1
/* Copyright (c) 2012, 2013 Nicira, Inc.
3
* Licensed under the Apache License, Version 2.0 (the "License");
4
* you may not use this file except in compliance with the License.
5
* You may obtain a copy of the License at:
7
* http://www.apache.org/licenses/LICENSE-2.0
9
* Unless required by applicable law or agreed to in writing, software
10
* distributed under the License is distributed on an "AS IS" BASIS,
11
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
* See the License for the specific language governing permissions and
13
* limitations under the License.
24
#include <sys/socket.h>
25
#include <sys/types.h>
30
#include "command-line.h"
33
#include "poll-loop.h"
34
#include "socket-util.h"
38
VLOG_DEFINE_THIS_MODULE(worker);
40
/* ovs_assert() logs the assertion message and logging sometimes goes through a
41
* worker, so using ovs_assert() in this source file could cause recursion. */
43
#define ovs_assert use_assert_instead_of_ovs_assert_in_this_module
45
/* Header for an RPC request. */
46
struct worker_request {
47
size_t request_len; /* Length of the payload in bytes. */
48
worker_request_func *request_cb; /* Function to call in worker process. */
49
worker_reply_func *reply_cb; /* Function to call in main process. */
50
void *reply_aux; /* Auxiliary data for 'reply_cb'. */
53
/* Header for an RPC reply. */
55
size_t reply_len; /* Length of the payload in bytes. */
56
worker_reply_func *reply_cb; /* Function to call in main process. */
57
void *reply_aux; /* Auxiliary data for 'reply_cb'. */
60
/* Receive buffer for a RPC request or reply. */
63
struct ofpbuf header; /* Header data. */
64
int fds[SOUTIL_MAX_FDS]; /* File descriptors. */
68
struct ofpbuf payload; /* Payload data. */
71
static int client_sock = -1;
72
static struct rxbuf client_rx;
74
static void rxbuf_init(struct rxbuf *);
75
static void rxbuf_clear(struct rxbuf *);
76
static int rxbuf_run(struct rxbuf *, int sock, size_t header_len);
78
static struct iovec *prefix_iov(void *data, size_t len,
79
const struct iovec *iovs, size_t n_iovs);
81
static void worker_broke(void);
83
static void worker_main(int fd) NO_RETURN;
85
/* Starts a worker process as a subprocess of the current process. Currently
86
* only a single worker process is supported, so this function may only be
89
* The client should call worker_run() and worker_wait() from its main loop.
91
* Call this function between daemonize_start() and daemonize_complete(). */
97
assert(client_sock < 0);
99
/* Create non-blocking socket pair. */
100
xsocketpair(AF_UNIX, SOCK_STREAM, 0, work_fds);
101
xset_nonblocking(work_fds[0]);
102
xset_nonblocking(work_fds[1]);
104
/* Don't let the worker process own the responsibility to delete
105
* the pidfile. Register it again after the fork. */
106
remove_pidfile_from_unlink();
107
if (!fork_and_clean_up()) {
108
/* In child (worker) process. */
109
daemonize_post_detach();
111
worker_main(work_fds[1]);
115
/* In parent (main) process. */
116
add_pidfile_to_unlink();
118
client_sock = work_fds[0];
119
rxbuf_init(&client_rx);
122
/* Returns true if this process has started a worker and the worker is not
123
* known to have malfunctioned. */
125
worker_is_running(void)
127
return client_sock >= 0;
130
/* If a worker process was started, processes RPC replies from it, calling the
131
* registered 'reply_cb' callbacks.
133
* If the worker process died or malfunctioned, aborts. */
137
if (worker_is_running()) {
140
error = rxbuf_run(&client_rx, client_sock,
141
sizeof(struct worker_reply));
143
struct worker_reply *reply = client_rx.header.data;
144
reply->reply_cb(&client_rx.payload, client_rx.fds,
145
client_rx.n_fds, reply->reply_aux);
146
rxbuf_clear(&client_rx);
147
} else if (error != EAGAIN) {
149
VLOG_ABORT("receive from worker failed (%s)",
150
ovs_retval_to_string(error));
155
/* Causes the poll loop to wake up if we need to process RPC replies. */
159
if (worker_is_running()) {
160
poll_fd_wait(client_sock, POLLIN);
164
/* Interface for main process to interact with the worker. */
166
/* Sends an RPC request to the worker process. The worker process will call
167
* 'request_cb' passing the 'size' (zero or more) bytes of data in 'data' as
168
* arguments as well as the 'n_fds' (SOUTIL_MAX_FDS or fewer) file descriptors
171
* If and only if 'reply_cb' is nonnull, 'request_cb' must call worker_reply()
172
* or worker_reply_iovec() with a reply. The main process will later call
173
* 'reply_cb' with the reply data (if any) and file descriptors (if any).
175
* 'request_cb' receives copies (as if by dup()) of the file descriptors in
176
* fds[]. 'request_cb' takes ownership of these copies, and the caller of
177
* worker_request() retains its ownership of the originals.
179
* This function may block until the RPC request has been sent (if the socket
180
* buffer fills up) but it does not wait for the reply (if any). If this
181
* function blocks, it may invoke reply callbacks for previous requests.
183
* The worker process executes RPC requests in strict order of submission and
184
* runs each request to completion before beginning the next request. The main
185
* process invokes reply callbacks in strict order of request submission. */
187
worker_request(const void *data, size_t size,
188
const int fds[], size_t n_fds,
189
worker_request_func *request_cb,
190
worker_reply_func *reply_cb, void *aux)
195
iov.iov_base = (void *) data;
197
worker_request_iovec(&iov, 1, fds, n_fds, request_cb, reply_cb, aux);
199
worker_request_iovec(NULL, 0, fds, n_fds, request_cb, reply_cb, aux);
204
worker_send_iovec(const struct iovec iovs[], size_t n_iovs,
205
const int fds[], size_t n_fds)
213
/* Try to send the rest of the request. */
214
error = send_iovec_and_fds_fully(client_sock, iovs, n_iovs,
215
fds, n_fds, sent, &sent);
216
if (error != EAGAIN) {
220
/* Process replies to avoid deadlock. */
223
/* Wait for 'client_sock' to become ready before trying again. We
224
* can't use poll_block() because it sometimes calls into vlog, which
225
* calls indirectly into worker_send_iovec(). To be usable here,
226
* poll_block() would therefore need to be reentrant, but it isn't
227
* (calling it recursively causes memory corruption and an eventual
229
pfd.fd = client_sock;
230
pfd.events = POLLIN | POLLOUT;
232
error = poll(&pfd, 1, -1) < 0 ? errno : 0;
233
} while (error == EINTR);
236
VLOG_ABORT("poll failed (%s)", strerror(error));
241
/* Same as worker_request() except that the data to send is specified as an
242
* array of iovecs. */
244
worker_request_iovec(const struct iovec iovs[], size_t n_iovs,
245
const int fds[], size_t n_fds,
246
worker_request_func *request_cb,
247
worker_reply_func *reply_cb, void *aux)
249
static bool recursing = false;
250
struct worker_request rq;
251
struct iovec *all_iovs;
254
assert(worker_is_running());
258
rq.request_len = iovec_len(iovs, n_iovs);
259
rq.request_cb = request_cb;
260
rq.reply_cb = reply_cb;
263
all_iovs = prefix_iov(&rq, sizeof rq, iovs, n_iovs);
264
error = worker_send_iovec(all_iovs, n_iovs + 1, fds, n_fds);
267
VLOG_ABORT("send failed (%s)", strerror(error));
274
/* Closes the client socket, if any, so that worker_is_running() will return
277
* The client does this just before aborting if the worker process dies or
278
* malfunctions, to prevent the logging subsystem from trying to use the
279
* worker to log the failure. */
283
if (client_sock >= 0) {
289
/* Interfaces for RPC implementations (running in the worker process). */
291
static int server_sock = -1;
292
static bool expect_reply;
293
static struct worker_request request;
295
/* When a call to worker_request() or worker_request_iovec() provides a
296
* 'reply_cb' callback, the 'request_cb' implementation must call this function
297
* to send its reply. The main process will call 'reply_cb' passing the
298
* 'size' (zero or more) bytes of data in 'data' as arguments as well as the
299
* 'n_fds' (SOUTIL_MAX_FDS or fewer) file descriptors in 'fds'.
301
* If a call to worker_request() or worker_request_iovec() provides no
302
* 'reply_cb' callback, the 'request_cb' implementation must not call this
305
* 'reply_cb' receives copies (as if by dup()) of the file descriptors in
306
* fds[]. 'reply_cb' takes ownership of these copies, and the caller of
307
* worker_reply() retains its ownership of the originals.
309
* This function blocks until the RPC reply has been sent (if the socket buffer
310
* fills up) but it does not wait for the main process to receive or to process
313
worker_reply(const void *data, size_t size, const int fds[], size_t n_fds)
318
iov.iov_base = (void *) data;
320
worker_reply_iovec(&iov, 1, fds, n_fds);
322
worker_reply_iovec(NULL, 0, fds, n_fds);
326
/* Same as worker_reply() except that the data to send is specified as an array
329
worker_reply_iovec(const struct iovec *iovs, size_t n_iovs,
330
const int fds[], size_t n_fds)
332
struct worker_reply reply;
333
struct iovec *all_iovs;
336
assert(expect_reply);
337
expect_reply = false;
339
reply.reply_len = iovec_len(iovs, n_iovs);
340
reply.reply_cb = request.reply_cb;
341
reply.reply_aux = request.reply_aux;
343
all_iovs = prefix_iov(&reply, sizeof reply, iovs, n_iovs);
345
error = send_iovec_and_fds_fully_block(server_sock, all_iovs, n_iovs + 1,
347
if (error == EPIPE) {
348
/* Parent probably died. Continue processing any RPCs still buffered,
349
* to avoid missing log messages. */
350
VLOG_INFO("send failed (%s)", strerror(error));
352
VLOG_FATAL("send failed (%s)", strerror(error));
365
subprogram_name = "worker";
366
proctitle_set("worker process for pid %lu", (unsigned long int) getppid());
367
VLOG_INFO("worker process started");
373
error = rxbuf_run(&rx, server_sock, sizeof(struct worker_request));
375
request = *(struct worker_request *) rx.header.data;
377
expect_reply = request.reply_cb != NULL;
378
request.request_cb(&rx.payload, rx.fds, rx.n_fds);
379
assert(!expect_reply);
382
} else if (error == EOF && !rx.header.size) {
383
/* Main process closed the IPC socket. Exit cleanly. */
385
} else if (error != EAGAIN) {
386
VLOG_FATAL("RPC receive failed (%s)", ovs_retval_to_string(error));
389
poll_fd_wait(server_sock, POLLIN);
393
VLOG_INFO("worker process exiting");
398
rxbuf_init(struct rxbuf *rx)
400
ofpbuf_init(&rx->header, 0);
402
ofpbuf_init(&rx->payload, 0);
406
rxbuf_clear(struct rxbuf *rx)
408
ofpbuf_clear(&rx->header);
410
ofpbuf_clear(&rx->payload);
414
rxbuf_run(struct rxbuf *rx, int sock, size_t header_len)
417
if (!rx->header.size) {
420
ofpbuf_clear(&rx->header);
421
ofpbuf_prealloc_tailroom(&rx->header, header_len);
423
retval = recv_data_and_fds(sock, rx->header.data, header_len,
424
rx->fds, &rx->n_fds);
426
return retval ? -retval : EOF;
428
rx->header.size += retval;
429
} else if (rx->header.size < header_len) {
433
error = read_fully(sock, ofpbuf_tail(&rx->header),
434
header_len - rx->header.size, &bytes_read);
435
rx->header.size += bytes_read;
440
size_t payload_len = *(size_t *) rx->header.data;
442
if (rx->payload.size < payload_len) {
443
size_t left = payload_len - rx->payload.size;
447
ofpbuf_prealloc_tailroom(&rx->payload, left);
448
error = read_fully(sock, ofpbuf_tail(&rx->payload), left,
450
rx->payload.size += bytes_read;
461
static struct iovec *
462
prefix_iov(void *data, size_t len, const struct iovec *iovs, size_t n_iovs)
466
dst = xmalloc((n_iovs + 1) * sizeof *dst);
467
dst[0].iov_base = data;
468
dst[0].iov_len = len;
469
memcpy(dst + 1, iovs, n_iovs * sizeof *iovs);