1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018 Intel Corporation
10 #include <rte_string_fns.h>
12 #include "eal_memalloc.h"
14 #include "malloc_elem.h"
15 #include "malloc_mp.h"
17 #define MP_ACTION_SYNC "mp_malloc_sync"
18 /**< request sent by primary process to notify of changes in memory map */
19 #define MP_ACTION_ROLLBACK "mp_malloc_rollback"
20 /**< request sent by primary process to notify of changes in memory map. this is
21 * essentially a regular sync request, but we cannot send sync requests while
22 * another one is in progress, and we might have to - therefore, we do this as
23 * a separate callback.
25 #define MP_ACTION_REQUEST "mp_malloc_request"
26 /**< request sent by secondary process to ask for allocation/deallocation */
27 #define MP_ACTION_RESPONSE "mp_malloc_response"
28 /**< response sent to secondary process to indicate result of request */
30 /* forward declarations */
32 handle_sync_response(const struct rte_mp_msg *request,
33 const struct rte_mp_reply *reply);
35 handle_rollback_response(const struct rte_mp_msg *request,
36 const struct rte_mp_reply *reply);
38 #define MP_TIMEOUT_S 5 /**< 5 seconds timeouts */
40 /* when we're allocating, we need to store some state to ensure that we can
43 struct primary_alloc_req_state {
44 struct malloc_heap *heap;
45 struct rte_memseg **ms;
47 struct malloc_elem *elem;
53 REQ_STATE_INACTIVE = 0,
59 TAILQ_ENTRY(mp_request) next;
60 struct malloc_mp_req user_req; /**< contents of request */
61 pthread_cond_t cond; /**< variable we use to time out on this request */
62 enum req_state state; /**< indicate status of this request */
63 struct primary_alloc_req_state alloc_state;
67 * We could've used just a single request, but it may be possible for
68 * secondaries to timeout earlier than the primary, and send a new request while
69 * primary is still expecting replies to the old one. Therefore, each new
70 * request will get assigned a new ID, which is how we will distinguish between
71 * expected and unexpected messages.
73 TAILQ_HEAD(mp_request_list, mp_request);
75 struct mp_request_list list;
78 .list = TAILQ_HEAD_INITIALIZER(mp_request_list.list),
79 .lock = PTHREAD_MUTEX_INITIALIZER
83 * General workflow is the following:
86 * S: send request to primary
87 * P: attempt to allocate memory
88 * if failed, sendmsg failure
89 * if success, send sync request
90 * S: if received msg of failure, quit
91 * if received sync request, synchronize memory map and reply with result
92 * P: if received sync request result
93 * if success, sendmsg success
94 * if failure, roll back allocation and send a rollback request
95 * S: if received msg of success, quit
96 * if received rollback request, synchronize memory map and reply with result
97 * P: if received sync request result
98 * sendmsg sync request result
99 * S: if received msg, quit
101 * Aside from timeouts, there are three points where we can quit:
102 * - if allocation failed straight away
103 * - if allocation and sync request succeeded
104 * - if allocation succeeded, sync request failed, allocation rolled back and
105 * rollback request received (irrespective of whether it succeeded or failed)
108 * S: send request to primary
109 * P: attempt to deallocate memory
110 * if failed, sendmsg failure
111 * if success, send sync request
112 * S: if received msg of failure, quit
113 * if received sync request, synchronize memory map and reply with result
114 * P: if received sync request result
115 * sendmsg sync request result
116 * S: if received msg, quit
118 * There is no "rollback" from deallocation, as it's safe to have some memory
119 * mapped in some processes - it's absent from the heap, so it won't get used.
122 static struct mp_request *
123 find_request_by_id(uint64_t id)
125 struct mp_request *req;
126 TAILQ_FOREACH(req, &mp_request_list.list, next) {
127 if (req->user_req.id == id)
133 /* this ID is, like, totally guaranteed to be absolutely unique. pinky swear. */
140 } while (find_request_by_id(id) != NULL);
144 /* secondary will respond to sync requests thusly */
146 handle_sync(const struct rte_mp_msg *msg, const void *peer)
148 struct rte_mp_msg reply;
149 const struct malloc_mp_req *req =
150 (const struct malloc_mp_req *)msg->param;
151 struct malloc_mp_req *resp =
152 (struct malloc_mp_req *)reply.param;
155 if (req->t != REQ_TYPE_SYNC) {
156 RTE_LOG(ERR, EAL, "Unexpected request from primary\n");
160 memset(&reply, 0, sizeof(reply));
163 strlcpy(reply.name, msg->name, sizeof(reply.name));
164 reply.len_param = sizeof(*resp);
166 ret = eal_memalloc_sync_with_primary();
168 resp->t = REQ_TYPE_SYNC;
170 resp->result = ret == 0 ? REQ_RESULT_SUCCESS : REQ_RESULT_FAIL;
172 rte_mp_reply(&reply, peer);
178 handle_alloc_request(const struct malloc_mp_req *m,
179 struct mp_request *req)
181 const struct malloc_req_alloc *ar = &m->alloc_req;
182 struct malloc_heap *heap;
183 struct malloc_elem *elem;
184 struct rte_memseg **ms;
189 alloc_sz = RTE_ALIGN_CEIL(ar->align + ar->elt_size +
190 MALLOC_ELEM_TRAILER_LEN, ar->page_sz);
191 n_segs = alloc_sz / ar->page_sz;
195 /* we can't know in advance how many pages we'll need, so we malloc */
196 ms = malloc(sizeof(*ms) * n_segs);
198 RTE_LOG(ERR, EAL, "Couldn't allocate memory for request state\n");
201 memset(ms, 0, sizeof(*ms) * n_segs);
203 elem = alloc_pages_on_heap(heap, ar->page_sz, ar->elt_size, ar->socket,
204 ar->flags, ar->align, ar->bound, ar->contig, ms,
210 map_addr = ms[0]->addr;
212 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_ALLOC, map_addr, alloc_sz);
214 /* we have succeeded in allocating memory, but we still need to sync
215 * with other processes. however, since DPDK IPC is single-threaded, we
216 * send an asynchronous request and exit this callback.
219 req->alloc_state.ms = ms;
220 req->alloc_state.ms_len = n_segs;
221 req->alloc_state.map_addr = map_addr;
222 req->alloc_state.map_len = alloc_sz;
223 req->alloc_state.elem = elem;
224 req->alloc_state.heap = heap;
232 /* first stage of primary handling requests from secondary */
234 handle_request(const struct rte_mp_msg *msg, const void *peer __rte_unused)
236 const struct malloc_mp_req *m =
237 (const struct malloc_mp_req *)msg->param;
238 struct mp_request *entry;
241 /* lock access to request */
242 pthread_mutex_lock(&mp_request_list.lock);
244 /* make sure it's not a dupe */
245 entry = find_request_by_id(m->id);
247 RTE_LOG(ERR, EAL, "Duplicate request id\n");
251 entry = malloc(sizeof(*entry));
253 RTE_LOG(ERR, EAL, "Unable to allocate memory for request\n");
258 memset(entry, 0, sizeof(*entry));
260 if (m->t == REQ_TYPE_ALLOC) {
261 ret = handle_alloc_request(m, entry);
262 } else if (m->t == REQ_TYPE_FREE) {
263 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
264 m->free_req.addr, m->free_req.len);
266 ret = malloc_heap_free_pages(m->free_req.addr,
269 RTE_LOG(ERR, EAL, "Unexpected request from secondary\n");
274 struct rte_mp_msg resp_msg;
275 struct malloc_mp_req *resp =
276 (struct malloc_mp_req *)resp_msg.param;
278 /* send failure message straight away */
279 resp_msg.num_fds = 0;
280 resp_msg.len_param = sizeof(*resp);
281 strlcpy(resp_msg.name, MP_ACTION_RESPONSE,
282 sizeof(resp_msg.name));
285 resp->result = REQ_RESULT_FAIL;
288 if (rte_mp_sendmsg(&resp_msg)) {
289 RTE_LOG(ERR, EAL, "Couldn't send response\n");
292 /* we did not modify the request */
295 struct rte_mp_msg sr_msg;
296 struct malloc_mp_req *sr =
297 (struct malloc_mp_req *)sr_msg.param;
300 memset(&sr_msg, 0, sizeof(sr_msg));
302 /* we can do something, so send sync request asynchronously */
304 sr_msg.len_param = sizeof(*sr);
305 strlcpy(sr_msg.name, MP_ACTION_SYNC, sizeof(sr_msg.name));
308 ts.tv_sec = MP_TIMEOUT_S;
310 /* sync requests carry no data */
311 sr->t = REQ_TYPE_SYNC;
314 /* there may be stray timeout still waiting */
316 ret = rte_mp_request_async(&sr_msg, &ts,
317 handle_sync_response);
318 } while (ret != 0 && rte_errno == EEXIST);
320 RTE_LOG(ERR, EAL, "Couldn't send sync request\n");
321 if (m->t == REQ_TYPE_ALLOC)
322 free(entry->alloc_state.ms);
326 /* mark request as in progress */
327 memcpy(&entry->user_req, m, sizeof(*m));
328 entry->state = REQ_STATE_ACTIVE;
330 TAILQ_INSERT_TAIL(&mp_request_list.list, entry, next);
332 pthread_mutex_unlock(&mp_request_list.lock);
335 pthread_mutex_unlock(&mp_request_list.lock);
340 /* callback for asynchronous sync requests for primary. this will either do a
341 * sendmsg with results, or trigger rollback request.
344 handle_sync_response(const struct rte_mp_msg *request,
345 const struct rte_mp_reply *reply)
347 enum malloc_req_result result;
348 struct mp_request *entry;
349 const struct malloc_mp_req *mpreq =
350 (const struct malloc_mp_req *)request->param;
353 /* lock the request */
354 pthread_mutex_lock(&mp_request_list.lock);
356 entry = find_request_by_id(mpreq->id);
358 RTE_LOG(ERR, EAL, "Wrong request ID\n");
362 result = REQ_RESULT_SUCCESS;
364 if (reply->nb_received != reply->nb_sent)
365 result = REQ_RESULT_FAIL;
367 for (i = 0; i < reply->nb_received; i++) {
368 struct malloc_mp_req *resp =
369 (struct malloc_mp_req *)reply->msgs[i].param;
371 if (resp->t != REQ_TYPE_SYNC) {
372 RTE_LOG(ERR, EAL, "Unexpected response to sync request\n");
373 result = REQ_RESULT_FAIL;
376 if (resp->id != entry->user_req.id) {
377 RTE_LOG(ERR, EAL, "Response to wrong sync request\n");
378 result = REQ_RESULT_FAIL;
381 if (resp->result == REQ_RESULT_FAIL) {
382 result = REQ_RESULT_FAIL;
387 if (entry->user_req.t == REQ_TYPE_FREE) {
388 struct rte_mp_msg msg;
389 struct malloc_mp_req *resp = (struct malloc_mp_req *)msg.param;
391 memset(&msg, 0, sizeof(msg));
393 /* this is a free request, just sendmsg result */
394 resp->t = REQ_TYPE_FREE;
395 resp->result = result;
396 resp->id = entry->user_req.id;
398 msg.len_param = sizeof(*resp);
399 strlcpy(msg.name, MP_ACTION_RESPONSE, sizeof(msg.name));
401 if (rte_mp_sendmsg(&msg))
402 RTE_LOG(ERR, EAL, "Could not send message to secondary process\n");
404 TAILQ_REMOVE(&mp_request_list.list, entry, next);
406 } else if (entry->user_req.t == REQ_TYPE_ALLOC &&
407 result == REQ_RESULT_SUCCESS) {
408 struct malloc_heap *heap = entry->alloc_state.heap;
409 struct rte_mp_msg msg;
410 struct malloc_mp_req *resp =
411 (struct malloc_mp_req *)msg.param;
413 memset(&msg, 0, sizeof(msg));
415 heap->total_size += entry->alloc_state.map_len;
417 /* result is success, so just notify secondary about this */
418 resp->t = REQ_TYPE_ALLOC;
419 resp->result = result;
420 resp->id = entry->user_req.id;
422 msg.len_param = sizeof(*resp);
423 strlcpy(msg.name, MP_ACTION_RESPONSE, sizeof(msg.name));
425 if (rte_mp_sendmsg(&msg))
426 RTE_LOG(ERR, EAL, "Could not send message to secondary process\n");
428 TAILQ_REMOVE(&mp_request_list.list, entry, next);
429 free(entry->alloc_state.ms);
431 } else if (entry->user_req.t == REQ_TYPE_ALLOC &&
432 result == REQ_RESULT_FAIL) {
433 struct rte_mp_msg rb_msg;
434 struct malloc_mp_req *rb =
435 (struct malloc_mp_req *)rb_msg.param;
437 struct primary_alloc_req_state *state =
441 memset(&rb_msg, 0, sizeof(rb_msg));
443 /* we've failed to sync, so do a rollback */
444 eal_memalloc_mem_event_notify(RTE_MEM_EVENT_FREE,
445 state->map_addr, state->map_len);
447 rollback_expand_heap(state->ms, state->ms_len, state->elem,
448 state->map_addr, state->map_len);
450 /* send rollback request */
452 rb_msg.len_param = sizeof(*rb);
453 strlcpy(rb_msg.name, MP_ACTION_ROLLBACK, sizeof(rb_msg.name));
456 ts.tv_sec = MP_TIMEOUT_S;
458 /* sync requests carry no data */
459 rb->t = REQ_TYPE_SYNC;
460 rb->id = entry->user_req.id;
462 /* there may be stray timeout still waiting */
464 ret = rte_mp_request_async(&rb_msg, &ts,
465 handle_rollback_response);
466 } while (ret != 0 && rte_errno == EEXIST);
468 RTE_LOG(ERR, EAL, "Could not send rollback request to secondary process\n");
470 /* we couldn't send rollback request, but that's OK -
471 * secondary will time out, and memory has been removed
474 TAILQ_REMOVE(&mp_request_list.list, entry, next);
480 RTE_LOG(ERR, EAL, " to sync request of unknown type\n");
484 pthread_mutex_unlock(&mp_request_list.lock);
487 pthread_mutex_unlock(&mp_request_list.lock);
492 handle_rollback_response(const struct rte_mp_msg *request,
493 const struct rte_mp_reply *reply __rte_unused)
495 struct rte_mp_msg msg;
496 struct malloc_mp_req *resp = (struct malloc_mp_req *)msg.param;
497 const struct malloc_mp_req *mpreq =
498 (const struct malloc_mp_req *)request->param;
499 struct mp_request *entry;
501 /* lock the request */
502 pthread_mutex_lock(&mp_request_list.lock);
504 memset(&msg, 0, sizeof(msg));
506 entry = find_request_by_id(mpreq->id);
508 RTE_LOG(ERR, EAL, "Wrong request ID\n");
512 if (entry->user_req.t != REQ_TYPE_ALLOC) {
513 RTE_LOG(ERR, EAL, "Unexpected active request\n");
517 /* we don't care if rollback succeeded, request still failed */
518 resp->t = REQ_TYPE_ALLOC;
519 resp->result = REQ_RESULT_FAIL;
520 resp->id = mpreq->id;
522 msg.len_param = sizeof(*resp);
523 strlcpy(msg.name, MP_ACTION_RESPONSE, sizeof(msg.name));
525 if (rte_mp_sendmsg(&msg))
526 RTE_LOG(ERR, EAL, "Could not send message to secondary process\n");
529 TAILQ_REMOVE(&mp_request_list.list, entry, next);
530 free(entry->alloc_state.ms);
533 pthread_mutex_unlock(&mp_request_list.lock);
536 pthread_mutex_unlock(&mp_request_list.lock);
540 /* final stage of the request from secondary */
542 handle_response(const struct rte_mp_msg *msg, const void *peer __rte_unused)
544 const struct malloc_mp_req *m =
545 (const struct malloc_mp_req *)msg->param;
546 struct mp_request *entry;
548 pthread_mutex_lock(&mp_request_list.lock);
550 entry = find_request_by_id(m->id);
552 /* update request status */
553 entry->user_req.result = m->result;
555 entry->state = REQ_STATE_COMPLETE;
557 /* trigger thread wakeup */
558 pthread_cond_signal(&entry->cond);
561 pthread_mutex_unlock(&mp_request_list.lock);
566 /* synchronously request memory map sync, this is only called whenever primary
567 * process initiates the allocation.
572 struct rte_mp_msg msg;
573 struct rte_mp_reply reply;
574 struct malloc_mp_req *req = (struct malloc_mp_req *)msg.param;
578 memset(&msg, 0, sizeof(msg));
579 memset(&reply, 0, sizeof(reply));
581 /* no need to create tailq entries as this is entirely synchronous */
584 msg.len_param = sizeof(*req);
585 strlcpy(msg.name, MP_ACTION_SYNC, sizeof(msg.name));
587 /* sync request carries no data */
588 req->t = REQ_TYPE_SYNC;
589 req->id = get_unique_id();
592 ts.tv_sec = MP_TIMEOUT_S;
594 /* there may be stray timeout still waiting */
596 ret = rte_mp_request_sync(&msg, &reply, &ts);
597 } while (ret != 0 && rte_errno == EEXIST);
599 RTE_LOG(ERR, EAL, "Could not send sync request to secondary process\n");
604 if (reply.nb_received != reply.nb_sent) {
605 RTE_LOG(ERR, EAL, "Not all secondaries have responded\n");
610 for (i = 0; i < reply.nb_received; i++) {
611 struct malloc_mp_req *resp =
612 (struct malloc_mp_req *)reply.msgs[i].param;
613 if (resp->t != REQ_TYPE_SYNC) {
614 RTE_LOG(ERR, EAL, "Unexpected response from secondary\n");
618 if (resp->id != req->id) {
619 RTE_LOG(ERR, EAL, "Wrong request ID\n");
623 if (resp->result != REQ_RESULT_SUCCESS) {
624 RTE_LOG(ERR, EAL, "Secondary process failed to synchronize\n");
636 /* this is a synchronous wrapper around a bunch of asynchronous requests to
637 * primary process. this will initiate a request and wait until responses come.
640 request_to_primary(struct malloc_mp_req *user_req)
642 struct rte_mp_msg msg;
643 struct malloc_mp_req *msg_req = (struct malloc_mp_req *)msg.param;
644 struct mp_request *entry;
649 memset(&msg, 0, sizeof(msg));
650 memset(&ts, 0, sizeof(ts));
652 pthread_mutex_lock(&mp_request_list.lock);
654 entry = malloc(sizeof(*entry));
656 RTE_LOG(ERR, EAL, "Cannot allocate memory for request\n");
660 memset(entry, 0, sizeof(*entry));
662 if (gettimeofday(&now, NULL) < 0) {
663 RTE_LOG(ERR, EAL, "Cannot get current time\n");
667 ts.tv_nsec = (now.tv_usec * 1000) % 1000000000;
668 ts.tv_sec = now.tv_sec + MP_TIMEOUT_S +
669 (now.tv_usec * 1000) / 1000000000;
671 /* initialize the request */
672 pthread_cond_init(&entry->cond, NULL);
675 msg.len_param = sizeof(*msg_req);
676 strlcpy(msg.name, MP_ACTION_REQUEST, sizeof(msg.name));
678 /* (attempt to) get a unique id */
679 user_req->id = get_unique_id();
681 /* copy contents of user request into the message */
682 memcpy(msg_req, user_req, sizeof(*msg_req));
684 if (rte_mp_sendmsg(&msg)) {
685 RTE_LOG(ERR, EAL, "Cannot send message to primary\n");
689 /* copy contents of user request into active request */
690 memcpy(&entry->user_req, user_req, sizeof(*user_req));
692 /* mark request as in progress */
693 entry->state = REQ_STATE_ACTIVE;
695 TAILQ_INSERT_TAIL(&mp_request_list.list, entry, next);
697 /* finally, wait on timeout */
699 ret = pthread_cond_timedwait(&entry->cond,
700 &mp_request_list.lock, &ts);
701 } while (ret != 0 && ret != ETIMEDOUT);
703 if (entry->state != REQ_STATE_COMPLETE) {
704 RTE_LOG(ERR, EAL, "Request timed out\n");
708 user_req->result = entry->user_req.result;
710 TAILQ_REMOVE(&mp_request_list.list, entry, next);
713 pthread_mutex_unlock(&mp_request_list.lock);
716 pthread_mutex_unlock(&mp_request_list.lock);
722 register_mp_requests(void)
724 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
725 if (rte_mp_action_register(MP_ACTION_REQUEST, handle_request)) {
726 RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
731 if (rte_mp_action_register(MP_ACTION_SYNC, handle_sync)) {
732 RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
736 if (rte_mp_action_register(MP_ACTION_ROLLBACK, handle_sync)) {
737 RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",
741 if (rte_mp_action_register(MP_ACTION_RESPONSE,
743 RTE_LOG(ERR, EAL, "Couldn't register '%s' action\n",