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(discard_const(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
)
158 struct msghdr msg
= { .msg_iov
= iov
, .msg_iovlen
= 1 };
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);
165 iov
[0].iov_base
= (void *)ptr
;
166 iov
[0].iov_len
= nbytes
;
169 n
= recvmsg(fd
, &msg
, 0);
170 } while ((n
== -1) && (errno
== EINTR
));
177 size_t num_fds
= msghdr_extract_fds(&msg
, NULL
, 0);
180 msghdr_extract_fds(&msg
, fds
, num_fds
);
185 for (i
=0; i
<num_fds
; i
++) {
199 static ssize_t
write_fd(int fd
, void *ptr
, size_t nbytes
, int sendfd
)
201 struct msghdr msg
= {0};
202 size_t bufsize
= msghdr_prep_fds(NULL
, NULL
, 0, &sendfd
, 1);
203 uint8_t buf
[bufsize
];
207 msghdr_prep_fds(&msg
, buf
, bufsize
, &sendfd
, 1);
209 iov
.iov_base
= (void *)ptr
;
210 iov
.iov_len
= nbytes
;
215 sent
= sendmsg(fd
, &msg
, 0);
216 } while ((sent
== -1) && (errno
== EINTR
));
221 static void aio_child_cleanup(struct tevent_context
*event_ctx
,
222 struct tevent_timer
*te
,
226 struct aio_child_list
*list
= talloc_get_type_abort(
227 private_data
, struct aio_child_list
);
228 struct aio_child
*child
, *next
;
230 TALLOC_FREE(list
->cleanup_event
);
232 for (child
= list
->children
; child
!= NULL
; child
= next
) {
236 DEBUG(10, ("child %d currently active\n",
241 if (child
->dont_delete
) {
242 DEBUG(10, ("Child %d was active since last cleanup\n",
244 child
->dont_delete
= false;
248 DEBUG(10, ("Child %d idle for more than 30 seconds, "
249 "deleting\n", (int)child
->pid
));
255 if (list
->children
!= NULL
) {
257 * Re-schedule the next cleanup round
259 list
->cleanup_event
= tevent_add_timer(server_event_context(), list
,
260 timeval_add(&now
, 30, 0),
261 aio_child_cleanup
, list
);
266 static struct aio_child_list
*init_aio_children(struct vfs_handle_struct
*handle
)
268 struct aio_child_list
*data
= NULL
;
270 if (SMB_VFS_HANDLE_TEST_DATA(handle
)) {
271 SMB_VFS_HANDLE_GET_DATA(handle
, data
, struct aio_child_list
,
276 data
= talloc_zero(NULL
, struct aio_child_list
);
283 * Regardless of whether the child_list had been around or not, make
284 * sure that we have a cleanup timed event. This timed event will
285 * delete itself when it finds that no children are around anymore.
288 if (data
->cleanup_event
== NULL
) {
289 data
->cleanup_event
= tevent_add_timer(server_event_context(), data
,
290 timeval_current_ofs(30, 0),
291 aio_child_cleanup
, data
);
292 if (data
->cleanup_event
== NULL
) {
298 if (!SMB_VFS_HANDLE_TEST_DATA(handle
)) {
299 SMB_VFS_HANDLE_SET_DATA(handle
, data
, free_aio_children
,
300 struct aio_child_list
, return False
);
306 static void aio_child_loop(int sockfd
, struct mmap_area
*map
)
311 struct rw_cmd cmd_struct
;
312 struct rw_ret ret_struct
;
314 ret
= read_fd(sockfd
, &cmd_struct
, sizeof(cmd_struct
), &fd
);
315 if (ret
!= sizeof(cmd_struct
)) {
316 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret
,
321 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
322 cmd_type_str(cmd_struct
.cmd
),
323 (int)cmd_struct
.n
, (int)cmd_struct
.offset
, fd
));
325 if (cmd_struct
.erratic_testing_mode
) {
327 * For developer testing, we want erratic behaviour for
333 * use generate_random_buffer, we just forked from a
334 * common parent state
336 generate_random_buffer(&randval
, sizeof(randval
));
337 msecs
= randval
+ 20;
338 DEBUG(10, ("delaying for %u msecs\n", msecs
));
342 ZERO_STRUCT(ret_struct
);
344 switch (cmd_struct
.cmd
) {
346 ret_struct
.size
= sys_pread(
347 fd
, discard_const(map
->ptr
), cmd_struct
.n
,
350 /* This breaks "make test" when run with aio_fork module. */
352 ret_struct
.size
= MAX(1, ret_struct
.size
* 0.9);
357 ret_struct
.size
= sys_pwrite(
358 fd
, discard_const(map
->ptr
), cmd_struct
.n
,
362 ret_struct
.size
= fsync(fd
);
365 ret_struct
.size
= -1;
369 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
370 (int)ret_struct
.size
));
372 if (ret_struct
.size
== -1) {
373 ret_struct
.ret_errno
= errno
;
377 * Close the fd before telling our parent we're done. The
378 * parent might close and re-open the file very quickly, and
379 * with system-level share modes (GPFS) we would get an
380 * unjustified SHARING_VIOLATION.
384 ret
= write_data(sockfd
, (char *)&ret_struct
,
386 if (ret
!= sizeof(ret_struct
)) {
387 DEBUG(10, ("could not write ret_struct: %s\n",
394 static int aio_child_destructor(struct aio_child
*child
)
398 SMB_ASSERT(!child
->busy
);
400 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
401 (int)child
->pid
, child
->sockfd
));
404 * closing the sockfd makes the child not return from recvmsg() on RHEL
405 * 5.5 so instead force the child to exit by writing bad data to it
407 write(child
->sockfd
, &c
, sizeof(c
));
408 close(child
->sockfd
);
409 DLIST_REMOVE(child
->list
->children
, child
);
414 * We have to close all fd's in open files, we might incorrectly hold a system
415 * level share mode on a file.
418 static struct files_struct
*close_fsp_fd(struct files_struct
*fsp
,
421 if ((fsp
->fh
!= NULL
) && (fsp
->fh
->fd
!= -1)) {
428 static int create_aio_child(struct smbd_server_connection
*sconn
,
429 struct aio_child_list
*children
,
431 struct aio_child
**presult
)
433 struct aio_child
*result
;
437 fdpair
[0] = fdpair
[1] = -1;
439 result
= talloc_zero(children
, struct aio_child
);
440 if (result
== NULL
) {
444 if (socketpair(AF_UNIX
, SOCK_STREAM
, 0, fdpair
) == -1) {
446 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno
)));
450 DEBUG(10, ("fdpair = %d/%d\n", fdpair
[0], fdpair
[1]));
452 result
->map
= mmap_area_init(result
, map_size
);
453 if (result
->map
== NULL
) {
455 DEBUG(0, ("Could not create mmap area\n"));
459 result
->pid
= fork();
460 if (result
->pid
== -1) {
462 DEBUG(0, ("fork failed: %s\n", strerror(errno
)));
466 if (result
->pid
== 0) {
468 result
->sockfd
= fdpair
[1];
469 files_forall(sconn
, close_fsp_fd
, NULL
);
470 aio_child_loop(result
->sockfd
, result
->map
);
473 DEBUG(10, ("Child %d created with sockfd %d\n",
474 (int)result
->pid
, fdpair
[0]));
476 result
->sockfd
= fdpair
[0];
479 result
->list
= children
;
480 DLIST_ADD(children
->children
, result
);
482 talloc_set_destructor(result
, aio_child_destructor
);
489 if (fdpair
[0] != -1) close(fdpair
[0]);
490 if (fdpair
[1] != -1) close(fdpair
[1]);
496 static int get_idle_child(struct vfs_handle_struct
*handle
,
497 struct aio_child
**pchild
)
499 struct aio_child_list
*children
;
500 struct aio_child
*child
;
502 children
= init_aio_children(handle
);
503 if (children
== NULL
) {
507 for (child
= children
->children
; child
!= NULL
; child
= child
->next
) {
516 DEBUG(10, ("no idle child found, creating new one\n"));
518 ret
= create_aio_child(handle
->conn
->sconn
, children
,
521 DEBUG(10, ("create_aio_child failed: %s\n",
527 child
->dont_delete
= true;
534 struct aio_fork_pread_state
{
535 struct aio_child
*child
;
540 static void aio_fork_pread_done(struct tevent_req
*subreq
);
542 static struct tevent_req
*aio_fork_pread_send(struct vfs_handle_struct
*handle
,
544 struct tevent_context
*ev
,
545 struct files_struct
*fsp
,
547 size_t n
, off_t offset
)
549 struct tevent_req
*req
, *subreq
;
550 struct aio_fork_pread_state
*state
;
554 struct aio_fork_config
*config
;
556 SMB_VFS_HANDLE_GET_DATA(handle
, config
,
557 struct aio_fork_config
,
560 req
= tevent_req_create(mem_ctx
, &state
, struct aio_fork_pread_state
);
566 /* TODO: support variable buffers */
567 tevent_req_error(req
, EINVAL
);
568 return tevent_req_post(req
, ev
);
571 err
= get_idle_child(handle
, &state
->child
);
573 tevent_req_error(req
, err
);
574 return tevent_req_post(req
, ev
);
581 cmd
.erratic_testing_mode
= config
->erratic_testing_mode
;
583 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
584 (int)state
->child
->pid
));
587 * Not making this async. We're writing into an empty unix
588 * domain socket. This should never block.
590 written
= write_fd(state
->child
->sockfd
, &cmd
, sizeof(cmd
),
595 TALLOC_FREE(state
->child
);
597 DEBUG(10, ("write_fd failed: %s\n", strerror(err
)));
598 tevent_req_error(req
, err
);
599 return tevent_req_post(req
, ev
);
602 subreq
= read_packet_send(state
, ev
, state
->child
->sockfd
,
603 sizeof(struct rw_ret
), NULL
, NULL
);
604 if (tevent_req_nomem(subreq
, req
)) {
605 TALLOC_FREE(state
->child
); /* we sent sth down */
606 return tevent_req_post(req
, ev
);
608 tevent_req_set_callback(subreq
, aio_fork_pread_done
, req
);
612 static void aio_fork_pread_done(struct tevent_req
*subreq
)
614 struct tevent_req
*req
= tevent_req_callback_data(
615 subreq
, struct tevent_req
);
616 struct aio_fork_pread_state
*state
= tevent_req_data(
617 req
, struct aio_fork_pread_state
);
621 struct rw_ret
*retbuf
;
623 nread
= read_packet_recv(subreq
, talloc_tos(), &buf
, &err
);
626 TALLOC_FREE(state
->child
);
627 tevent_req_error(req
, err
);
631 state
->child
->busy
= false;
633 retbuf
= (struct rw_ret
*)buf
;
634 state
->ret
= retbuf
->size
;
635 state
->err
= retbuf
->ret_errno
;
636 tevent_req_done(req
);
639 static ssize_t
aio_fork_pread_recv(struct tevent_req
*req
, int *err
)
641 struct aio_fork_pread_state
*state
= tevent_req_data(
642 req
, struct aio_fork_pread_state
);
644 if (tevent_req_is_unix_error(req
, err
)) {
647 if (state
->ret
== -1) {
653 struct aio_fork_pwrite_state
{
654 struct aio_child
*child
;
659 static void aio_fork_pwrite_done(struct tevent_req
*subreq
);
661 static struct tevent_req
*aio_fork_pwrite_send(
662 struct vfs_handle_struct
*handle
, TALLOC_CTX
*mem_ctx
,
663 struct tevent_context
*ev
, struct files_struct
*fsp
,
664 const void *data
, size_t n
, off_t offset
)
666 struct tevent_req
*req
, *subreq
;
667 struct aio_fork_pwrite_state
*state
;
671 struct aio_fork_config
*config
;
672 SMB_VFS_HANDLE_GET_DATA(handle
, config
,
673 struct aio_fork_config
,
676 req
= tevent_req_create(mem_ctx
, &state
, struct aio_fork_pwrite_state
);
682 /* TODO: support variable buffers */
683 tevent_req_error(req
, EINVAL
);
684 return tevent_req_post(req
, ev
);
687 err
= get_idle_child(handle
, &state
->child
);
689 tevent_req_error(req
, err
);
690 return tevent_req_post(req
, ev
);
697 cmd
.erratic_testing_mode
= config
->erratic_testing_mode
;
699 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
700 (int)state
->child
->pid
));
703 * Not making this async. We're writing into an empty unix
704 * domain socket. This should never block.
706 written
= write_fd(state
->child
->sockfd
, &cmd
, sizeof(cmd
),
711 TALLOC_FREE(state
->child
);
713 DEBUG(10, ("write_fd failed: %s\n", strerror(err
)));
714 tevent_req_error(req
, err
);
715 return tevent_req_post(req
, ev
);
718 subreq
= read_packet_send(state
, ev
, state
->child
->sockfd
,
719 sizeof(struct rw_ret
), NULL
, NULL
);
720 if (tevent_req_nomem(subreq
, req
)) {
721 TALLOC_FREE(state
->child
); /* we sent sth down */
722 return tevent_req_post(req
, ev
);
724 tevent_req_set_callback(subreq
, aio_fork_pwrite_done
, req
);
728 static void aio_fork_pwrite_done(struct tevent_req
*subreq
)
730 struct tevent_req
*req
= tevent_req_callback_data(
731 subreq
, struct tevent_req
);
732 struct aio_fork_pwrite_state
*state
= tevent_req_data(
733 req
, struct aio_fork_pwrite_state
);
737 struct rw_ret
*retbuf
;
739 nread
= read_packet_recv(subreq
, talloc_tos(), &buf
, &err
);
742 TALLOC_FREE(state
->child
);
743 tevent_req_error(req
, err
);
747 state
->child
->busy
= false;
749 retbuf
= (struct rw_ret
*)buf
;
750 state
->ret
= retbuf
->size
;
751 state
->err
= retbuf
->ret_errno
;
752 tevent_req_done(req
);
755 static ssize_t
aio_fork_pwrite_recv(struct tevent_req
*req
, int *err
)
757 struct aio_fork_pwrite_state
*state
= tevent_req_data(
758 req
, struct aio_fork_pwrite_state
);
760 if (tevent_req_is_unix_error(req
, err
)) {
763 if (state
->ret
== -1) {
769 struct aio_fork_fsync_state
{
770 struct aio_child
*child
;
775 static void aio_fork_fsync_done(struct tevent_req
*subreq
);
777 static struct tevent_req
*aio_fork_fsync_send(
778 struct vfs_handle_struct
*handle
, TALLOC_CTX
*mem_ctx
,
779 struct tevent_context
*ev
, struct files_struct
*fsp
)
781 struct tevent_req
*req
, *subreq
;
782 struct aio_fork_fsync_state
*state
;
786 struct aio_fork_config
*config
;
788 SMB_VFS_HANDLE_GET_DATA(handle
, config
,
789 struct aio_fork_config
,
792 req
= tevent_req_create(mem_ctx
, &state
, struct aio_fork_fsync_state
);
797 err
= get_idle_child(handle
, &state
->child
);
799 tevent_req_error(req
, err
);
800 return tevent_req_post(req
, ev
);
805 cmd
.erratic_testing_mode
= config
->erratic_testing_mode
;
807 DEBUG(10, ("sending fd %d to child %d\n", fsp
->fh
->fd
,
808 (int)state
->child
->pid
));
811 * Not making this async. We're writing into an empty unix
812 * domain socket. This should never block.
814 written
= write_fd(state
->child
->sockfd
, &cmd
, sizeof(cmd
),
819 TALLOC_FREE(state
->child
);
821 DEBUG(10, ("write_fd failed: %s\n", strerror(err
)));
822 tevent_req_error(req
, err
);
823 return tevent_req_post(req
, ev
);
826 subreq
= read_packet_send(state
, ev
, state
->child
->sockfd
,
827 sizeof(struct rw_ret
), NULL
, NULL
);
828 if (tevent_req_nomem(subreq
, req
)) {
829 TALLOC_FREE(state
->child
); /* we sent sth down */
830 return tevent_req_post(req
, ev
);
832 tevent_req_set_callback(subreq
, aio_fork_fsync_done
, req
);
836 static void aio_fork_fsync_done(struct tevent_req
*subreq
)
838 struct tevent_req
*req
= tevent_req_callback_data(
839 subreq
, struct tevent_req
);
840 struct aio_fork_fsync_state
*state
= tevent_req_data(
841 req
, struct aio_fork_fsync_state
);
845 struct rw_ret
*retbuf
;
847 nread
= read_packet_recv(subreq
, talloc_tos(), &buf
, &err
);
850 TALLOC_FREE(state
->child
);
851 tevent_req_error(req
, err
);
855 state
->child
->busy
= false;
857 retbuf
= (struct rw_ret
*)buf
;
858 state
->ret
= retbuf
->size
;
859 state
->err
= retbuf
->ret_errno
;
860 tevent_req_done(req
);
863 static int aio_fork_fsync_recv(struct tevent_req
*req
, int *err
)
865 struct aio_fork_fsync_state
*state
= tevent_req_data(
866 req
, struct aio_fork_fsync_state
);
868 if (tevent_req_is_unix_error(req
, err
)) {
871 if (state
->ret
== -1) {
877 static int aio_fork_connect(vfs_handle_struct
*handle
, const char *service
,
881 struct aio_fork_config
*config
;
882 ret
= SMB_VFS_NEXT_CONNECT(handle
, service
, user
);
888 config
= talloc_zero(handle
->conn
, struct aio_fork_config
);
890 SMB_VFS_NEXT_DISCONNECT(handle
);
891 DEBUG(0, ("talloc_zero() failed\n"));
895 config
->erratic_testing_mode
= lp_parm_bool(SNUM(handle
->conn
), "vfs_aio_fork",
896 "erratic_testing_mode", false);
898 SMB_VFS_HANDLE_SET_DATA(handle
, config
,
899 NULL
, struct aio_fork_config
,
902 /*********************************************************************
903 * How many threads to initialize ?
904 * 100 per process seems insane as a default until you realize that
905 * (a) Threads terminate after 1 second when idle.
906 * (b) Throttling is done in SMB2 via the crediting algorithm.
907 * (c) SMB1 clients are limited to max_mux (50) outstanding
908 * requests and Windows clients don't use this anyway.
909 * Essentially we want this to be unlimited unless smb.conf
911 *********************************************************************/
912 aio_pending_size
= 100;
916 static struct vfs_fn_pointers vfs_aio_fork_fns
= {
917 .connect_fn
= aio_fork_connect
,
918 .pread_send_fn
= aio_fork_pread_send
,
919 .pread_recv_fn
= aio_fork_pread_recv
,
920 .pwrite_send_fn
= aio_fork_pwrite_send
,
921 .pwrite_recv_fn
= aio_fork_pwrite_recv
,
922 .fsync_send_fn
= aio_fork_fsync_send
,
923 .fsync_recv_fn
= aio_fork_fsync_recv
,
926 NTSTATUS
vfs_aio_fork_init(void);
927 NTSTATUS
vfs_aio_fork_init(void)
929 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION
,
930 "aio_fork", &vfs_aio_fork_fns
);