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"
37 static int mmap_area_destructor(struct mmap_area *area)
39 munmap((void *)area->ptr, area->size);
43 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
45 struct mmap_area *result;
48 result = talloc(mem_ctx, struct mmap_area);
50 DEBUG(0, ("talloc failed\n"));
54 fd = open("/dev/zero", O_RDWR);
56 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
61 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
62 MAP_SHARED|MAP_FILE, fd, 0);
63 if (result->ptr == MAP_FAILED) {
64 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
71 talloc_set_destructor(result, mmap_area_destructor);
91 struct aio_child_list;
94 struct aio_child *prev, *next;
95 struct aio_child_list *list;
96 SMB_STRUCT_AIOCB *aiocb;
99 struct fd_event *sock_event;
100 struct rw_ret retval;
101 struct mmap_area *map; /* ==NULL means write request */
102 bool dont_delete; /* Marked as in use since last cleanup */
105 bool called_from_suspend;
106 bool completion_done;
109 struct aio_child_list {
110 struct aio_child *children;
111 struct timed_event *cleanup_event;
114 static void free_aio_children(void **p)
119 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
124 #ifndef HAVE_MSGHDR_MSG_CONTROL
128 #ifdef HAVE_MSGHDR_MSG_CONTROL
131 char control[CMSG_SPACE(sizeof(int))];
133 struct cmsghdr *cmptr;
135 msg.msg_control = control_un.control;
136 msg.msg_controllen = sizeof(control_un.control);
138 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
139 msg.msg_accrights = (caddr_t) &newfd;
140 msg.msg_accrightslen = sizeof(int);
142 #error Can not pass file descriptors
149 iov[0].iov_base = (void *)ptr;
150 iov[0].iov_len = nbytes;
154 if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
158 #ifdef HAVE_MSGHDR_MSG_CONTROL
159 if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
160 && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
161 if (cmptr->cmsg_level != SOL_SOCKET) {
162 DEBUG(10, ("control level != SOL_SOCKET"));
166 if (cmptr->cmsg_type != SCM_RIGHTS) {
167 DEBUG(10, ("control type != SCM_RIGHTS"));
171 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
173 *recvfd = -1; /* descriptor was not passed */
176 if (msg.msg_accrightslen == sizeof(int)) {
180 *recvfd = -1; /* descriptor was not passed */
187 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
192 #ifdef HAVE_MSGHDR_MSG_CONTROL
195 char control[CMSG_SPACE(sizeof(int))];
197 struct cmsghdr *cmptr;
200 ZERO_STRUCT(control_un);
202 msg.msg_control = control_un.control;
203 msg.msg_controllen = sizeof(control_un.control);
205 cmptr = CMSG_FIRSTHDR(&msg);
206 cmptr->cmsg_len = CMSG_LEN(sizeof(int));
207 cmptr->cmsg_level = SOL_SOCKET;
208 cmptr->cmsg_type = SCM_RIGHTS;
209 memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
212 msg.msg_accrights = (caddr_t) &sendfd;
213 msg.msg_accrightslen = sizeof(int);
220 iov[0].iov_base = (void *)ptr;
221 iov[0].iov_len = nbytes;
225 return (sendmsg(fd, &msg, 0));
228 static void aio_child_cleanup(struct event_context *event_ctx,
229 struct timed_event *te,
233 struct aio_child_list *list = talloc_get_type_abort(
234 private_data, struct aio_child_list);
235 struct aio_child *child, *next;
237 TALLOC_FREE(list->cleanup_event);
239 for (child = list->children; child != NULL; child = next) {
242 if (child->aiocb != NULL) {
243 DEBUG(10, ("child %d currently active\n",
248 if (child->dont_delete) {
249 DEBUG(10, ("Child %d was active since last cleanup\n",
251 child->dont_delete = false;
255 DEBUG(10, ("Child %d idle for more than 30 seconds, "
256 "deleting\n", (int)child->pid));
262 if (list->children != NULL) {
264 * Re-schedule the next cleanup round
266 list->cleanup_event = event_add_timed(server_event_context(), list,
267 timeval_add(&now, 30, 0),
268 aio_child_cleanup, list);
273 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
275 struct aio_child_list *data = NULL;
277 if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
278 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
283 data = talloc_zero(NULL, struct aio_child_list);
290 * Regardless of whether the child_list had been around or not, make
291 * sure that we have a cleanup timed event. This timed event will
292 * delete itself when it finds that no children are around anymore.
295 if (data->cleanup_event == NULL) {
296 data->cleanup_event = event_add_timed(server_event_context(), data,
297 timeval_current_ofs(30, 0),
298 aio_child_cleanup, data);
299 if (data->cleanup_event == NULL) {
305 if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
306 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
307 struct aio_child_list, return False);
313 static void aio_child_loop(int sockfd, struct mmap_area *map)
318 struct rw_cmd cmd_struct;
319 struct rw_ret ret_struct;
321 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
322 if (ret != sizeof(cmd_struct)) {
323 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
328 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
329 cmd_struct.read_cmd ? "read" : "write",
330 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
332 #ifdef ENABLE_BUILD_FARM_HACKS
335 * In the build farm, we want erratic behaviour for
341 * use generate_random_buffer, we just forked from a
342 * common parent state
344 generate_random_buffer(&randval, sizeof(randval));
345 msecs = randval + 20;
346 DEBUG(10, ("delaying for %u msecs\n", msecs));
352 ZERO_STRUCT(ret_struct);
354 if (cmd_struct.read_cmd) {
355 ret_struct.size = sys_pread(
356 fd, (void *)map->ptr, cmd_struct.n,
359 /* This breaks "make test" when run with aio_fork module. */
360 #ifdef ENABLE_BUILD_FARM_HACKS
361 ret_struct.size = MAX(1, ret_struct.size * 0.9);
366 ret_struct.size = sys_pwrite(
367 fd, (void *)map->ptr, cmd_struct.n,
371 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
372 (int)ret_struct.size));
374 if (ret_struct.size == -1) {
375 ret_struct.ret_errno = errno;
379 * Close the fd before telling our parent we're done. The
380 * parent might close and re-open the file very quickly, and
381 * with system-level share modes (GPFS) we would get an
382 * unjustified SHARING_VIOLATION.
386 ret = write_data(sockfd, (char *)&ret_struct,
388 if (ret != sizeof(ret_struct)) {
389 DEBUG(10, ("could not write ret_struct: %s\n",
396 static void handle_aio_completion(struct event_context *event_ctx,
397 struct fd_event *event, uint16 flags,
400 struct aio_extra *aio_ex = NULL;
401 struct aio_child *child = (struct aio_child *)p;
404 DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
406 if ((flags & EVENT_FD_READ) == 0) {
410 status = read_data(child->sockfd, (char *)&child->retval,
411 sizeof(child->retval));
413 if (!NT_STATUS_IS_OK(status)) {
414 DEBUG(1, ("aio child %d died: %s\n", (int)child->pid,
416 child->retval.size = -1;
417 child->retval.ret_errno = EIO;
420 if (child->aiocb == NULL) {
421 DEBUG(1, ("Inactive child died\n"));
426 if (child->cancelled) {
428 child->cancelled = false;
432 if (child->read_cmd && (child->retval.size > 0)) {
433 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
434 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
438 if (child->called_from_suspend) {
439 child->completion_done = true;
442 aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
443 smbd_aio_complete_aio_ex(aio_ex);
447 static int aio_child_destructor(struct aio_child *child)
451 SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
453 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
454 child->pid, child->sockfd));
457 * closing the sockfd makes the child not return from recvmsg() on RHEL
458 * 5.5 so instead force the child to exit by writing bad data to it
460 write(child->sockfd, &c, sizeof(c));
461 close(child->sockfd);
462 DLIST_REMOVE(child->list->children, child);
467 * We have to close all fd's in open files, we might incorrectly hold a system
468 * level share mode on a file.
471 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
474 if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
481 static NTSTATUS create_aio_child(struct smbd_server_connection *sconn,
482 struct aio_child_list *children,
484 struct aio_child **presult)
486 struct aio_child *result;
490 fdpair[0] = fdpair[1] = -1;
492 result = talloc_zero(children, struct aio_child);
493 NT_STATUS_HAVE_NO_MEMORY(result);
495 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
496 status = map_nt_error_from_unix(errno);
497 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
501 DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
503 result->map = mmap_area_init(result, map_size);
504 if (result->map == NULL) {
505 status = map_nt_error_from_unix(errno);
506 DEBUG(0, ("Could not create mmap area\n"));
510 result->pid = fork();
511 if (result->pid == -1) {
512 status = map_nt_error_from_unix(errno);
513 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
517 if (result->pid == 0) {
519 result->sockfd = fdpair[1];
520 files_forall(sconn, close_fsp_fd, NULL);
521 aio_child_loop(result->sockfd, result->map);
524 DEBUG(10, ("Child %d created with sockfd %d\n",
525 result->pid, fdpair[0]));
527 result->sockfd = fdpair[0];
530 result->sock_event = event_add_fd(server_event_context(), result,
531 result->sockfd, EVENT_FD_READ,
532 handle_aio_completion,
534 if (result->sock_event == NULL) {
535 status = NT_STATUS_NO_MEMORY;
536 DEBUG(0, ("event_add_fd failed\n"));
540 result->list = children;
541 DLIST_ADD(children->children, result);
543 talloc_set_destructor(result, aio_child_destructor);
550 if (fdpair[0] != -1) close(fdpair[0]);
551 if (fdpair[1] != -1) close(fdpair[1]);
557 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
558 struct aio_child **pchild)
560 struct aio_child_list *children;
561 struct aio_child *child;
564 children = init_aio_children(handle);
565 if (children == NULL) {
566 return NT_STATUS_NO_MEMORY;
569 for (child = children->children; child != NULL; child = child->next) {
570 if (child->aiocb == NULL) {
577 DEBUG(10, ("no idle child found, creating new one\n"));
579 status = create_aio_child(handle->conn->sconn, children,
581 if (!NT_STATUS_IS_OK(status)) {
582 DEBUG(10, ("create_aio_child failed: %s\n",
588 child->dont_delete = true;
594 static int aio_fork_read(struct vfs_handle_struct *handle,
595 struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
597 struct aio_child *child;
602 if (aiocb->aio_nbytes > 128*1024) {
603 /* TODO: support variable buffers */
608 status = get_idle_child(handle, &child);
609 if (!NT_STATUS_IS_OK(status)) {
610 DEBUG(10, ("Could not get an idle child\n"));
614 child->read_cmd = true;
615 child->aiocb = aiocb;
616 child->retval.ret_errno = EINPROGRESS;
619 cmd.n = aiocb->aio_nbytes;
620 cmd.offset = aiocb->aio_offset;
621 cmd.read_cmd = child->read_cmd;
623 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
626 ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
628 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
635 static int aio_fork_write(struct vfs_handle_struct *handle,
636 struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
638 struct aio_child *child;
643 if (aiocb->aio_nbytes > 128*1024) {
644 /* TODO: support variable buffers */
649 status = get_idle_child(handle, &child);
650 if (!NT_STATUS_IS_OK(status)) {
651 DEBUG(10, ("Could not get an idle child\n"));
655 child->read_cmd = false;
656 child->aiocb = aiocb;
657 child->retval.ret_errno = EINPROGRESS;
659 memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
663 cmd.n = aiocb->aio_nbytes;
664 cmd.offset = aiocb->aio_offset;
665 cmd.read_cmd = child->read_cmd;
667 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
670 ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
672 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
679 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
680 SMB_STRUCT_AIOCB *aiocb)
682 struct aio_child_list *children;
683 struct aio_child *child;
685 children = init_aio_children(handle);
686 if (children == NULL) {
690 for (child = children->children; child != NULL; child = child->next) {
691 if (child->aiocb == aiocb) {
699 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
700 struct files_struct *fsp,
701 SMB_STRUCT_AIOCB *aiocb)
703 struct aio_child *child = aio_fork_find_child(handle, aiocb);
707 DEBUG(0, ("returning EINVAL\n"));
713 if (child->cancelled) {
718 if (child->retval.size == -1) {
719 errno = child->retval.ret_errno;
722 return child->retval.size;
725 static int aio_fork_cancel(struct vfs_handle_struct *handle,
726 struct files_struct *fsp,
727 SMB_STRUCT_AIOCB *aiocb)
729 struct aio_child_list *children;
730 struct aio_child *child;
732 children = init_aio_children(handle);
733 if (children == NULL) {
738 for (child = children->children; child != NULL; child = child->next) {
739 if (child->aiocb == NULL) {
742 if (child->aiocb->aio_fildes != fsp->fh->fd) {
745 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
750 * We let the child do its job, but we discard the result when
754 child->cancelled = true;
760 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
761 struct files_struct *fsp,
762 SMB_STRUCT_AIOCB *aiocb)
764 struct aio_child *child = aio_fork_find_child(handle, aiocb);
771 return child->retval.ret_errno;
774 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
775 struct tevent_timer *te,
779 bool *timed_out = (bool *)private_data;
780 /* Remove this timed event handler. */
785 static int aio_fork_suspend(struct vfs_handle_struct *handle,
786 struct files_struct *fsp,
787 const SMB_STRUCT_AIOCB * const aiocb_array[],
789 const struct timespec *timeout)
791 struct aio_child_list *children = NULL;
792 TALLOC_CTX *frame = talloc_stackframe();
793 struct event_context *ev = NULL;
796 bool timed_out = false;
799 children = init_aio_children(handle);
800 if (children == NULL) {
805 /* This is a blocking call, and has to use a sub-event loop. */
806 ev = event_context_init(frame);
813 struct timeval tv = convert_timespec_to_timeval(*timeout);
814 struct tevent_timer *te = tevent_add_timer(ev,
816 timeval_current_ofs(tv.tv_sec,
818 aio_fork_suspend_timed_out,
826 for (i = 0; i < n; i++) {
827 struct aio_child *child = NULL;
828 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
835 * We're going to cheat here. We know that smbd/aio.c
836 * only calls this when it's waiting for every single
837 * outstanding call to finish on a close, so just wait
838 * individually for each IO to complete. We don't care
839 * what order they finish - only that they all do. JRA.
842 for (child = children->children; child != NULL; child = child->next) {
843 struct tevent_fd *event;
845 if (child->aiocb == NULL) {
848 if (child->aiocb->aio_fildes != fsp->fh->fd) {
851 if (child->aiocb != aiocb) {
855 if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
859 event = event_add_fd(ev,
863 handle_aio_completion,
870 child->called_from_suspend = true;
872 while (!child->completion_done) {
873 if (tevent_loop_once(ev) == -1) {
895 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
898 /*********************************************************************
899 * How many threads to initialize ?
900 * 100 per process seems insane as a default until you realize that
901 * (a) Threads terminate after 1 second when idle.
902 * (b) Throttling is done in SMB2 via the crediting algorithm.
903 * (c) SMB1 clients are limited to max_mux (50) outstanding
904 * requests and Windows clients don't use this anyway.
905 * Essentially we want this to be unlimited unless smb.conf
907 *********************************************************************/
908 aio_pending_size = 100;
909 return SMB_VFS_NEXT_CONNECT(handle, service, user);
912 static struct vfs_fn_pointers vfs_aio_fork_fns = {
913 .connect_fn = aio_fork_connect,
914 .aio_read_fn = aio_fork_read,
915 .aio_write_fn = aio_fork_write,
916 .aio_return_fn = aio_fork_return_fn,
917 .aio_cancel_fn = aio_fork_cancel,
918 .aio_error_fn = aio_fork_error_fn,
919 .aio_suspend_fn = aio_fork_suspend,
922 NTSTATUS vfs_aio_fork_init(void);
923 NTSTATUS vfs_aio_fork_init(void)
925 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
926 "aio_fork", &vfs_aio_fork_fns);