2 * Simulate the Posix AIO using mmap/fork
4 * Copyright (C) Volker Lendecke 2008
5 * Copyright (C) Jeremy Allison 2010
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #include "system/filesys.h"
24 #include "system/shmem.h"
25 #include "smbd/smbd.h"
26 #include "smbd/globals.h"
27 #include "lib/async_req/async_sock.h"
28 #include "lib/util/tevent_unix.h"
29 #include "lib/sys_rw.h"
30 #include "lib/sys_rw_data.h"
32 #if !defined(HAVE_STRUCT_MSGHDR_MSG_CONTROL) && !defined(HAVE_STRUCT_MSGHDR_MSG_ACCRIGHTS)
33 # error Can not pass file descriptors
42 struct aio_fork_config {
43 bool erratic_testing_mode;
51 static int mmap_area_destructor(struct mmap_area *area)
53 munmap((void *)area->ptr, area->size);
57 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
59 struct mmap_area *result;
62 result = talloc(mem_ctx, struct mmap_area);
64 DEBUG(0, ("talloc failed\n"));
68 fd = open("/dev/zero", O_RDWR);
70 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
75 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
76 MAP_SHARED|MAP_FILE, fd, 0);
78 if (result->ptr == MAP_FAILED) {
79 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
84 talloc_set_destructor(result, mmap_area_destructor);
99 static const char *cmd_type_str(enum cmd_type cmd)
114 result = "<UNKNOWN>";
124 bool erratic_testing_mode;
132 struct aio_child_list;
135 struct aio_child *prev, *next;
136 struct aio_child_list *list;
139 struct mmap_area *map;
140 bool dont_delete; /* Marked as in use since last cleanup */
144 struct aio_child_list {
145 struct aio_child *children;
146 struct tevent_timer *cleanup_event;
149 static void free_aio_children(void **p)
154 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
159 #ifndef HAVE_STRUCT_MSGHDR_MSG_CONTROL
163 msg.msg_accrights = (caddr_t) &newfd;
164 msg.msg_accrightslen = sizeof(int);
169 char control[CMSG_SPACE(sizeof(int))];
171 struct cmsghdr *cmptr;
174 ZERO_STRUCT(control_un);
176 msg.msg_control = control_un.control;
177 msg.msg_controllen = sizeof(control_un.control);
183 iov[0].iov_base = (void *)ptr;
184 iov[0].iov_len = nbytes;
188 if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
192 #ifdef HAVE_STRUCT_MSGHDR_MSG_CONTROL
193 if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
194 && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
195 if (cmptr->cmsg_level != SOL_SOCKET) {
196 DEBUG(10, ("control level != SOL_SOCKET"));
200 if (cmptr->cmsg_type != SCM_RIGHTS) {
201 DEBUG(10, ("control type != SCM_RIGHTS"));
205 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
207 *recvfd = -1; /* descriptor was not passed */
210 if (msg.msg_accrightslen == sizeof(int)) {
214 *recvfd = -1; /* descriptor was not passed */
221 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
226 #ifdef HAVE_STRUCT_MSGHDR_MSG_CONTROL
229 char control[CMSG_SPACE(sizeof(int))];
231 struct cmsghdr *cmptr;
234 ZERO_STRUCT(control_un);
236 msg.msg_control = control_un.control;
237 msg.msg_controllen = sizeof(control_un.control);
239 cmptr = CMSG_FIRSTHDR(&msg);
240 cmptr->cmsg_len = CMSG_LEN(sizeof(int));
241 cmptr->cmsg_level = SOL_SOCKET;
242 cmptr->cmsg_type = SCM_RIGHTS;
243 memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
246 msg.msg_accrights = (caddr_t) &sendfd;
247 msg.msg_accrightslen = sizeof(int);
254 iov[0].iov_base = (void *)ptr;
255 iov[0].iov_len = nbytes;
259 return (sendmsg(fd, &msg, 0));
262 static void aio_child_cleanup(struct tevent_context *event_ctx,
263 struct tevent_timer *te,
267 struct aio_child_list *list = talloc_get_type_abort(
268 private_data, struct aio_child_list);
269 struct aio_child *child, *next;
271 TALLOC_FREE(list->cleanup_event);
273 for (child = list->children; child != NULL; child = next) {
277 DEBUG(10, ("child %d currently active\n",
282 if (child->dont_delete) {
283 DEBUG(10, ("Child %d was active since last cleanup\n",
285 child->dont_delete = false;
289 DEBUG(10, ("Child %d idle for more than 30 seconds, "
290 "deleting\n", (int)child->pid));
296 if (list->children != NULL) {
298 * Re-schedule the next cleanup round
300 list->cleanup_event = tevent_add_timer(server_event_context(), list,
301 timeval_add(&now, 30, 0),
302 aio_child_cleanup, list);
307 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
309 struct aio_child_list *data = NULL;
311 if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
312 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
317 data = talloc_zero(NULL, struct aio_child_list);
324 * Regardless of whether the child_list had been around or not, make
325 * sure that we have a cleanup timed event. This timed event will
326 * delete itself when it finds that no children are around anymore.
329 if (data->cleanup_event == NULL) {
330 data->cleanup_event = tevent_add_timer(server_event_context(), data,
331 timeval_current_ofs(30, 0),
332 aio_child_cleanup, data);
333 if (data->cleanup_event == NULL) {
339 if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
340 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
341 struct aio_child_list, return False);
347 static void aio_child_loop(int sockfd, struct mmap_area *map)
352 struct rw_cmd cmd_struct;
353 struct rw_ret ret_struct;
355 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
356 if (ret != sizeof(cmd_struct)) {
357 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
362 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
363 cmd_type_str(cmd_struct.cmd),
364 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
366 if (cmd_struct.erratic_testing_mode) {
368 * For developer testing, we want erratic behaviour for
374 * use generate_random_buffer, we just forked from a
375 * common parent state
377 generate_random_buffer(&randval, sizeof(randval));
378 msecs = randval + 20;
379 DEBUG(10, ("delaying for %u msecs\n", msecs));
383 ZERO_STRUCT(ret_struct);
385 switch (cmd_struct.cmd) {
387 ret_struct.size = sys_pread(
388 fd, (void *)map->ptr, cmd_struct.n,
391 /* This breaks "make test" when run with aio_fork module. */
393 ret_struct.size = MAX(1, ret_struct.size * 0.9);
398 ret_struct.size = sys_pwrite(
399 fd, (void *)map->ptr, cmd_struct.n,
403 ret_struct.size = fsync(fd);
406 ret_struct.size = -1;
410 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
411 (int)ret_struct.size));
413 if (ret_struct.size == -1) {
414 ret_struct.ret_errno = errno;
418 * Close the fd before telling our parent we're done. The
419 * parent might close and re-open the file very quickly, and
420 * with system-level share modes (GPFS) we would get an
421 * unjustified SHARING_VIOLATION.
425 ret = write_data(sockfd, (char *)&ret_struct,
427 if (ret != sizeof(ret_struct)) {
428 DEBUG(10, ("could not write ret_struct: %s\n",
435 static int aio_child_destructor(struct aio_child *child)
439 SMB_ASSERT(!child->busy);
441 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
442 child->pid, child->sockfd));
445 * closing the sockfd makes the child not return from recvmsg() on RHEL
446 * 5.5 so instead force the child to exit by writing bad data to it
448 write(child->sockfd, &c, sizeof(c));
449 close(child->sockfd);
450 DLIST_REMOVE(child->list->children, child);
455 * We have to close all fd's in open files, we might incorrectly hold a system
456 * level share mode on a file.
459 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
462 if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
469 static int create_aio_child(struct smbd_server_connection *sconn,
470 struct aio_child_list *children,
472 struct aio_child **presult)
474 struct aio_child *result;
478 fdpair[0] = fdpair[1] = -1;
480 result = talloc_zero(children, struct aio_child);
481 if (result == NULL) {
485 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
487 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
491 DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
493 result->map = mmap_area_init(result, map_size);
494 if (result->map == NULL) {
496 DEBUG(0, ("Could not create mmap area\n"));
500 result->pid = fork();
501 if (result->pid == -1) {
503 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
507 if (result->pid == 0) {
509 result->sockfd = fdpair[1];
510 files_forall(sconn, close_fsp_fd, NULL);
511 aio_child_loop(result->sockfd, result->map);
514 DEBUG(10, ("Child %d created with sockfd %d\n",
515 result->pid, fdpair[0]));
517 result->sockfd = fdpair[0];
520 result->list = children;
521 DLIST_ADD(children->children, result);
523 talloc_set_destructor(result, aio_child_destructor);
530 if (fdpair[0] != -1) close(fdpair[0]);
531 if (fdpair[1] != -1) close(fdpair[1]);
537 static int get_idle_child(struct vfs_handle_struct *handle,
538 struct aio_child **pchild)
540 struct aio_child_list *children;
541 struct aio_child *child;
543 children = init_aio_children(handle);
544 if (children == NULL) {
548 for (child = children->children; child != NULL; child = child->next) {
557 DEBUG(10, ("no idle child found, creating new one\n"));
559 ret = create_aio_child(handle->conn->sconn, children,
562 DEBUG(10, ("create_aio_child failed: %s\n",
568 child->dont_delete = true;
575 struct aio_fork_pread_state {
576 struct aio_child *child;
581 static void aio_fork_pread_done(struct tevent_req *subreq);
583 static struct tevent_req *aio_fork_pread_send(struct vfs_handle_struct *handle,
585 struct tevent_context *ev,
586 struct files_struct *fsp,
588 size_t n, off_t offset)
590 struct tevent_req *req, *subreq;
591 struct aio_fork_pread_state *state;
595 struct aio_fork_config *config;
597 SMB_VFS_HANDLE_GET_DATA(handle, config,
598 struct aio_fork_config,
601 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
607 /* TODO: support variable buffers */
608 tevent_req_error(req, EINVAL);
609 return tevent_req_post(req, ev);
612 err = get_idle_child(handle, &state->child);
614 tevent_req_error(req, err);
615 return tevent_req_post(req, ev);
622 cmd.erratic_testing_mode = config->erratic_testing_mode;
624 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
625 (int)state->child->pid));
628 * Not making this async. We're writing into an empty unix
629 * domain socket. This should never block.
631 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
636 TALLOC_FREE(state->child);
638 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
639 tevent_req_error(req, err);
640 return tevent_req_post(req, ev);
643 subreq = read_packet_send(state, ev, state->child->sockfd,
644 sizeof(struct rw_ret), NULL, NULL);
645 if (tevent_req_nomem(subreq, req)) {
646 TALLOC_FREE(state->child); /* we sent sth down */
647 return tevent_req_post(req, ev);
649 tevent_req_set_callback(subreq, aio_fork_pread_done, req);
653 static void aio_fork_pread_done(struct tevent_req *subreq)
655 struct tevent_req *req = tevent_req_callback_data(
656 subreq, struct tevent_req);
657 struct aio_fork_pread_state *state = tevent_req_data(
658 req, struct aio_fork_pread_state);
662 struct rw_ret *retbuf;
664 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
667 TALLOC_FREE(state->child);
668 tevent_req_error(req, err);
672 state->child->busy = false;
674 retbuf = (struct rw_ret *)buf;
675 state->ret = retbuf->size;
676 state->err = retbuf->ret_errno;
677 tevent_req_done(req);
680 static ssize_t aio_fork_pread_recv(struct tevent_req *req, int *err)
682 struct aio_fork_pread_state *state = tevent_req_data(
683 req, struct aio_fork_pread_state);
685 if (tevent_req_is_unix_error(req, err)) {
688 if (state->ret == -1) {
694 struct aio_fork_pwrite_state {
695 struct aio_child *child;
700 static void aio_fork_pwrite_done(struct tevent_req *subreq);
702 static struct tevent_req *aio_fork_pwrite_send(
703 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
704 struct tevent_context *ev, struct files_struct *fsp,
705 const void *data, size_t n, off_t offset)
707 struct tevent_req *req, *subreq;
708 struct aio_fork_pwrite_state *state;
712 struct aio_fork_config *config;
713 SMB_VFS_HANDLE_GET_DATA(handle, config,
714 struct aio_fork_config,
717 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
723 /* TODO: support variable buffers */
724 tevent_req_error(req, EINVAL);
725 return tevent_req_post(req, ev);
728 err = get_idle_child(handle, &state->child);
730 tevent_req_error(req, err);
731 return tevent_req_post(req, ev);
738 cmd.erratic_testing_mode = config->erratic_testing_mode;
740 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
741 (int)state->child->pid));
744 * Not making this async. We're writing into an empty unix
745 * domain socket. This should never block.
747 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
752 TALLOC_FREE(state->child);
754 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
755 tevent_req_error(req, err);
756 return tevent_req_post(req, ev);
759 subreq = read_packet_send(state, ev, state->child->sockfd,
760 sizeof(struct rw_ret), NULL, NULL);
761 if (tevent_req_nomem(subreq, req)) {
762 TALLOC_FREE(state->child); /* we sent sth down */
763 return tevent_req_post(req, ev);
765 tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
769 static void aio_fork_pwrite_done(struct tevent_req *subreq)
771 struct tevent_req *req = tevent_req_callback_data(
772 subreq, struct tevent_req);
773 struct aio_fork_pwrite_state *state = tevent_req_data(
774 req, struct aio_fork_pwrite_state);
778 struct rw_ret *retbuf;
780 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
783 TALLOC_FREE(state->child);
784 tevent_req_error(req, err);
788 state->child->busy = false;
790 retbuf = (struct rw_ret *)buf;
791 state->ret = retbuf->size;
792 state->err = retbuf->ret_errno;
793 tevent_req_done(req);
796 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req, int *err)
798 struct aio_fork_pwrite_state *state = tevent_req_data(
799 req, struct aio_fork_pwrite_state);
801 if (tevent_req_is_unix_error(req, err)) {
804 if (state->ret == -1) {
810 struct aio_fork_fsync_state {
811 struct aio_child *child;
816 static void aio_fork_fsync_done(struct tevent_req *subreq);
818 static struct tevent_req *aio_fork_fsync_send(
819 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
820 struct tevent_context *ev, struct files_struct *fsp)
822 struct tevent_req *req, *subreq;
823 struct aio_fork_fsync_state *state;
827 struct aio_fork_config *config;
829 SMB_VFS_HANDLE_GET_DATA(handle, config,
830 struct aio_fork_config,
833 req = tevent_req_create(mem_ctx, &state, struct aio_fork_fsync_state);
838 err = get_idle_child(handle, &state->child);
840 tevent_req_error(req, err);
841 return tevent_req_post(req, ev);
846 cmd.erratic_testing_mode = config->erratic_testing_mode;
848 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
849 (int)state->child->pid));
852 * Not making this async. We're writing into an empty unix
853 * domain socket. This should never block.
855 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
860 TALLOC_FREE(state->child);
862 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
863 tevent_req_error(req, err);
864 return tevent_req_post(req, ev);
867 subreq = read_packet_send(state, ev, state->child->sockfd,
868 sizeof(struct rw_ret), NULL, NULL);
869 if (tevent_req_nomem(subreq, req)) {
870 TALLOC_FREE(state->child); /* we sent sth down */
871 return tevent_req_post(req, ev);
873 tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
877 static void aio_fork_fsync_done(struct tevent_req *subreq)
879 struct tevent_req *req = tevent_req_callback_data(
880 subreq, struct tevent_req);
881 struct aio_fork_fsync_state *state = tevent_req_data(
882 req, struct aio_fork_fsync_state);
886 struct rw_ret *retbuf;
888 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
891 TALLOC_FREE(state->child);
892 tevent_req_error(req, err);
896 state->child->busy = false;
898 retbuf = (struct rw_ret *)buf;
899 state->ret = retbuf->size;
900 state->err = retbuf->ret_errno;
901 tevent_req_done(req);
904 static int aio_fork_fsync_recv(struct tevent_req *req, int *err)
906 struct aio_fork_fsync_state *state = tevent_req_data(
907 req, struct aio_fork_fsync_state);
909 if (tevent_req_is_unix_error(req, err)) {
912 if (state->ret == -1) {
918 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
922 struct aio_fork_config *config;
923 ret = SMB_VFS_NEXT_CONNECT(handle, service, user);
929 config = talloc_zero(handle->conn, struct aio_fork_config);
931 SMB_VFS_NEXT_DISCONNECT(handle);
932 DEBUG(0, ("talloc_zero() failed\n"));
936 config->erratic_testing_mode = lp_parm_bool(SNUM(handle->conn), "vfs_aio_fork",
937 "erratic_testing_mode", false);
939 SMB_VFS_HANDLE_SET_DATA(handle, config,
940 NULL, struct aio_fork_config,
943 /*********************************************************************
944 * How many threads to initialize ?
945 * 100 per process seems insane as a default until you realize that
946 * (a) Threads terminate after 1 second when idle.
947 * (b) Throttling is done in SMB2 via the crediting algorithm.
948 * (c) SMB1 clients are limited to max_mux (50) outstanding
949 * requests and Windows clients don't use this anyway.
950 * Essentially we want this to be unlimited unless smb.conf
952 *********************************************************************/
953 aio_pending_size = 100;
957 static struct vfs_fn_pointers vfs_aio_fork_fns = {
958 .connect_fn = aio_fork_connect,
959 .pread_send_fn = aio_fork_pread_send,
960 .pread_recv_fn = aio_fork_pread_recv,
961 .pwrite_send_fn = aio_fork_pwrite_send,
962 .pwrite_recv_fn = aio_fork_pwrite_recv,
963 .fsync_send_fn = aio_fork_fsync_send,
964 .fsync_recv_fn = aio_fork_fsync_recv,
967 NTSTATUS vfs_aio_fork_init(void);
968 NTSTATUS vfs_aio_fork_init(void)
970 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
971 "aio_fork", &vfs_aio_fork_fns);