2 * Simulate the Posix AIO using mmap/fork
4 * Copyright (C) Volker Lendecke 2008
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 static int mmap_area_destructor(struct mmap_area *area)
30 munmap((void *)area->ptr, area->size);
34 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
36 struct mmap_area *result;
39 result = talloc(mem_ctx, struct mmap_area);
41 DEBUG(0, ("talloc failed\n"));
45 fd = open("/dev/zero", O_RDWR);
47 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
52 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
53 MAP_SHARED|MAP_FILE, fd, 0);
54 if (result->ptr == MAP_FAILED) {
55 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
62 talloc_set_destructor(result, mmap_area_destructor);
82 struct aio_child_list;
85 struct aio_child *prev, *next;
86 struct aio_child_list *list;
87 SMB_STRUCT_AIOCB *aiocb;
90 struct fd_event *sock_event;
92 struct mmap_area *map; /* ==NULL means write request */
93 bool dont_delete; /* Marked as in use since last cleanup */
98 struct aio_child_list {
99 struct aio_child *children;
100 struct timed_event *cleanup_event;
103 static void free_aio_children(void **p)
108 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
113 #ifndef HAVE_MSGHDR_MSG_CONTROL
117 #ifdef HAVE_MSGHDR_MSG_CONTROL
120 char control[CMSG_SPACE(sizeof(int))];
122 struct cmsghdr *cmptr;
124 msg.msg_control = control_un.control;
125 msg.msg_controllen = sizeof(control_un.control);
127 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
128 msg.msg_accrights = (caddr_t) &newfd;
129 msg.msg_accrightslen = sizeof(int);
131 #error Can not pass file descriptors
138 iov[0].iov_base = (void *)ptr;
139 iov[0].iov_len = nbytes;
143 if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
147 #ifdef HAVE_MSGHDR_MSG_CONTROL
148 if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
149 && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
150 if (cmptr->cmsg_level != SOL_SOCKET) {
151 DEBUG(10, ("control level != SOL_SOCKET"));
155 if (cmptr->cmsg_type != SCM_RIGHTS) {
156 DEBUG(10, ("control type != SCM_RIGHTS"));
160 *recvfd = *((int *) CMSG_DATA(cmptr));
162 *recvfd = -1; /* descriptor was not passed */
165 if (msg.msg_accrightslen == sizeof(int)) {
169 *recvfd = -1; /* descriptor was not passed */
176 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
181 #ifdef HAVE_MSGHDR_MSG_CONTROL
184 char control[CMSG_SPACE(sizeof(int))];
186 struct cmsghdr *cmptr;
189 ZERO_STRUCT(control_un);
191 msg.msg_control = control_un.control;
192 msg.msg_controllen = sizeof(control_un.control);
194 cmptr = CMSG_FIRSTHDR(&msg);
195 cmptr->cmsg_len = CMSG_LEN(sizeof(int));
196 cmptr->cmsg_level = SOL_SOCKET;
197 cmptr->cmsg_type = SCM_RIGHTS;
198 *((int *) CMSG_DATA(cmptr)) = sendfd;
201 msg.msg_accrights = (caddr_t) &sendfd;
202 msg.msg_accrightslen = sizeof(int);
209 iov[0].iov_base = (void *)ptr;
210 iov[0].iov_len = nbytes;
214 return (sendmsg(fd, &msg, 0));
217 static void aio_child_cleanup(struct event_context *event_ctx,
218 struct timed_event *te,
222 struct aio_child_list *list = talloc_get_type_abort(
223 private_data, struct aio_child_list);
224 struct aio_child *child, *next;
226 TALLOC_FREE(list->cleanup_event);
228 for (child = list->children; child != NULL; child = next) {
231 if (child->aiocb != NULL) {
232 DEBUG(10, ("child %d currently active\n",
237 if (child->dont_delete) {
238 DEBUG(10, ("Child %d was active since last cleanup\n",
240 child->dont_delete = false;
244 DEBUG(10, ("Child %d idle for more than 30 seconds, "
245 "deleting\n", (int)child->pid));
250 if (list->children != NULL) {
252 * Re-schedule the next cleanup round
254 list->cleanup_event = event_add_timed(smbd_event_context(), list,
255 timeval_add(&now, 30, 0),
256 aio_child_cleanup, list);
261 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
263 struct aio_child_list *data = NULL;
265 if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
266 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
271 data = TALLOC_ZERO_P(NULL, struct aio_child_list);
278 * Regardless of whether the child_list had been around or not, make
279 * sure that we have a cleanup timed event. This timed event will
280 * delete itself when it finds that no children are around anymore.
283 if (data->cleanup_event == NULL) {
284 data->cleanup_event = event_add_timed(smbd_event_context(), data,
285 timeval_current_ofs(30, 0),
286 aio_child_cleanup, data);
287 if (data->cleanup_event == NULL) {
293 if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
294 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
295 struct aio_child_list, return False);
301 static void aio_child_loop(int sockfd, struct mmap_area *map)
306 struct rw_cmd cmd_struct;
307 struct rw_ret ret_struct;
309 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
310 if (ret != sizeof(cmd_struct)) {
311 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
316 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
317 cmd_struct.read_cmd ? "read" : "write",
318 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
320 #ifdef ENABLE_BUILD_FARM_HACKS
323 * In the build farm, we want erratic behaviour for
329 * use generate_random_buffer, we just forked from a
330 * common parent state
332 generate_random_buffer(&randval, sizeof(randval));
333 msecs = randval + 20;
334 DEBUG(10, ("delaying for %u msecs\n", msecs));
340 ZERO_STRUCT(ret_struct);
342 if (cmd_struct.read_cmd) {
343 ret_struct.size = sys_pread(
344 fd, (void *)map->ptr, cmd_struct.n,
348 ret_struct.size = sys_pwrite(
349 fd, (void *)map->ptr, cmd_struct.n,
353 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
354 (int)ret_struct.size));
356 if (ret_struct.size == -1) {
357 ret_struct.ret_errno = errno;
361 * Close the fd before telling our parent we're done. The
362 * parent might close and re-open the file very quickly, and
363 * with system-level share modes (GPFS) we would get an
364 * unjustified SHARING_VIOLATION.
368 ret = write_data(sockfd, (char *)&ret_struct,
370 if (ret != sizeof(ret_struct)) {
371 DEBUG(10, ("could not write ret_struct: %s\n",
378 static void handle_aio_completion(struct event_context *event_ctx,
379 struct fd_event *event, uint16 flags,
382 struct aio_child *child = (struct aio_child *)p;
385 DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
387 if ((flags & EVENT_FD_READ) == 0) {
391 if (!NT_STATUS_IS_OK(read_data(child->sockfd,
392 (char *)&child->retval,
393 sizeof(child->retval)))) {
394 DEBUG(0, ("aio child %d died\n", (int)child->pid));
395 child->retval.size = -1;
396 child->retval.ret_errno = EIO;
399 if (child->cancelled) {
401 child->cancelled = false;
405 if (child->read_cmd && (child->retval.size > 0)) {
406 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
407 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
411 mid = child->aiocb->aio_sigevent.sigev_value.sival_int;
413 DEBUG(10, ("mid %d finished\n", (int)mid));
415 smbd_aio_complete_mid(mid);
418 static int aio_child_destructor(struct aio_child *child)
420 SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
421 close(child->sockfd);
422 DLIST_REMOVE(child->list->children, child);
427 * We have to close all fd's in open files, we might incorrectly hold a system
428 * level share mode on a file.
431 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
434 if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
441 static NTSTATUS create_aio_child(struct aio_child_list *children,
443 struct aio_child **presult)
445 struct aio_child *result;
449 fdpair[0] = fdpair[1] = -1;
451 result = TALLOC_ZERO_P(children, struct aio_child);
452 NT_STATUS_HAVE_NO_MEMORY(result);
454 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
455 status = map_nt_error_from_unix(errno);
456 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
460 DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
462 result->map = mmap_area_init(result, map_size);
463 if (result->map == NULL) {
464 status = map_nt_error_from_unix(errno);
465 DEBUG(0, ("Could not create mmap area\n"));
469 result->pid = sys_fork();
470 if (result->pid == -1) {
471 status = map_nt_error_from_unix(errno);
472 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
476 if (result->pid == 0) {
478 result->sockfd = fdpair[1];
479 file_walk_table(close_fsp_fd, NULL);
480 aio_child_loop(result->sockfd, result->map);
483 DEBUG(10, ("Child %d created\n", result->pid));
485 result->sockfd = fdpair[0];
488 result->sock_event = event_add_fd(smbd_event_context(), result,
489 result->sockfd, EVENT_FD_READ,
490 handle_aio_completion,
492 if (result->sock_event == NULL) {
493 status = NT_STATUS_NO_MEMORY;
494 DEBUG(0, ("event_add_fd failed\n"));
498 result->list = children;
499 DLIST_ADD(children->children, result);
501 talloc_set_destructor(result, aio_child_destructor);
508 if (fdpair[0] != -1) close(fdpair[0]);
509 if (fdpair[1] != -1) close(fdpair[1]);
515 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
516 struct aio_child **pchild)
518 struct aio_child_list *children;
519 struct aio_child *child;
522 children = init_aio_children(handle);
523 if (children == NULL) {
524 return NT_STATUS_NO_MEMORY;
527 for (child = children->children; child != NULL; child = child->next) {
528 if (child->aiocb == NULL) {
535 DEBUG(10, ("no idle child found, creating new one\n"));
537 status = create_aio_child(children, 128*1024, &child);
538 if (!NT_STATUS_IS_OK(status)) {
539 DEBUG(10, ("create_aio_child failed: %s\n",
545 child->dont_delete = true;
551 static int aio_fork_read(struct vfs_handle_struct *handle,
552 struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
554 struct aio_child *child;
559 if (aiocb->aio_nbytes > 128*1024) {
560 /* TODO: support variable buffers */
565 status = get_idle_child(handle, &child);
566 if (!NT_STATUS_IS_OK(status)) {
567 DEBUG(10, ("Could not get an idle child\n"));
571 child->read_cmd = true;
572 child->aiocb = aiocb;
573 child->retval.ret_errno = EINPROGRESS;
576 cmd.n = aiocb->aio_nbytes;
577 cmd.offset = aiocb->aio_offset;
578 cmd.read_cmd = child->read_cmd;
580 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
583 ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
585 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
592 static int aio_fork_write(struct vfs_handle_struct *handle,
593 struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
595 struct aio_child *child;
600 if (aiocb->aio_nbytes > 128*1024) {
601 /* TODO: support variable buffers */
606 status = get_idle_child(handle, &child);
607 if (!NT_STATUS_IS_OK(status)) {
608 DEBUG(10, ("Could not get an idle child\n"));
612 child->read_cmd = false;
613 child->aiocb = aiocb;
614 child->retval.ret_errno = EINPROGRESS;
616 memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
620 cmd.n = aiocb->aio_nbytes;
621 cmd.offset = aiocb->aio_offset;
622 cmd.read_cmd = child->read_cmd;
624 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
627 ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
629 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
636 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
637 SMB_STRUCT_AIOCB *aiocb)
639 struct aio_child_list *children;
640 struct aio_child *child;
642 children = init_aio_children(handle);
643 if (children == NULL) {
647 for (child = children->children; child != NULL; child = child->next) {
648 if (child->aiocb == aiocb) {
656 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
657 struct files_struct *fsp,
658 SMB_STRUCT_AIOCB *aiocb)
660 struct aio_child *child = aio_fork_find_child(handle, aiocb);
664 DEBUG(0, ("returning EINVAL\n"));
670 if (child->retval.size == -1) {
671 errno = child->retval.ret_errno;
674 return child->retval.size;
677 static int aio_fork_cancel(struct vfs_handle_struct *handle,
678 struct files_struct *fsp,
679 SMB_STRUCT_AIOCB *aiocb)
681 struct aio_child_list *children;
682 struct aio_child *child;
684 children = init_aio_children(handle);
685 if (children == NULL) {
690 for (child = children->children; child != NULL; child = child->next) {
691 if (child->aiocb == NULL) {
694 if (child->aiocb->aio_fildes != fsp->fh->fd) {
697 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
702 * We let the child do its job, but we discard the result when
706 child->cancelled = true;
712 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
713 struct files_struct *fsp,
714 SMB_STRUCT_AIOCB *aiocb)
716 struct aio_child *child = aio_fork_find_child(handle, aiocb);
723 return child->retval.ret_errno;
726 /* VFS operations structure */
728 static vfs_op_tuple aio_fork_ops[] = {
729 {SMB_VFS_OP(aio_fork_read), SMB_VFS_OP_AIO_READ,
730 SMB_VFS_LAYER_TRANSPARENT},
731 {SMB_VFS_OP(aio_fork_write), SMB_VFS_OP_AIO_WRITE,
732 SMB_VFS_LAYER_TRANSPARENT},
733 {SMB_VFS_OP(aio_fork_return_fn), SMB_VFS_OP_AIO_RETURN,
734 SMB_VFS_LAYER_TRANSPARENT},
735 {SMB_VFS_OP(aio_fork_cancel), SMB_VFS_OP_AIO_CANCEL,
736 SMB_VFS_LAYER_TRANSPARENT},
737 {SMB_VFS_OP(aio_fork_error_fn), SMB_VFS_OP_AIO_ERROR,
738 SMB_VFS_LAYER_TRANSPARENT},
739 {SMB_VFS_OP(NULL), SMB_VFS_OP_NOOP,
743 NTSTATUS vfs_aio_fork_init(void);
744 NTSTATUS vfs_aio_fork_init(void)
746 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
747 "aio_fork", aio_fork_ops);