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"
27 #include "lib/async_req/async_sock.h"
28 #include "lib/util/tevent_unix.h"
29 #include "lib/sys_rw.h"
30 #include "lib/sys_rw_data.h"
31 #include "lib/msghdr.h"
33 #if !defined(HAVE_STRUCT_MSGHDR_MSG_CONTROL) && !defined(HAVE_STRUCT_MSGHDR_MSG_ACCRIGHTS)
34 # error Can not pass file descriptors
43 struct aio_fork_config
{
44 bool erratic_testing_mode
;
52 static int mmap_area_destructor(struct mmap_area
*area
)
54 munmap((void *)area
->ptr
, area
->size
);
58 static struct mmap_area
*mmap_area_init(TALLOC_CTX
*mem_ctx
, size_t size
)
60 struct mmap_area
*result
;
63 result
= talloc(mem_ctx
, struct mmap_area
);
65 DEBUG(0, ("talloc failed\n"));
69 fd
= open("/dev/zero", O_RDWR
);
71 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
76 result
->ptr
= mmap(NULL
, size
, PROT_READ
|PROT_WRITE
,
77 MAP_SHARED
|MAP_FILE
, fd
, 0);
79 if (result
->ptr
== MAP_FAILED
) {
80 DEBUG(1, ("mmap failed: %s\n", strerror(errno
)));
85 talloc_set_destructor(result
, mmap_area_destructor
);
100 static const char *cmd_type_str(enum cmd_type cmd
)
115 result
= "<UNKNOWN>";
125 bool erratic_testing_mode
;
133 struct aio_child_list
;
136 struct aio_child
*prev
, *next
;
137 struct aio_child_list
*list
;
140 struct mmap_area
*map
;
141 bool dont_delete
; /* Marked as in use since last cleanup */
145 struct aio_child_list
{
146 struct aio_child
*children
;
147 struct tevent_timer
*cleanup_event
;
150 static void free_aio_children(void **p
)
155 static ssize_t
read_fd(int fd
, void *ptr
, size_t nbytes
, int *recvfd
)
160 size_t bufsize
= msghdr_prep_recv_fds(NULL
, NULL
, 0, 1);
161 uint8_t buf
[bufsize
];
163 msghdr_prep_recv_fds(&msg
, buf
, bufsize
, 1);
168 iov
[0].iov_base
= (void *)ptr
;
169 iov
[0].iov_len
= nbytes
;
174 n
= recvmsg(fd
, &msg
, 0);
175 } while ((n
== -1) && (errno
== EINTR
));
182 size_t num_fds
= msghdr_extract_fds(&msg
, NULL
, 0);
185 msghdr_extract_fds(&msg
, fds
, num_fds
);
190 for (i
=0; i
<num_fds
; i
++) {
204 static ssize_t
write_fd(int fd
, void *ptr
, size_t nbytes
, int sendfd
)
206 struct msghdr msg
= {0};
207 size_t bufsize
= msghdr_prep_fds(NULL
, NULL
, 0, &sendfd
, 1);
208 uint8_t buf
[bufsize
];
212 msghdr_prep_fds(&msg
, buf
, bufsize
, &sendfd
, 1);
214 iov
.iov_base
= (void *)ptr
;
215 iov
.iov_len
= nbytes
;
220 sent
= sendmsg(fd
, &msg
, 0);
221 } while ((sent
== -1) && (errno
== EINTR
));
226 static void aio_child_cleanup(struct tevent_context
*event_ctx
,
227 struct tevent_timer
*te
,
231 struct aio_child_list
*list
= talloc_get_type_abort(
232 private_data
, struct aio_child_list
);
233 struct aio_child
*child
, *next
;
235 TALLOC_FREE(list
->cleanup_event
);
237 for (child
= list
->children
; child
!= NULL
; child
= next
) {
241 DEBUG(10, ("child %d currently active\n",
246 if (child
->dont_delete
) {
247 DEBUG(10, ("Child %d was active since last cleanup\n",
249 child
->dont_delete
= false;
253 DEBUG(10, ("Child %d idle for more than 30 seconds, "
254 "deleting\n", (int)child
->pid
));
260 if (list
->children
!= NULL
) {
262 * Re-schedule the next cleanup round
264 list
->cleanup_event
= tevent_add_timer(server_event_context(), list
,
265 timeval_add(&now
, 30, 0),
266 aio_child_cleanup
, list
);
271 static struct aio_child_list
*init_aio_children(struct vfs_handle_struct
*handle
)
273 struct aio_child_list
*data
= NULL
;
275 if (SMB_VFS_HANDLE_TEST_DATA(handle
)) {
276 SMB_VFS_HANDLE_GET_DATA(handle
, data
, struct aio_child_list
,
281 data
= talloc_zero(NULL
, struct aio_child_list
);
288 * Regardless of whether the child_list had been around or not, make
289 * sure that we have a cleanup timed event. This timed event will
290 * delete itself when it finds that no children are around anymore.
293 if (data
->cleanup_event
== NULL
) {
294 data
->cleanup_event
= tevent_add_timer(server_event_context(), data
,
295 timeval_current_ofs(30, 0),
296 aio_child_cleanup
, data
);
297 if (data
->cleanup_event
== NULL
) {
303 if (!SMB_VFS_HANDLE_TEST_DATA(handle
)) {
304 SMB_VFS_HANDLE_SET_DATA(handle
, data
, free_aio_children
,
305 struct aio_child_list
, return False
);
311 static void aio_child_loop(int sockfd
, struct mmap_area
*map
)
316 struct rw_cmd cmd_struct
;
317 struct rw_ret ret_struct
;
319 ret
= read_fd(sockfd
, &cmd_struct
, sizeof(cmd_struct
), &fd
);
320 if (ret
!= sizeof(cmd_struct
)) {
321 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret
,
326 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
327 cmd_type_str(cmd_struct
.cmd
),
328 (int)cmd_struct
.n
, (int)cmd_struct
.offset
, fd
));
330 if (cmd_struct
.erratic_testing_mode
) {
332 * For developer testing, we want erratic behaviour for
338 * use generate_random_buffer, we just forked from a
339 * common parent state
341 generate_random_buffer(&randval
, sizeof(randval
));
342 msecs
= randval
+ 20;
343 DEBUG(10, ("delaying for %u msecs\n", msecs
));
347 ZERO_STRUCT(ret_struct
);
349 switch (cmd_struct
.cmd
) {
351 ret_struct
.size
= sys_pread(
352 fd
, (void *)map
->ptr
, cmd_struct
.n
,
355 /* This breaks "make test" when run with aio_fork module. */
357 ret_struct
.size
= MAX(1, ret_struct
.size
* 0.9);
362 ret_struct
.size
= sys_pwrite(
363 fd
, (void *)map
->ptr
, cmd_struct
.n
,
367 ret_struct
.size
= fsync(fd
);
370 ret_struct
.size
= -1;
374 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
375 (int)ret_struct
.size
));
377 if (ret_struct
.size
== -1) {
378 ret_struct
.ret_errno
= errno
;
382 * Close the fd before telling our parent we're done. The
383 * parent might close and re-open the file very quickly, and
384 * with system-level share modes (GPFS) we would get an
385 * unjustified SHARING_VIOLATION.
389 ret
= write_data(sockfd
, (char *)&ret_struct
,
391 if (ret
!= sizeof(ret_struct
)) {
392 DEBUG(10, ("could not write ret_struct: %s\n",
399 static int aio_child_destructor(struct aio_child
*child
)
403 SMB_ASSERT(!child
->busy
);
405 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
406 child
->pid
, child
->sockfd
));
409 * closing the sockfd makes the child not return from recvmsg() on RHEL
410 * 5.5 so instead force the child to exit by writing bad data to it
412 write(child
->sockfd
, &c
, sizeof(c
));
413 close(child
->sockfd
);
414 DLIST_REMOVE(child
->list
->children
, child
);
419 * We have to close all fd's in open files, we might incorrectly hold a system
420 * level share mode on a file.
423 static struct files_struct
*close_fsp_fd(struct files_struct
*fsp
,
426 if ((fsp
->fh
!= NULL
) && (fsp
->fh
->fd
!= -1)) {
433 static int create_aio_child(struct smbd_server_connection
*sconn
,
434 struct aio_child_list
*children
,
436 struct aio_child
**presult
)
438 struct aio_child
*result
;
442 fdpair
[0] = fdpair
[1] = -1;
444 result
= talloc_zero(children
, struct aio_child
);
445 if (result
== NULL
) {
449 if (socketpair(AF_UNIX
, SOCK_STREAM
, 0, fdpair
) == -1) {
451 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno
)));
455 DEBUG(10, ("fdpair = %d/%d\n", fdpair
[0], fdpair
[1]));
457 result
->map
= mmap_area_init(result
, map_size
);
458 if (result
->map
== NULL
) {
460 DEBUG(0, ("Could not create mmap area\n"));
464 result
->pid
= fork();
465 if (result
->pid
== -1) {
467 DEBUG(0, ("fork failed: %s\n", strerror(errno
)));
471 if (result
->pid
== 0) {
473 result
->sockfd
= fdpair
[1];
474 files_forall(sconn
, close_fsp_fd
, NULL
);
475 aio_child_loop(result
->sockfd
, result
->map
);
478 DEBUG(10, ("Child %d created with sockfd %d\n",
479 result
->pid
, fdpair
[0]));
481 result
->sockfd
= fdpair
[0];
484 result
->list
= children
;
485 DLIST_ADD(children
->children
, result
);
487 talloc_set_destructor(result
, aio_child_destructor
);
494 if (fdpair
[0] != -1) close(fdpair
[0]);
495 if (fdpair
[1] != -1) close(fdpair
[1]);
501 static int get_idle_child(struct vfs_handle_struct
*handle
,
502 struct aio_child
**pchild
)
504 struct aio_child_list
*children
;
505 struct aio_child
*child
;
507 children
= init_aio_children(handle
);
508 if (children
== NULL
) {
512 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
521 DEBUG(10, ("no idle child found, creating new one\n"));
523 ret
= create_aio_child(handle
->conn
->sconn
, children
,
526 DEBUG(10, ("create_aio_child failed: %s\n",
532 child
->dont_delete
= true;
539 struct aio_fork_pread_state
{
540 struct aio_child
*child
;
545 static void aio_fork_pread_done(struct tevent_req
*subreq
);
547 static struct tevent_req
*aio_fork_pread_send(struct vfs_handle_struct
*handle
,
549 struct tevent_context
*ev
,
550 struct files_struct
*fsp
,
552 size_t n
, off_t offset
)
554 struct tevent_req
*req
, *subreq
;
555 struct aio_fork_pread_state
*state
;
559 struct aio_fork_config
*config
;
561 SMB_VFS_HANDLE_GET_DATA(handle
, config
,
562 struct aio_fork_config
,
565 req
= tevent_req_create(mem_ctx
, &state
, struct aio_fork_pread_state
);
571 /* TODO: support variable buffers */
572 tevent_req_error(req
, EINVAL
);
573 return tevent_req_post(req
, ev
);
576 err
= get_idle_child(handle
, &state
->child
);
578 tevent_req_error(req
, err
);
579 return tevent_req_post(req
, ev
);
586 cmd
.erratic_testing_mode
= config
->erratic_testing_mode
;
588 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
589 (int)state
->child
->pid
));
592 * Not making this async. We're writing into an empty unix
593 * domain socket. This should never block.
595 written
= write_fd(state
->child
->sockfd
, &cmd
, sizeof(cmd
),
600 TALLOC_FREE(state
->child
);
602 DEBUG(10, ("write_fd failed: %s\n", strerror(err
)));
603 tevent_req_error(req
, err
);
604 return tevent_req_post(req
, ev
);
607 subreq
= read_packet_send(state
, ev
, state
->child
->sockfd
,
608 sizeof(struct rw_ret
), NULL
, NULL
);
609 if (tevent_req_nomem(subreq
, req
)) {
610 TALLOC_FREE(state
->child
); /* we sent sth down */
611 return tevent_req_post(req
, ev
);
613 tevent_req_set_callback(subreq
, aio_fork_pread_done
, req
);
617 static void aio_fork_pread_done(struct tevent_req
*subreq
)
619 struct tevent_req
*req
= tevent_req_callback_data(
620 subreq
, struct tevent_req
);
621 struct aio_fork_pread_state
*state
= tevent_req_data(
622 req
, struct aio_fork_pread_state
);
626 struct rw_ret
*retbuf
;
628 nread
= read_packet_recv(subreq
, talloc_tos(), &buf
, &err
);
631 TALLOC_FREE(state
->child
);
632 tevent_req_error(req
, err
);
636 state
->child
->busy
= false;
638 retbuf
= (struct rw_ret
*)buf
;
639 state
->ret
= retbuf
->size
;
640 state
->err
= retbuf
->ret_errno
;
641 tevent_req_done(req
);
644 static ssize_t
aio_fork_pread_recv(struct tevent_req
*req
, int *err
)
646 struct aio_fork_pread_state
*state
= tevent_req_data(
647 req
, struct aio_fork_pread_state
);
649 if (tevent_req_is_unix_error(req
, err
)) {
652 if (state
->ret
== -1) {
658 struct aio_fork_pwrite_state
{
659 struct aio_child
*child
;
664 static void aio_fork_pwrite_done(struct tevent_req
*subreq
);
666 static struct tevent_req
*aio_fork_pwrite_send(
667 struct vfs_handle_struct
*handle
, TALLOC_CTX
*mem_ctx
,
668 struct tevent_context
*ev
, struct files_struct
*fsp
,
669 const void *data
, size_t n
, off_t offset
)
671 struct tevent_req
*req
, *subreq
;
672 struct aio_fork_pwrite_state
*state
;
676 struct aio_fork_config
*config
;
677 SMB_VFS_HANDLE_GET_DATA(handle
, config
,
678 struct aio_fork_config
,
681 req
= tevent_req_create(mem_ctx
, &state
, struct aio_fork_pwrite_state
);
687 /* TODO: support variable buffers */
688 tevent_req_error(req
, EINVAL
);
689 return tevent_req_post(req
, ev
);
692 err
= get_idle_child(handle
, &state
->child
);
694 tevent_req_error(req
, err
);
695 return tevent_req_post(req
, ev
);
702 cmd
.erratic_testing_mode
= config
->erratic_testing_mode
;
704 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
705 (int)state
->child
->pid
));
708 * Not making this async. We're writing into an empty unix
709 * domain socket. This should never block.
711 written
= write_fd(state
->child
->sockfd
, &cmd
, sizeof(cmd
),
716 TALLOC_FREE(state
->child
);
718 DEBUG(10, ("write_fd failed: %s\n", strerror(err
)));
719 tevent_req_error(req
, err
);
720 return tevent_req_post(req
, ev
);
723 subreq
= read_packet_send(state
, ev
, state
->child
->sockfd
,
724 sizeof(struct rw_ret
), NULL
, NULL
);
725 if (tevent_req_nomem(subreq
, req
)) {
726 TALLOC_FREE(state
->child
); /* we sent sth down */
727 return tevent_req_post(req
, ev
);
729 tevent_req_set_callback(subreq
, aio_fork_pwrite_done
, req
);
733 static void aio_fork_pwrite_done(struct tevent_req
*subreq
)
735 struct tevent_req
*req
= tevent_req_callback_data(
736 subreq
, struct tevent_req
);
737 struct aio_fork_pwrite_state
*state
= tevent_req_data(
738 req
, struct aio_fork_pwrite_state
);
742 struct rw_ret
*retbuf
;
744 nread
= read_packet_recv(subreq
, talloc_tos(), &buf
, &err
);
747 TALLOC_FREE(state
->child
);
748 tevent_req_error(req
, err
);
752 state
->child
->busy
= false;
754 retbuf
= (struct rw_ret
*)buf
;
755 state
->ret
= retbuf
->size
;
756 state
->err
= retbuf
->ret_errno
;
757 tevent_req_done(req
);
760 static ssize_t
aio_fork_pwrite_recv(struct tevent_req
*req
, int *err
)
762 struct aio_fork_pwrite_state
*state
= tevent_req_data(
763 req
, struct aio_fork_pwrite_state
);
765 if (tevent_req_is_unix_error(req
, err
)) {
768 if (state
->ret
== -1) {
774 struct aio_fork_fsync_state
{
775 struct aio_child
*child
;
780 static void aio_fork_fsync_done(struct tevent_req
*subreq
);
782 static struct tevent_req
*aio_fork_fsync_send(
783 struct vfs_handle_struct
*handle
, TALLOC_CTX
*mem_ctx
,
784 struct tevent_context
*ev
, struct files_struct
*fsp
)
786 struct tevent_req
*req
, *subreq
;
787 struct aio_fork_fsync_state
*state
;
791 struct aio_fork_config
*config
;
793 SMB_VFS_HANDLE_GET_DATA(handle
, config
,
794 struct aio_fork_config
,
797 req
= tevent_req_create(mem_ctx
, &state
, struct aio_fork_fsync_state
);
802 err
= get_idle_child(handle
, &state
->child
);
804 tevent_req_error(req
, err
);
805 return tevent_req_post(req
, ev
);
810 cmd
.erratic_testing_mode
= config
->erratic_testing_mode
;
812 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
813 (int)state
->child
->pid
));
816 * Not making this async. We're writing into an empty unix
817 * domain socket. This should never block.
819 written
= write_fd(state
->child
->sockfd
, &cmd
, sizeof(cmd
),
824 TALLOC_FREE(state
->child
);
826 DEBUG(10, ("write_fd failed: %s\n", strerror(err
)));
827 tevent_req_error(req
, err
);
828 return tevent_req_post(req
, ev
);
831 subreq
= read_packet_send(state
, ev
, state
->child
->sockfd
,
832 sizeof(struct rw_ret
), NULL
, NULL
);
833 if (tevent_req_nomem(subreq
, req
)) {
834 TALLOC_FREE(state
->child
); /* we sent sth down */
835 return tevent_req_post(req
, ev
);
837 tevent_req_set_callback(subreq
, aio_fork_fsync_done
, req
);
841 static void aio_fork_fsync_done(struct tevent_req
*subreq
)
843 struct tevent_req
*req
= tevent_req_callback_data(
844 subreq
, struct tevent_req
);
845 struct aio_fork_fsync_state
*state
= tevent_req_data(
846 req
, struct aio_fork_fsync_state
);
850 struct rw_ret
*retbuf
;
852 nread
= read_packet_recv(subreq
, talloc_tos(), &buf
, &err
);
855 TALLOC_FREE(state
->child
);
856 tevent_req_error(req
, err
);
860 state
->child
->busy
= false;
862 retbuf
= (struct rw_ret
*)buf
;
863 state
->ret
= retbuf
->size
;
864 state
->err
= retbuf
->ret_errno
;
865 tevent_req_done(req
);
868 static int aio_fork_fsync_recv(struct tevent_req
*req
, int *err
)
870 struct aio_fork_fsync_state
*state
= tevent_req_data(
871 req
, struct aio_fork_fsync_state
);
873 if (tevent_req_is_unix_error(req
, err
)) {
876 if (state
->ret
== -1) {
882 static int aio_fork_connect(vfs_handle_struct
*handle
, const char *service
,
886 struct aio_fork_config
*config
;
887 ret
= SMB_VFS_NEXT_CONNECT(handle
, service
, user
);
893 config
= talloc_zero(handle
->conn
, struct aio_fork_config
);
895 SMB_VFS_NEXT_DISCONNECT(handle
);
896 DEBUG(0, ("talloc_zero() failed\n"));
900 config
->erratic_testing_mode
= lp_parm_bool(SNUM(handle
->conn
), "vfs_aio_fork",
901 "erratic_testing_mode", false);
903 SMB_VFS_HANDLE_SET_DATA(handle
, config
,
904 NULL
, struct aio_fork_config
,
907 /*********************************************************************
908 * How many threads to initialize ?
909 * 100 per process seems insane as a default until you realize that
910 * (a) Threads terminate after 1 second when idle.
911 * (b) Throttling is done in SMB2 via the crediting algorithm.
912 * (c) SMB1 clients are limited to max_mux (50) outstanding
913 * requests and Windows clients don't use this anyway.
914 * Essentially we want this to be unlimited unless smb.conf
916 *********************************************************************/
917 aio_pending_size
= 100;
921 static struct vfs_fn_pointers vfs_aio_fork_fns
= {
922 .connect_fn
= aio_fork_connect
,
923 .pread_send_fn
= aio_fork_pread_send
,
924 .pread_recv_fn
= aio_fork_pread_recv
,
925 .pwrite_send_fn
= aio_fork_pwrite_send
,
926 .pwrite_recv_fn
= aio_fork_pwrite_recv
,
927 .fsync_send_fn
= aio_fork_fsync_send
,
928 .fsync_recv_fn
= aio_fork_fsync_recv
,
931 NTSTATUS
vfs_aio_fork_init(void);
932 NTSTATUS
vfs_aio_fork_init(void)
934 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION
,
935 "aio_fork", &vfs_aio_fork_fns
);