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 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.
26 /* The signal we'll use to signify aio done. */
28 #define RT_SIGNAL_AIO (SIGRTMIN+3)
31 /****************************************************************************
32 The buffer we keep around whilst an aio request is in process.
33 *****************************************************************************/
36 struct aio_extra *next, *prev;
45 static struct aio_extra *aio_list_head;
47 /****************************************************************************
48 Create the extended aio struct we must keep around for the lifetime
50 *****************************************************************************/
52 static struct aio_extra *create_aio_ex_read(files_struct *fsp, size_t buflen,
55 struct aio_extra *aio_ex = SMB_MALLOC_P(struct aio_extra);
61 /* The output buffer stored in the aio_ex is the start of
62 the smb return buffer. The buffer used in the acb
63 is the start of the reply data portion of that buffer. */
64 aio_ex->outbuf = SMB_MALLOC_ARRAY(char, buflen);
65 if (!aio_ex->outbuf) {
69 DLIST_ADD(aio_list_head, aio_ex);
71 aio_ex->read_req = True;
76 /****************************************************************************
77 Create the extended aio struct we must keep around for the lifetime
78 of the aio_write call.
79 *****************************************************************************/
81 static struct aio_extra *create_aio_ex_write(files_struct *fsp,
86 struct aio_extra *aio_ex = SMB_MALLOC_P(struct aio_extra);
93 /* We need space for an output reply of outbuflen bytes. */
94 aio_ex->outbuf = SMB_MALLOC_ARRAY(char, outbuflen);
95 if (!aio_ex->outbuf) {
100 if (!(aio_ex->inbuf = SMB_MALLOC_ARRAY(char, inbuflen))) {
101 SAFE_FREE(aio_ex->outbuf);
106 DLIST_ADD(aio_list_head, aio_ex);
108 aio_ex->read_req = False;
113 /****************************************************************************
114 Delete the extended aio struct.
115 *****************************************************************************/
117 static void delete_aio_ex(struct aio_extra *aio_ex)
119 DLIST_REMOVE(aio_list_head, aio_ex);
120 SAFE_FREE(aio_ex->inbuf);
121 SAFE_FREE(aio_ex->outbuf);
125 /****************************************************************************
126 Given the aiocb struct find the extended aio struct containing it.
127 *****************************************************************************/
129 static struct aio_extra *find_aio_ex(uint16 mid)
133 for( p = aio_list_head; p; p = p->next) {
141 /****************************************************************************
142 We can have these many aio buffers in flight.
143 *****************************************************************************/
145 #define AIO_PENDING_SIZE 10
146 static sig_atomic_t signals_received;
147 static int outstanding_aio_calls;
148 static uint16 aio_pending_array[AIO_PENDING_SIZE];
150 /****************************************************************************
151 Signal handler when an aio request completes.
152 *****************************************************************************/
154 static void signal_handler(int sig, siginfo_t *info, void *unused)
156 if (signals_received < AIO_PENDING_SIZE) {
157 aio_pending_array[signals_received] = info->si_value.sival_int;
159 } /* Else signal is lost. */
160 sys_select_signal(RT_SIGNAL_AIO);
163 /****************************************************************************
164 Is there a signal waiting ?
165 *****************************************************************************/
167 BOOL aio_finished(void)
169 return (signals_received != 0);
172 /****************************************************************************
173 Initialize the signal handler for aio read/write.
174 *****************************************************************************/
176 void initialize_async_io_handler(void)
178 struct sigaction act;
181 act.sa_sigaction = signal_handler;
182 act.sa_flags = SA_SIGINFO;
183 sigemptyset( &act.sa_mask );
184 if (sigaction(RT_SIGNAL_AIO, &act, NULL) != 0) {
185 DEBUG(0,("Failed to setup RT_SIGNAL_AIO handler\n"));
188 /* the signal can start off blocked due to a bug in bash */
189 BlockSignals(False, RT_SIGNAL_AIO);
192 /****************************************************************************
193 Set up an aio request from a SMBreadX call.
194 *****************************************************************************/
196 BOOL schedule_aio_read_and_X(connection_struct *conn,
197 char *inbuf, char *outbuf,
198 int length, int len_outbuf,
199 files_struct *fsp, SMB_OFF_T startpos,
202 struct aio_extra *aio_ex;
205 size_t min_aio_read_size = lp_aio_read_size(SNUM(conn));
207 if (!min_aio_read_size || (smb_maxcnt < min_aio_read_size)) {
208 /* Too small a read for aio request. */
209 DEBUG(10,("schedule_aio_read_and_X: read size (%u) too small "
210 "for minimum aio_read of %u\n",
211 (unsigned int)smb_maxcnt,
212 (unsigned int)min_aio_read_size ));
216 /* Only do this on non-chained and non-chaining reads not using the
218 if (chain_size !=0 || (CVAL(inbuf,smb_vwv0) != 0xFF)
219 || (lp_write_cache_size(SNUM(conn)) != 0) ) {
223 if (outstanding_aio_calls >= AIO_PENDING_SIZE) {
224 DEBUG(10,("schedule_aio_read_and_X: Already have %d aio "
225 "activities outstanding.\n",
226 outstanding_aio_calls ));
230 /* The following is safe from integer wrap as we've already
231 checked smb_maxcnt is 128k or less. */
232 bufsize = PTR_DIFF(smb_buf(outbuf),outbuf) + smb_maxcnt;
234 if ((aio_ex = create_aio_ex_read(fsp, bufsize,
235 SVAL(inbuf,smb_mid))) == NULL) {
236 DEBUG(10,("schedule_aio_read_and_X: malloc fail.\n"));
240 /* Copy the SMB header already setup in outbuf. */
241 memcpy(aio_ex->outbuf, outbuf, smb_buf(outbuf) - outbuf);
242 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
246 /* Now set up the aio record for the read call. */
248 a->aio_fildes = fsp->fh->fd;
249 a->aio_buf = smb_buf(aio_ex->outbuf);
250 a->aio_nbytes = smb_maxcnt;
251 a->aio_offset = startpos;
252 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
253 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
254 a->aio_sigevent.sigev_value.sival_int = aio_ex->mid;
256 if (SMB_VFS_AIO_READ(fsp,a) == -1) {
257 DEBUG(0,("schedule_aio_read_and_X: aio_read failed. "
258 "Error %s\n", strerror(errno) ));
259 delete_aio_ex(aio_ex);
263 DEBUG(10,("schedule_aio_read_and_X: scheduled aio_read for file %s, "
264 "offset %.0f, len = %u (mid = %u)\n",
265 fsp->fsp_name, (double)startpos, (unsigned int)smb_maxcnt,
266 (unsigned int)aio_ex->mid ));
268 srv_defer_sign_response(aio_ex->mid);
269 outstanding_aio_calls++;
273 /****************************************************************************
274 Set up an aio request from a SMBwriteX call.
275 *****************************************************************************/
277 BOOL schedule_aio_write_and_X(connection_struct *conn,
278 char *inbuf, char *outbuf,
279 int length, int len_outbuf,
280 files_struct *fsp, char *data,
284 struct aio_extra *aio_ex;
286 size_t inbufsize, outbufsize;
287 BOOL write_through = BITSETW(inbuf+smb_vwv7,0);
288 size_t min_aio_write_size = lp_aio_write_size(SNUM(conn));
290 if (!min_aio_write_size || (numtowrite < min_aio_write_size)) {
291 /* Too small a write for aio request. */
292 DEBUG(10,("schedule_aio_write_and_X: write size (%u) too "
293 "small for minimum aio_write of %u\n",
294 (unsigned int)numtowrite,
295 (unsigned int)min_aio_write_size ));
299 /* Only do this on non-chained and non-chaining reads not using the
301 if (chain_size !=0 || (CVAL(inbuf,smb_vwv0) != 0xFF)
302 || (lp_write_cache_size(SNUM(conn)) != 0) ) {
306 if (outstanding_aio_calls >= AIO_PENDING_SIZE) {
307 DEBUG(3,("schedule_aio_write_and_X: Already have %d aio "
308 "activities outstanding.\n",
309 outstanding_aio_calls ));
310 DEBUG(10,("schedule_aio_write_and_X: failed to schedule "
311 "aio_write for file %s, offset %.0f, len = %u "
313 fsp->fsp_name, (double)startpos,
314 (unsigned int)numtowrite,
315 (unsigned int)SVAL(inbuf,smb_mid) ));
319 inbufsize = smb_len(inbuf) + 4;
320 outbufsize = smb_len(outbuf) + 4;
321 if (!(aio_ex = create_aio_ex_write(fsp, inbufsize, outbufsize,
322 SVAL(inbuf,smb_mid)))) {
323 DEBUG(0,("schedule_aio_write_and_X: malloc fail.\n"));
327 /* Copy the SMB header already setup in outbuf. */
328 memcpy(aio_ex->inbuf, inbuf, inbufsize);
330 /* Copy the SMB header already setup in outbuf. */
331 memcpy(aio_ex->outbuf, outbuf, outbufsize);
332 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
336 /* Now set up the aio record for the write call. */
338 a->aio_fildes = fsp->fh->fd;
339 a->aio_buf = aio_ex->inbuf + (PTR_DIFF(data, inbuf));
340 a->aio_nbytes = numtowrite;
341 a->aio_offset = startpos;
342 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
343 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
344 a->aio_sigevent.sigev_value.sival_int = aio_ex->mid;
346 if (SMB_VFS_AIO_WRITE(fsp,a) == -1) {
347 DEBUG(3,("schedule_aio_wrote_and_X: aio_write failed. "
348 "Error %s\n", strerror(errno) ));
349 delete_aio_ex(aio_ex);
353 if (!write_through && !lp_syncalways(SNUM(fsp->conn))
354 && fsp->aio_write_behind) {
355 /* Lie to the client and immediately claim we finished the
357 SSVAL(aio_ex->outbuf,smb_vwv2,numtowrite);
358 SSVAL(aio_ex->outbuf,smb_vwv4,(numtowrite>>16)&1);
359 show_msg(aio_ex->outbuf);
360 if (!send_smb(smbd_server_fd(),aio_ex->outbuf)) {
361 exit_server_cleanly("handle_aio_write: 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) ));
411 outsize = (UNIXERROR(ERRDOS,ERRnoaccess));
414 outsize = set_message(outbuf,12,nread,False);
415 SSVAL(outbuf,smb_vwv2,0xFFFF); /* Remaining - must be * -1. */
416 SSVAL(outbuf,smb_vwv5,nread);
417 SSVAL(outbuf,smb_vwv6,smb_offset(data,outbuf));
418 SSVAL(outbuf,smb_vwv7,((nread >> 16) & 1));
419 SSVAL(smb_buf(outbuf),-2,nread);
421 DEBUG( 3, ( "handle_aio_read_complete file %s max=%d "
423 aio_ex->fsp->fsp_name,
424 aio_ex->acb.aio_nbytes, (int)nread ) );
427 smb_setlen(outbuf,outsize - 4);
429 if (!send_smb(smbd_server_fd(),outbuf)) {
430 exit_server_cleanly("handle_aio_read_complete: send_smb "
434 DEBUG(10,("handle_aio_read_complete: scheduled aio_read completed "
435 "for file %s, offset %.0f, len = %u\n",
436 aio_ex->fsp->fsp_name, (double)aio_ex->acb.aio_offset,
437 (unsigned int)nread ));
442 /****************************************************************************
443 Complete the write and return the data or error back to the client.
444 Returns errno or zero if all ok.
445 *****************************************************************************/
447 static int handle_aio_write_complete(struct aio_extra *aio_ex)
450 files_struct *fsp = aio_ex->fsp;
451 char *outbuf = aio_ex->outbuf;
452 ssize_t numtowrite = aio_ex->acb.aio_nbytes;
453 ssize_t nwritten = SMB_VFS_AIO_RETURN(fsp,&aio_ex->acb);
455 if (fsp->aio_write_behind) {
456 if (nwritten != numtowrite) {
457 if (nwritten == -1) {
458 DEBUG(5,("handle_aio_write_complete: "
459 "aio_write_behind failed ! File %s "
460 "is corrupt ! Error %s\n",
461 fsp->fsp_name, strerror(errno) ));
464 DEBUG(0,("handle_aio_write_complete: "
465 "aio_write_behind failed ! File %s "
466 "is corrupt ! Wanted %u bytes but "
467 "only wrote %d\n", fsp->fsp_name,
468 (unsigned int)numtowrite,
473 DEBUG(10,("handle_aio_write_complete: "
474 "aio_write_behind completed for file %s\n",
480 /* We don't need outsize or set_message here as we've already set the
481 fixed size length when we set up the aio call. */
484 DEBUG( 3,( "handle_aio_write: file %s wanted %u bytes. "
485 "nwritten == %d. Error = %s\n",
486 fsp->fsp_name, (unsigned int)numtowrite,
487 (int)nwritten, strerror(errno) ));
489 /* If errno is ECANCELED then don't return anything to the
491 if (errno == ECANCELED) {
492 srv_cancel_sign_response(aio_ex->mid);
496 UNIXERROR(ERRHRD,ERRdiskfull);
499 BOOL write_through = BITSETW(aio_ex->inbuf+smb_vwv7,0);
501 SSVAL(outbuf,smb_vwv2,nwritten);
502 SSVAL(outbuf,smb_vwv4,(nwritten>>16)&1);
503 if (nwritten < (ssize_t)numtowrite) {
504 SCVAL(outbuf,smb_rcls,ERRHRD);
505 SSVAL(outbuf,smb_err,ERRdiskfull);
508 DEBUG(3,("handle_aio_write: fnum=%d num=%d wrote=%d\n",
509 fsp->fnum, (int)numtowrite, (int)nwritten));
510 sync_file(fsp->conn,fsp, write_through);
514 if (!send_smb(smbd_server_fd(),outbuf)) {
515 exit_server_cleanly("handle_aio_write: send_smb failed.");
518 DEBUG(10,("handle_aio_write_complete: scheduled aio_write completed "
519 "for file %s, offset %.0f, requested %u, written = %u\n",
520 fsp->fsp_name, (double)aio_ex->acb.aio_offset,
521 (unsigned int)numtowrite, (unsigned int)nwritten ));
526 /****************************************************************************
527 Handle any aio completion. Returns True if finished (and sets *perr if err
528 was non-zero), False if not.
529 *****************************************************************************/
531 static BOOL handle_aio_completed(struct aio_extra *aio_ex, int *perr)
535 /* Ensure the operation has really completed. */
536 if (SMB_VFS_AIO_ERROR(aio_ex->fsp, &aio_ex->acb) == EINPROGRESS) {
537 DEBUG(10,( "handle_aio_completed: operation mid %u still in "
538 "process for file %s\n",
539 aio_ex->mid, aio_ex->fsp->fsp_name ));
543 if (aio_ex->read_req) {
544 err = handle_aio_read_complete(aio_ex);
546 err = handle_aio_write_complete(aio_ex);
550 *perr = err; /* Only save non-zero errors. */
556 /****************************************************************************
557 Handle any aio completion inline.
558 Returns non-zero errno if fail or zero if all ok.
559 *****************************************************************************/
561 int process_aio_queue(void)
566 BlockSignals(True, RT_SIGNAL_AIO);
568 DEBUG(10,("process_aio_queue: signals_received = %d\n",
569 (int)signals_received));
570 DEBUG(10,("process_aio_queue: outstanding_aio_calls = %d\n",
571 outstanding_aio_calls));
573 if (!signals_received) {
574 BlockSignals(False, RT_SIGNAL_AIO);
578 /* Drain all the complete aio_reads. */
579 for (i = 0; i < signals_received; i++) {
580 uint16 mid = aio_pending_array[i];
581 files_struct *fsp = NULL;
582 struct aio_extra *aio_ex = find_aio_ex(mid);
585 DEBUG(3,("process_aio_queue: Can't find record to "
586 "match mid %u.\n", (unsigned int)mid));
587 srv_cancel_sign_response(mid);
593 /* file was closed whilst I/O was outstanding. Just
595 DEBUG( 3,( "process_aio_queue: file closed whilst "
596 "aio outstanding.\n"));
597 srv_cancel_sign_response(mid);
601 if (!handle_aio_completed(aio_ex, &ret)) {
605 delete_aio_ex(aio_ex);
608 outstanding_aio_calls -= signals_received;
609 signals_received = 0;
610 BlockSignals(False, RT_SIGNAL_AIO);
614 /****************************************************************************
615 We're doing write behind and the client closed the file. Wait up to 30
616 seconds (my arbitrary choice) for the aio to complete. Return 0 if all writes
617 completed, errno to return if not.
618 *****************************************************************************/
620 #define SMB_TIME_FOR_AIO_COMPLETE_WAIT 29
622 int wait_for_aio_completion(files_struct *fsp)
624 struct aio_extra *aio_ex;
625 const SMB_STRUCT_AIOCB **aiocb_list;
626 int aio_completion_count = 0;
627 time_t start_time = time(NULL);
630 for (seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT;
631 seconds_left >= 0;) {
636 aio_completion_count = 0;
637 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
638 if (aio_ex->fsp == fsp) {
639 aio_completion_count++;
643 if (!aio_completion_count) {
647 DEBUG(3,("wait_for_aio_completion: waiting for %d aio events "
648 "to complete.\n", aio_completion_count ));
650 aiocb_list = SMB_MALLOC_ARRAY(const SMB_STRUCT_AIOCB *,
651 aio_completion_count);
656 for( i = 0, aio_ex = aio_list_head;
658 aio_ex = aio_ex->next) {
659 if (aio_ex->fsp == fsp) {
660 aiocb_list[i++] = &aio_ex->acb;
664 /* Now wait up to seconds_left for completion. */
665 ts.tv_sec = seconds_left;
668 DEBUG(10,("wait_for_aio_completion: %d events, doing a wait "
670 aio_completion_count, seconds_left ));
672 err = SMB_VFS_AIO_SUSPEND(fsp, aiocb_list,
673 aio_completion_count, &ts);
675 DEBUG(10,("wait_for_aio_completion: returned err = %d, "
676 "errno = %s\n", err, strerror(errno) ));
678 if (err == -1 && errno == EAGAIN) {
679 DEBUG(0,("wait_for_aio_completion: aio_suspend timed "
680 "out waiting for %d events after a wait of "
681 "%d seconds\n", aio_completion_count,
684 cancel_aio_by_fsp(fsp);
685 SAFE_FREE(aiocb_list);
689 /* One or more events might have completed - process them if
691 for( i = 0; i < aio_completion_count; i++) {
692 uint16 mid = aiocb_list[i]->aio_sigevent.sigev_value.sival_int;
694 aio_ex = find_aio_ex(mid);
697 DEBUG(0, ("wait_for_aio_completion: mid %u "
698 "doesn't match an aio record\n",
699 (unsigned int)mid ));
703 if (!handle_aio_completed(aio_ex, &err)) {
706 delete_aio_ex(aio_ex);
709 SAFE_FREE(aiocb_list);
710 seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT
711 - (time(NULL) - start_time);
714 /* We timed out - we don't know why. Return ret if already an error,
716 DEBUG(10,("wait_for_aio_completion: aio_suspend timed out waiting "
718 aio_completion_count));
723 /****************************************************************************
724 Cancel any outstanding aio requests. The client doesn't care about the reply.
725 *****************************************************************************/
727 void cancel_aio_by_fsp(files_struct *fsp)
729 struct aio_extra *aio_ex;
731 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
732 if (aio_ex->fsp == fsp) {
733 /* Don't delete the aio_extra record as we may have
734 completed and don't yet know it. Just do the
735 aio_cancel call and return. */
736 SMB_VFS_AIO_CANCEL(fsp,fsp->fh->fd, &aio_ex->acb);
737 aio_ex->fsp = NULL; /* fsp will be closed when we
744 BOOL aio_finished(void)
749 void initialize_async_io_handler(void)
753 int process_aio_queue(void)
758 BOOL schedule_aio_read_and_X(connection_struct *conn,
759 char *inbuf, char *outbuf,
760 int length, int len_outbuf,
761 files_struct *fsp, SMB_OFF_T startpos,
767 BOOL schedule_aio_write_and_X(connection_struct *conn,
768 char *inbuf, char *outbuf,
769 int length, int len_outbuf,
770 files_struct *fsp, char *data,
777 void cancel_aio_by_fsp(files_struct *fsp)
781 BOOL wait_for_aio_completion(files_struct *fsp)