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/util/sys_rw.h"
30 #include "lib/util/sys_rw_data.h"
31 #include "lib/util/msghdr.h"
32 #include "smbprofile.h"
33 #include "lib/global_contexts.h"
35 #if !defined(HAVE_STRUCT_MSGHDR_MSG_CONTROL) && !defined(HAVE_STRUCT_MSGHDR_MSG_ACCRIGHTS)
36 # error Can not pass file descriptors
45 struct aio_child_list;
47 struct aio_fork_config {
48 bool erratic_testing_mode;
49 struct aio_child_list *children;
57 static int mmap_area_destructor(struct mmap_area *area)
59 munmap(discard_const(area->ptr), area->size);
63 static struct mmap_area *mmap_area_init(TALLOC_CTX *mem_ctx, size_t size)
65 struct mmap_area *result;
68 result = talloc(mem_ctx, struct mmap_area);
70 DEBUG(0, ("talloc failed\n"));
74 fd = open("/dev/zero", O_RDWR);
76 DEBUG(3, ("open(\"/dev/zero\") failed: %s\n",
81 result->ptr = mmap(NULL, size, PROT_READ|PROT_WRITE,
82 MAP_SHARED|MAP_FILE, fd, 0);
84 if (result->ptr == MAP_FAILED) {
85 DEBUG(1, ("mmap failed: %s\n", strerror(errno)));
90 talloc_set_destructor(result, mmap_area_destructor);
105 static const char *cmd_type_str(enum cmd_type cmd)
120 result = "<UNKNOWN>";
130 bool erratic_testing_mode;
139 struct aio_child_list;
142 struct aio_child *prev, *next;
143 struct aio_child_list *list;
146 struct mmap_area *map;
147 bool dont_delete; /* Marked as in use since last cleanup */
151 struct aio_child_list {
152 struct aio_child *children;
153 struct tevent_timer *cleanup_event;
156 static ssize_t read_fd(int fd, void *ptr, size_t nbytes, int *recvfd)
159 struct msghdr msg = { .msg_iov = iov, .msg_iovlen = 1 };
161 size_t bufsize = msghdr_prep_recv_fds(NULL, NULL, 0, 1);
162 uint8_t buf[bufsize];
164 msghdr_prep_recv_fds(&msg, buf, bufsize, 1);
166 iov[0].iov_base = (void *)ptr;
167 iov[0].iov_len = nbytes;
170 n = recvmsg(fd, &msg, 0);
171 } while ((n == -1) && (errno == EINTR));
178 size_t num_fds = msghdr_extract_fds(&msg, NULL, 0);
181 msghdr_extract_fds(&msg, fds, num_fds);
186 for (i=0; i<num_fds; i++) {
200 static ssize_t write_fd(int fd, void *ptr, size_t nbytes, int sendfd)
202 struct msghdr msg = {0};
203 size_t bufsize = msghdr_prep_fds(NULL, NULL, 0, &sendfd, 1);
204 uint8_t buf[bufsize];
208 msghdr_prep_fds(&msg, buf, bufsize, &sendfd, 1);
210 iov.iov_base = (void *)ptr;
211 iov.iov_len = nbytes;
216 sent = sendmsg(fd, &msg, 0);
217 } while ((sent == -1) && (errno == EINTR));
222 static void aio_child_cleanup(struct tevent_context *event_ctx,
223 struct tevent_timer *te,
227 struct aio_child_list *list = talloc_get_type_abort(
228 private_data, struct aio_child_list);
229 struct aio_child *child, *next;
231 TALLOC_FREE(list->cleanup_event);
233 for (child = list->children; child != NULL; child = next) {
237 DEBUG(10, ("child %d currently active\n",
242 if (child->dont_delete) {
243 DEBUG(10, ("Child %d was active since last cleanup\n",
245 child->dont_delete = false;
249 DEBUG(10, ("Child %d idle for more than 30 seconds, "
250 "deleting\n", (int)child->pid));
256 if (list->children != NULL) {
258 * Re-schedule the next cleanup round
260 list->cleanup_event = tevent_add_timer(global_event_context(), list,
261 timeval_add(&now, 30, 0),
262 aio_child_cleanup, list);
267 static struct aio_child_list *init_aio_children(struct vfs_handle_struct *handle)
269 struct aio_fork_config *config;
270 struct aio_child_list *children;
272 SMB_VFS_HANDLE_GET_DATA(handle, config, struct aio_fork_config,
275 if (config->children == NULL) {
276 config->children = talloc_zero(config, struct aio_child_list);
277 if (config->children == NULL) {
281 children = config->children;
284 * Regardless of whether the child_list had been around or not, make
285 * sure that we have a cleanup timed event. This timed event will
286 * delete itself when it finds that no children are around anymore.
289 if (children->cleanup_event == NULL) {
290 children->cleanup_event =
291 tevent_add_timer(global_event_context(), children,
292 timeval_current_ofs(30, 0),
293 aio_child_cleanup, children);
294 if (children->cleanup_event == NULL) {
295 TALLOC_FREE(config->children);
303 static void aio_child_loop(int sockfd, struct mmap_area *map)
308 struct rw_cmd cmd_struct;
309 struct rw_ret ret_struct;
310 struct timespec start, end;
312 ret = read_fd(sockfd, &cmd_struct, sizeof(cmd_struct), &fd);
313 if (ret != sizeof(cmd_struct)) {
314 DEBUG(10, ("read_fd returned %d: %s\n", (int)ret,
319 DEBUG(10, ("aio_child_loop: %s %d bytes at %d from fd %d\n",
320 cmd_type_str(cmd_struct.cmd),
321 (int)cmd_struct.n, (int)cmd_struct.offset, fd));
323 if (cmd_struct.erratic_testing_mode) {
325 * For developer testing, we want erratic behaviour for
331 * use generate_random_buffer, we just forked from a
332 * common parent state
334 generate_random_buffer(&randval, sizeof(randval));
335 msecs = (randval%20)+1;
336 DEBUG(10, ("delaying for %u msecs\n", msecs));
340 ZERO_STRUCT(ret_struct);
342 PROFILE_TIMESTAMP(&start);
344 switch (cmd_struct.cmd) {
346 ret_struct.size = sys_pread_full(
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_full(
358 fd, discard_const(map->ptr), cmd_struct.n,
362 ret_struct.size = fsync(fd);
365 ret_struct.size = -1;
369 PROFILE_TIMESTAMP(&end);
370 ret_struct.duration = nsec_time_diff(&end, &start);
371 DEBUG(10, ("aio_child_loop: syscall returned %d\n",
372 (int)ret_struct.size));
374 if (ret_struct.size == -1) {
375 ret_struct.ret_errno = errno;
379 * Close the fd before telling our parent we're done. The
380 * parent might close and re-open the file very quickly, and
381 * with system-level share modes (GPFS) we would get an
382 * unjustified SHARING_VIOLATION.
386 ret = write_data(sockfd, (char *)&ret_struct,
388 if (ret != sizeof(ret_struct)) {
389 DEBUG(10, ("could not write ret_struct: %s\n",
396 static int aio_child_destructor(struct aio_child *child)
400 SMB_ASSERT(!child->busy);
402 DEBUG(10, ("aio_child_destructor: removing child %d on fd %d\n",
403 (int)child->pid, child->sockfd));
406 * closing the sockfd makes the child not return from recvmsg() on RHEL
407 * 5.5 so instead force the child to exit by writing bad data to it
409 sys_write_v(child->sockfd, &c, sizeof(c));
410 close(child->sockfd);
411 DLIST_REMOVE(child->list->children, child);
416 * We have to close all fd's in open files, we might incorrectly hold a system
417 * level share mode on a file.
420 static struct files_struct *close_fsp_fd(struct files_struct *fsp,
423 if ((fsp->fh != NULL) && (fsp_get_pathref_fd(fsp) != -1)) {
424 close(fsp_get_pathref_fd(fsp));
430 static int create_aio_child(struct smbd_server_connection *sconn,
431 struct aio_child_list *children,
433 struct aio_child **presult)
435 struct aio_child *result;
439 fdpair[0] = fdpair[1] = -1;
441 result = talloc_zero(children, struct aio_child);
442 if (result == NULL) {
446 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdpair) == -1) {
448 DEBUG(10, ("socketpair() failed: %s\n", strerror(errno)));
452 DEBUG(10, ("fdpair = %d/%d\n", fdpair[0], fdpair[1]));
454 result->map = mmap_area_init(result, map_size);
455 if (result->map == NULL) {
457 DEBUG(0, ("Could not create mmap area\n"));
461 result->pid = fork();
462 if (result->pid == -1) {
464 DEBUG(0, ("fork failed: %s\n", strerror(errno)));
468 if (result->pid == 0) {
470 result->sockfd = fdpair[1];
471 files_forall(sconn, close_fsp_fd, NULL);
472 aio_child_loop(result->sockfd, result->map);
475 DEBUG(10, ("Child %d created with sockfd %d\n",
476 (int)result->pid, fdpair[0]));
478 result->sockfd = fdpair[0];
481 result->list = children;
482 DLIST_ADD(children->children, result);
484 talloc_set_destructor(result, aio_child_destructor);
491 if (fdpair[0] != -1) close(fdpair[0]);
492 if (fdpair[1] != -1) close(fdpair[1]);
498 static int get_idle_child(struct vfs_handle_struct *handle,
499 struct aio_child **pchild)
501 struct aio_child_list *children;
502 struct aio_child *child;
504 children = init_aio_children(handle);
505 if (children == NULL) {
509 for (child = children->children; child != NULL; child = child->next) {
518 DEBUG(10, ("no idle child found, creating new one\n"));
520 ret = create_aio_child(handle->conn->sconn, children,
523 DEBUG(10, ("create_aio_child failed: %s\n",
529 child->dont_delete = true;
536 struct aio_fork_pread_state {
537 struct aio_child *child;
541 struct vfs_aio_state vfs_aio_state;
544 static void aio_fork_pread_done(struct tevent_req *subreq);
546 static struct tevent_req *aio_fork_pread_send(struct vfs_handle_struct *handle,
548 struct tevent_context *ev,
549 struct files_struct *fsp,
551 size_t n, off_t offset)
553 struct tevent_req *req, *subreq;
554 struct aio_fork_pread_state *state;
558 struct aio_fork_config *config;
560 SMB_VFS_HANDLE_GET_DATA(handle, config,
561 struct aio_fork_config,
564 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pread_state);
572 /* TODO: support variable buffers */
573 tevent_req_error(req, EINVAL);
574 return tevent_req_post(req, ev);
577 err = get_idle_child(handle, &state->child);
579 tevent_req_error(req, err);
580 return tevent_req_post(req, ev);
587 cmd.erratic_testing_mode = config->erratic_testing_mode;
589 DEBUG(10, ("sending fd %d to child %d\n", fsp_get_io_fd(fsp),
590 (int)state->child->pid));
593 * Not making this async. We're writing into an empty unix
594 * domain socket. This should never block.
596 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
601 TALLOC_FREE(state->child);
603 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
604 tevent_req_error(req, err);
605 return tevent_req_post(req, ev);
608 subreq = read_packet_send(state, ev, state->child->sockfd,
609 sizeof(struct rw_ret), NULL, NULL);
610 if (tevent_req_nomem(subreq, req)) {
611 TALLOC_FREE(state->child); /* we sent sth down */
612 return tevent_req_post(req, ev);
614 tevent_req_set_callback(subreq, aio_fork_pread_done, req);
618 static void aio_fork_pread_done(struct tevent_req *subreq)
620 struct tevent_req *req = tevent_req_callback_data(
621 subreq, struct tevent_req);
622 struct aio_fork_pread_state *state = tevent_req_data(
623 req, struct aio_fork_pread_state);
627 struct rw_ret *retbuf;
629 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
632 TALLOC_FREE(state->child);
633 tevent_req_error(req, err);
637 retbuf = (struct rw_ret *)buf;
638 state->ret = retbuf->size;
639 state->vfs_aio_state.error = retbuf->ret_errno;
640 state->vfs_aio_state.duration = retbuf->duration;
642 if ((size_t)state->ret > state->n) {
643 tevent_req_error(req, EIO);
644 state->child->busy = false;
647 memcpy(state->data, state->child->map->ptr, state->ret);
649 state->child->busy = false;
651 tevent_req_done(req);
654 static ssize_t aio_fork_pread_recv(struct tevent_req *req,
655 struct vfs_aio_state *vfs_aio_state)
657 struct aio_fork_pread_state *state = tevent_req_data(
658 req, struct aio_fork_pread_state);
660 if (tevent_req_is_unix_error(req, &vfs_aio_state->error)) {
663 *vfs_aio_state = state->vfs_aio_state;
667 struct aio_fork_pwrite_state {
668 struct aio_child *child;
670 struct vfs_aio_state vfs_aio_state;
673 static void aio_fork_pwrite_done(struct tevent_req *subreq);
675 static struct tevent_req *aio_fork_pwrite_send(
676 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
677 struct tevent_context *ev, struct files_struct *fsp,
678 const void *data, size_t n, off_t offset)
680 struct tevent_req *req, *subreq;
681 struct aio_fork_pwrite_state *state;
685 struct aio_fork_config *config;
686 SMB_VFS_HANDLE_GET_DATA(handle, config,
687 struct aio_fork_config,
690 req = tevent_req_create(mem_ctx, &state, struct aio_fork_pwrite_state);
696 /* TODO: support variable buffers */
697 tevent_req_error(req, EINVAL);
698 return tevent_req_post(req, ev);
701 err = get_idle_child(handle, &state->child);
703 tevent_req_error(req, err);
704 return tevent_req_post(req, ev);
707 memcpy(state->child->map->ptr, data, n);
713 cmd.erratic_testing_mode = config->erratic_testing_mode;
715 DEBUG(10, ("sending fd %d to child %d\n", fsp_get_io_fd(fsp),
716 (int)state->child->pid));
719 * Not making this async. We're writing into an empty unix
720 * domain socket. This should never block.
722 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
727 TALLOC_FREE(state->child);
729 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
730 tevent_req_error(req, err);
731 return tevent_req_post(req, ev);
734 subreq = read_packet_send(state, ev, state->child->sockfd,
735 sizeof(struct rw_ret), NULL, NULL);
736 if (tevent_req_nomem(subreq, req)) {
737 TALLOC_FREE(state->child); /* we sent sth down */
738 return tevent_req_post(req, ev);
740 tevent_req_set_callback(subreq, aio_fork_pwrite_done, req);
744 static void aio_fork_pwrite_done(struct tevent_req *subreq)
746 struct tevent_req *req = tevent_req_callback_data(
747 subreq, struct tevent_req);
748 struct aio_fork_pwrite_state *state = tevent_req_data(
749 req, struct aio_fork_pwrite_state);
753 struct rw_ret *retbuf;
755 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
758 TALLOC_FREE(state->child);
759 tevent_req_error(req, err);
763 state->child->busy = false;
765 retbuf = (struct rw_ret *)buf;
766 state->ret = retbuf->size;
767 state->vfs_aio_state.error = retbuf->ret_errno;
768 state->vfs_aio_state.duration = retbuf->duration;
769 tevent_req_done(req);
772 static ssize_t aio_fork_pwrite_recv(struct tevent_req *req,
773 struct vfs_aio_state *vfs_aio_state)
775 struct aio_fork_pwrite_state *state = tevent_req_data(
776 req, struct aio_fork_pwrite_state);
778 if (tevent_req_is_unix_error(req, &vfs_aio_state->error)) {
781 *vfs_aio_state = state->vfs_aio_state;
785 struct aio_fork_fsync_state {
786 struct aio_child *child;
788 struct vfs_aio_state vfs_aio_state;
791 static void aio_fork_fsync_done(struct tevent_req *subreq);
793 static struct tevent_req *aio_fork_fsync_send(
794 struct vfs_handle_struct *handle, TALLOC_CTX *mem_ctx,
795 struct tevent_context *ev, struct files_struct *fsp)
797 struct tevent_req *req, *subreq;
798 struct aio_fork_fsync_state *state;
802 struct aio_fork_config *config;
804 SMB_VFS_HANDLE_GET_DATA(handle, config,
805 struct aio_fork_config,
808 req = tevent_req_create(mem_ctx, &state, struct aio_fork_fsync_state);
813 err = get_idle_child(handle, &state->child);
815 tevent_req_error(req, err);
816 return tevent_req_post(req, ev);
821 cmd.erratic_testing_mode = config->erratic_testing_mode;
823 DEBUG(10, ("sending fd %d to child %d\n", fsp_get_io_fd(fsp),
824 (int)state->child->pid));
827 * Not making this async. We're writing into an empty unix
828 * domain socket. This should never block.
830 written = write_fd(state->child->sockfd, &cmd, sizeof(cmd),
835 TALLOC_FREE(state->child);
837 DEBUG(10, ("write_fd failed: %s\n", strerror(err)));
838 tevent_req_error(req, err);
839 return tevent_req_post(req, ev);
842 subreq = read_packet_send(state, ev, state->child->sockfd,
843 sizeof(struct rw_ret), NULL, NULL);
844 if (tevent_req_nomem(subreq, req)) {
845 TALLOC_FREE(state->child); /* we sent sth down */
846 return tevent_req_post(req, ev);
848 tevent_req_set_callback(subreq, aio_fork_fsync_done, req);
852 static void aio_fork_fsync_done(struct tevent_req *subreq)
854 struct tevent_req *req = tevent_req_callback_data(
855 subreq, struct tevent_req);
856 struct aio_fork_fsync_state *state = tevent_req_data(
857 req, struct aio_fork_fsync_state);
861 struct rw_ret *retbuf;
863 nread = read_packet_recv(subreq, talloc_tos(), &buf, &err);
866 TALLOC_FREE(state->child);
867 tevent_req_error(req, err);
871 state->child->busy = false;
873 retbuf = (struct rw_ret *)buf;
874 state->ret = retbuf->size;
875 state->vfs_aio_state.error = retbuf->ret_errno;
876 state->vfs_aio_state.duration = retbuf->duration;
877 tevent_req_done(req);
880 static int aio_fork_fsync_recv(struct tevent_req *req,
881 struct vfs_aio_state *vfs_aio_state)
883 struct aio_fork_fsync_state *state = tevent_req_data(
884 req, struct aio_fork_fsync_state);
886 if (tevent_req_is_unix_error(req, &vfs_aio_state->error)) {
889 *vfs_aio_state = state->vfs_aio_state;
893 static int aio_fork_connect(vfs_handle_struct *handle, const char *service,
897 struct aio_fork_config *config;
898 ret = SMB_VFS_NEXT_CONNECT(handle, service, user);
904 config = talloc_zero(handle->conn, struct aio_fork_config);
906 SMB_VFS_NEXT_DISCONNECT(handle);
907 DEBUG(0, ("talloc_zero() failed\n"));
911 config->erratic_testing_mode = lp_parm_bool(SNUM(handle->conn), "vfs_aio_fork",
912 "erratic_testing_mode", false);
914 SMB_VFS_HANDLE_SET_DATA(handle, config,
915 NULL, struct aio_fork_config,
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,
932 NTSTATUS vfs_aio_fork_init(TALLOC_CTX *ctx)
934 return smb_register_vfs(SMB_VFS_INTERFACE_VERSION,
935 "aio_fork", &vfs_aio_fork_fns);