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"
41 static int mmap_area_destructor(struct mmap_area *area)
43 munmap((void *)area->ptr, area->size);
47 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
49 struct mmap_area *result;
52 result = talloc(mem_ctx, struct mmap_area);
54 DEBUG(0, ("talloc failed\n"));
58 fd = open("/dev/zero", O_RDWR);
60 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
65 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
66 MAP_SHARED|MAP_FILE, fd, 0);
67 if (result->ptr == MAP_FAILED) {
68 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
75 talloc_set_destructor(result, mmap_area_destructor);
90 static const char *cmd_type_str(enum cmd_type cmd)
105 result = "<UNKNOWN>";
122 struct aio_child_list;
125 struct aio_child *prev, *next;
126 struct aio_child_list *list;
129 struct mmap_area *map;
130 bool dont_delete; /* Marked as in use since last cleanup */
134 struct aio_child_list {
135 struct aio_child *children;
136 struct timed_event *cleanup_event;
139 static void free_aio_children(void **p)
144 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
149 #ifndef HAVE_MSGHDR_MSG_CONTROL
153 #ifdef HAVE_MSGHDR_MSG_CONTROL
156 char control[CMSG_SPACE(sizeof(int))];
158 struct cmsghdr *cmptr;
160 msg.msg_control = control_un.control;
161 msg.msg_controllen = sizeof(control_un.control);
163 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
164 msg.msg_accrights = (caddr_t) &newfd;
165 msg.msg_accrightslen = sizeof(int);
167 #error Can not pass file descriptors
175 iov[0].iov_base = (void *)ptr;
176 iov[0].iov_len = nbytes;
180 if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
184 #ifdef HAVE_MSGHDR_MSG_CONTROL
185 if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
186 && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
187 if (cmptr->cmsg_level != SOL_SOCKET) {
188 DEBUG(10, ("control level != SOL_SOCKET"));
192 if (cmptr->cmsg_type != SCM_RIGHTS) {
193 DEBUG(10, ("control type != SCM_RIGHTS"));
197 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
199 *recvfd = -1; /* descriptor was not passed */
202 if (msg.msg_accrightslen == sizeof(int)) {
206 *recvfd = -1; /* descriptor was not passed */
213 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
218 #ifdef HAVE_MSGHDR_MSG_CONTROL
221 char control[CMSG_SPACE(sizeof(int))];
223 struct cmsghdr *cmptr;
226 ZERO_STRUCT(control_un);
228 msg.msg_control = control_un.control;
229 msg.msg_controllen = sizeof(control_un.control);
231 cmptr = CMSG_FIRSTHDR(&msg);
232 cmptr->cmsg_len = CMSG_LEN(sizeof(int));
233 cmptr->cmsg_level = SOL_SOCKET;
234 cmptr->cmsg_type = SCM_RIGHTS;
235 memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
238 msg.msg_accrights = (caddr_t) &sendfd;
239 msg.msg_accrightslen = sizeof(int);
246 iov[0].iov_base = (void *)ptr;
247 iov[0].iov_len = nbytes;
251 return (sendmsg(fd, &msg, 0));
254 static void aio_child_cleanup(struct event_context *event_ctx,
255 struct timed_event *te,
259 struct aio_child_list *list = talloc_get_type_abort(
260 private_data, struct aio_child_list);
261 struct aio_child *child, *next;
263 TALLOC_FREE(list->cleanup_event);
265 for (child = list->children; child != NULL; child = next) {
269 DEBUG(10, ("child %d currently active\n",
274 if (child->dont_delete) {
275 DEBUG(10, ("Child %d was active since last cleanup\n",
277 child->dont_delete = false;
281 DEBUG(10, ("Child %d idle for more than 30 seconds, "
282 "deleting\n", (int)child->pid));
288 if (list->children != NULL) {
290 * Re-schedule the next cleanup round
292 list->cleanup_event = event_add_timed(server_event_context(), list,
293 timeval_add(&now, 30, 0),
294 aio_child_cleanup, list);
299 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
301 struct aio_child_list *data = NULL;
303 if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
304 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
309 data = talloc_zero(NULL, struct aio_child_list);
316 * Regardless of whether the child_list had been around or not, make
317 * sure that we have a cleanup timed event. This timed event will
318 * delete itself when it finds that no children are around anymore.
321 if (data->cleanup_event == NULL) {
322 data->cleanup_event = event_add_timed(server_event_context(), data,
323 timeval_current_ofs(30, 0),
324 aio_child_cleanup, data);
325 if (data->cleanup_event == NULL) {
331 if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
332 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
333 struct aio_child_list, return False);
339 static void aio_child_loop(int sockfd, struct mmap_area *map)
344 struct rw_cmd cmd_struct;
345 struct rw_ret ret_struct;
347 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
348 if (ret != sizeof(cmd_struct)) {
349 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
354 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
355 cmd_type_str(cmd_struct.cmd),
356 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
361 * For developer testing, we want erratic behaviour for
367 * use generate_random_buffer, we just forked from a
368 * common parent state
370 generate_random_buffer(&randval, sizeof(randval));
371 msecs = randval + 20;
372 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;
591 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
597 /* TODO: support variable buffers */
598 tevent_req_error(req, EINVAL);
599 return tevent_req_post(req, ev);
602 err = get_idle_child(handle, &state->child);
604 tevent_req_error(req, err);
605 return tevent_req_post(req, ev);
613 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
614 (int)state->child->pid));
617 * Not making this async. We're writing into an empty unix
618 * domain socket. This should never block.
620 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
625 TALLOC_FREE(state->child);
627 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
628 tevent_req_error(req, err);
629 return tevent_req_post(req, ev);
632 subreq = read_packet_send(state, ev, state->child->sockfd,
633 sizeof(struct rw_ret), NULL, NULL);
634 if (tevent_req_nomem(subreq, req)) {
635 TALLOC_FREE(state->child); /* we sent sth down */
636 return tevent_req_post(req, ev);
638 tevent_req_set_callback(subreq, aio_fork_pread_done, req);
642 static void aio_fork_pread_done(struct tevent_req *subreq)
644 struct tevent_req *req = tevent_req_callback_data(
645 subreq, struct tevent_req);
646 struct aio_fork_pread_state *state = tevent_req_data(
647 req, struct aio_fork_pread_state);
651 struct rw_ret *retbuf;
653 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
656 TALLOC_FREE(state->child);
657 tevent_req_error(req, err);
661 state->child->busy = false;
663 retbuf = (struct rw_ret *)buf;
664 state->ret = retbuf->size;
665 state->err = retbuf->ret_errno;
666 tevent_req_done(req);
669 static ssize_t aio_fork_pread_recv(struct tevent_req *req, int *err)
671 struct aio_fork_pread_state *state = tevent_req_data(
672 req, struct aio_fork_pread_state);
674 if (tevent_req_is_unix_error(req, err)) {
677 if (state->ret == -1) {
683 struct aio_fork_pwrite_state {
684 struct aio_child *child;
689 static void aio_fork_pwrite_done(struct tevent_req *subreq);
691 static struct tevent_req *aio_fork_pwrite_send(
692 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
693 struct tevent_context *ev, struct files_struct *fsp,
694 const void *data, size_t n, off_t offset)
696 struct tevent_req *req, *subreq;
697 struct aio_fork_pwrite_state *state;
702 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
708 /* TODO: support variable buffers */
709 tevent_req_error(req, EINVAL);
710 return tevent_req_post(req, ev);
713 err = get_idle_child(handle, &state->child);
715 tevent_req_error(req, err);
716 return tevent_req_post(req, ev);
724 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
725 (int)state->child->pid));
728 * Not making this async. We're writing into an empty unix
729 * domain socket. This should never block.
731 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
736 TALLOC_FREE(state->child);
738 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
739 tevent_req_error(req, err);
740 return tevent_req_post(req, ev);
743 subreq = read_packet_send(state, ev, state->child->sockfd,
744 sizeof(struct rw_ret), NULL, NULL);
745 if (tevent_req_nomem(subreq, req)) {
746 TALLOC_FREE(state->child); /* we sent sth down */
747 return tevent_req_post(req, ev);
749 tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
753 static void aio_fork_pwrite_done(struct tevent_req *subreq)
755 struct tevent_req *req = tevent_req_callback_data(
756 subreq, struct tevent_req);
757 struct aio_fork_pwrite_state *state = tevent_req_data(
758 req, struct aio_fork_pwrite_state);
762 struct rw_ret *retbuf;
764 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
767 TALLOC_FREE(state->child);
768 tevent_req_error(req, err);
772 state->child->busy = false;
774 retbuf = (struct rw_ret *)buf;
775 state->ret = retbuf->size;
776 state->err = retbuf->ret_errno;
777 tevent_req_done(req);
780 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req, int *err)
782 struct aio_fork_pwrite_state *state = tevent_req_data(
783 req, struct aio_fork_pwrite_state);
785 if (tevent_req_is_unix_error(req, err)) {
788 if (state->ret == -1) {
794 struct aio_fork_fsync_state {
795 struct aio_child *child;
800 static void aio_fork_fsync_done(struct tevent_req *subreq);
802 static struct tevent_req *aio_fork_fsync_send(
803 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
804 struct tevent_context *ev, struct files_struct *fsp)
806 struct tevent_req *req, *subreq;
807 struct aio_fork_fsync_state *state;
812 req = tevent_req_create(mem_ctx, &state, struct aio_fork_fsync_state);
817 err = get_idle_child(handle, &state->child);
819 tevent_req_error(req, err);
820 return tevent_req_post(req, ev);
826 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
827 (int)state->child->pid));
830 * Not making this async. We're writing into an empty unix
831 * domain socket. This should never block.
833 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
838 TALLOC_FREE(state->child);
840 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
841 tevent_req_error(req, err);
842 return tevent_req_post(req, ev);
845 subreq = read_packet_send(state, ev, state->child->sockfd,
846 sizeof(struct rw_ret), NULL, NULL);
847 if (tevent_req_nomem(subreq, req)) {
848 TALLOC_FREE(state->child); /* we sent sth down */
849 return tevent_req_post(req, ev);
851 tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
855 static void aio_fork_fsync_done(struct tevent_req *subreq)
857 struct tevent_req *req = tevent_req_callback_data(
858 subreq, struct tevent_req);
859 struct aio_fork_fsync_state *state = tevent_req_data(
860 req, struct aio_fork_fsync_state);
864 struct rw_ret *retbuf;
866 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
869 TALLOC_FREE(state->child);
870 tevent_req_error(req, err);
874 state->child->busy = false;
876 retbuf = (struct rw_ret *)buf;
877 state->ret = retbuf->size;
878 state->err = retbuf->ret_errno;
879 tevent_req_done(req);
882 static int aio_fork_fsync_recv(struct tevent_req *req, int *err)
884 struct aio_fork_fsync_state *state = tevent_req_data(
885 req, struct aio_fork_fsync_state);
887 if (tevent_req_is_unix_error(req, err)) {
890 if (state->ret == -1) {
896 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
899 /*********************************************************************
900 * How many threads to initialize ?
901 * 100 per process seems insane as a default until you realize that
902 * (a) Threads terminate after 1 second when idle.
903 * (b) Throttling is done in SMB2 via the crediting algorithm.
904 * (c) SMB1 clients are limited to max_mux (50) outstanding
905 * requests and Windows clients don't use this anyway.
906 * Essentially we want this to be unlimited unless smb.conf
908 *********************************************************************/
909 aio_pending_size = 100;
910 return SMB_VFS_NEXT_CONNECT(handle, service, user);
913 static struct vfs_fn_pointers vfs_aio_fork_fns = {
914 .connect_fn = aio_fork_connect,
915 .pread_send_fn = aio_fork_pread_send,
916 .pread_recv_fn = aio_fork_pread_recv,
917 .pwrite_send_fn = aio_fork_pwrite_send,
918 .pwrite_recv_fn = aio_fork_pwrite_recv,
919 .fsync_send_fn = aio_fork_fsync_send,
920 .fsync_recv_fn = aio_fork_fsync_recv,
923 NTSTATUS vfs_aio_fork_init(void);
924 NTSTATUS vfs_aio_fork_init(void)
926 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
927 "aio_fork", &vfs_aio_fork_fns);