks.nb_blocks, | |
3410 | sizeof(RDMALocalBlock), dest_ram_sort_func); | |
71cd7306 LC |
3411 | for (i = 0; i < local->nb_blocks; i++) { |
3412 | local->block[i].index = i; | |
3413 | } | |
3414 | ||
2da776db MH |
3415 | if (rdma->pin_all) { |
3416 | ret = qemu_rdma_reg_whole_ram_blocks(rdma); | |
3417 | if (ret) { | |
733252de DDAG |
3418 | error_report("rdma migration: error dest " |
3419 | "registering ram blocks"); | |
2da776db MH |
3420 | goto out; |
3421 | } | |
3422 | } | |
3423 | ||
3424 | /* | |
3425 | * Dest uses this to prepare to transmit the RAMBlock descriptions | |
3426 | * to the source VM after connection setup. | |
3427 | * Both sides use the "remote" structure to communicate and update | |
3428 | * their "local" descriptions with what was sent. | |
3429 | */ | |
3430 | for (i = 0; i < local->nb_blocks; i++) { | |
a97270ad | 3431 | rdma->dest_blocks[i].remote_host_addr = |
fbce8c25 | 3432 | (uintptr_t)(local->block[i].local_host_addr); |
2da776db MH |
3433 | |
3434 | if (rdma->pin_all) { | |
a97270ad | 3435 | rdma->dest_blocks[i].remote_rkey = local->block[i].mr->rkey; |
2da776db MH |
3436 | } |
3437 | ||
a97270ad DDAG |
3438 | rdma->dest_blocks[i].offset = local->block[i].offset; |
3439 | rdma->dest_blocks[i].length = local->block[i].length; | |
2da776db | 3440 | |
a97270ad | 3441 | dest_block_to_network(&rdma->dest_blocks[i]); |
e4d63320 DDAG |
3442 | trace_qemu_rdma_registration_handle_ram_blocks_loop( |
3443 | local->block[i].block_name, | |
3444 | local->block[i].offset, | |
3445 | local->block[i].length, | |
3446 | local->block[i].local_host_addr, | |
3447 | local->block[i].src_index); | |
2da776db MH |
3448 | } |
3449 | ||
3450 | blocks.len = rdma->local_ram_blocks.nb_blocks | |
a97270ad | 3451 | * sizeof(RDMADestBlock); |
2da776db MH |
3452 | |
3453 | ||
3454 | ret = qemu_rdma_post_send_control(rdma, | |
a97270ad | 3455 | (uint8_t *) rdma->dest_blocks, &blocks); |
2da776db MH |
3456 | |
3457 | if (ret < 0) { | |
733252de | 3458 | error_report("rdma migration: error sending remote info"); |
2da776db MH |
3459 | goto out; |
3460 | } | |
3461 | ||
3462 | break; | |
3463 | case RDMA_CONTROL_REGISTER_REQUEST: | |
733252de | 3464 | trace_qemu_rdma_registration_handle_register(head.repeat); |
2da776db MH |
3465 | |
3466 | reg_resp.repeat = head.repeat; | |
3467 | registers = (RDMARegister *) rdma->wr_data[idx].control_curr; | |
3468 | ||
3469 | for (count = 0; count < head.repeat; count++) { | |
3470 | uint64_t chunk; | |
3471 | uint8_t *chunk_start, *chunk_end; | |
3472 | ||
3473 | reg = ®isters[count]; | |
3474 | network_to_register(reg); | |
3475 | ||
3476 | reg_result = &results[count]; | |
3477 | ||
733252de | 3478 | trace_qemu_rdma_registration_handle_register_loop(count, |
2da776db MH |
3479 | reg->current_index, reg->key.current_addr, reg->chunks); |
3480 | ||
afcddefd DDAG |
3481 | if (reg->current_index >= rdma->local_ram_blocks.nb_blocks) { |
3482 | error_report("rdma: 'register' bad block index %u (vs %d)", | |
3483 | (unsigned int)reg->current_index, | |
3484 | rdma->local_ram_blocks.nb_blocks); | |
3485 | ret = -ENOENT; | |
24b41d66 | 3486 | goto out; |
afcddefd | 3487 | } |
2da776db MH |
3488 | block = &(rdma->local_ram_blocks.block[reg->current_index]); |
3489 | if (block->is_ram_block) { | |
afcddefd DDAG |
3490 | if (block->offset > reg->key.current_addr) { |
3491 | error_report("rdma: bad register address for block %s" | |
3492 | " offset: %" PRIx64 " current_addr: %" PRIx64, | |
3493 | block->block_name, block->offset, | |
3494 | reg->key.current_addr); | |
3495 | ret = -ERANGE; | |
24b41d66 | 3496 | goto out; |
afcddefd | 3497 | } |
2da776db MH |
3498 | host_addr = (block->local_host_addr + |
3499 | (reg->key.current_addr - block->offset)); | |
3500 | chunk = ram_chunk_index(block->local_host_addr, | |
3501 | (uint8_t *) host_addr); | |
3502 | } else { | |
3503 | chunk = reg->key.chunk; | |
3504 | host_addr = block->local_host_addr + | |
3505 | (reg->key.chunk * (1UL << RDMA_REG_CHUNK_SHIFT)); | |
afcddefd DDAG |
3506 | /* Check for particularly bad chunk value */ |
3507 | if (host_addr < (void *)block->local_host_addr) { | |
3508 | error_report("rdma: bad chunk for block %s" | |
3509 | " chunk: %" PRIx64, | |
3510 | block->block_name, reg->key.chunk); | |
3511 | ret = -ERANGE; | |
24b41d66 | 3512 | goto out; |
afcddefd | 3513 | } |
2da776db MH |
3514 | } |
3515 | chunk_start = ram_chunk_start(block, chunk); | |
3516 | chunk_end = ram_chunk_end(block, chunk + reg->chunks); | |
3517 | if (qemu_rdma_register_and_get_keys(rdma, block, | |
3ac040c0 | 3518 | (uintptr_t)host_addr, NULL, ®_result->rkey, |
2da776db | 3519 | chunk, chunk_start, chunk_end)) { |
733252de | 3520 | error_report("cannot get rkey"); |
2da776db MH |
3521 | ret = -EINVAL; |
3522 | goto out; | |
3523 | } | |
3524 | ||
fbce8c25 | 3525 | reg_result->host_addr = (uintptr_t)block->local_host_addr; |
2da776db | 3526 | |
733252de DDAG |
3527 | trace_qemu_rdma_registration_handle_register_rkey( |
3528 | reg_result->rkey); | |
2da776db MH |
3529 | |
3530 | result_to_network(reg_result); | |
3531 | } | |
3532 | ||
3533 | ret = qemu_rdma_post_send_control(rdma, | |
3534 | (uint8_t *) results, ®_resp); | |
3535 | ||
3536 | if (ret < 0) { | |
733252de | 3537 | error_report("Failed to send control buffer"); |
2da776db MH |
3538 | goto out; |
3539 | } | |
3540 | break; | |
3541 | case RDMA_CONTROL_UNREGISTER_REQUEST: | |
733252de | 3542 | trace_qemu_rdma_registration_handle_unregister(head.repeat); |
2da776db MH |
3543 | unreg_resp.repeat = head.repeat; |
3544 | registers = (RDMARegister *) rdma->wr_data[idx].control_curr; | |
3545 | ||
3546 | for (count = 0; count < head.repeat; count++) { | |
3547 | reg = ®isters[count]; | |
3548 | network_to_register(reg); | |
3549 | ||
733252de DDAG |
3550 | trace_qemu_rdma_registration_handle_unregister_loop(count, |
3551 | reg->current_index, reg->key.chunk); | |
2da776db MH |
3552 | |
3553 | block = &(rdma->local_ram_blocks.block[reg->current_index]); | |
3554 | ||
3555 | ret = ibv_dereg_mr(block->pmr[reg->key.chunk]); | |
3556 | block->pmr[reg->key.chunk] = NULL; | |
3557 | ||
3558 | if (ret != 0) { | |
3559 | perror("rdma unregistration chunk failed"); | |
3560 | ret = -ret; | |
3561 | goto out; | |
3562 | } | |
3563 | ||
3564 | rdma->total_registrations--; | |
3565 | ||
733252de DDAG |
3566 | trace_qemu_rdma_registration_handle_unregister_success( |
3567 | reg->key.chunk); | |
2da776db MH |
3568 | } |
3569 | ||
3570 | ret = qemu_rdma_post_send_control(rdma, NULL, &unreg_resp); | |
3571 | ||
3572 | if (ret < 0) { | |
733252de | 3573 | error_report("Failed to send control buffer"); |
2da776db MH |
3574 | goto out; |
3575 | } | |
3576 | break; | |
3577 | case RDMA_CONTROL_REGISTER_RESULT: | |
733252de | 3578 | error_report("Invalid RESULT message at dest."); |
2da776db MH |
3579 | ret = -EIO; |
3580 | goto out; | |
3581 | default: | |
482a33c5 | 3582 | error_report("Unknown control message %s", control_desc(head.type)); |
2da776db MH |
3583 | ret = -EIO; |
3584 | goto out; | |
3585 | } | |
3586 | } while (1); | |
3587 | out: | |
3588 | if (ret < 0) { | |
3589 | rdma->error_state = ret; | |
3590 | } | |
74637e6f | 3591 | rcu_read_unlock(); |
2da776db MH |
3592 | return ret; |
3593 | } | |
3594 | ||
e4d63320 DDAG |
3595 | /* Destination: |
3596 | * Called via a ram_control_load_hook during the initial RAM load section which | |
3597 | * lists the RAMBlocks by name. This lets us know the order of the RAMBlocks | |
3598 | * on the source. | |
3599 | * We've already built our local RAMBlock list, but not yet sent the list to | |
3600 | * the source. | |
3601 | */ | |
6ddd2d76 DB |
3602 | static int |
3603 | rdma_block_notification_handle(QIOChannelRDMA *rioc, const char *name) | |
e4d63320 | 3604 | { |
74637e6f | 3605 | RDMAContext *rdma; |
e4d63320 DDAG |
3606 | int curr; |
3607 | int found = -1; | |
3608 | ||
74637e6f LC |
3609 | rcu_read_lock(); |
3610 | rdma = atomic_rcu_read(&rioc->rdmain); | |
3611 | ||
3612 | if (!rdma) { | |
3613 | rcu_read_unlock(); | |
3614 | return -EIO; | |
3615 | } | |
3616 | ||
e4d63320 DDAG |
3617 | /* Find the matching RAMBlock in our local list */ |
3618 | for (curr = 0; curr < rdma->local_ram_blocks.nb_blocks; curr++) { | |
3619 | if (!strcmp(rdma->local_ram_blocks.block[curr].block_name, name)) { | |
3620 | found = curr; | |
3621 | break; | |
3622 | } | |
3623 | } | |
3624 | ||
3625 | if (found == -1) { | |
3626 | error_report("RAMBlock '%s' not found on destination", name); | |
74637e6f | 3627 | rcu_read_unlock(); |
e4d63320 DDAG |
3628 | return -ENOENT; |
3629 | } | |
3630 | ||
3631 | rdma->local_ram_blocks.block[curr].src_index = rdma->next_src_index; | |
3632 | trace_rdma_block_notification_handle(name, rdma->next_src_index); | |
3633 | rdma->next_src_index++; | |
3634 | ||
74637e6f | 3635 | rcu_read_unlock(); |
e4d63320 DDAG |
3636 | return 0; |
3637 | } | |
3638 | ||
632e3a5c DDAG |
3639 | static int rdma_load_hook(QEMUFile *f, void *opaque, uint64_t flags, void *data) |
3640 | { | |
3641 | switch (flags) { | |
3642 | case RAM_CONTROL_BLOCK_REG: | |
e4d63320 | 3643 | return rdma_block_notification_handle(opaque, data); |
632e3a5c DDAG |
3644 | |
3645 | case RAM_CONTROL_HOOK: | |
3646 | return qemu_rdma_registration_handle(f, opaque); | |
3647 | ||
3648 | default: | |
3649 | /* Shouldn't be called with any other values */ | |
3650 | abort(); | |
3651 | } | |
3652 | } | |
3653 | ||
2da776db | 3654 | static int qemu_rdma_registration_start(QEMUFile *f, void *opaque, |
632e3a5c | 3655 | uint64_t flags, void *data) |
2da776db | 3656 | { |
6ddd2d76 | 3657 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(opaque); |
74637e6f LC |
3658 | RDMAContext *rdma; |
3659 | ||
3660 | rcu_read_lock(); | |
3661 | rdma = atomic_rcu_read(&rioc->rdmaout); | |
3662 | if (!rdma) { | |
3663 | rcu_read_unlock(); | |
3664 | return -EIO; | |
3665 | } | |
2da776db MH |
3666 | |
3667 | CHECK_ERROR_STATE(); | |
3668 | ||
ccb7e1b5 | 3669 | if (migrate_get_current()->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { |
74637e6f | 3670 | rcu_read_unlock(); |
ccb7e1b5 LC |
3671 | return 0; |
3672 | } | |
3673 | ||
733252de | 3674 | trace_qemu_rdma_registration_start(flags); |
2da776db MH |
3675 | qemu_put_be64(f, RAM_SAVE_FLAG_HOOK); |
3676 | qemu_fflush(f); | |
3677 | ||
74637e6f | 3678 | rcu_read_unlock(); |
2da776db MH |
3679 | return 0; |
3680 | } | |
3681 | ||
3682 | /* | |
3683 | * Inform dest that dynamic registrations are done for now. | |
3684 | * First, flush writes, if any. | |
3685 | */ | |
3686 | static int qemu_rdma_registration_stop(QEMUFile *f, void *opaque, | |
632e3a5c | 3687 | uint64_t flags, void *data) |
2da776db MH |
3688 | { |
3689 | Error *local_err = NULL, **errp = &local_err; | |
6ddd2d76 | 3690 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(opaque); |
74637e6f | 3691 | RDMAContext *rdma; |
2da776db MH |
3692 | RDMAControlHeader head = { .len = 0, .repeat = 1 }; |
3693 | int ret = 0; | |
3694 | ||
74637e6f LC |
3695 | rcu_read_lock(); |
3696 | rdma = atomic_rcu_read(&rioc->rdmaout); | |
3697 | if (!rdma) { | |
3698 | rcu_read_unlock(); | |
3699 | return -EIO; | |
3700 | } | |
3701 | ||
2da776db MH |
3702 | CHECK_ERROR_STATE(); |
3703 | ||
ccb7e1b5 | 3704 | if (migrate_get_current()->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) { |
74637e6f | 3705 | rcu_read_unlock(); |
ccb7e1b5 LC |
3706 | return 0; |
3707 | } | |
3708 | ||
2da776db MH |
3709 | qemu_fflush(f); |
3710 | ret = qemu_rdma_drain_cq(f, rdma); | |
3711 | ||
3712 | if (ret < 0) { | |
3713 | goto err; | |
3714 | } | |
3715 | ||
3716 | if (flags == RAM_CONTROL_SETUP) { | |
3717 | RDMAControlHeader resp = {.type = RDMA_CONTROL_RAM_BLOCKS_RESULT }; | |
3718 | RDMALocalBlocks *local = &rdma->local_ram_blocks; | |
e4d63320 | 3719 | int reg_result_idx, i, nb_dest_blocks; |
2da776db MH |
3720 | |
3721 | head.type = RDMA_CONTROL_RAM_BLOCKS_REQUEST; | |
733252de | 3722 | trace_qemu_rdma_registration_stop_ram(); |
2da776db MH |
3723 | |
3724 | /* | |
3725 | * Make sure that we parallelize the pinning on both sides. | |
3726 | * For very large guests, doing this serially takes a really | |
3727 | * long time, so we have to 'interleave' the pinning locally | |
3728 | * with the control messages by performing the pinning on this | |
3729 | * side before we receive the control response from the other | |
3730 | * side that the pinning has completed. | |
3731 | */ | |
3732 | ret = qemu_rdma_exchange_send(rdma, &head, NULL, &resp, | |
3733 | ®_result_idx, rdma->pin_all ? | |
3734 | qemu_rdma_reg_whole_ram_blocks : NULL); | |
3735 | if (ret < 0) { | |
66988941 | 3736 | ERROR(errp, "receiving remote info!"); |
74637e6f | 3737 | rcu_read_unlock(); |
2da776db MH |
3738 | return ret; |
3739 | } | |
3740 | ||
a97270ad | 3741 | nb_dest_blocks = resp.len / sizeof(RDMADestBlock); |
2da776db MH |
3742 | |
3743 | /* | |
3744 | * The protocol uses two different sets of rkeys (mutually exclusive): | |
3745 | * 1. One key to represent the virtual address of the entire ram block. | |
3746 | * (dynamic chunk registration disabled - pin everything with one rkey.) | |
3747 | * 2. One to represent individual chunks within a ram block. | |
3748 | * (dynamic chunk registration enabled - pin individual chunks.) | |
3749 | * | |
3750 | * Once the capability is successfully negotiated, the destination transmits | |
3751 | * the keys to use (or sends them later) including the virtual addresses | |
3752 | * and then propagates the remote ram block descriptions to his local copy. | |
3753 | */ | |
3754 | ||
a97270ad | 3755 | if (local->nb_blocks != nb_dest_blocks) { |
e4d63320 | 3756 | ERROR(errp, "ram blocks mismatch (Number of blocks %d vs %d) " |
2da776db | 3757 | "Your QEMU command line parameters are probably " |
e4d63320 DDAG |
3758 | "not identical on both the source and destination.", |
3759 | local->nb_blocks, nb_dest_blocks); | |
ef4b722d | 3760 | rdma->error_state = -EINVAL; |
74637e6f | 3761 | rcu_read_unlock(); |
2da776db MH |
3762 | return -EINVAL; |
3763 | } | |
3764 | ||
885e8f98 | 3765 | qemu_rdma_move_header(rdma, reg_result_idx, &resp); |
a97270ad | 3766 | memcpy(rdma->dest_blocks, |
885e8f98 | 3767 | rdma->wr_data[reg_result_idx].control_curr, resp.len); |
a97270ad DDAG |
3768 | for (i = 0; i < nb_dest_blocks; i++) { |
3769 | network_to_dest_block(&rdma->dest_blocks[i]); | |
2da776db | 3770 | |
e4d63320 DDAG |
3771 | /* We require that the blocks are in the same order */ |
3772 | if (rdma->dest_blocks[i].length != local->block[i].length) { | |
3773 | ERROR(errp, "Block %s/%d has a different length %" PRIu64 | |
3774 | "vs %" PRIu64, local->block[i].block_name, i, | |
3775 | local->block[i].length, | |
3776 | rdma->dest_blocks[i].length); | |
ef4b722d | 3777 | rdma->error_state = -EINVAL; |
74637e6f | 3778 | rcu_read_unlock(); |
2da776db MH |
3779 | return -EINVAL; |
3780 | } | |
e4d63320 DDAG |
3781 | local->block[i].remote_host_addr = |
3782 | rdma->dest_blocks[i].remote_host_addr; | |
3783 | local->block[i].remote_rkey = rdma->dest_blocks[i].remote_rkey; | |
2da776db MH |
3784 | } |
3785 | } | |
3786 | ||
733252de | 3787 | trace_qemu_rdma_registration_stop(flags); |
2da776db MH |
3788 | |
3789 | head.type = RDMA_CONTROL_REGISTER_FINISHED; | |
3790 | ret = qemu_rdma_exchange_send(rdma, &head, NULL, NULL, NULL, NULL); | |
3791 | ||
3792 | if (ret < 0) { | |
3793 | goto err; | |
3794 | } | |
3795 | ||
74637e6f | 3796 | rcu_read_unlock(); |
2da776db MH |
3797 | return 0; |
3798 | err: | |
3799 | rdma->error_state = ret; | |
74637e6f | 3800 | rcu_read_unlock(); |
2da776db MH |
3801 | return ret; |
3802 | } | |
3803 | ||
0436e09f | 3804 | static const QEMUFileHooks rdma_read_hooks = { |
632e3a5c | 3805 | .hook_ram_load = rdma_load_hook, |
2da776db MH |
3806 | }; |
3807 | ||
0436e09f | 3808 | static const QEMUFileHooks rdma_write_hooks = { |
2da776db MH |
3809 | .before_ram_iterate = qemu_rdma_registration_start, |
3810 | .after_ram_iterate = qemu_rdma_registration_stop, | |
3811 | .save_page = qemu_rdma_save_page, | |
3812 | }; | |
3813 | ||
6ddd2d76 DB |
3814 | |
3815 | static void qio_channel_rdma_finalize(Object *obj) | |
3816 | { | |
3817 | QIOChannelRDMA *rioc = QIO_CHANNEL_RDMA(obj); | |
74637e6f LC |
3818 | if (rioc->rdmain) { |
3819 | qemu_rdma_cleanup(rioc->rdmain); | |
3820 | g_free(rioc->rdmain); | |
3821 | rioc->rdmain = NULL; | |
3822 | } | |
3823 | if (rioc->rdmaout) { | |
3824 | qemu_rdma_cleanup(rioc->rdmaout); | |
3825 | g_free(rioc->rdmaout); | |
3826 | rioc->rdmaout = NULL; | |
6ddd2d76 DB |
3827 | } |
3828 | } | |
3829 | ||
3830 | static void qio_channel_rdma_class_init(ObjectClass *klass, | |
3831 | void *class_data G_GNUC_UNUSED) | |
3832 | { | |
3833 | QIOChannelClass *ioc_klass = QIO_CHANNEL_CLASS(klass); | |
3834 | ||
3835 | ioc_klass->io_writev = qio_channel_rdma_writev; | |
3836 | ioc_klass->io_readv = qio_channel_rdma_readv; | |
3837 | ioc_klass->io_set_blocking = qio_channel_rdma_set_blocking; | |
3838 | ioc_klass->io_close = qio_channel_rdma_close; | |
3839 | ioc_klass->io_create_watch = qio_channel_rdma_create_watch; | |
4d9f675b | 3840 | ioc_klass->io_set_aio_fd_handler = qio_channel_rdma_set_aio_fd_handler; |
6ddd2d76 DB |
3841 | } |
3842 | ||
3843 | static const TypeInfo qio_channel_rdma_info = { | |
3844 | .parent = TYPE_QIO_CHANNEL, | |
3845 | .name = TYPE_QIO_CHANNEL_RDMA, | |
3846 | .instance_size = sizeof(QIOChannelRDMA), | |
3847 | .instance_finalize = qio_channel_rdma_finalize, | |
3848 | .class_init = qio_channel_rdma_class_init, | |
3849 | }; | |
3850 | ||
3851 | static void qio_channel_rdma_register_types(void) | |
3852 | { | |
3853 | type_register_static(&qio_channel_rdma_info); | |
3854 | } | |
3855 | ||
3856 | type_init(qio_channel_rdma_register_types); | |
3857 | ||
3858 | static QEMUFile *qemu_fopen_rdma(RDMAContext *rdma, const char *mode) | |
2da776db | 3859 | { |
6ddd2d76 | 3860 | QIOChannelRDMA *rioc; |
2da776db MH |
3861 | |
3862 | if (qemu_file_mode_is_not_valid(mode)) { | |
3863 | return NULL; | |
3864 | } | |
3865 | ||
6ddd2d76 | 3866 | rioc = QIO_CHANNEL_RDMA(object_new(TYPE_QIO_CHANNEL_RDMA)); |
2da776db MH |
3867 | |
3868 | if (mode[0] == 'w') { | |
6ddd2d76 | 3869 | rioc->file = qemu_fopen_channel_output(QIO_CHANNEL(rioc)); |
74637e6f LC |
3870 | rioc->rdmaout = rdma; |
3871 | rioc->rdmain = rdma->return_path; | |
6ddd2d76 | 3872 | qemu_file_set_hooks(rioc->file, &rdma_write_hooks); |
2da776db | 3873 | } else { |
6ddd2d76 | 3874 | rioc->file = qemu_fopen_channel_input(QIO_CHANNEL(rioc)); |
74637e6f LC |
3875 | rioc->rdmain = rdma; |
3876 | rioc->rdmaout = rdma->return_path; | |
6ddd2d76 | 3877 | qemu_file_set_hooks(rioc->file, &rdma_read_hooks); |
2da776db MH |
3878 | } |
3879 | ||
6ddd2d76 | 3880 | return rioc->file; |
2da776db MH |
3881 | } |
3882 | ||
3883 | static void rdma_accept_incoming_migration(void *opaque) | |
3884 | { | |
3885 | RDMAContext *rdma = opaque; | |
3886 | int ret; | |
3887 | QEMUFile *f; | |
3888 | Error *local_err = NULL, **errp = &local_err; | |
3889 | ||
24ec68ef | 3890 | trace_qemu_rdma_accept_incoming_migration(); |
2da776db MH |
3891 | ret = qemu_rdma_accept(rdma); |
3892 | ||
3893 | if (ret) { | |
66988941 | 3894 | ERROR(errp, "RDMA Migration initialization failed!"); |
2da776db MH |
3895 | return; |
3896 | } | |
3897 | ||
24ec68ef | 3898 | trace_qemu_rdma_accept_incoming_migration_accepted(); |
2da776db | 3899 | |
55cc1b59 LC |
3900 | if (rdma->is_return_path) { |
3901 | return; | |
3902 | } | |
3903 | ||
2da776db MH |
3904 | f = qemu_fopen_rdma(rdma, "rb"); |
3905 | if (f == NULL) { | |
66988941 | 3906 | ERROR(errp, "could not qemu_fopen_rdma!"); |
2da776db MH |
3907 | qemu_rdma_cleanup(rdma); |
3908 | return; | |
3909 | } | |
3910 | ||
3911 | rdma->migration_started_on_destination = 1; | |
22724f49 | 3912 | migration_fd_process_incoming(f); |
2da776db MH |
3913 | } |
3914 | ||
3915 | void rdma_start_incoming_migration(const char *host_port, Error **errp) | |
3916 | { | |
3917 | int ret; | |
55cc1b59 | 3918 | RDMAContext *rdma, *rdma_return_path; |
2da776db MH |
3919 | Error *local_err = NULL; |
3920 | ||
733252de | 3921 | trace_rdma_start_incoming_migration(); |
2da776db MH |
3922 | rdma = qemu_rdma_data_init(host_port, &local_err); |
3923 | ||
3924 | if (rdma == NULL) { | |
3925 | goto err; | |
3926 | } | |
3927 | ||
3928 | ret = qemu_rdma_dest_init(rdma, &local_err); | |
3929 | ||
3930 | if (ret) { | |
3931 | goto err; | |
3932 | } | |
3933 | ||
733252de | 3934 | trace_rdma_start_incoming_migration_after_dest_init(); |
2da776db MH |
3935 | |
3936 | ret = rdma_listen(rdma->listen_id, 5); | |
3937 | ||
3938 | if (ret) { | |
66988941 | 3939 | ERROR(errp, "listening on socket!"); |
2da776db MH |
3940 | goto err; |
3941 | } | |
3942 | ||
733252de | 3943 | trace_rdma_start_incoming_migration_after_rdma_listen(); |
2da776db | 3944 | |
55cc1b59 LC |
3945 | /* initialize the RDMAContext for return path */ |
3946 | if (migrate_postcopy()) { | |
3947 | rdma_return_path = qemu_rdma_data_init(host_port, &local_err); | |
3948 | ||
3949 | if (rdma_return_path == NULL) { | |
3950 | goto err; | |
3951 | } | |
3952 | ||
3953 | qemu_rdma_return_path_dest_init(rdma_return_path, rdma); | |
3954 | } | |
3955 | ||
82e1cc4b FZ |
3956 | qemu_set_fd_handler(rdma->channel->fd, rdma_accept_incoming_migration, |
3957 | NULL, (void *)(intptr_t)rdma); | |
2da776db MH |
3958 | return; |
3959 | err: | |
3960 | error_propagate(errp, local_err); | |
3961 | g_free(rdma); | |
55cc1b59 | 3962 | g_free(rdma_return_path); |
2da776db MH |
3963 | } |
3964 | ||
3965 | void rdma_start_outgoing_migration(void *opaque, | |
3966 | const char *host_port, Error **errp) | |
3967 | { | |
3968 | MigrationState *s = opaque; | |
d59ce6f3 | 3969 | RDMAContext *rdma = qemu_rdma_data_init(host_port, errp); |
55cc1b59 | 3970 | RDMAContext *rdma_return_path = NULL; |
2da776db MH |
3971 | int ret = 0; |
3972 | ||
3973 | if (rdma == NULL) { | |
2da776db MH |
3974 | goto err; |
3975 | } | |
3976 | ||
bbfb89e3 FZ |
3977 | ret = qemu_rdma_source_init(rdma, |
3978 | s->enabled_capabilities[MIGRATION_CAPABILITY_RDMA_PIN_ALL], errp); | |
2da776db MH |
3979 | |
3980 | if (ret) { | |
3981 | goto err; | |
3982 | } | |
3983 | ||
733252de | 3984 | trace_rdma_start_outgoing_migration_after_rdma_source_init(); |
d59ce6f3 | 3985 | ret = qemu_rdma_connect(rdma, errp); |
2da776db MH |
3986 | |
3987 | if (ret) { | |
3988 | goto err; | |
3989 | } | |
3990 | ||
55cc1b59 LC |
3991 | /* RDMA postcopy need a seprate queue pair for return path */ |
3992 | if (migrate_postcopy()) { | |
3993 | rdma_return_path = qemu_rdma_data_init(host_port, errp); | |
3994 | ||
3995 | if (rdma_return_path == NULL) { | |
3996 | goto err; | |
3997 | } | |
3998 | ||
3999 | ret = qemu_rdma_source_init(rdma_return_path, | |
4000 | s->enabled_capabilities[MIGRATION_CAPABILITY_RDMA_PIN_ALL], errp); | |
4001 | ||
4002 | if (ret) { | |
4003 | goto err; | |
4004 | } | |
4005 | ||
4006 | ret = qemu_rdma_connect(rdma_return_path, errp); | |
4007 | ||
4008 | if (ret) { | |
4009 | goto err; | |
4010 | } | |
4011 | ||
4012 | rdma->return_path = rdma_return_path; | |
4013 | rdma_return_path->return_path = rdma; | |
4014 | rdma_return_path->is_return_path = true; | |
4015 | } | |
4016 | ||
733252de | 4017 | trace_rdma_start_outgoing_migration_after_rdma_connect(); |
2da776db | 4018 | |
89a02a9f | 4019 | s->to_dst_file = qemu_fopen_rdma(rdma, "wb"); |
cce8040b | 4020 | migrate_fd_connect(s, NULL); |
2da776db MH |
4021 | return; |
4022 | err: | |
2da776db | 4023 | g_free(rdma); |
55cc1b59 | 4024 | g_free(rdma_return_path); |
2da776db | 4025 | } |