2 Unix SMB/Netbios implementation.
4 async_io read handling using POSIX async io.
5 Copyright (C) Jeremy Allison 2005.
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 3 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, see <http://www.gnu.org/licenses/>.
22 #include "smbd/globals.h"
26 /* The signal we'll use to signify aio done. */
31 #define RT_SIGNAL_AIO (SIGRTMIN+3)
34 #ifndef HAVE_STRUCT_SIGEVENT_SIGEV_VALUE_SIVAL_PTR
35 #ifdef HAVE_STRUCT_SIGEVENT_SIGEV_VALUE_SIGVAL_PTR
36 #define sival_int sigval_int
37 #define sival_ptr sigval_ptr
41 /****************************************************************************
42 The buffer we keep around whilst an aio request is in process.
43 *****************************************************************************/
46 struct aio_extra *next, *prev;
49 struct smb_request *req;
51 int (*handle_completion)(struct aio_extra *ex);
54 static int handle_aio_read_complete(struct aio_extra *aio_ex);
55 static int handle_aio_write_complete(struct aio_extra *aio_ex);
57 static int aio_extra_destructor(struct aio_extra *aio_ex)
59 DLIST_REMOVE(aio_list_head, aio_ex);
63 /****************************************************************************
64 Create the extended aio struct we must keep around for the lifetime
66 *****************************************************************************/
68 static struct aio_extra *create_aio_extra(files_struct *fsp, size_t buflen)
70 struct aio_extra *aio_ex = TALLOC_ZERO_P(NULL, struct aio_extra);
76 /* The output buffer stored in the aio_ex is the start of
77 the smb return buffer. The buffer used in the acb
78 is the start of the reply data portion of that buffer. */
80 aio_ex->outbuf = TALLOC_ARRAY(aio_ex, char, buflen);
81 if (!aio_ex->outbuf) {
85 DLIST_ADD(aio_list_head, aio_ex);
86 talloc_set_destructor(aio_ex, aio_extra_destructor);
91 /****************************************************************************
92 Given the mid find the extended aio struct containing it.
93 *****************************************************************************/
95 static struct aio_extra *find_aio_ex(uint16 mid)
99 for( p = aio_list_head; p; p = p->next) {
100 if (mid == p->req->mid) {
107 /****************************************************************************
108 We can have these many aio buffers in flight.
109 *****************************************************************************/
111 /****************************************************************************
112 Set up an aio request from a SMBreadX call.
113 *****************************************************************************/
115 bool schedule_aio_read_and_X(connection_struct *conn,
116 struct smb_request *req,
117 files_struct *fsp, SMB_OFF_T startpos,
120 struct aio_extra *aio_ex;
123 size_t min_aio_read_size = lp_aio_read_size(SNUM(conn));
126 if (fsp->base_fsp != NULL) {
127 /* No AIO on streams yet */
128 DEBUG(10, ("AIO on streams not yet supported\n"));
132 if ((!min_aio_read_size || (smb_maxcnt < min_aio_read_size))
133 && !SMB_VFS_AIO_FORCE(fsp)) {
134 /* Too small a read for aio request. */
135 DEBUG(10,("schedule_aio_read_and_X: read size (%u) too small "
136 "for minimum aio_read of %u\n",
137 (unsigned int)smb_maxcnt,
138 (unsigned int)min_aio_read_size ));
142 /* Only do this on non-chained and non-chaining reads not using the
144 if (chain_size !=0 || (CVAL(req->vwv+0, 0) != 0xFF)
145 || (lp_write_cache_size(SNUM(conn)) != 0) ) {
149 if (outstanding_aio_calls >= aio_pending_size) {
150 DEBUG(10,("schedule_aio_read_and_X: Already have %d aio "
151 "activities outstanding.\n",
152 outstanding_aio_calls ));
156 /* The following is safe from integer wrap as we've already checked
157 smb_maxcnt is 128k or less. Wct is 12 for read replies */
159 bufsize = smb_size + 12 * 2 + smb_maxcnt;
161 if ((aio_ex = create_aio_extra(fsp, bufsize)) == NULL) {
162 DEBUG(10,("schedule_aio_read_and_X: malloc fail.\n"));
165 aio_ex->handle_completion = handle_aio_read_complete;
167 construct_reply_common_req(req, aio_ex->outbuf);
168 srv_set_message(aio_ex->outbuf, 12, 0, True);
169 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
173 /* Now set up the aio record for the read call. */
175 a->aio_fildes = fsp->fh->fd;
176 a->aio_buf = smb_buf(aio_ex->outbuf);
177 a->aio_nbytes = smb_maxcnt;
178 a->aio_offset = startpos;
179 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
180 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
181 a->aio_sigevent.sigev_value.sival_int = req->mid;
184 ret = SMB_VFS_AIO_READ(fsp, a);
188 DEBUG(0,("schedule_aio_read_and_X: aio_read failed. "
189 "Error %s\n", strerror(errno) ));
194 aio_ex->req = talloc_move(aio_ex, &req);
196 DEBUG(10,("schedule_aio_read_and_X: scheduled aio_read for file %s, "
197 "offset %.0f, len = %u (mid = %u)\n",
198 fsp->fsp_name, (double)startpos, (unsigned int)smb_maxcnt,
199 (unsigned int)aio_ex->req->mid ));
201 srv_defer_sign_response(aio_ex->req->mid);
202 outstanding_aio_calls++;
206 /****************************************************************************
207 Set up an aio request from a SMBwriteX call.
208 *****************************************************************************/
210 bool schedule_aio_write_and_X(connection_struct *conn,
211 struct smb_request *req,
212 files_struct *fsp, char *data,
216 struct aio_extra *aio_ex;
219 bool write_through = BITSETW(req->vwv+7,0);
220 size_t min_aio_write_size = lp_aio_write_size(SNUM(conn));
223 if (fsp->base_fsp != NULL) {
224 /* No AIO on streams yet */
225 DEBUG(10, ("AIO on streams not yet supported\n"));
229 if ((!min_aio_write_size || (numtowrite < min_aio_write_size))
230 && !SMB_VFS_AIO_FORCE(fsp)) {
231 /* Too small a write for aio request. */
232 DEBUG(10,("schedule_aio_write_and_X: write size (%u) too "
233 "small for minimum aio_write of %u\n",
234 (unsigned int)numtowrite,
235 (unsigned int)min_aio_write_size ));
239 /* Only do this on non-chained and non-chaining reads not using the
241 if (chain_size !=0 || (CVAL(req->vwv+0, 0) != 0xFF)
242 || (lp_write_cache_size(SNUM(conn)) != 0) ) {
246 if (outstanding_aio_calls >= aio_pending_size) {
247 DEBUG(3,("schedule_aio_write_and_X: Already have %d aio "
248 "activities outstanding.\n",
249 outstanding_aio_calls ));
250 DEBUG(10,("schedule_aio_write_and_X: failed to schedule "
251 "aio_write for file %s, offset %.0f, len = %u "
253 fsp->fsp_name, (double)startpos,
254 (unsigned int)numtowrite,
255 (unsigned int)req->mid ));
259 bufsize = smb_size + 6*2;
261 if (!(aio_ex = create_aio_extra(fsp, bufsize))) {
262 DEBUG(0,("schedule_aio_write_and_X: malloc fail.\n"));
265 aio_ex->handle_completion = handle_aio_write_complete;
267 construct_reply_common_req(req, aio_ex->outbuf);
268 srv_set_message(aio_ex->outbuf, 6, 0, True);
269 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
273 /* Now set up the aio record for the write call. */
275 a->aio_fildes = fsp->fh->fd;
277 a->aio_nbytes = numtowrite;
278 a->aio_offset = startpos;
279 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
280 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
281 a->aio_sigevent.sigev_value.sival_int = req->mid;
284 ret = SMB_VFS_AIO_WRITE(fsp, a);
288 DEBUG(3,("schedule_aio_wrote_and_X: aio_write failed. "
289 "Error %s\n", strerror(errno) ));
294 aio_ex->req = talloc_move(aio_ex, &req);
296 release_level_2_oplocks_on_change(fsp);
298 if (!write_through && !lp_syncalways(SNUM(fsp->conn))
299 && fsp->aio_write_behind) {
300 /* Lie to the client and immediately claim we finished the
302 SSVAL(aio_ex->outbuf,smb_vwv2,numtowrite);
303 SSVAL(aio_ex->outbuf,smb_vwv4,(numtowrite>>16)&1);
304 show_msg(aio_ex->outbuf);
305 if (!srv_send_smb(smbd_server_fd(),aio_ex->outbuf,
306 IS_CONN_ENCRYPTED(fsp->conn))) {
307 exit_server_cleanly("handle_aio_write: srv_send_smb "
310 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write "
311 "behind for file %s\n", fsp->fsp_name ));
313 srv_defer_sign_response(aio_ex->req->mid);
315 outstanding_aio_calls++;
317 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write for file "
318 "%s, offset %.0f, len = %u (mid = %u) "
319 "outstanding_aio_calls = %d\n",
320 fsp->fsp_name, (double)startpos, (unsigned int)numtowrite,
321 (unsigned int)aio_ex->req->mid, outstanding_aio_calls ));
327 /****************************************************************************
328 Complete the read and return the data or error back to the client.
329 Returns errno or zero if all ok.
330 *****************************************************************************/
332 static int handle_aio_read_complete(struct aio_extra *aio_ex)
336 char *outbuf = aio_ex->outbuf;
337 char *data = smb_buf(outbuf);
338 ssize_t nread = SMB_VFS_AIO_RETURN(aio_ex->fsp,&aio_ex->acb);
341 /* We're relying here on the fact that if the fd is
342 closed then the aio will complete and aio_return
343 will return an error. Hopefully this is
346 /* If errno is ECANCELED then don't return anything to the
348 if (errno == ECANCELED) {
349 srv_cancel_sign_response(aio_ex->req->mid);
353 DEBUG( 3,( "handle_aio_read_complete: file %s nread == -1. "
355 aio_ex->fsp->fsp_name, strerror(errno) ));
358 ERROR_NT(map_nt_error_from_unix(ret));
359 outsize = srv_set_message(outbuf,0,0,true);
361 outsize = srv_set_message(outbuf,12,nread,False);
362 SSVAL(outbuf,smb_vwv2,0xFFFF); /* Remaining - must be * -1. */
363 SSVAL(outbuf,smb_vwv5,nread);
364 SSVAL(outbuf,smb_vwv6,smb_offset(data,outbuf));
365 SSVAL(outbuf,smb_vwv7,((nread >> 16) & 1));
366 SSVAL(smb_buf(outbuf),-2,nread);
368 aio_ex->fsp->fh->pos = aio_ex->acb.aio_offset + nread;
369 aio_ex->fsp->fh->position_information = aio_ex->fsp->fh->pos;
371 DEBUG( 3, ( "handle_aio_read_complete file %s max=%d "
373 aio_ex->fsp->fsp_name,
374 (int)aio_ex->acb.aio_nbytes, (int)nread ) );
377 smb_setlen(outbuf,outsize - 4);
379 if (!srv_send_smb(smbd_server_fd(),outbuf,
380 IS_CONN_ENCRYPTED(aio_ex->fsp->conn))) {
381 exit_server_cleanly("handle_aio_read_complete: srv_send_smb "
385 DEBUG(10,("handle_aio_read_complete: scheduled aio_read completed "
386 "for file %s, offset %.0f, len = %u\n",
387 aio_ex->fsp->fsp_name, (double)aio_ex->acb.aio_offset,
388 (unsigned int)nread ));
393 /****************************************************************************
394 Complete the write and return the data or error back to the client.
395 Returns errno or zero if all ok.
396 *****************************************************************************/
398 static int handle_aio_write_complete(struct aio_extra *aio_ex)
401 files_struct *fsp = aio_ex->fsp;
402 char *outbuf = aio_ex->outbuf;
403 ssize_t numtowrite = aio_ex->acb.aio_nbytes;
404 ssize_t nwritten = SMB_VFS_AIO_RETURN(fsp,&aio_ex->acb);
406 if (fsp->aio_write_behind) {
407 if (nwritten != numtowrite) {
408 if (nwritten == -1) {
409 DEBUG(5,("handle_aio_write_complete: "
410 "aio_write_behind failed ! File %s "
411 "is corrupt ! Error %s\n",
412 fsp->fsp_name, strerror(errno) ));
415 DEBUG(0,("handle_aio_write_complete: "
416 "aio_write_behind failed ! File %s "
417 "is corrupt ! Wanted %u bytes but "
418 "only wrote %d\n", fsp->fsp_name,
419 (unsigned int)numtowrite,
424 DEBUG(10,("handle_aio_write_complete: "
425 "aio_write_behind completed for file %s\n",
431 /* We don't need outsize or set_message here as we've already set the
432 fixed size length when we set up the aio call. */
435 DEBUG( 3,( "handle_aio_write: file %s wanted %u bytes. "
436 "nwritten == %d. Error = %s\n",
437 fsp->fsp_name, (unsigned int)numtowrite,
438 (int)nwritten, strerror(errno) ));
440 /* If errno is ECANCELED then don't return anything to the
442 if (errno == ECANCELED) {
443 srv_cancel_sign_response(aio_ex->req->mid);
448 ERROR_BOTH(map_nt_error_from_unix(ret), ERRHRD, ERRdiskfull);
449 srv_set_message(outbuf,0,0,true);
451 bool write_through = BITSETW(aio_ex->req->vwv+7,0);
454 SSVAL(outbuf,smb_vwv2,nwritten);
455 SSVAL(outbuf,smb_vwv4,(nwritten>>16)&1);
456 if (nwritten < (ssize_t)numtowrite) {
457 SCVAL(outbuf,smb_rcls,ERRHRD);
458 SSVAL(outbuf,smb_err,ERRdiskfull);
461 DEBUG(3,("handle_aio_write: fnum=%d num=%d wrote=%d\n",
462 fsp->fnum, (int)numtowrite, (int)nwritten));
463 status = sync_file(fsp->conn,fsp, write_through);
464 if (!NT_STATUS_IS_OK(status)) {
466 ERROR_BOTH(map_nt_error_from_unix(ret),
467 ERRHRD, ERRdiskfull);
468 srv_set_message(outbuf,0,0,true);
469 DEBUG(5,("handle_aio_write: sync_file for %s returned %s\n",
470 fsp->fsp_name, nt_errstr(status) ));
473 aio_ex->fsp->fh->pos = aio_ex->acb.aio_offset + nwritten;
477 if (!srv_send_smb(smbd_server_fd(),outbuf,IS_CONN_ENCRYPTED(fsp->conn))) {
478 exit_server_cleanly("handle_aio_write: srv_send_smb failed.");
481 DEBUG(10,("handle_aio_write_complete: scheduled aio_write completed "
482 "for file %s, offset %.0f, requested %u, written = %u\n",
483 fsp->fsp_name, (double)aio_ex->acb.aio_offset,
484 (unsigned int)numtowrite, (unsigned int)nwritten ));
489 /****************************************************************************
490 Handle any aio completion. Returns True if finished (and sets *perr if err
491 was non-zero), False if not.
492 *****************************************************************************/
494 static bool handle_aio_completed(struct aio_extra *aio_ex, int *perr)
499 DEBUG(3, ("handle_aio_completed: Non-existing aio_ex passed\n"));
503 /* Ensure the operation has really completed. */
504 if (SMB_VFS_AIO_ERROR(aio_ex->fsp, &aio_ex->acb) == EINPROGRESS) {
505 DEBUG(10,( "handle_aio_completed: operation mid %u still in "
506 "process for file %s\n",
507 aio_ex->req->mid, aio_ex->fsp->fsp_name ));
511 err = aio_ex->handle_completion(aio_ex);
513 *perr = err; /* Only save non-zero errors. */
519 /****************************************************************************
520 Handle any aio completion inline.
521 Returns non-zero errno if fail or zero if all ok.
522 *****************************************************************************/
524 void smbd_aio_complete_mid(unsigned int mid)
526 files_struct *fsp = NULL;
527 struct aio_extra *aio_ex = find_aio_ex(mid);
530 DEBUG(10,("smbd_aio_complete_mid: mid[%u]\n", mid));
533 DEBUG(3,("smbd_aio_complete_mid: Can't find record to "
534 "match mid %u.\n", mid));
535 srv_cancel_sign_response(mid);
541 /* file was closed whilst I/O was outstanding. Just
543 DEBUG( 3,( "smbd_aio_complete_mid: file closed whilst "
544 "aio outstanding (mid[%u]).\n", mid));
545 srv_cancel_sign_response(mid);
549 if (!handle_aio_completed(aio_ex, &ret)) {
556 static void smbd_aio_signal_handler(struct tevent_context *ev_ctx,
557 struct tevent_signal *se,
558 int signum, int count,
559 void *_info, void *private_data)
561 siginfo_t *info = (siginfo_t *)_info;
562 unsigned int mid = (unsigned int)info->si_value.sival_int;
564 smbd_aio_complete_mid(mid);
567 /****************************************************************************
568 We're doing write behind and the client closed the file. Wait up to 30
569 seconds (my arbitrary choice) for the aio to complete. Return 0 if all writes
570 completed, errno to return if not.
571 *****************************************************************************/
573 #define SMB_TIME_FOR_AIO_COMPLETE_WAIT 29
575 int wait_for_aio_completion(files_struct *fsp)
577 struct aio_extra *aio_ex;
578 const SMB_STRUCT_AIOCB **aiocb_list;
579 int aio_completion_count = 0;
580 time_t start_time = time(NULL);
583 for (seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT;
584 seconds_left >= 0;) {
589 aio_completion_count = 0;
590 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
591 if (aio_ex->fsp == fsp) {
592 aio_completion_count++;
596 if (!aio_completion_count) {
600 DEBUG(3,("wait_for_aio_completion: waiting for %d aio events "
601 "to complete.\n", aio_completion_count ));
603 aiocb_list = SMB_MALLOC_ARRAY(const SMB_STRUCT_AIOCB *,
604 aio_completion_count);
609 for( i = 0, aio_ex = aio_list_head;
611 aio_ex = aio_ex->next) {
612 if (aio_ex->fsp == fsp) {
613 aiocb_list[i++] = &aio_ex->acb;
617 /* Now wait up to seconds_left for completion. */
618 ts.tv_sec = seconds_left;
621 DEBUG(10,("wait_for_aio_completion: %d events, doing a wait "
623 aio_completion_count, seconds_left ));
625 err = SMB_VFS_AIO_SUSPEND(fsp, aiocb_list,
626 aio_completion_count, &ts);
628 DEBUG(10,("wait_for_aio_completion: returned err = %d, "
629 "errno = %s\n", err, strerror(errno) ));
631 if (err == -1 && errno == EAGAIN) {
632 DEBUG(0,("wait_for_aio_completion: aio_suspend timed "
633 "out waiting for %d events after a wait of "
634 "%d seconds\n", aio_completion_count,
637 cancel_aio_by_fsp(fsp);
638 SAFE_FREE(aiocb_list);
642 /* One or more events might have completed - process them if
644 for( i = 0; i < aio_completion_count; i++) {
645 uint16 mid = aiocb_list[i]->aio_sigevent.sigev_value.sival_int;
647 aio_ex = find_aio_ex(mid);
650 DEBUG(0, ("wait_for_aio_completion: mid %u "
651 "doesn't match an aio record\n",
652 (unsigned int)mid ));
656 if (!handle_aio_completed(aio_ex, &err)) {
662 SAFE_FREE(aiocb_list);
663 seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT
664 - (time(NULL) - start_time);
667 /* We timed out - we don't know why. Return ret if already an error,
669 DEBUG(10,("wait_for_aio_completion: aio_suspend timed out waiting "
671 aio_completion_count));
676 /****************************************************************************
677 Cancel any outstanding aio requests. The client doesn't care about the reply.
678 *****************************************************************************/
680 void cancel_aio_by_fsp(files_struct *fsp)
682 struct aio_extra *aio_ex;
684 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
685 if (aio_ex->fsp == fsp) {
686 /* Don't delete the aio_extra record as we may have
687 completed and don't yet know it. Just do the
688 aio_cancel call and return. */
689 SMB_VFS_AIO_CANCEL(fsp, &aio_ex->acb);
690 aio_ex->fsp = NULL; /* fsp will be closed when we
696 /****************************************************************************
697 Initialize the signal handler for aio read/write.
698 *****************************************************************************/
700 void initialize_async_io_handler(void)
702 aio_signal_event = tevent_add_signal(smbd_event_context(),
703 smbd_event_context(),
704 RT_SIGNAL_AIO, SA_SIGINFO,
705 smbd_aio_signal_handler,
707 if (!aio_signal_event) {
708 exit_server("Failed to setup RT_SIGNAL_AIO handler");
711 /* tevent supports 100 signal with SA_SIGINFO */
712 aio_pending_size = 100;
716 void initialize_async_io_handler(void)
720 bool schedule_aio_read_and_X(connection_struct *conn,
721 struct smb_request *req,
722 files_struct *fsp, SMB_OFF_T startpos,
728 bool schedule_aio_write_and_X(connection_struct *conn,
729 struct smb_request *req,
730 files_struct *fsp, char *data,
737 void cancel_aio_by_fsp(files_struct *fsp)
741 int wait_for_aio_completion(files_struct *fsp)
746 void smbd_aio_complete_mid(unsigned int mid);