thebestvibedev/tromcho.net
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Wait for RDMA-CM EVENT_ESTABLISHED after rdma_accept(), handle rdma_accept() before acking the event

Browse Source
This commit is contained in:
Vitaliy Filippov
2025-04-15 15:19:36 +03:00
parent 327c98a4b6
commit df2847df2d
2 changed files with 63 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/client/messenger.h
+1
View File
@@ -287,6 +287,7 @@ protected:
msgr_rdma_context_t* rdmacm_create_qp(rdma_cm_id *cmid); msgr_rdma_context_t* rdmacm_create_qp(rdma_cm_id *cmid);
void rdmacm_accept(rdma_cm_event *ev); void rdmacm_accept(rdma_cm_event *ev);
void rdmacm_try_connect_peer(uint64_t peer_osd, const std::string & addr, int rdmacm_port, int fallback_tcp_port); void rdmacm_try_connect_peer(uint64_t peer_osd, const std::string & addr, int rdmacm_port, int fallback_tcp_port);
void rdmacm_set_conn_timeout(rdmacm_connecting_t *conn);
void rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res); void rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res);
void rdmacm_address_resolved(rdma_cm_event *ev); void rdmacm_address_resolved(rdma_cm_event *ev);
void rdmacm_route_resolved(rdma_cm_event *ev); void rdmacm_route_resolved(rdma_cm_event *ev);
src/client/msgr_rdmacm.cpp
+62 -49
View File
@@ -70,7 +70,7 @@ void osd_messenger_t::rdmacm_destroy_listener(rdma_cm_id *listener)
void osd_messenger_t::handle_rdmacm_events() void osd_messenger_t::handle_rdmacm_events()
{ {
// rdma_destroy_id infinitely waits for pthread_cond if called before all events are acked :-( // rdma_destroy_id infinitely waits for pthread_cond if called before all events are acked :-(...
std::vector<rdma_cm_event> events_copy; std::vector<rdma_cm_event> events_copy;
while (1) while (1)
{ {
@@ -83,7 +83,15 @@ void osd_messenger_t::handle_rdmacm_events()
fprintf(stderr, "Failed to get RDMA-CM event: %s (code %d)\n", strerror(errno), errno); fprintf(stderr, "Failed to get RDMA-CM event: %s (code %d)\n", strerror(errno), errno);
exit(1); exit(1);
} }
events_copy.push_back(*ev); // ...so we save a copy of all events EXCEPT connection requests, otherwise they sometimes fail with EVENT_DISCONNECT
if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
{
rdmacm_accept(ev);
}
else
{
events_copy.push_back(*ev);
}
r = rdma_ack_cm_event(ev); r = rdma_ack_cm_event(ev);
if (r != 0) if (r != 0)
{ {
@@ -96,7 +104,7 @@ void osd_messenger_t::handle_rdmacm_events()
auto ev = &evl; auto ev = &evl;
if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST) if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
{ {
rdmacm_accept(ev); // Do nothing, handled above
} }
else if (ev->event == RDMA_CM_EVENT_CONNECT_ERROR || else if (ev->event == RDMA_CM_EVENT_CONNECT_ERROR ||
ev->event == RDMA_CM_EVENT_REJECTED || ev->event == RDMA_CM_EVENT_REJECTED ||
@@ -287,29 +295,34 @@ void osd_messenger_t::rdmacm_accept(rdma_cm_event *ev)
rdma_destroy_id(ev->id); rdma_destroy_id(ev->id);
return; return;
} }
rdma_context->cm_refs++; // Wait for RDMA_CM_ESTABLISHED, and enable the connection only after it
// Wrap into a new msgr_rdma_connection_t auto conn = new rdmacm_connecting_t;
msgr_rdma_connection_t *conn = new msgr_rdma_connection_t;
conn->ctx = rdma_context;
conn->max_send = rdma_max_send;
conn->max_recv = rdma_max_recv;
conn->max_sge = rdma_max_sge > rdma_context->attrx.orig_attr.max_sge
? rdma_context->attrx.orig_attr.max_sge : rdma_max_sge;
conn->max_msg = rdma_max_msg;
conn->cmid = ev->id; conn->cmid = ev->id;
conn->qp = ev->id->qp; conn->peer_fd = fake_fd;
auto cl = new osd_client_t(); conn->parsed_addr = *(sockaddr_storage*)rdma_get_peer_addr(ev->id);
cl->peer_fd = fake_fd; conn->rdma_context = rdma_context;
cl->peer_state = PEER_RDMA; rdmacm_set_conn_timeout(conn);
cl->peer_addr = *(sockaddr_storage*)rdma_get_peer_addr(ev->id); rdmacm_connecting[ev->id] = conn;
cl->in_buf = malloc_or_die(receive_buffer_size); fprintf(stderr, "[OSD %ju] new client %d: connection from %s via RDMA-CM\n", this->osd_num, conn->peer_fd,
cl->rdma_conn = conn; addr_to_string(conn->parsed_addr).c_str());
clients[fake_fd] = cl; }
rdmacm_connections[ev->id] = cl;
// Add initial receive request(s) void osd_messenger_t::rdmacm_set_conn_timeout(rdmacm_connecting_t *conn)
try_recv_rdma(cl); {
fprintf(stderr, "[OSD %ju] new client %d: connection from %s via RDMA-CM\n", this->osd_num, fake_fd, conn->timeout_ms = peer_connect_timeout*1000;
addr_to_string(cl->peer_addr).c_str()); if (peer_connect_timeout > 0)
{
conn->timeout_id = tfd->set_timer(1000*peer_connect_timeout, false, [this, cmid = conn->cmid](int timer_id)
{
auto conn = rdmacm_connecting.at(cmid);
conn->timeout_id = -1;
if (conn->peer_osd)
fprintf(stderr, "RDMA-CM connection to %s timed out\n", conn->addr.c_str());
else
fprintf(stderr, "Incoming RDMA-CM connection from %s timed out\n", addr_to_string(conn->parsed_addr).c_str());
rdmacm_on_connect_peer_error(cmid, -EPIPE);
});
}
} }
void osd_messenger_t::rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res) void osd_messenger_t::rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res)
@@ -332,15 +345,18 @@ void osd_messenger_t::rdmacm_on_connect_peer_error(rdma_cm_id *cmid, int res)
} }
rdmacm_connecting.erase(cmid); rdmacm_connecting.erase(cmid);
delete conn; delete conn;
if (!disable_tcp) if (peer_osd)
{ {
// Fall back to TCP instead of just reporting the error to on_connect_peer() if (!disable_tcp)
try_connect_peer_tcp(peer_osd, addr.c_str(), tcp_port); {
} // Fall back to TCP instead of just reporting the error to on_connect_peer()
else try_connect_peer_tcp(peer_osd, addr.c_str(), tcp_port);
{ }
// TCP is disabled else
on_connect_peer(peer_osd, res == 0 ? -EINVAL : (res > 0 ? -res : res)); {
// TCP is disabled
on_connect_peer(peer_osd, res == 0 ? -EINVAL : (res > 0 ? -res : res));
}
} }
} }
@@ -374,6 +390,8 @@ void osd_messenger_t::rdmacm_try_connect_peer(uint64_t peer_osd, const std::stri
on_connect_peer(peer_osd, res); on_connect_peer(peer_osd, res);
return; return;
} }
if (log_level > 0)
fprintf(stderr, "Trying to connect to OSD %ju at %s:%d via RDMA-CM\n", peer_osd, addr.c_str(), rdmacm_port);
auto conn = new rdmacm_connecting_t; auto conn = new rdmacm_connecting_t;
rdmacm_connecting[cmid] = conn; rdmacm_connecting[cmid] = conn;
conn->cmid = cmid; conn->cmid = cmid;
@@ -383,19 +401,7 @@ void osd_messenger_t::rdmacm_try_connect_peer(uint64_t peer_osd, const std::stri
conn->parsed_addr = sa; conn->parsed_addr = sa;
conn->rdmacm_port = rdmacm_port; conn->rdmacm_port = rdmacm_port;
conn->tcp_port = fallback_tcp_port; conn->tcp_port = fallback_tcp_port;
conn->timeout_ms = peer_connect_timeout*1000; rdmacm_set_conn_timeout(conn);
conn->timeout_id = -1;
if (peer_connect_timeout > 0)
{
conn->timeout_id = tfd->set_timer(1000*peer_connect_timeout, false, [this, cmid](int timer_id)
{
auto conn = rdmacm_connecting.at(cmid);
conn->timeout_id = -1;
fprintf(stderr, "RDMA-CM connection to %s timed out\n", conn->addr.c_str());
rdmacm_on_connect_peer_error(cmid, -EPIPE);
return;
});
}
if (rdma_resolve_addr(cmid, NULL, (sockaddr*)&conn->parsed_addr, conn->timeout_ms) != 0) if (rdma_resolve_addr(cmid, NULL, (sockaddr*)&conn->parsed_addr, conn->timeout_ms) != 0)
{ {
auto res = -errno; auto res = -errno;
@@ -494,7 +500,7 @@ void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
// Wrap into a new msgr_rdma_connection_t // Wrap into a new msgr_rdma_connection_t
msgr_rdma_connection_t *rc = new msgr_rdma_connection_t; msgr_rdma_connection_t *rc = new msgr_rdma_connection_t;
rc->ctx = conn->rdma_context; rc->ctx = conn->rdma_context;
rc->ctx->cm_refs++; rc->ctx->cm_refs++; // FIXME now unused, count also connecting_t's when used
rc->max_send = rdma_max_send; rc->max_send = rdma_max_send;
rc->max_recv = rdma_max_recv; rc->max_recv = rdma_max_recv;
rc->max_sge = rdma_max_sge > rc->ctx->attrx.orig_attr.max_sge rc->max_sge = rdma_max_sge > rc->ctx->attrx.orig_attr.max_sge
@@ -514,13 +520,20 @@ void osd_messenger_t::rdmacm_established(rdma_cm_event *ev)
cl->rdma_conn = rc; cl->rdma_conn = rc;
clients[conn->peer_fd] = cl; clients[conn->peer_fd] = cl;
if (conn->timeout_id >= 0) if (conn->timeout_id >= 0)
{
tfd->clear_timer(conn->timeout_id); tfd->clear_timer(conn->timeout_id);
}
delete conn; delete conn;
rdmacm_connecting.erase(cmid); rdmacm_connecting.erase(cmid);
rdmacm_connections[cmid] = cl; rdmacm_connections[cmid] = cl;
if (log_level > 0) if (log_level > 0 && peer_osd)
{
fprintf(stderr, "Successfully connected with OSD %ju using RDMA-CM\n", peer_osd); fprintf(stderr, "Successfully connected with OSD %ju using RDMA-CM\n", peer_osd);
}
// Add initial receive request(s) // Add initial receive request(s)
try_recv_rdma(cl); try_recv_rdma(cl);
check_peer_config(cl); if (peer_osd)
{
check_peer_config(cl);
}
} }