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/>.
25 /* The signal we'll use to signify aio done. */
27 #define RT_SIGNAL_AIO (SIGRTMIN+3)
30 /****************************************************************************
31 The buffer we keep around whilst an aio request is in process.
32 *****************************************************************************/
35 struct aio_extra *next, *prev;
44 static struct aio_extra *aio_list_head;
46 /****************************************************************************
47 Create the extended aio struct we must keep around for the lifetime
49 *****************************************************************************/
51 static struct aio_extra *create_aio_ex_read(files_struct *fsp, size_t buflen,
54 struct aio_extra *aio_ex = SMB_MALLOC_P(struct aio_extra);
60 /* The output buffer stored in the aio_ex is the start of
61 the smb return buffer. The buffer used in the acb
62 is the start of the reply data portion of that buffer. */
63 aio_ex->outbuf = SMB_MALLOC_ARRAY(char, buflen);
64 if (!aio_ex->outbuf) {
68 DLIST_ADD(aio_list_head, aio_ex);
70 aio_ex->read_req = True;
75 /****************************************************************************
76 Create the extended aio struct we must keep around for the lifetime
77 of the aio_write call.
78 *****************************************************************************/
80 static struct aio_extra *create_aio_ex_write(files_struct *fsp,
85 struct aio_extra *aio_ex = SMB_MALLOC_P(struct aio_extra);
92 /* We need space for an output reply of outbuflen bytes. */
93 aio_ex->outbuf = SMB_MALLOC_ARRAY(char, outbuflen);
94 if (!aio_ex->outbuf) {
99 if (!(aio_ex->inbuf = SMB_MALLOC_ARRAY(char, inbuflen))) {
100 SAFE_FREE(aio_ex->outbuf);
105 DLIST_ADD(aio_list_head, aio_ex);
107 aio_ex->read_req = False;
112 /****************************************************************************
113 Delete the extended aio struct.
114 *****************************************************************************/
116 static void delete_aio_ex(struct aio_extra *aio_ex)
118 DLIST_REMOVE(aio_list_head, aio_ex);
119 SAFE_FREE(aio_ex->inbuf);
120 SAFE_FREE(aio_ex->outbuf);
124 /****************************************************************************
125 Given the aiocb struct find the extended aio struct containing it.
126 *****************************************************************************/
128 static struct aio_extra *find_aio_ex(uint16 mid)
132 for( p = aio_list_head; p; p = p->next) {
140 /****************************************************************************
141 We can have these many aio buffers in flight.
142 *****************************************************************************/
144 #define AIO_PENDING_SIZE 10
145 static sig_atomic_t signals_received;
146 static int outstanding_aio_calls;
147 static uint16 aio_pending_array[AIO_PENDING_SIZE];
149 /****************************************************************************
150 Signal handler when an aio request completes.
151 *****************************************************************************/
153 static void signal_handler(int sig, siginfo_t *info, void *unused)
155 if (signals_received < AIO_PENDING_SIZE) {
156 aio_pending_array[signals_received] = info->si_value.sival_int;
158 } /* Else signal is lost. */
159 sys_select_signal(RT_SIGNAL_AIO);
162 /****************************************************************************
163 Is there a signal waiting ?
164 *****************************************************************************/
166 bool aio_finished(void)
168 return (signals_received != 0);
171 /****************************************************************************
172 Initialize the signal handler for aio read/write.
173 *****************************************************************************/
175 void initialize_async_io_handler(void)
177 struct sigaction act;
180 act.sa_sigaction = signal_handler;
181 act.sa_flags = SA_SIGINFO;
182 sigemptyset( &act.sa_mask );
183 if (sigaction(RT_SIGNAL_AIO, &act, NULL) != 0) {
184 DEBUG(0,("Failed to setup RT_SIGNAL_AIO handler\n"));
187 /* the signal can start off blocked due to a bug in bash */
188 BlockSignals(False, RT_SIGNAL_AIO);
191 /****************************************************************************
192 Set up an aio request from a SMBreadX call.
193 *****************************************************************************/
195 bool schedule_aio_read_and_X(connection_struct *conn,
196 struct smb_request *req,
197 files_struct *fsp, SMB_OFF_T startpos,
200 struct aio_extra *aio_ex;
203 size_t min_aio_read_size = lp_aio_read_size(SNUM(conn));
205 if (!min_aio_read_size || (smb_maxcnt < min_aio_read_size)) {
206 /* Too small a read for aio request. */
207 DEBUG(10,("schedule_aio_read_and_X: read size (%u) too small "
208 "for minimum aio_read of %u\n",
209 (unsigned int)smb_maxcnt,
210 (unsigned int)min_aio_read_size ));
214 /* Only do this on non-chained and non-chaining reads not using the
216 if (chain_size !=0 || (CVAL(req->inbuf,smb_vwv0) != 0xFF)
217 || (lp_write_cache_size(SNUM(conn)) != 0) ) {
221 if (outstanding_aio_calls >= AIO_PENDING_SIZE) {
222 DEBUG(10,("schedule_aio_read_and_X: Already have %d aio "
223 "activities outstanding.\n",
224 outstanding_aio_calls ));
228 /* The following is safe from integer wrap as we've already checked
229 smb_maxcnt is 128k or less. Wct is 12 for read replies */
231 bufsize = smb_size + 12 * 2 + smb_maxcnt;
233 if ((aio_ex = create_aio_ex_read(fsp, bufsize, req->mid)) == NULL) {
234 DEBUG(10,("schedule_aio_read_and_X: malloc fail.\n"));
238 construct_reply_common((char *)req->inbuf, aio_ex->outbuf);
239 srv_set_message(aio_ex->outbuf, 12, 0, True);
240 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
244 /* Now set up the aio record for the read call. */
246 a->aio_fildes = fsp->fh->fd;
247 a->aio_buf = smb_buf(aio_ex->outbuf);
248 a->aio_nbytes = smb_maxcnt;
249 a->aio_offset = startpos;
250 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
251 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
252 a->aio_sigevent.sigev_value.sival_int = aio_ex->mid;
254 if (SMB_VFS_AIO_READ(fsp,a) == -1) {
255 DEBUG(0,("schedule_aio_read_and_X: aio_read failed. "
256 "Error %s\n", strerror(errno) ));
257 delete_aio_ex(aio_ex);
261 DEBUG(10,("schedule_aio_read_and_X: scheduled aio_read for file %s, "
262 "offset %.0f, len = %u (mid = %u)\n",
263 fsp->fsp_name, (double)startpos, (unsigned int)smb_maxcnt,
264 (unsigned int)aio_ex->mid ));
266 srv_defer_sign_response(aio_ex->mid);
267 outstanding_aio_calls++;
271 /****************************************************************************
272 Set up an aio request from a SMBwriteX call.
273 *****************************************************************************/
275 bool schedule_aio_write_and_X(connection_struct *conn,
276 struct smb_request *req,
277 files_struct *fsp, char *data,
281 struct aio_extra *aio_ex;
283 size_t inbufsize, outbufsize;
284 bool write_through = BITSETW(req->inbuf+smb_vwv7,0);
285 size_t min_aio_write_size = lp_aio_write_size(SNUM(conn));
287 if (!min_aio_write_size || (numtowrite < min_aio_write_size)) {
288 /* Too small a write for aio request. */
289 DEBUG(10,("schedule_aio_write_and_X: write size (%u) too "
290 "small for minimum aio_write of %u\n",
291 (unsigned int)numtowrite,
292 (unsigned int)min_aio_write_size ));
296 /* Only do this on non-chained and non-chaining reads not using the
298 if (chain_size !=0 || (CVAL(req->inbuf,smb_vwv0) != 0xFF)
299 || (lp_write_cache_size(SNUM(conn)) != 0) ) {
303 if (outstanding_aio_calls >= AIO_PENDING_SIZE) {
304 DEBUG(3,("schedule_aio_write_and_X: Already have %d aio "
305 "activities outstanding.\n",
306 outstanding_aio_calls ));
307 DEBUG(10,("schedule_aio_write_and_X: failed to schedule "
308 "aio_write for file %s, offset %.0f, len = %u "
310 fsp->fsp_name, (double)startpos,
311 (unsigned int)numtowrite,
312 (unsigned int)req->mid ));
316 inbufsize = smb_len(req->inbuf) + 4;
317 reply_outbuf(req, 6, 0);
318 outbufsize = smb_len(req->outbuf) + 4;
319 if (!(aio_ex = create_aio_ex_write(fsp, inbufsize, outbufsize,
321 DEBUG(0,("schedule_aio_write_and_X: malloc fail.\n"));
325 /* Copy the SMB header already setup in outbuf. */
326 memcpy(aio_ex->inbuf, req->inbuf, inbufsize);
328 /* Copy the SMB header already setup in outbuf. */
329 memcpy(aio_ex->outbuf, req->outbuf, outbufsize);
330 TALLOC_FREE(req->outbuf);
331 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
335 /* Now set up the aio record for the write call. */
337 a->aio_fildes = fsp->fh->fd;
338 a->aio_buf = aio_ex->inbuf + (PTR_DIFF(data, req->inbuf));
339 a->aio_nbytes = numtowrite;
340 a->aio_offset = startpos;
341 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
342 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
343 a->aio_sigevent.sigev_value.sival_int = aio_ex->mid;
345 if (SMB_VFS_AIO_WRITE(fsp,a) == -1) {
346 DEBUG(3,("schedule_aio_wrote_and_X: aio_write failed. "
347 "Error %s\n", strerror(errno) ));
348 delete_aio_ex(aio_ex);
352 if (!write_through && !lp_syncalways(SNUM(fsp->conn))
353 && fsp->aio_write_behind) {
354 /* Lie to the client and immediately claim we finished the
356 SSVAL(aio_ex->outbuf,smb_vwv2,numtowrite);
357 SSVAL(aio_ex->outbuf,smb_vwv4,(numtowrite>>16)&1);
358 show_msg(aio_ex->outbuf);
359 if (!srv_send_smb(smbd_server_fd(),aio_ex->outbuf,
360 IS_CONN_ENCRYPTED(fsp->conn))) {
361 exit_server_cleanly("handle_aio_write: srv_send_smb "
364 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write "
365 "behind for file %s\n", fsp->fsp_name ));
367 srv_defer_sign_response(aio_ex->mid);
369 outstanding_aio_calls++;
371 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write for file "
372 "%s, offset %.0f, len = %u (mid = %u) "
373 "outstanding_aio_calls = %d\n",
374 fsp->fsp_name, (double)startpos, (unsigned int)numtowrite,
375 (unsigned int)aio_ex->mid, outstanding_aio_calls ));
381 /****************************************************************************
382 Complete the read and return the data or error back to the client.
383 Returns errno or zero if all ok.
384 *****************************************************************************/
386 static int handle_aio_read_complete(struct aio_extra *aio_ex)
390 char *outbuf = aio_ex->outbuf;
391 char *data = smb_buf(outbuf);
392 ssize_t nread = SMB_VFS_AIO_RETURN(aio_ex->fsp,&aio_ex->acb);
395 /* We're relying here on the fact that if the fd is
396 closed then the aio will complete and aio_return
397 will return an error. Hopefully this is
400 /* If errno is ECANCELED then don't return anything to the
402 if (errno == ECANCELED) {
403 srv_cancel_sign_response(aio_ex->mid);
407 DEBUG( 3,( "handle_aio_read_complete: file %s nread == -1. "
409 aio_ex->fsp->fsp_name, strerror(errno) ));
412 ERROR_NT(map_nt_error_from_unix(ret));
413 outsize = srv_set_message(outbuf,0,0,true);
415 outsize = srv_set_message(outbuf,12,nread,False);
416 SSVAL(outbuf,smb_vwv2,0xFFFF); /* Remaining - must be * -1. */
417 SSVAL(outbuf,smb_vwv5,nread);
418 SSVAL(outbuf,smb_vwv6,smb_offset(data,outbuf));
419 SSVAL(outbuf,smb_vwv7,((nread >> 16) & 1));
420 SSVAL(smb_buf(outbuf),-2,nread);
422 DEBUG( 3, ( "handle_aio_read_complete file %s max=%d "
424 aio_ex->fsp->fsp_name,
425 (int)aio_ex->acb.aio_nbytes, (int)nread ) );
428 smb_setlen(outbuf,outsize - 4);
430 if (!srv_send_smb(smbd_server_fd(),outbuf,
431 IS_CONN_ENCRYPTED(aio_ex->fsp->conn))) {
432 exit_server_cleanly("handle_aio_read_complete: srv_send_smb "
436 DEBUG(10,("handle_aio_read_complete: scheduled aio_read completed "
437 "for file %s, offset %.0f, len = %u\n",
438 aio_ex->fsp->fsp_name, (double)aio_ex->acb.aio_offset,
439 (unsigned int)nread ));
444 /****************************************************************************
445 Complete the write and return the data or error back to the client.
446 Returns errno or zero if all ok.
447 *****************************************************************************/
449 static int handle_aio_write_complete(struct aio_extra *aio_ex)
452 files_struct *fsp = aio_ex->fsp;
453 char *outbuf = aio_ex->outbuf;
454 ssize_t numtowrite = aio_ex->acb.aio_nbytes;
455 ssize_t nwritten = SMB_VFS_AIO_RETURN(fsp,&aio_ex->acb);
457 if (fsp->aio_write_behind) {
458 if (nwritten != numtowrite) {
459 if (nwritten == -1) {
460 DEBUG(5,("handle_aio_write_complete: "
461 "aio_write_behind failed ! File %s "
462 "is corrupt ! Error %s\n",
463 fsp->fsp_name, strerror(errno) ));
466 DEBUG(0,("handle_aio_write_complete: "
467 "aio_write_behind failed ! File %s "
468 "is corrupt ! Wanted %u bytes but "
469 "only wrote %d\n", fsp->fsp_name,
470 (unsigned int)numtowrite,
475 DEBUG(10,("handle_aio_write_complete: "
476 "aio_write_behind completed for file %s\n",
482 /* We don't need outsize or set_message here as we've already set the
483 fixed size length when we set up the aio call. */
486 DEBUG( 3,( "handle_aio_write: file %s wanted %u bytes. "
487 "nwritten == %d. Error = %s\n",
488 fsp->fsp_name, (unsigned int)numtowrite,
489 (int)nwritten, strerror(errno) ));
491 /* If errno is ECANCELED then don't return anything to the
493 if (errno == ECANCELED) {
494 srv_cancel_sign_response(aio_ex->mid);
499 ERROR_BOTH(map_nt_error_from_unix(ret), ERRHRD, ERRdiskfull);
500 srv_set_message(outbuf,0,0,true);
502 bool write_through = BITSETW(aio_ex->inbuf+smb_vwv7,0);
505 SSVAL(outbuf,smb_vwv2,nwritten);
506 SSVAL(outbuf,smb_vwv4,(nwritten>>16)&1);
507 if (nwritten < (ssize_t)numtowrite) {
508 SCVAL(outbuf,smb_rcls,ERRHRD);
509 SSVAL(outbuf,smb_err,ERRdiskfull);
512 DEBUG(3,("handle_aio_write: fnum=%d num=%d wrote=%d\n",
513 fsp->fnum, (int)numtowrite, (int)nwritten));
514 status = sync_file(fsp->conn,fsp, write_through);
515 if (!NT_STATUS_IS_OK(status)) {
517 ERROR_BOTH(map_nt_error_from_unix(ret),
518 ERRHRD, ERRdiskfull);
519 srv_set_message(outbuf,0,0,true);
520 DEBUG(5,("handle_aio_write: sync_file for %s returned %s\n",
521 fsp->fsp_name, nt_errstr(status) ));
526 if (!srv_send_smb(smbd_server_fd(),outbuf,IS_CONN_ENCRYPTED(fsp->conn))) {
527 exit_server_cleanly("handle_aio_write: srv_send_smb failed.");
530 DEBUG(10,("handle_aio_write_complete: scheduled aio_write completed "
531 "for file %s, offset %.0f, requested %u, written = %u\n",
532 fsp->fsp_name, (double)aio_ex->acb.aio_offset,
533 (unsigned int)numtowrite, (unsigned int)nwritten ));
538 /****************************************************************************
539 Handle any aio completion. Returns True if finished (and sets *perr if err
540 was non-zero), False if not.
541 *****************************************************************************/
543 static bool handle_aio_completed(struct aio_extra *aio_ex, int *perr)
547 /* Ensure the operation has really completed. */
548 if (SMB_VFS_AIO_ERROR(aio_ex->fsp, &aio_ex->acb) == EINPROGRESS) {
549 DEBUG(10,( "handle_aio_completed: operation mid %u still in "
550 "process for file %s\n",
551 aio_ex->mid, aio_ex->fsp->fsp_name ));
555 if (aio_ex->read_req) {
556 err = handle_aio_read_complete(aio_ex);
558 err = handle_aio_write_complete(aio_ex);
562 *perr = err; /* Only save non-zero errors. */
568 /****************************************************************************
569 Handle any aio completion inline.
570 Returns non-zero errno if fail or zero if all ok.
571 *****************************************************************************/
573 int process_aio_queue(void)
578 BlockSignals(True, RT_SIGNAL_AIO);
580 DEBUG(10,("process_aio_queue: signals_received = %d\n",
581 (int)signals_received));
582 DEBUG(10,("process_aio_queue: outstanding_aio_calls = %d\n",
583 outstanding_aio_calls));
585 if (!signals_received) {
586 BlockSignals(False, RT_SIGNAL_AIO);
590 /* Drain all the complete aio_reads. */
591 for (i = 0; i < signals_received; i++) {
592 uint16 mid = aio_pending_array[i];
593 files_struct *fsp = NULL;
594 struct aio_extra *aio_ex = find_aio_ex(mid);
597 DEBUG(3,("process_aio_queue: Can't find record to "
598 "match mid %u.\n", (unsigned int)mid));
599 srv_cancel_sign_response(mid);
605 /* file was closed whilst I/O was outstanding. Just
607 DEBUG( 3,( "process_aio_queue: file closed whilst "
608 "aio outstanding.\n"));
609 srv_cancel_sign_response(mid);
613 if (!handle_aio_completed(aio_ex, &ret)) {
617 delete_aio_ex(aio_ex);
620 outstanding_aio_calls -= signals_received;
621 signals_received = 0;
622 BlockSignals(False, RT_SIGNAL_AIO);
626 /****************************************************************************
627 We're doing write behind and the client closed the file. Wait up to 30
628 seconds (my arbitrary choice) for the aio to complete. Return 0 if all writes
629 completed, errno to return if not.
630 *****************************************************************************/
632 #define SMB_TIME_FOR_AIO_COMPLETE_WAIT 29
634 int wait_for_aio_completion(files_struct *fsp)
636 struct aio_extra *aio_ex;
637 const SMB_STRUCT_AIOCB **aiocb_list;
638 int aio_completion_count = 0;
639 time_t start_time = time(NULL);
642 for (seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT;
643 seconds_left >= 0;) {
648 aio_completion_count = 0;
649 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
650 if (aio_ex->fsp == fsp) {
651 aio_completion_count++;
655 if (!aio_completion_count) {
659 DEBUG(3,("wait_for_aio_completion: waiting for %d aio events "
660 "to complete.\n", aio_completion_count ));
662 aiocb_list = SMB_MALLOC_ARRAY(const SMB_STRUCT_AIOCB *,
663 aio_completion_count);
668 for( i = 0, aio_ex = aio_list_head;
670 aio_ex = aio_ex->next) {
671 if (aio_ex->fsp == fsp) {
672 aiocb_list[i++] = &aio_ex->acb;
676 /* Now wait up to seconds_left for completion. */
677 ts.tv_sec = seconds_left;
680 DEBUG(10,("wait_for_aio_completion: %d events, doing a wait "
682 aio_completion_count, seconds_left ));
684 err = SMB_VFS_AIO_SUSPEND(fsp, aiocb_list,
685 aio_completion_count, &ts);
687 DEBUG(10,("wait_for_aio_completion: returned err = %d, "
688 "errno = %s\n", err, strerror(errno) ));
690 if (err == -1 && errno == EAGAIN) {
691 DEBUG(0,("wait_for_aio_completion: aio_suspend timed "
692 "out waiting for %d events after a wait of "
693 "%d seconds\n", aio_completion_count,
696 cancel_aio_by_fsp(fsp);
697 SAFE_FREE(aiocb_list);
701 /* One or more events might have completed - process them if
703 for( i = 0; i < aio_completion_count; i++) {
704 uint16 mid = aiocb_list[i]->aio_sigevent.sigev_value.sival_int;
706 aio_ex = find_aio_ex(mid);
709 DEBUG(0, ("wait_for_aio_completion: mid %u "
710 "doesn't match an aio record\n",
711 (unsigned int)mid ));
715 if (!handle_aio_completed(aio_ex, &err)) {
718 delete_aio_ex(aio_ex);
721 SAFE_FREE(aiocb_list);
722 seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT
723 - (time(NULL) - start_time);
726 /* We timed out - we don't know why. Return ret if already an error,
728 DEBUG(10,("wait_for_aio_completion: aio_suspend timed out waiting "
730 aio_completion_count));
735 /****************************************************************************
736 Cancel any outstanding aio requests. The client doesn't care about the reply.
737 *****************************************************************************/
739 void cancel_aio_by_fsp(files_struct *fsp)
741 struct aio_extra *aio_ex;
743 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
744 if (aio_ex->fsp == fsp) {
745 /* Don't delete the aio_extra record as we may have
746 completed and don't yet know it. Just do the
747 aio_cancel call and return. */
748 SMB_VFS_AIO_CANCEL(fsp, &aio_ex->acb);
749 aio_ex->fsp = NULL; /* fsp will be closed when we
756 bool aio_finished(void)
761 void initialize_async_io_handler(void)
765 int process_aio_queue(void)
770 bool schedule_aio_read_and_X(connection_struct *conn,
771 struct smb_request *req,
772 files_struct *fsp, SMB_OFF_T startpos,
778 bool schedule_aio_write_and_X(connection_struct *conn,
779 struct smb_request *req,
780 files_struct *fsp, char *data,
787 void cancel_aio_by_fsp(files_struct *fsp)
791 int wait_for_aio_completion(files_struct *fsp)