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.
29 static int mmap_area_destructor(struct mmap_area
*area
)
31 munmap((void *)area
->ptr
, area
->size
);
35 static struct mmap_area
*mmap_area_init(TALLOC_CTX
*mem_ctx
, size_t size
)
37 struct mmap_area
*result
;
40 result
= talloc(mem_ctx
, struct mmap_area
);
42 DEBUG(0, ("talloc failed\n"));
46 fd
= open("/dev/zero", O_RDWR
);
48 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
53 result
->ptr
= mmap(NULL
, size
, PROT_READ
|PROT_WRITE
,
54 MAP_SHARED
|MAP_FILE
, fd
, 0);
55 if (result
->ptr
== MAP_FAILED
) {
56 DEBUG(1, ("mmap failed: %s\n", strerror(errno
)));
63 talloc_set_destructor(result
, mmap_area_destructor
);
83 struct aio_child_list
;
86 struct aio_child
*prev
, *next
;
87 struct aio_child_list
*list
;
88 SMB_STRUCT_AIOCB
*aiocb
;
91 struct fd_event
*sock_event
;
93 struct mmap_area
*map
; /* ==NULL means write request */
94 bool dont_delete
; /* Marked as in use since last cleanup */
99 struct aio_child_list
{
100 struct aio_child
*children
;
101 struct timed_event
*cleanup_event
;
104 static void free_aio_children(void **p
)
109 static ssize_t
read_fd(int fd
, void *ptr
, size_t nbytes
, int *recvfd
)
114 #ifndef HAVE_MSGHDR_MSG_CONTROL
118 #ifdef HAVE_MSGHDR_MSG_CONTROL
121 char control
[CMSG_SPACE(sizeof(int))];
123 struct cmsghdr
*cmptr
;
125 msg
.msg_control
= control_un
.control
;
126 msg
.msg_controllen
= sizeof(control_un
.control
);
128 #if HAVE_MSGHDR_MSG_ACCTRIGHTS
129 msg
.msg_accrights
= (caddr_t
) &newfd
;
130 msg
.msg_accrightslen
= sizeof(int);
132 #error Can not pass file descriptors
139 iov
[0].iov_base
= (void *)ptr
;
140 iov
[0].iov_len
= nbytes
;
144 if ( (n
= recvmsg(fd
, &msg
, 0)) <= 0) {
148 #ifdef HAVE_MSGHDR_MSG_CONTROL
149 if ((cmptr
= CMSG_FIRSTHDR(&msg
)) != NULL
150 && cmptr
->cmsg_len
== CMSG_LEN(sizeof(int))) {
151 if (cmptr
->cmsg_level
!= SOL_SOCKET
) {
152 DEBUG(10, ("control level != SOL_SOCKET"));
156 if (cmptr
->cmsg_type
!= SCM_RIGHTS
) {
157 DEBUG(10, ("control type != SCM_RIGHTS"));
161 *recvfd
= *((int *) CMSG_DATA(cmptr
));
163 *recvfd
= -1; /* descriptor was not passed */
166 if (msg
.msg_accrightslen
== sizeof(int)) {
170 *recvfd
= -1; /* descriptor was not passed */
177 static ssize_t
write_fd(int fd
, void *ptr
, size_t nbytes
, int sendfd
)
182 #ifdef HAVE_MSGHDR_MSG_CONTROL
185 char control
[CMSG_SPACE(sizeof(int))];
187 struct cmsghdr
*cmptr
;
190 ZERO_STRUCT(control_un
);
192 msg
.msg_control
= control_un
.control
;
193 msg
.msg_controllen
= sizeof(control_un
.control
);
195 cmptr
= CMSG_FIRSTHDR(&msg
);
196 cmptr
->cmsg_len
= CMSG_LEN(sizeof(int));
197 cmptr
->cmsg_level
= SOL_SOCKET
;
198 cmptr
->cmsg_type
= SCM_RIGHTS
;
199 *((int *) CMSG_DATA(cmptr
)) = sendfd
;
202 msg
.msg_accrights
= (caddr_t
) &sendfd
;
203 msg
.msg_accrightslen
= sizeof(int);
210 iov
[0].iov_base
= (void *)ptr
;
211 iov
[0].iov_len
= nbytes
;
215 return (sendmsg(fd
, &msg
, 0));
218 static void aio_child_cleanup(struct event_context
*event_ctx
,
219 struct timed_event
*te
,
223 struct aio_child_list
*list
= talloc_get_type_abort(
224 private_data
, struct aio_child_list
);
225 struct aio_child
*child
, *next
;
227 TALLOC_FREE(list
->cleanup_event
);
229 for (child
= list
->children
; child
!= NULL
; child
= next
) {
232 if (child
->aiocb
!= NULL
) {
233 DEBUG(10, ("child %d currently active\n",
238 if (child
->dont_delete
) {
239 DEBUG(10, ("Child %d was active since last cleanup\n",
241 child
->dont_delete
= false;
245 DEBUG(10, ("Child %d idle for more than 30 seconds, "
246 "deleting\n", (int)child
->pid
));
251 if (list
->children
!= NULL
) {
253 * Re-schedule the next cleanup round
255 list
->cleanup_event
= event_add_timed(smbd_event_context(), list
,
256 timeval_add(&now
, 30, 0),
257 aio_child_cleanup
, list
);
262 static struct aio_child_list
*init_aio_children(struct vfs_handle_struct
*handle
)
264 struct aio_child_list
*data
= NULL
;
266 if (SMB_VFS_HANDLE_TEST_DATA(handle
)) {
267 SMB_VFS_HANDLE_GET_DATA(handle
, data
, struct aio_child_list
,
272 data
= TALLOC_ZERO_P(NULL
, struct aio_child_list
);
279 * Regardless of whether the child_list had been around or not, make
280 * sure that we have a cleanup timed event. This timed event will
281 * delete itself when it finds that no children are around anymore.
284 if (data
->cleanup_event
== NULL
) {
285 data
->cleanup_event
= event_add_timed(smbd_event_context(), data
,
286 timeval_current_ofs(30, 0),
287 aio_child_cleanup
, data
);
288 if (data
->cleanup_event
== NULL
) {
294 if (!SMB_VFS_HANDLE_TEST_DATA(handle
)) {
295 SMB_VFS_HANDLE_SET_DATA(handle
, data
, free_aio_children
,
296 struct aio_child_list
, return False
);
302 static void aio_child_loop(int sockfd
, struct mmap_area
*map
)
307 struct rw_cmd cmd_struct
;
308 struct rw_ret ret_struct
;
310 ret
= read_fd(sockfd
, &cmd_struct
, sizeof(cmd_struct
), &fd
);
311 if (ret
!= sizeof(cmd_struct
)) {
312 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret
,
317 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
318 cmd_struct
.read_cmd
? "read" : "write",
319 (int)cmd_struct
.n
, (int)cmd_struct
.offset
, fd
));
321 #ifdef ENABLE_BUILD_FARM_HACKS
324 * In the build farm, we want erratic behaviour for
330 * use generate_random_buffer, we just forked from a
331 * common parent state
333 generate_random_buffer(&randval
, sizeof(randval
));
334 msecs
= randval
+ 20;
335 DEBUG(10, ("delaying for %u msecs\n", msecs
));
341 ZERO_STRUCT(ret_struct
);
343 if (cmd_struct
.read_cmd
) {
344 ret_struct
.size
= sys_pread(
345 fd
, (void *)map
->ptr
, cmd_struct
.n
,
348 /* This breaks "make test" when run with aio_fork module. */
349 #ifdef ENABLE_BUILD_FARM_HACKS
350 ret_struct
.size
= MAX(1, ret_struct
.size
* 0.9);
355 ret_struct
.size
= sys_pwrite(
356 fd
, (void *)map
->ptr
, cmd_struct
.n
,
360 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
361 (int)ret_struct
.size
));
363 if (ret_struct
.size
== -1) {
364 ret_struct
.ret_errno
= errno
;
368 * Close the fd before telling our parent we're done. The
369 * parent might close and re-open the file very quickly, and
370 * with system-level share modes (GPFS) we would get an
371 * unjustified SHARING_VIOLATION.
375 ret
= write_data(sockfd
, (char *)&ret_struct
,
377 if (ret
!= sizeof(ret_struct
)) {
378 DEBUG(10, ("could not write ret_struct: %s\n",
385 static void handle_aio_completion(struct event_context
*event_ctx
,
386 struct fd_event
*event
, uint16 flags
,
389 struct aio_extra
*aio_ex
= NULL
;
390 struct aio_child
*child
= (struct aio_child
*)p
;
392 DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags
));
394 if ((flags
& EVENT_FD_READ
) == 0) {
398 if (!NT_STATUS_IS_OK(read_data(child
->sockfd
,
399 (char *)&child
->retval
,
400 sizeof(child
->retval
)))) {
401 DEBUG(0, ("aio child %d died\n", (int)child
->pid
));
402 child
->retval
.size
= -1;
403 child
->retval
.ret_errno
= EIO
;
406 if (child
->cancelled
) {
408 child
->cancelled
= false;
412 if (child
->read_cmd
&& (child
->retval
.size
> 0)) {
413 SMB_ASSERT(child
->retval
.size
<= child
->aiocb
->aio_nbytes
);
414 memcpy((void *)child
->aiocb
->aio_buf
, (void *)child
->map
->ptr
,
418 aio_ex
= (struct aio_extra
*)child
->aiocb
->aio_sigevent
.sigev_value
.sival_ptr
;
419 smbd_aio_complete_aio_ex(aio_ex
);
422 static int aio_child_destructor(struct aio_child
*child
)
424 SMB_ASSERT((child
->aiocb
== NULL
) || child
->cancelled
);
425 close(child
->sockfd
);
426 DLIST_REMOVE(child
->list
->children
, child
);
431 * We have to close all fd's in open files, we might incorrectly hold a system
432 * level share mode on a file.
435 static struct files_struct
*close_fsp_fd(struct files_struct
*fsp
,
438 if ((fsp
->fh
!= NULL
) && (fsp
->fh
->fd
!= -1)) {
445 static NTSTATUS
create_aio_child(struct smbd_server_connection
*sconn
,
446 struct aio_child_list
*children
,
448 struct aio_child
**presult
)
450 struct aio_child
*result
;
454 fdpair
[0] = fdpair
[1] = -1;
456 result
= TALLOC_ZERO_P(children
, struct aio_child
);
457 NT_STATUS_HAVE_NO_MEMORY(result
);
459 if (socketpair(AF_UNIX
, SOCK_STREAM
, 0, fdpair
) == -1) {
460 status
= map_nt_error_from_unix(errno
);
461 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno
)));
465 DEBUG(10, ("fdpair = %d/%d\n", fdpair
[0], fdpair
[1]));
467 result
->map
= mmap_area_init(result
, map_size
);
468 if (result
->map
== NULL
) {
469 status
= map_nt_error_from_unix(errno
);
470 DEBUG(0, ("Could not create mmap area\n"));
474 result
->pid
= sys_fork();
475 if (result
->pid
== -1) {
476 status
= map_nt_error_from_unix(errno
);
477 DEBUG(0, ("fork failed: %s\n", strerror(errno
)));
481 if (result
->pid
== 0) {
483 result
->sockfd
= fdpair
[1];
484 files_forall(sconn
, close_fsp_fd
, NULL
);
485 aio_child_loop(result
->sockfd
, result
->map
);
488 DEBUG(10, ("Child %d created\n", result
->pid
));
490 result
->sockfd
= fdpair
[0];
493 result
->sock_event
= event_add_fd(smbd_event_context(), result
,
494 result
->sockfd
, EVENT_FD_READ
,
495 handle_aio_completion
,
497 if (result
->sock_event
== NULL
) {
498 status
= NT_STATUS_NO_MEMORY
;
499 DEBUG(0, ("event_add_fd failed\n"));
503 result
->list
= children
;
504 DLIST_ADD(children
->children
, result
);
506 talloc_set_destructor(result
, aio_child_destructor
);
513 if (fdpair
[0] != -1) close(fdpair
[0]);
514 if (fdpair
[1] != -1) close(fdpair
[1]);
520 static NTSTATUS
get_idle_child(struct vfs_handle_struct
*handle
,
521 struct aio_child
**pchild
)
523 struct aio_child_list
*children
;
524 struct aio_child
*child
;
527 children
= init_aio_children(handle
);
528 if (children
== NULL
) {
529 return NT_STATUS_NO_MEMORY
;
532 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
533 if (child
->aiocb
== NULL
) {
540 DEBUG(10, ("no idle child found, creating new one\n"));
542 status
= create_aio_child(handle
->conn
->sconn
, children
,
544 if (!NT_STATUS_IS_OK(status
)) {
545 DEBUG(10, ("create_aio_child failed: %s\n",
551 child
->dont_delete
= true;
557 static int aio_fork_read(struct vfs_handle_struct
*handle
,
558 struct files_struct
*fsp
, SMB_STRUCT_AIOCB
*aiocb
)
560 struct aio_child
*child
;
565 if (aiocb
->aio_nbytes
> 128*1024) {
566 /* TODO: support variable buffers */
571 status
= get_idle_child(handle
, &child
);
572 if (!NT_STATUS_IS_OK(status
)) {
573 DEBUG(10, ("Could not get an idle child\n"));
577 child
->read_cmd
= true;
578 child
->aiocb
= aiocb
;
579 child
->retval
.ret_errno
= EINPROGRESS
;
582 cmd
.n
= aiocb
->aio_nbytes
;
583 cmd
.offset
= aiocb
->aio_offset
;
584 cmd
.read_cmd
= child
->read_cmd
;
586 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
589 ret
= write_fd(child
->sockfd
, &cmd
, sizeof(cmd
), fsp
->fh
->fd
);
591 DEBUG(10, ("write_fd failed: %s\n", strerror(errno
)));
598 static int aio_fork_write(struct vfs_handle_struct
*handle
,
599 struct files_struct
*fsp
, SMB_STRUCT_AIOCB
*aiocb
)
601 struct aio_child
*child
;
606 if (aiocb
->aio_nbytes
> 128*1024) {
607 /* TODO: support variable buffers */
612 status
= get_idle_child(handle
, &child
);
613 if (!NT_STATUS_IS_OK(status
)) {
614 DEBUG(10, ("Could not get an idle child\n"));
618 child
->read_cmd
= false;
619 child
->aiocb
= aiocb
;
620 child
->retval
.ret_errno
= EINPROGRESS
;
622 memcpy((void *)child
->map
->ptr
, (void *)aiocb
->aio_buf
,
626 cmd
.n
= aiocb
->aio_nbytes
;
627 cmd
.offset
= aiocb
->aio_offset
;
628 cmd
.read_cmd
= child
->read_cmd
;
630 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
633 ret
= write_fd(child
->sockfd
, &cmd
, sizeof(cmd
), fsp
->fh
->fd
);
635 DEBUG(10, ("write_fd failed: %s\n", strerror(errno
)));
642 static struct aio_child
*aio_fork_find_child(struct vfs_handle_struct
*handle
,
643 SMB_STRUCT_AIOCB
*aiocb
)
645 struct aio_child_list
*children
;
646 struct aio_child
*child
;
648 children
= init_aio_children(handle
);
649 if (children
== NULL
) {
653 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
654 if (child
->aiocb
== aiocb
) {
662 static ssize_t
aio_fork_return_fn(struct vfs_handle_struct
*handle
,
663 struct files_struct
*fsp
,
664 SMB_STRUCT_AIOCB
*aiocb
)
666 struct aio_child
*child
= aio_fork_find_child(handle
, aiocb
);
670 DEBUG(0, ("returning EINVAL\n"));
676 if (child
->retval
.size
== -1) {
677 errno
= child
->retval
.ret_errno
;
680 return child
->retval
.size
;
683 static int aio_fork_cancel(struct vfs_handle_struct
*handle
,
684 struct files_struct
*fsp
,
685 SMB_STRUCT_AIOCB
*aiocb
)
687 struct aio_child_list
*children
;
688 struct aio_child
*child
;
690 children
= init_aio_children(handle
);
691 if (children
== NULL
) {
696 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
697 if (child
->aiocb
== NULL
) {
700 if (child
->aiocb
->aio_fildes
!= fsp
->fh
->fd
) {
703 if ((aiocb
!= NULL
) && (child
->aiocb
!= aiocb
)) {
708 * We let the child do its job, but we discard the result when
712 child
->cancelled
= true;
718 static int aio_fork_error_fn(struct vfs_handle_struct
*handle
,
719 struct files_struct
*fsp
,
720 SMB_STRUCT_AIOCB
*aiocb
)
722 struct aio_child
*child
= aio_fork_find_child(handle
, aiocb
);
729 return child
->retval
.ret_errno
;
732 static void aio_fork_suspend_timed_out(struct tevent_context
*event_ctx
,
733 struct tevent_timer
*te
,
737 bool *timed_out
= (bool *)private_data
;
738 /* Remove this timed event handler. */
743 static int aio_fork_suspend(struct vfs_handle_struct
*handle
,
744 struct files_struct
*fsp
,
745 const SMB_STRUCT_AIOCB
* const aiocb_array
[],
747 const struct timespec
*timeout
)
749 struct aio_child_list
*children
= NULL
;
750 TALLOC_CTX
*frame
= talloc_stackframe();
751 struct event_context
*ev
= NULL
;
754 bool timed_out
= false;
756 children
= init_aio_children(handle
);
757 if (children
== NULL
) {
762 /* This is a blocking call, and has to use a sub-event loop. */
763 ev
= event_context_init(frame
);
770 struct timeval tv
= convert_timespec_to_timeval(*timeout
);
771 struct tevent_timer
*te
= tevent_add_timer(ev
,
773 timeval_current_ofs(tv
.tv_sec
,
775 aio_fork_suspend_timed_out
,
783 for (i
= 0; i
< n
; i
++) {
784 struct aio_child
*child
= NULL
;
785 const SMB_STRUCT_AIOCB
*aiocb
= aiocb_array
[i
];
792 * We're going to cheat here. We know that smbd/aio.c
793 * only calls this when it's waiting for every single
794 * outstanding call to finish on a close, so just wait
795 * individually for each IO to complete. We don't care
796 * what order they finish - only that they all do. JRA.
799 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
800 if (child
->aiocb
== NULL
) {
803 if (child
->aiocb
->aio_fildes
!= fsp
->fh
->fd
) {
806 if (child
->aiocb
!= aiocb
) {
810 if (child
->aiocb
->aio_sigevent
.sigev_value
.sival_ptr
== NULL
) {
814 /* We're never using this event on the
815 * main event context again... */
816 TALLOC_FREE(child
->sock_event
);
818 child
->sock_event
= event_add_fd(ev
,
822 handle_aio_completion
,
826 if (tevent_loop_once(ev
) == -1) {
835 /* We set child->aiocb to NULL in our hooked
837 if (child
->aiocb
== NULL
) {
852 static struct vfs_fn_pointers vfs_aio_fork_fns
= {
853 .aio_read
= aio_fork_read
,
854 .aio_write
= aio_fork_write
,
855 .aio_return_fn
= aio_fork_return_fn
,
856 .aio_cancel
= aio_fork_cancel
,
857 .aio_error_fn
= aio_fork_error_fn
,
858 .aio_suspend
= aio_fork_suspend
,
861 NTSTATUS
vfs_aio_fork_init(void);
862 NTSTATUS
vfs_aio_fork_init(void)
864 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION
,
865 "aio_fork", &vfs_aio_fork_fns
);