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
150 iov
[0].iov_base
= (void *)ptr
;
151 iov
[0].iov_len
= nbytes
;
155 if ( (n
= recvmsg(fd
, &msg
, 0)) <= 0) {
159 #ifdef HAVE_MSGHDR_MSG_CONTROL
160 if ((cmptr
= CMSG_FIRSTHDR(&msg
)) != NULL
161 && cmptr
->cmsg_len
== CMSG_LEN(sizeof(int))) {
162 if (cmptr
->cmsg_level
!= SOL_SOCKET
) {
163 DEBUG(10, ("control level != SOL_SOCKET"));
167 if (cmptr
->cmsg_type
!= SCM_RIGHTS
) {
168 DEBUG(10, ("control type != SCM_RIGHTS"));
172 memcpy(recvfd
, CMSG_DATA(cmptr
), sizeof(*recvfd
));
174 *recvfd
= -1; /* descriptor was not passed */
177 if (msg
.msg_accrightslen
== sizeof(int)) {
181 *recvfd
= -1; /* descriptor was not passed */
188 static ssize_t
write_fd(int fd
, void *ptr
, size_t nbytes
, int sendfd
)
193 #ifdef HAVE_MSGHDR_MSG_CONTROL
196 char control
[CMSG_SPACE(sizeof(int))];
198 struct cmsghdr
*cmptr
;
201 ZERO_STRUCT(control_un
);
203 msg
.msg_control
= control_un
.control
;
204 msg
.msg_controllen
= sizeof(control_un
.control
);
206 cmptr
= CMSG_FIRSTHDR(&msg
);
207 cmptr
->cmsg_len
= CMSG_LEN(sizeof(int));
208 cmptr
->cmsg_level
= SOL_SOCKET
;
209 cmptr
->cmsg_type
= SCM_RIGHTS
;
210 memcpy(CMSG_DATA(cmptr
), &sendfd
, sizeof(sendfd
));
213 msg
.msg_accrights
= (caddr_t
) &sendfd
;
214 msg
.msg_accrightslen
= sizeof(int);
221 iov
[0].iov_base
= (void *)ptr
;
222 iov
[0].iov_len
= nbytes
;
226 return (sendmsg(fd
, &msg
, 0));
229 static void aio_child_cleanup(struct event_context
*event_ctx
,
230 struct timed_event
*te
,
234 struct aio_child_list
*list
= talloc_get_type_abort(
235 private_data
, struct aio_child_list
);
236 struct aio_child
*child
, *next
;
238 TALLOC_FREE(list
->cleanup_event
);
240 for (child
= list
->children
; child
!= NULL
; child
= next
) {
243 if (child
->aiocb
!= NULL
) {
244 DEBUG(10, ("child %d currently active\n",
249 if (child
->dont_delete
) {
250 DEBUG(10, ("Child %d was active since last cleanup\n",
252 child
->dont_delete
= false;
256 DEBUG(10, ("Child %d idle for more than 30 seconds, "
257 "deleting\n", (int)child
->pid
));
263 if (list
->children
!= NULL
) {
265 * Re-schedule the next cleanup round
267 list
->cleanup_event
= event_add_timed(server_event_context(), list
,
268 timeval_add(&now
, 30, 0),
269 aio_child_cleanup
, list
);
274 static struct aio_child_list
*init_aio_children(struct vfs_handle_struct
*handle
)
276 struct aio_child_list
*data
= NULL
;
278 if (SMB_VFS_HANDLE_TEST_DATA(handle
)) {
279 SMB_VFS_HANDLE_GET_DATA(handle
, data
, struct aio_child_list
,
284 data
= talloc_zero(NULL
, struct aio_child_list
);
291 * Regardless of whether the child_list had been around or not, make
292 * sure that we have a cleanup timed event. This timed event will
293 * delete itself when it finds that no children are around anymore.
296 if (data
->cleanup_event
== NULL
) {
297 data
->cleanup_event
= event_add_timed(server_event_context(), data
,
298 timeval_current_ofs(30, 0),
299 aio_child_cleanup
, data
);
300 if (data
->cleanup_event
== NULL
) {
306 if (!SMB_VFS_HANDLE_TEST_DATA(handle
)) {
307 SMB_VFS_HANDLE_SET_DATA(handle
, data
, free_aio_children
,
308 struct aio_child_list
, return False
);
314 static void aio_child_loop(int sockfd
, struct mmap_area
*map
)
319 struct rw_cmd cmd_struct
;
320 struct rw_ret ret_struct
;
322 ret
= read_fd(sockfd
, &cmd_struct
, sizeof(cmd_struct
), &fd
);
323 if (ret
!= sizeof(cmd_struct
)) {
324 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret
,
329 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
330 cmd_struct
.read_cmd
? "read" : "write",
331 (int)cmd_struct
.n
, (int)cmd_struct
.offset
, fd
));
333 #ifdef ENABLE_BUILD_FARM_HACKS
336 * In the build farm, we want erratic behaviour for
342 * use generate_random_buffer, we just forked from a
343 * common parent state
345 generate_random_buffer(&randval
, sizeof(randval
));
346 msecs
= randval
+ 20;
347 DEBUG(10, ("delaying for %u msecs\n", msecs
));
353 ZERO_STRUCT(ret_struct
);
355 if (cmd_struct
.read_cmd
) {
356 ret_struct
.size
= sys_pread(
357 fd
, (void *)map
->ptr
, cmd_struct
.n
,
360 /* This breaks "make test" when run with aio_fork module. */
361 #ifdef ENABLE_BUILD_FARM_HACKS
362 ret_struct
.size
= MAX(1, ret_struct
.size
* 0.9);
367 ret_struct
.size
= sys_pwrite(
368 fd
, (void *)map
->ptr
, cmd_struct
.n
,
372 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
373 (int)ret_struct
.size
));
375 if (ret_struct
.size
== -1) {
376 ret_struct
.ret_errno
= errno
;
380 * Close the fd before telling our parent we're done. The
381 * parent might close and re-open the file very quickly, and
382 * with system-level share modes (GPFS) we would get an
383 * unjustified SHARING_VIOLATION.
387 ret
= write_data(sockfd
, (char *)&ret_struct
,
389 if (ret
!= sizeof(ret_struct
)) {
390 DEBUG(10, ("could not write ret_struct: %s\n",
397 static void handle_aio_completion(struct event_context
*event_ctx
,
398 struct fd_event
*event
, uint16 flags
,
401 struct aio_extra
*aio_ex
= NULL
;
402 struct aio_child
*child
= (struct aio_child
*)p
;
405 DEBUG(10, ("handle_aio_completion called with flags=%d\n", flags
));
407 if ((flags
& EVENT_FD_READ
) == 0) {
411 status
= read_data(child
->sockfd
, (char *)&child
->retval
,
412 sizeof(child
->retval
));
414 if (!NT_STATUS_IS_OK(status
)) {
415 DEBUG(1, ("aio child %d died: %s\n", (int)child
->pid
,
417 child
->retval
.size
= -1;
418 child
->retval
.ret_errno
= EIO
;
421 if (child
->aiocb
== NULL
) {
422 DEBUG(1, ("Inactive child died\n"));
427 if (child
->cancelled
) {
429 child
->cancelled
= false;
433 if (child
->read_cmd
&& (child
->retval
.size
> 0)) {
434 SMB_ASSERT(child
->retval
.size
<= child
->aiocb
->aio_nbytes
);
435 memcpy((void *)child
->aiocb
->aio_buf
, (void *)child
->map
->ptr
,
439 if (child
->called_from_suspend
) {
440 child
->completion_done
= true;
443 aio_ex
= (struct aio_extra
*)child
->aiocb
->aio_sigevent
.sigev_value
.sival_ptr
;
444 smbd_aio_complete_aio_ex(aio_ex
);
448 static int aio_child_destructor(struct aio_child
*child
)
452 SMB_ASSERT((child
->aiocb
== NULL
) || child
->cancelled
);
454 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
455 child
->pid
, child
->sockfd
));
458 * closing the sockfd makes the child not return from recvmsg() on RHEL
459 * 5.5 so instead force the child to exit by writing bad data to it
461 write(child
->sockfd
, &c
, sizeof(c
));
462 close(child
->sockfd
);
463 DLIST_REMOVE(child
->list
->children
, child
);
468 * We have to close all fd's in open files, we might incorrectly hold a system
469 * level share mode on a file.
472 static struct files_struct
*close_fsp_fd(struct files_struct
*fsp
,
475 if ((fsp
->fh
!= NULL
) && (fsp
->fh
->fd
!= -1)) {
482 static NTSTATUS
create_aio_child(struct smbd_server_connection
*sconn
,
483 struct aio_child_list
*children
,
485 struct aio_child
**presult
)
487 struct aio_child
*result
;
491 fdpair
[0] = fdpair
[1] = -1;
493 result
= talloc_zero(children
, struct aio_child
);
494 NT_STATUS_HAVE_NO_MEMORY(result
);
496 if (socketpair(AF_UNIX
, SOCK_STREAM
, 0, fdpair
) == -1) {
497 status
= map_nt_error_from_unix(errno
);
498 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno
)));
502 DEBUG(10, ("fdpair = %d/%d\n", fdpair
[0], fdpair
[1]));
504 result
->map
= mmap_area_init(result
, map_size
);
505 if (result
->map
== NULL
) {
506 status
= map_nt_error_from_unix(errno
);
507 DEBUG(0, ("Could not create mmap area\n"));
511 result
->pid
= fork();
512 if (result
->pid
== -1) {
513 status
= map_nt_error_from_unix(errno
);
514 DEBUG(0, ("fork failed: %s\n", strerror(errno
)));
518 if (result
->pid
== 0) {
520 result
->sockfd
= fdpair
[1];
521 files_forall(sconn
, close_fsp_fd
, NULL
);
522 aio_child_loop(result
->sockfd
, result
->map
);
525 DEBUG(10, ("Child %d created with sockfd %d\n",
526 result
->pid
, fdpair
[0]));
528 result
->sockfd
= fdpair
[0];
531 result
->sock_event
= event_add_fd(server_event_context(), result
,
532 result
->sockfd
, EVENT_FD_READ
,
533 handle_aio_completion
,
535 if (result
->sock_event
== NULL
) {
536 status
= NT_STATUS_NO_MEMORY
;
537 DEBUG(0, ("event_add_fd failed\n"));
541 result
->list
= children
;
542 DLIST_ADD(children
->children
, result
);
544 talloc_set_destructor(result
, aio_child_destructor
);
551 if (fdpair
[0] != -1) close(fdpair
[0]);
552 if (fdpair
[1] != -1) close(fdpair
[1]);
558 static NTSTATUS
get_idle_child(struct vfs_handle_struct
*handle
,
559 struct aio_child
**pchild
)
561 struct aio_child_list
*children
;
562 struct aio_child
*child
;
565 children
= init_aio_children(handle
);
566 if (children
== NULL
) {
567 return NT_STATUS_NO_MEMORY
;
570 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
571 if (child
->aiocb
== NULL
) {
578 DEBUG(10, ("no idle child found, creating new one\n"));
580 status
= create_aio_child(handle
->conn
->sconn
, children
,
582 if (!NT_STATUS_IS_OK(status
)) {
583 DEBUG(10, ("create_aio_child failed: %s\n",
589 child
->dont_delete
= true;
595 static int aio_fork_read(struct vfs_handle_struct
*handle
,
596 struct files_struct
*fsp
, SMB_STRUCT_AIOCB
*aiocb
)
598 struct aio_child
*child
;
603 if (aiocb
->aio_nbytes
> 128*1024) {
604 /* TODO: support variable buffers */
609 status
= get_idle_child(handle
, &child
);
610 if (!NT_STATUS_IS_OK(status
)) {
611 DEBUG(10, ("Could not get an idle child\n"));
615 child
->read_cmd
= true;
616 child
->aiocb
= aiocb
;
617 child
->retval
.ret_errno
= EINPROGRESS
;
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 int aio_fork_write(struct vfs_handle_struct
*handle
,
637 struct files_struct
*fsp
, SMB_STRUCT_AIOCB
*aiocb
)
639 struct aio_child
*child
;
644 if (aiocb
->aio_nbytes
> 128*1024) {
645 /* TODO: support variable buffers */
650 status
= get_idle_child(handle
, &child
);
651 if (!NT_STATUS_IS_OK(status
)) {
652 DEBUG(10, ("Could not get an idle child\n"));
656 child
->read_cmd
= false;
657 child
->aiocb
= aiocb
;
658 child
->retval
.ret_errno
= EINPROGRESS
;
660 memcpy((void *)child
->map
->ptr
, (void *)aiocb
->aio_buf
,
664 cmd
.n
= aiocb
->aio_nbytes
;
665 cmd
.offset
= aiocb
->aio_offset
;
666 cmd
.read_cmd
= child
->read_cmd
;
668 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
671 ret
= write_fd(child
->sockfd
, &cmd
, sizeof(cmd
), fsp
->fh
->fd
);
673 DEBUG(10, ("write_fd failed: %s\n", strerror(errno
)));
680 static struct aio_child
*aio_fork_find_child(struct vfs_handle_struct
*handle
,
681 SMB_STRUCT_AIOCB
*aiocb
)
683 struct aio_child_list
*children
;
684 struct aio_child
*child
;
686 children
= init_aio_children(handle
);
687 if (children
== NULL
) {
691 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
692 if (child
->aiocb
== aiocb
) {
700 static ssize_t
aio_fork_return_fn(struct vfs_handle_struct
*handle
,
701 struct files_struct
*fsp
,
702 SMB_STRUCT_AIOCB
*aiocb
)
704 struct aio_child
*child
= aio_fork_find_child(handle
, aiocb
);
708 DEBUG(0, ("returning EINVAL\n"));
714 if (child
->cancelled
) {
719 if (child
->retval
.size
== -1) {
720 errno
= child
->retval
.ret_errno
;
723 return child
->retval
.size
;
726 static int aio_fork_cancel(struct vfs_handle_struct
*handle
,
727 struct files_struct
*fsp
,
728 SMB_STRUCT_AIOCB
*aiocb
)
730 struct aio_child_list
*children
;
731 struct aio_child
*child
;
733 children
= init_aio_children(handle
);
734 if (children
== NULL
) {
739 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
740 if (child
->aiocb
== NULL
) {
743 if (child
->aiocb
->aio_fildes
!= fsp
->fh
->fd
) {
746 if ((aiocb
!= NULL
) && (child
->aiocb
!= aiocb
)) {
751 * We let the child do its job, but we discard the result when
755 child
->cancelled
= true;
761 static int aio_fork_error_fn(struct vfs_handle_struct
*handle
,
762 struct files_struct
*fsp
,
763 SMB_STRUCT_AIOCB
*aiocb
)
765 struct aio_child
*child
= aio_fork_find_child(handle
, aiocb
);
772 return child
->retval
.ret_errno
;
775 static void aio_fork_suspend_timed_out(struct tevent_context
*event_ctx
,
776 struct tevent_timer
*te
,
780 bool *timed_out
= (bool *)private_data
;
781 /* Remove this timed event handler. */
786 static int aio_fork_suspend(struct vfs_handle_struct
*handle
,
787 struct files_struct
*fsp
,
788 const SMB_STRUCT_AIOCB
* const aiocb_array
[],
790 const struct timespec
*timeout
)
792 struct aio_child_list
*children
= NULL
;
793 TALLOC_CTX
*frame
= talloc_stackframe();
794 struct event_context
*ev
= NULL
;
797 bool timed_out
= false;
800 children
= init_aio_children(handle
);
801 if (children
== NULL
) {
806 /* This is a blocking call, and has to use a sub-event loop. */
807 ev
= event_context_init(frame
);
814 struct timeval tv
= convert_timespec_to_timeval(*timeout
);
815 struct tevent_timer
*te
= tevent_add_timer(ev
,
817 timeval_current_ofs(tv
.tv_sec
,
819 aio_fork_suspend_timed_out
,
827 for (i
= 0; i
< n
; i
++) {
828 struct aio_child
*child
= NULL
;
829 const SMB_STRUCT_AIOCB
*aiocb
= aiocb_array
[i
];
836 * We're going to cheat here. We know that smbd/aio.c
837 * only calls this when it's waiting for every single
838 * outstanding call to finish on a close, so just wait
839 * individually for each IO to complete. We don't care
840 * what order they finish - only that they all do. JRA.
843 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
844 struct tevent_fd
*event
;
846 if (child
->aiocb
== NULL
) {
849 if (child
->aiocb
->aio_fildes
!= fsp
->fh
->fd
) {
852 if (child
->aiocb
!= aiocb
) {
856 if (child
->aiocb
->aio_sigevent
.sigev_value
.sival_ptr
== NULL
) {
860 event
= event_add_fd(ev
,
864 handle_aio_completion
,
871 child
->called_from_suspend
= true;
873 while (!child
->completion_done
) {
874 if (tevent_loop_once(ev
) == -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 .aio_read_fn
= aio_fork_read
,
916 .aio_write_fn
= aio_fork_write
,
917 .aio_return_fn
= aio_fork_return_fn
,
918 .aio_cancel_fn
= aio_fork_cancel
,
919 .aio_error_fn
= aio_fork_error_fn
,
920 .aio_suspend_fn
= aio_fork_suspend
,
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
);