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"
30 #if !defined(HAVE_STRUCT_MSGHDR_MSG_CONTROL) && !defined(HAVE_STRUCT_MSGHDR_MSG_ACCRIGHTS)
31 # error Can not pass file descriptors
40 struct aio_fork_config {
41 bool erratic_testing_mode;
49 static int mmap_area_destructor(struct mmap_area *area)
51 munmap((void *)area->ptr, area->size);
55 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
57 struct mmap_area *result;
60 result = talloc(mem_ctx, struct mmap_area);
62 DEBUG(0, ("talloc failed\n"));
66 fd = open("/dev/zero", O_RDWR);
68 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
73 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
74 MAP_SHARED|MAP_FILE, fd, 0);
76 if (result->ptr == MAP_FAILED) {
77 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
82 talloc_set_destructor(result, mmap_area_destructor);
97 static const char *cmd_type_str(enum cmd_type cmd)
112 result = "<UNKNOWN>";
122 bool erratic_testing_mode;
130 struct aio_child_list;
133 struct aio_child *prev, *next;
134 struct aio_child_list *list;
137 struct mmap_area *map;
138 bool dont_delete; /* Marked as in use since last cleanup */
142 struct aio_child_list {
143 struct aio_child *children;
144 struct tevent_timer *cleanup_event;
147 static void free_aio_children(void **p)
152 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
157 #ifndef HAVE_STRUCT_MSGHDR_MSG_CONTROL
160 msg.msg_accrights = (caddr_t) &newfd;
161 msg.msg_accrightslen = sizeof(int);
166 char control[CMSG_SPACE(sizeof(int))];
168 struct cmsghdr *cmptr;
170 msg.msg_control = control_un.control;
171 msg.msg_controllen = sizeof(control_un.control);
178 iov[0].iov_base = (void *)ptr;
179 iov[0].iov_len = nbytes;
183 if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
187 #ifdef HAVE_STRUCT_MSGHDR_MSG_CONTROL
188 if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
189 && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
190 if (cmptr->cmsg_level != SOL_SOCKET) {
191 DEBUG(10, ("control level != SOL_SOCKET"));
195 if (cmptr->cmsg_type != SCM_RIGHTS) {
196 DEBUG(10, ("control type != SCM_RIGHTS"));
200 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
202 *recvfd = -1; /* descriptor was not passed */
205 if (msg.msg_accrightslen == sizeof(int)) {
209 *recvfd = -1; /* descriptor was not passed */
216 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
221 #ifdef HAVE_STRUCT_MSGHDR_MSG_CONTROL
224 char control[CMSG_SPACE(sizeof(int))];
226 struct cmsghdr *cmptr;
229 ZERO_STRUCT(control_un);
231 msg.msg_control = control_un.control;
232 msg.msg_controllen = sizeof(control_un.control);
234 cmptr = CMSG_FIRSTHDR(&msg);
235 cmptr->cmsg_len = CMSG_LEN(sizeof(int));
236 cmptr->cmsg_level = SOL_SOCKET;
237 cmptr->cmsg_type = SCM_RIGHTS;
238 memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
241 msg.msg_accrights = (caddr_t) &sendfd;
242 msg.msg_accrightslen = sizeof(int);
249 iov[0].iov_base = (void *)ptr;
250 iov[0].iov_len = nbytes;
254 return (sendmsg(fd, &msg, 0));
257 static void aio_child_cleanup(struct tevent_context *event_ctx,
258 struct tevent_timer *te,
262 struct aio_child_list *list = talloc_get_type_abort(
263 private_data, struct aio_child_list);
264 struct aio_child *child, *next;
266 TALLOC_FREE(list->cleanup_event);
268 for (child = list->children; child != NULL; child = next) {
272 DEBUG(10, ("child %d currently active\n",
277 if (child->dont_delete) {
278 DEBUG(10, ("Child %d was active since last cleanup\n",
280 child->dont_delete = false;
284 DEBUG(10, ("Child %d idle for more than 30 seconds, "
285 "deleting\n", (int)child->pid));
291 if (list->children != NULL) {
293 * Re-schedule the next cleanup round
295 list->cleanup_event = tevent_add_timer(server_event_context(), list,
296 timeval_add(&now, 30, 0),
297 aio_child_cleanup, list);
302 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
304 struct aio_child_list *data = NULL;
306 if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
307 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
312 data = talloc_zero(NULL, struct aio_child_list);
319 * Regardless of whether the child_list had been around or not, make
320 * sure that we have a cleanup timed event. This timed event will
321 * delete itself when it finds that no children are around anymore.
324 if (data->cleanup_event == NULL) {
325 data->cleanup_event = tevent_add_timer(server_event_context(), data,
326 timeval_current_ofs(30, 0),
327 aio_child_cleanup, data);
328 if (data->cleanup_event == NULL) {
334 if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
335 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
336 struct aio_child_list, return False);
342 static void aio_child_loop(int sockfd, struct mmap_area *map)
347 struct rw_cmd cmd_struct;
348 struct rw_ret ret_struct;
350 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
351 if (ret != sizeof(cmd_struct)) {
352 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
357 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
358 cmd_type_str(cmd_struct.cmd),
359 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
361 if (cmd_struct.erratic_testing_mode) {
363 * For developer testing, we want erratic behaviour for
369 * use generate_random_buffer, we just forked from a
370 * common parent state
372 generate_random_buffer(&randval, sizeof(randval));
373 msecs = randval + 20;
374 DEBUG(10, ("delaying for %u msecs\n", msecs));
378 ZERO_STRUCT(ret_struct);
380 switch (cmd_struct.cmd) {
382 ret_struct.size = sys_pread(
383 fd, (void *)map->ptr, cmd_struct.n,
386 /* This breaks "make test" when run with aio_fork module. */
388 ret_struct.size = MAX(1, ret_struct.size * 0.9);
393 ret_struct.size = sys_pwrite(
394 fd, (void *)map->ptr, cmd_struct.n,
398 ret_struct.size = fsync(fd);
401 ret_struct.size = -1;
405 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
406 (int)ret_struct.size));
408 if (ret_struct.size == -1) {
409 ret_struct.ret_errno = errno;
413 * Close the fd before telling our parent we're done. The
414 * parent might close and re-open the file very quickly, and
415 * with system-level share modes (GPFS) we would get an
416 * unjustified SHARING_VIOLATION.
420 ret = write_data(sockfd, (char *)&ret_struct,
422 if (ret != sizeof(ret_struct)) {
423 DEBUG(10, ("could not write ret_struct: %s\n",
430 static int aio_child_destructor(struct aio_child *child)
434 SMB_ASSERT(!child->busy);
436 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
437 child->pid, child->sockfd));
440 * closing the sockfd makes the child not return from recvmsg() on RHEL
441 * 5.5 so instead force the child to exit by writing bad data to it
443 write(child->sockfd, &c, sizeof(c));
444 close(child->sockfd);
445 DLIST_REMOVE(child->list->children, child);
450 * We have to close all fd's in open files, we might incorrectly hold a system
451 * level share mode on a file.
454 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
457 if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
464 static int create_aio_child(struct smbd_server_connection *sconn,
465 struct aio_child_list *children,
467 struct aio_child **presult)
469 struct aio_child *result;
473 fdpair[0] = fdpair[1] = -1;
475 result = talloc_zero(children, struct aio_child);
476 if (result == NULL) {
480 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
482 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
486 DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
488 result->map = mmap_area_init(result, map_size);
489 if (result->map == NULL) {
491 DEBUG(0, ("Could not create mmap area\n"));
495 result->pid = fork();
496 if (result->pid == -1) {
498 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
502 if (result->pid == 0) {
504 result->sockfd = fdpair[1];
505 files_forall(sconn, close_fsp_fd, NULL);
506 aio_child_loop(result->sockfd, result->map);
509 DEBUG(10, ("Child %d created with sockfd %d\n",
510 result->pid, fdpair[0]));
512 result->sockfd = fdpair[0];
515 result->list = children;
516 DLIST_ADD(children->children, result);
518 talloc_set_destructor(result, aio_child_destructor);
525 if (fdpair[0] != -1) close(fdpair[0]);
526 if (fdpair[1] != -1) close(fdpair[1]);
532 static int get_idle_child(struct vfs_handle_struct *handle,
533 struct aio_child **pchild)
535 struct aio_child_list *children;
536 struct aio_child *child;
538 children = init_aio_children(handle);
539 if (children == NULL) {
543 for (child = children->children; child != NULL; child = child->next) {
552 DEBUG(10, ("no idle child found, creating new one\n"));
554 ret = create_aio_child(handle->conn->sconn, children,
557 DEBUG(10, ("create_aio_child failed: %s\n",
563 child->dont_delete = true;
570 struct aio_fork_pread_state {
571 struct aio_child *child;
576 static void aio_fork_pread_done(struct tevent_req *subreq);
578 static struct tevent_req *aio_fork_pread_send(struct vfs_handle_struct *handle,
580 struct tevent_context *ev,
581 struct files_struct *fsp,
583 size_t n, off_t offset)
585 struct tevent_req *req, *subreq;
586 struct aio_fork_pread_state *state;
590 struct aio_fork_config *config;
592 SMB_VFS_HANDLE_GET_DATA(handle, config,
593 struct aio_fork_config,
596 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
602 /* TODO: support variable buffers */
603 tevent_req_error(req, EINVAL);
604 return tevent_req_post(req, ev);
607 err = get_idle_child(handle, &state->child);
609 tevent_req_error(req, err);
610 return tevent_req_post(req, ev);
617 cmd.erratic_testing_mode = config->erratic_testing_mode;
619 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
620 (int)state->child->pid));
623 * Not making this async. We're writing into an empty unix
624 * domain socket. This should never block.
626 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
631 TALLOC_FREE(state->child);
633 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
634 tevent_req_error(req, err);
635 return tevent_req_post(req, ev);
638 subreq = read_packet_send(state, ev, state->child->sockfd,
639 sizeof(struct rw_ret), NULL, NULL);
640 if (tevent_req_nomem(subreq, req)) {
641 TALLOC_FREE(state->child); /* we sent sth down */
642 return tevent_req_post(req, ev);
644 tevent_req_set_callback(subreq, aio_fork_pread_done, req);
648 static void aio_fork_pread_done(struct tevent_req *subreq)
650 struct tevent_req *req = tevent_req_callback_data(
651 subreq, struct tevent_req);
652 struct aio_fork_pread_state *state = tevent_req_data(
653 req, struct aio_fork_pread_state);
657 struct rw_ret *retbuf;
659 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
662 TALLOC_FREE(state->child);
663 tevent_req_error(req, err);
667 state->child->busy = false;
669 retbuf = (struct rw_ret *)buf;
670 state->ret = retbuf->size;
671 state->err = retbuf->ret_errno;
672 tevent_req_done(req);
675 static ssize_t aio_fork_pread_recv(struct tevent_req *req, int *err)
677 struct aio_fork_pread_state *state = tevent_req_data(
678 req, struct aio_fork_pread_state);
680 if (tevent_req_is_unix_error(req, err)) {
683 if (state->ret == -1) {
689 struct aio_fork_pwrite_state {
690 struct aio_child *child;
695 static void aio_fork_pwrite_done(struct tevent_req *subreq);
697 static struct tevent_req *aio_fork_pwrite_send(
698 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
699 struct tevent_context *ev, struct files_struct *fsp,
700 const void *data, size_t n, off_t offset)
702 struct tevent_req *req, *subreq;
703 struct aio_fork_pwrite_state *state;
707 struct aio_fork_config *config;
708 SMB_VFS_HANDLE_GET_DATA(handle, config,
709 struct aio_fork_config,
712 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
718 /* TODO: support variable buffers */
719 tevent_req_error(req, EINVAL);
720 return tevent_req_post(req, ev);
723 err = get_idle_child(handle, &state->child);
725 tevent_req_error(req, err);
726 return tevent_req_post(req, ev);
733 cmd.erratic_testing_mode = config->erratic_testing_mode;
735 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
736 (int)state->child->pid));
739 * Not making this async. We're writing into an empty unix
740 * domain socket. This should never block.
742 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
747 TALLOC_FREE(state->child);
749 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
750 tevent_req_error(req, err);
751 return tevent_req_post(req, ev);
754 subreq = read_packet_send(state, ev, state->child->sockfd,
755 sizeof(struct rw_ret), NULL, NULL);
756 if (tevent_req_nomem(subreq, req)) {
757 TALLOC_FREE(state->child); /* we sent sth down */
758 return tevent_req_post(req, ev);
760 tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
764 static void aio_fork_pwrite_done(struct tevent_req *subreq)
766 struct tevent_req *req = tevent_req_callback_data(
767 subreq, struct tevent_req);
768 struct aio_fork_pwrite_state *state = tevent_req_data(
769 req, struct aio_fork_pwrite_state);
773 struct rw_ret *retbuf;
775 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
778 TALLOC_FREE(state->child);
779 tevent_req_error(req, err);
783 state->child->busy = false;
785 retbuf = (struct rw_ret *)buf;
786 state->ret = retbuf->size;
787 state->err = retbuf->ret_errno;
788 tevent_req_done(req);
791 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req, int *err)
793 struct aio_fork_pwrite_state *state = tevent_req_data(
794 req, struct aio_fork_pwrite_state);
796 if (tevent_req_is_unix_error(req, err)) {
799 if (state->ret == -1) {
805 struct aio_fork_fsync_state {
806 struct aio_child *child;
811 static void aio_fork_fsync_done(struct tevent_req *subreq);
813 static struct tevent_req *aio_fork_fsync_send(
814 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
815 struct tevent_context *ev, struct files_struct *fsp)
817 struct tevent_req *req, *subreq;
818 struct aio_fork_fsync_state *state;
822 struct aio_fork_config *config;
824 SMB_VFS_HANDLE_GET_DATA(handle, config,
825 struct aio_fork_config,
828 req = tevent_req_create(mem_ctx, &state, struct aio_fork_fsync_state);
833 err = get_idle_child(handle, &state->child);
835 tevent_req_error(req, err);
836 return tevent_req_post(req, ev);
841 cmd.erratic_testing_mode = config->erratic_testing_mode;
843 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
844 (int)state->child->pid));
847 * Not making this async. We're writing into an empty unix
848 * domain socket. This should never block.
850 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
855 TALLOC_FREE(state->child);
857 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
858 tevent_req_error(req, err);
859 return tevent_req_post(req, ev);
862 subreq = read_packet_send(state, ev, state->child->sockfd,
863 sizeof(struct rw_ret), NULL, NULL);
864 if (tevent_req_nomem(subreq, req)) {
865 TALLOC_FREE(state->child); /* we sent sth down */
866 return tevent_req_post(req, ev);
868 tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
872 static void aio_fork_fsync_done(struct tevent_req *subreq)
874 struct tevent_req *req = tevent_req_callback_data(
875 subreq, struct tevent_req);
876 struct aio_fork_fsync_state *state = tevent_req_data(
877 req, struct aio_fork_fsync_state);
881 struct rw_ret *retbuf;
883 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
886 TALLOC_FREE(state->child);
887 tevent_req_error(req, err);
891 state->child->busy = false;
893 retbuf = (struct rw_ret *)buf;
894 state->ret = retbuf->size;
895 state->err = retbuf->ret_errno;
896 tevent_req_done(req);
899 static int aio_fork_fsync_recv(struct tevent_req *req, int *err)
901 struct aio_fork_fsync_state *state = tevent_req_data(
902 req, struct aio_fork_fsync_state);
904 if (tevent_req_is_unix_error(req, err)) {
907 if (state->ret == -1) {
913 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
917 struct aio_fork_config *config;
918 ret = SMB_VFS_NEXT_CONNECT(handle, service, user);
924 config = talloc_zero(handle->conn, struct aio_fork_config);
926 SMB_VFS_NEXT_DISCONNECT(handle);
927 DEBUG(0, ("talloc_zero() failed\n"));
931 config->erratic_testing_mode = lp_parm_bool(SNUM(handle->conn), "vfs_aio_fork",
932 "erratic_testing_mode", false);
934 SMB_VFS_HANDLE_SET_DATA(handle, config,
935 NULL, struct aio_fork_config,
938 /*********************************************************************
939 * How many threads to initialize ?
940 * 100 per process seems insane as a default until you realize that
941 * (a) Threads terminate after 1 second when idle.
942 * (b) Throttling is done in SMB2 via the crediting algorithm.
943 * (c) SMB1 clients are limited to max_mux (50) outstanding
944 * requests and Windows clients don't use this anyway.
945 * Essentially we want this to be unlimited unless smb.conf
947 *********************************************************************/
948 aio_pending_size = 100;
952 static struct vfs_fn_pointers vfs_aio_fork_fns = {
953 .connect_fn = aio_fork_connect,
954 .pread_send_fn = aio_fork_pread_send,
955 .pread_recv_fn = aio_fork_pread_recv,
956 .pwrite_send_fn = aio_fork_pwrite_send,
957 .pwrite_recv_fn = aio_fork_pwrite_recv,
958 .fsync_send_fn = aio_fork_fsync_send,
959 .fsync_recv_fn = aio_fork_fsync_recv,
962 NTSTATUS vfs_aio_fork_init(void);
963 NTSTATUS vfs_aio_fork_init(void)
965 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
966 "aio_fork", &vfs_aio_fork_fns);