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"
38 static int mmap_area_destructor(struct mmap_area *area)
40 munmap((void *)area->ptr, area->size);
44 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
46 struct mmap_area *result;
49 result = talloc(mem_ctx, struct mmap_area);
51 DEBUG(0, ("talloc failed\n"));
55 fd = open("/dev/zero", O_RDWR);
57 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
62 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
63 MAP_SHARED|MAP_FILE, fd, 0);
64 if (result->ptr == MAP_FAILED) {
65 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
72 talloc_set_destructor(result, mmap_area_destructor);
92 struct aio_child_list;
95 struct aio_child *prev, *next;
96 struct aio_child_list *list;
97 SMB_STRUCT_AIOCB *aiocb;
100 struct fd_event *sock_event;
101 struct rw_ret retval;
102 struct mmap_area *map; /* ==NULL means write request */
103 bool dont_delete; /* Marked as in use since last cleanup */
106 bool called_from_suspend;
107 bool completion_done;
110 struct aio_child_list {
111 struct aio_child *children;
112 struct timed_event *cleanup_event;
115 static void free_aio_children(void **p)
120 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
125 #ifndef HAVE_MSGHDR_MSG_CONTROL
129 #ifdef HAVE_MSGHDR_MSG_CONTROL
132 char control[CMSG_SPACE(sizeof(int))];
134 struct cmsghdr *cmptr;
136 msg.msg_control = control_un.control;
137 msg.msg_controllen = sizeof(control_un.control);
139 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
140 msg.msg_accrights = (caddr_t) &newfd;
141 msg.msg_accrightslen = sizeof(int);
143 #error Can not pass file descriptors
151 iov[0].iov_base = (void *)ptr;
152 iov[0].iov_len = nbytes;
156 if ( (n = recvmsg(fd, &msg, 0)) <= 0) {
160 #ifdef HAVE_MSGHDR_MSG_CONTROL
161 if ((cmptr = CMSG_FIRSTHDR(&msg)) != NULL
162 && cmptr->cmsg_len == CMSG_LEN(sizeof(int))) {
163 if (cmptr->cmsg_level != SOL_SOCKET) {
164 DEBUG(10, ("control level != SOL_SOCKET"));
168 if (cmptr->cmsg_type != SCM_RIGHTS) {
169 DEBUG(10, ("control type != SCM_RIGHTS"));
173 memcpy(recvfd, CMSG_DATA(cmptr), sizeof(*recvfd));
175 *recvfd = -1; /* descriptor was not passed */
178 if (msg.msg_accrightslen == sizeof(int)) {
182 *recvfd = -1; /* descriptor was not passed */
189 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
194 #ifdef HAVE_MSGHDR_MSG_CONTROL
197 char control[CMSG_SPACE(sizeof(int))];
199 struct cmsghdr *cmptr;
202 ZERO_STRUCT(control_un);
204 msg.msg_control = control_un.control;
205 msg.msg_controllen = sizeof(control_un.control);
207 cmptr = CMSG_FIRSTHDR(&msg);
208 cmptr->cmsg_len = CMSG_LEN(sizeof(int));
209 cmptr->cmsg_level = SOL_SOCKET;
210 cmptr->cmsg_type = SCM_RIGHTS;
211 memcpy(CMSG_DATA(cmptr), &sendfd, sizeof(sendfd));
214 msg.msg_accrights = (caddr_t) &sendfd;
215 msg.msg_accrightslen = sizeof(int);
222 iov[0].iov_base = (void *)ptr;
223 iov[0].iov_len = nbytes;
227 return (sendmsg(fd, &msg, 0));
230 static void aio_child_cleanup(struct event_context *event_ctx,
231 struct timed_event *te,
235 struct aio_child_list *list = talloc_get_type_abort(
236 private_data, struct aio_child_list);
237 struct aio_child *child, *next;
239 TALLOC_FREE(list->cleanup_event);
241 for (child = list->children; child != NULL; child = next) {
244 if (child->aiocb != NULL) {
245 DEBUG(10, ("child %d currently active\n",
250 if (child->dont_delete) {
251 DEBUG(10, ("Child %d was active since last cleanup\n",
253 child->dont_delete = false;
257 DEBUG(10, ("Child %d idle for more than 30 seconds, "
258 "deleting\n", (int)child->pid));
264 if (list->children != NULL) {
266 * Re-schedule the next cleanup round
268 list->cleanup_event = event_add_timed(server_event_context(), list,
269 timeval_add(&now, 30, 0),
270 aio_child_cleanup, list);
275 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
277 struct aio_child_list *data = NULL;
279 if (SMB_VFS_HANDLE_TEST_DATA(handle)) {
280 SMB_VFS_HANDLE_GET_DATA(handle, data, struct aio_child_list,
285 data = talloc_zero(NULL, struct aio_child_list);
292 * Regardless of whether the child_list had been around or not, make
293 * sure that we have a cleanup timed event. This timed event will
294 * delete itself when it finds that no children are around anymore.
297 if (data->cleanup_event == NULL) {
298 data->cleanup_event = event_add_timed(server_event_context(), data,
299 timeval_current_ofs(30, 0),
300 aio_child_cleanup, data);
301 if (data->cleanup_event == NULL) {
307 if (!SMB_VFS_HANDLE_TEST_DATA(handle)) {
308 SMB_VFS_HANDLE_SET_DATA(handle, data, free_aio_children,
309 struct aio_child_list, return False);
315 static void aio_child_loop(int sockfd, struct mmap_area *map)
320 struct rw_cmd cmd_struct;
321 struct rw_ret ret_struct;
323 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
324 if (ret != sizeof(cmd_struct)) {
325 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
330 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
331 cmd_struct.read_cmd ? "read" : "write",
332 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
334 #ifdef ENABLE_BUILD_FARM_HACKS
337 * In the build farm, we want erratic behaviour for
343 * use generate_random_buffer, we just forked from a
344 * common parent state
346 generate_random_buffer(&randval, sizeof(randval));
347 msecs = randval + 20;
348 DEBUG(10, ("delaying for %u msecs\n", msecs));
354 ZERO_STRUCT(ret_struct);
356 if (cmd_struct.read_cmd) {
357 ret_struct.size = sys_pread(
358 fd, (void *)map->ptr, cmd_struct.n,
361 /* This breaks "make test" when run with aio_fork module. */
362 #ifdef ENABLE_BUILD_FARM_HACKS
363 ret_struct.size = MAX(1, ret_struct.size * 0.9);
368 ret_struct.size = sys_pwrite(
369 fd, (void *)map->ptr, cmd_struct.n,
373 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
374 (int)ret_struct.size));
376 if (ret_struct.size == -1) {
377 ret_struct.ret_errno = errno;
381 * Close the fd before telling our parent we're done. The
382 * parent might close and re-open the file very quickly, and
383 * with system-level share modes (GPFS) we would get an
384 * unjustified SHARING_VIOLATION.
388 ret = write_data(sockfd, (char *)&ret_struct,
390 if (ret != sizeof(ret_struct)) {
391 DEBUG(10, ("could not write ret_struct: %s\n",
398 static void handle_aio_completion(struct event_context *event_ctx,
399 struct fd_event *event, uint16 flags,
402 struct aio_extra *aio_ex = NULL;
403 struct aio_child *child = (struct aio_child *)p;
406 DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags));
408 if ((flags & EVENT_FD_READ) == 0) {
412 status = read_data(child->sockfd, (char *)&child->retval,
413 sizeof(child->retval));
415 if (!NT_STATUS_IS_OK(status)) {
416 DEBUG(1, ("aio child %d died: %s\n", (int)child->pid,
418 child->retval.size = -1;
419 child->retval.ret_errno = EIO;
422 if (child->aiocb == NULL) {
423 DEBUG(1, ("Inactive child died\n"));
428 if (child->cancelled) {
430 child->cancelled = false;
434 if (child->read_cmd && (child->retval.size > 0)) {
435 SMB_ASSERT(child->retval.size <= child->aiocb->aio_nbytes);
436 memcpy((void *)child->aiocb->aio_buf, (void *)child->map->ptr,
440 if (child->called_from_suspend) {
441 child->completion_done = true;
444 aio_ex = (struct aio_extra *)child->aiocb->aio_sigevent.sigev_value.sival_ptr;
445 smbd_aio_complete_aio_ex(aio_ex);
449 static int aio_child_destructor(struct aio_child *child)
453 SMB_ASSERT((child->aiocb == NULL) || child->cancelled);
455 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
456 child->pid, child->sockfd));
459 * closing the sockfd makes the child not return from recvmsg() on RHEL
460 * 5.5 so instead force the child to exit by writing bad data to it
462 write(child->sockfd, &c, sizeof(c));
463 close(child->sockfd);
464 DLIST_REMOVE(child->list->children, child);
469 * We have to close all fd's in open files, we might incorrectly hold a system
470 * level share mode on a file.
473 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
476 if ((fsp->fh != NULL) && (fsp->fh->fd != -1)) {
483 static NTSTATUS create_aio_child(struct smbd_server_connection *sconn,
484 struct aio_child_list *children,
486 struct aio_child **presult)
488 struct aio_child *result;
492 fdpair[0] = fdpair[1] = -1;
494 result = talloc_zero(children, struct aio_child);
495 NT_STATUS_HAVE_NO_MEMORY(result);
497 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
498 status = map_nt_error_from_unix(errno);
499 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
503 DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
505 result->map = mmap_area_init(result, map_size);
506 if (result->map == NULL) {
507 status = map_nt_error_from_unix(errno);
508 DEBUG(0, ("Could not create mmap area\n"));
512 result->pid = fork();
513 if (result->pid == -1) {
514 status = map_nt_error_from_unix(errno);
515 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
519 if (result->pid == 0) {
521 result->sockfd = fdpair[1];
522 files_forall(sconn, close_fsp_fd, NULL);
523 aio_child_loop(result->sockfd, result->map);
526 DEBUG(10, ("Child %d created with sockfd %d\n",
527 result->pid, fdpair[0]));
529 result->sockfd = fdpair[0];
532 result->sock_event = event_add_fd(server_event_context(), result,
533 result->sockfd, EVENT_FD_READ,
534 handle_aio_completion,
536 if (result->sock_event == NULL) {
537 status = NT_STATUS_NO_MEMORY;
538 DEBUG(0, ("event_add_fd failed\n"));
542 result->list = children;
543 DLIST_ADD(children->children, result);
545 talloc_set_destructor(result, aio_child_destructor);
552 if (fdpair[0] != -1) close(fdpair[0]);
553 if (fdpair[1] != -1) close(fdpair[1]);
559 static NTSTATUS get_idle_child(struct vfs_handle_struct *handle,
560 struct aio_child **pchild)
562 struct aio_child_list *children;
563 struct aio_child *child;
566 children = init_aio_children(handle);
567 if (children == NULL) {
568 return NT_STATUS_NO_MEMORY;
571 for (child = children->children; child != NULL; child = child->next) {
572 if (child->aiocb == NULL) {
579 DEBUG(10, ("no idle child found, creating new one\n"));
581 status = create_aio_child(handle->conn->sconn, children,
583 if (!NT_STATUS_IS_OK(status)) {
584 DEBUG(10, ("create_aio_child failed: %s\n",
590 child->dont_delete = true;
596 static int aio_fork_read(struct vfs_handle_struct *handle,
597 struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
599 struct aio_child *child;
604 if (aiocb->aio_nbytes > 128*1024) {
605 /* TODO: support variable buffers */
610 status = get_idle_child(handle, &child);
611 if (!NT_STATUS_IS_OK(status)) {
612 DEBUG(10, ("Could not get an idle child\n"));
616 child->read_cmd = true;
617 child->aiocb = aiocb;
618 child->retval.ret_errno = EINPROGRESS;
621 cmd.n = aiocb->aio_nbytes;
622 cmd.offset = aiocb->aio_offset;
623 cmd.read_cmd = child->read_cmd;
625 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
628 ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
630 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
637 static int aio_fork_write(struct vfs_handle_struct *handle,
638 struct files_struct *fsp, SMB_STRUCT_AIOCB *aiocb)
640 struct aio_child *child;
645 if (aiocb->aio_nbytes > 128*1024) {
646 /* TODO: support variable buffers */
651 status = get_idle_child(handle, &child);
652 if (!NT_STATUS_IS_OK(status)) {
653 DEBUG(10, ("Could not get an idle child\n"));
657 child->read_cmd = false;
658 child->aiocb = aiocb;
659 child->retval.ret_errno = EINPROGRESS;
661 memcpy((void *)child->map->ptr, (void *)aiocb->aio_buf,
665 cmd.n = aiocb->aio_nbytes;
666 cmd.offset = aiocb->aio_offset;
667 cmd.read_cmd = child->read_cmd;
669 DEBUG(10, ("sending fd %d to child %d\n", fsp->fh->fd,
672 ret = write_fd(child->sockfd, &cmd, sizeof(cmd), fsp->fh->fd);
674 DEBUG(10, ("write_fd failed: %s\n", strerror(errno)));
681 static struct aio_child *aio_fork_find_child(struct vfs_handle_struct *handle,
682 SMB_STRUCT_AIOCB *aiocb)
684 struct aio_child_list *children;
685 struct aio_child *child;
687 children = init_aio_children(handle);
688 if (children == NULL) {
692 for (child = children->children; child != NULL; child = child->next) {
693 if (child->aiocb == aiocb) {
701 static ssize_t aio_fork_return_fn(struct vfs_handle_struct *handle,
702 struct files_struct *fsp,
703 SMB_STRUCT_AIOCB *aiocb)
705 struct aio_child *child = aio_fork_find_child(handle, aiocb);
709 DEBUG(0, ("returning EINVAL\n"));
715 if (child->cancelled) {
720 if (child->retval.size == -1) {
721 errno = child->retval.ret_errno;
724 return child->retval.size;
727 static int aio_fork_cancel(struct vfs_handle_struct *handle,
728 struct files_struct *fsp,
729 SMB_STRUCT_AIOCB *aiocb)
731 struct aio_child_list *children;
732 struct aio_child *child;
734 children = init_aio_children(handle);
735 if (children == NULL) {
740 for (child = children->children; child != NULL; child = child->next) {
741 if (child->aiocb == NULL) {
744 if (child->aiocb->aio_fildes != fsp->fh->fd) {
747 if ((aiocb != NULL) && (child->aiocb != aiocb)) {
752 * We let the child do its job, but we discard the result when
756 child->cancelled = true;
762 static int aio_fork_error_fn(struct vfs_handle_struct *handle,
763 struct files_struct *fsp,
764 SMB_STRUCT_AIOCB *aiocb)
766 struct aio_child *child = aio_fork_find_child(handle, aiocb);
773 return child->retval.ret_errno;
776 static void aio_fork_suspend_timed_out(struct tevent_context *event_ctx,
777 struct tevent_timer *te,
781 bool *timed_out = (bool *)private_data;
782 /* Remove this timed event handler. */
787 static int aio_fork_suspend(struct vfs_handle_struct *handle,
788 struct files_struct *fsp,
789 const SMB_STRUCT_AIOCB * const aiocb_array[],
791 const struct timespec *timeout)
793 struct aio_child_list *children = NULL;
794 TALLOC_CTX *frame = talloc_stackframe();
795 struct event_context *ev = NULL;
798 bool timed_out = false;
801 children = init_aio_children(handle);
802 if (children == NULL) {
807 /* This is a blocking call, and has to use a sub-event loop. */
808 ev = event_context_init(frame);
815 struct timeval tv = convert_timespec_to_timeval(*timeout);
816 struct tevent_timer *te = tevent_add_timer(ev,
818 timeval_current_ofs(tv.tv_sec,
820 aio_fork_suspend_timed_out,
828 for (i = 0; i < n; i++) {
829 struct aio_child *child = NULL;
830 const SMB_STRUCT_AIOCB *aiocb = aiocb_array[i];
837 * We're going to cheat here. We know that smbd/aio.c
838 * only calls this when it's waiting for every single
839 * outstanding call to finish on a close, so just wait
840 * individually for each IO to complete. We don't care
841 * what order they finish - only that they all do. JRA.
844 for (child = children->children; child != NULL; child = child->next) {
845 struct tevent_fd *event;
847 if (child->aiocb == NULL) {
850 if (child->aiocb->aio_fildes != fsp->fh->fd) {
853 if (child->aiocb != aiocb) {
857 if (child->aiocb->aio_sigevent.sigev_value.sival_ptr == NULL) {
861 event = event_add_fd(ev,
865 handle_aio_completion,
872 child->called_from_suspend = true;
874 while (!child->completion_done) {
875 if (tevent_loop_once(ev) == -1) {
897 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
900 /*********************************************************************
901 * How many threads to initialize ?
902 * 100 per process seems insane as a default until you realize that
903 * (a) Threads terminate after 1 second when idle.
904 * (b) Throttling is done in SMB2 via the crediting algorithm.
905 * (c) SMB1 clients are limited to max_mux (50) outstanding
906 * requests and Windows clients don't use this anyway.
907 * Essentially we want this to be unlimited unless smb.conf
909 *********************************************************************/
910 aio_pending_size = 100;
911 return SMB_VFS_NEXT_CONNECT(handle, service, user);
914 static struct vfs_fn_pointers vfs_aio_fork_fns = {
915 .connect_fn = aio_fork_connect,
916 .aio_read_fn = aio_fork_read,
917 .aio_write_fn = aio_fork_write,
918 .aio_return_fn = aio_fork_return_fn,
919 .aio_cancel_fn = aio_fork_cancel,
920 .aio_error_fn = aio_fork_error_fn,
921 .aio_suspend_fn = aio_fork_suspend,
924 NTSTATUS vfs_aio_fork_init(void);
925 NTSTATUS vfs_aio_fork_init(void)
927 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
928 "aio_fork", &vfs_aio_fork_fns);