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 (req_is_in_chain(req
) || (lp_write_cache_size(SNUM(conn
)) != 0)) {
148 if (outstanding_aio_calls
>= aio_pending_size
) {
149 DEBUG(10,("schedule_aio_read_and_X: Already have %d aio "
150 "activities outstanding.\n",
151 outstanding_aio_calls
));
155 /* The following is safe from integer wrap as we've already checked
156 smb_maxcnt is 128k or less. Wct is 12 for read replies */
158 bufsize
= smb_size
+ 12 * 2 + smb_maxcnt
;
160 if ((aio_ex
= create_aio_extra(fsp
, bufsize
)) == NULL
) {
161 DEBUG(10,("schedule_aio_read_and_X: malloc fail.\n"));
164 aio_ex
->handle_completion
= handle_aio_read_complete
;
166 construct_reply_common_req(req
, aio_ex
->outbuf
);
167 srv_set_message(aio_ex
->outbuf
, 12, 0, True
);
168 SCVAL(aio_ex
->outbuf
,smb_vwv0
,0xFF); /* Never a chained reply. */
172 /* Now set up the aio record for the read call. */
174 a
->aio_fildes
= fsp
->fh
->fd
;
175 a
->aio_buf
= smb_buf(aio_ex
->outbuf
);
176 a
->aio_nbytes
= smb_maxcnt
;
177 a
->aio_offset
= startpos
;
178 a
->aio_sigevent
.sigev_notify
= SIGEV_SIGNAL
;
179 a
->aio_sigevent
.sigev_signo
= RT_SIGNAL_AIO
;
180 a
->aio_sigevent
.sigev_value
.sival_int
= req
->mid
;
182 ret
= SMB_VFS_AIO_READ(fsp
, a
);
184 DEBUG(0,("schedule_aio_read_and_X: aio_read failed. "
185 "Error %s\n", strerror(errno
) ));
190 aio_ex
->req
= talloc_move(aio_ex
, &req
);
192 DEBUG(10,("schedule_aio_read_and_X: scheduled aio_read for file %s, "
193 "offset %.0f, len = %u (mid = %u)\n",
194 fsp
->fsp_name
, (double)startpos
, (unsigned int)smb_maxcnt
,
195 (unsigned int)aio_ex
->req
->mid
));
197 srv_defer_sign_response(aio_ex
->req
->mid
);
198 outstanding_aio_calls
++;
202 /****************************************************************************
203 Set up an aio request from a SMBwriteX call.
204 *****************************************************************************/
206 bool schedule_aio_write_and_X(connection_struct
*conn
,
207 struct smb_request
*req
,
208 files_struct
*fsp
, char *data
,
212 struct aio_extra
*aio_ex
;
215 bool write_through
= BITSETW(req
->vwv
+7,0);
216 size_t min_aio_write_size
= lp_aio_write_size(SNUM(conn
));
219 if (fsp
->base_fsp
!= NULL
) {
220 /* No AIO on streams yet */
221 DEBUG(10, ("AIO on streams not yet supported\n"));
225 if ((!min_aio_write_size
|| (numtowrite
< min_aio_write_size
))
226 && !SMB_VFS_AIO_FORCE(fsp
)) {
227 /* Too small a write for aio request. */
228 DEBUG(10,("schedule_aio_write_and_X: write size (%u) too "
229 "small for minimum aio_write of %u\n",
230 (unsigned int)numtowrite
,
231 (unsigned int)min_aio_write_size
));
235 /* Only do this on non-chained and non-chaining reads not using the
237 if (req_is_in_chain(req
) || (lp_write_cache_size(SNUM(conn
)) != 0)) {
241 if (outstanding_aio_calls
>= aio_pending_size
) {
242 DEBUG(3,("schedule_aio_write_and_X: Already have %d aio "
243 "activities outstanding.\n",
244 outstanding_aio_calls
));
245 DEBUG(10,("schedule_aio_write_and_X: failed to schedule "
246 "aio_write for file %s, offset %.0f, len = %u "
248 fsp
->fsp_name
, (double)startpos
,
249 (unsigned int)numtowrite
,
250 (unsigned int)req
->mid
));
254 bufsize
= smb_size
+ 6*2;
256 if (!(aio_ex
= create_aio_extra(fsp
, bufsize
))) {
257 DEBUG(0,("schedule_aio_write_and_X: malloc fail.\n"));
260 aio_ex
->handle_completion
= handle_aio_write_complete
;
262 construct_reply_common_req(req
, aio_ex
->outbuf
);
263 srv_set_message(aio_ex
->outbuf
, 6, 0, True
);
264 SCVAL(aio_ex
->outbuf
,smb_vwv0
,0xFF); /* Never a chained reply. */
268 /* Now set up the aio record for the write call. */
270 a
->aio_fildes
= fsp
->fh
->fd
;
272 a
->aio_nbytes
= numtowrite
;
273 a
->aio_offset
= startpos
;
274 a
->aio_sigevent
.sigev_notify
= SIGEV_SIGNAL
;
275 a
->aio_sigevent
.sigev_signo
= RT_SIGNAL_AIO
;
276 a
->aio_sigevent
.sigev_value
.sival_int
= req
->mid
;
278 ret
= SMB_VFS_AIO_WRITE(fsp
, a
);
280 DEBUG(3,("schedule_aio_wrote_and_X: aio_write failed. "
281 "Error %s\n", strerror(errno
) ));
286 aio_ex
->req
= talloc_move(aio_ex
, &req
);
288 /* This should actually be improved to span the write. */
289 contend_level2_oplocks_begin(fsp
, LEVEL2_CONTEND_WRITE
);
290 contend_level2_oplocks_end(fsp
, LEVEL2_CONTEND_WRITE
);
292 if (!write_through
&& !lp_syncalways(SNUM(fsp
->conn
))
293 && fsp
->aio_write_behind
) {
294 /* Lie to the client and immediately claim we finished the
296 SSVAL(aio_ex
->outbuf
,smb_vwv2
,numtowrite
);
297 SSVAL(aio_ex
->outbuf
,smb_vwv4
,(numtowrite
>>16)&1);
298 show_msg(aio_ex
->outbuf
);
299 if (!srv_send_smb(smbd_server_fd(),aio_ex
->outbuf
,
300 IS_CONN_ENCRYPTED(fsp
->conn
),
302 exit_server_cleanly("handle_aio_write: srv_send_smb "
305 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write "
306 "behind for file %s\n", fsp
->fsp_name
));
308 srv_defer_sign_response(aio_ex
->req
->mid
);
310 outstanding_aio_calls
++;
312 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write for file "
313 "%s, offset %.0f, len = %u (mid = %u) "
314 "outstanding_aio_calls = %d\n",
315 fsp
->fsp_name
, (double)startpos
, (unsigned int)numtowrite
,
316 (unsigned int)aio_ex
->req
->mid
, outstanding_aio_calls
));
322 /****************************************************************************
323 Complete the read and return the data or error back to the client.
324 Returns errno or zero if all ok.
325 *****************************************************************************/
327 static int handle_aio_read_complete(struct aio_extra
*aio_ex
)
331 char *outbuf
= aio_ex
->outbuf
;
332 char *data
= smb_buf(outbuf
);
333 ssize_t nread
= SMB_VFS_AIO_RETURN(aio_ex
->fsp
,&aio_ex
->acb
);
336 /* We're relying here on the fact that if the fd is
337 closed then the aio will complete and aio_return
338 will return an error. Hopefully this is
341 /* If errno is ECANCELED then don't return anything to the
343 if (errno
== ECANCELED
) {
344 srv_cancel_sign_response(aio_ex
->req
->mid
, false);
348 DEBUG( 3,( "handle_aio_read_complete: file %s nread == -1. "
350 aio_ex
->fsp
->fsp_name
, strerror(errno
) ));
353 ERROR_NT(map_nt_error_from_unix(ret
));
354 outsize
= srv_set_message(outbuf
,0,0,true);
356 outsize
= srv_set_message(outbuf
,12,nread
,False
);
357 SSVAL(outbuf
,smb_vwv2
,0xFFFF); /* Remaining - must be * -1. */
358 SSVAL(outbuf
,smb_vwv5
,nread
);
359 SSVAL(outbuf
,smb_vwv6
,smb_offset(data
,outbuf
));
360 SSVAL(outbuf
,smb_vwv7
,((nread
>> 16) & 1));
361 SSVAL(smb_buf(outbuf
),-2,nread
);
363 aio_ex
->fsp
->fh
->pos
= aio_ex
->acb
.aio_offset
+ nread
;
364 aio_ex
->fsp
->fh
->position_information
= aio_ex
->fsp
->fh
->pos
;
366 DEBUG( 3, ( "handle_aio_read_complete file %s max=%d "
368 aio_ex
->fsp
->fsp_name
,
369 (int)aio_ex
->acb
.aio_nbytes
, (int)nread
) );
372 smb_setlen(outbuf
,outsize
- 4);
374 if (!srv_send_smb(smbd_server_fd(),outbuf
,
375 IS_CONN_ENCRYPTED(aio_ex
->fsp
->conn
), NULL
)) {
376 exit_server_cleanly("handle_aio_read_complete: srv_send_smb "
380 DEBUG(10,("handle_aio_read_complete: scheduled aio_read completed "
381 "for file %s, offset %.0f, len = %u\n",
382 aio_ex
->fsp
->fsp_name
, (double)aio_ex
->acb
.aio_offset
,
383 (unsigned int)nread
));
388 /****************************************************************************
389 Complete the write and return the data or error back to the client.
390 Returns errno or zero if all ok.
391 *****************************************************************************/
393 static int handle_aio_write_complete(struct aio_extra
*aio_ex
)
396 files_struct
*fsp
= aio_ex
->fsp
;
397 char *outbuf
= aio_ex
->outbuf
;
398 ssize_t numtowrite
= aio_ex
->acb
.aio_nbytes
;
399 ssize_t nwritten
= SMB_VFS_AIO_RETURN(fsp
,&aio_ex
->acb
);
401 if (fsp
->aio_write_behind
) {
402 if (nwritten
!= numtowrite
) {
403 if (nwritten
== -1) {
404 DEBUG(5,("handle_aio_write_complete: "
405 "aio_write_behind failed ! File %s "
406 "is corrupt ! Error %s\n",
407 fsp
->fsp_name
, strerror(errno
) ));
410 DEBUG(0,("handle_aio_write_complete: "
411 "aio_write_behind failed ! File %s "
412 "is corrupt ! Wanted %u bytes but "
413 "only wrote %d\n", fsp
->fsp_name
,
414 (unsigned int)numtowrite
,
419 DEBUG(10,("handle_aio_write_complete: "
420 "aio_write_behind completed for file %s\n",
426 /* We don't need outsize or set_message here as we've already set the
427 fixed size length when we set up the aio call. */
430 DEBUG( 3,( "handle_aio_write: file %s wanted %u bytes. "
431 "nwritten == %d. Error = %s\n",
432 fsp
->fsp_name
, (unsigned int)numtowrite
,
433 (int)nwritten
, strerror(errno
) ));
435 /* If errno is ECANCELED then don't return anything to the
437 if (errno
== ECANCELED
) {
438 srv_cancel_sign_response(aio_ex
->req
->mid
, false);
443 ERROR_BOTH(map_nt_error_from_unix(ret
), ERRHRD
, ERRdiskfull
);
444 srv_set_message(outbuf
,0,0,true);
446 bool write_through
= BITSETW(aio_ex
->req
->vwv
+7,0);
449 SSVAL(outbuf
,smb_vwv2
,nwritten
);
450 SSVAL(outbuf
,smb_vwv4
,(nwritten
>>16)&1);
451 if (nwritten
< (ssize_t
)numtowrite
) {
452 SCVAL(outbuf
,smb_rcls
,ERRHRD
);
453 SSVAL(outbuf
,smb_err
,ERRdiskfull
);
456 DEBUG(3,("handle_aio_write: fnum=%d num=%d wrote=%d\n",
457 fsp
->fnum
, (int)numtowrite
, (int)nwritten
));
458 status
= sync_file(fsp
->conn
,fsp
, write_through
);
459 if (!NT_STATUS_IS_OK(status
)) {
461 ERROR_BOTH(map_nt_error_from_unix(ret
),
462 ERRHRD
, ERRdiskfull
);
463 srv_set_message(outbuf
,0,0,true);
464 DEBUG(5,("handle_aio_write: sync_file for %s returned %s\n",
465 fsp
->fsp_name
, nt_errstr(status
) ));
468 aio_ex
->fsp
->fh
->pos
= aio_ex
->acb
.aio_offset
+ nwritten
;
472 if (!srv_send_smb(smbd_server_fd(),outbuf
,IS_CONN_ENCRYPTED(fsp
->conn
),
474 exit_server_cleanly("handle_aio_write: srv_send_smb failed.");
477 DEBUG(10,("handle_aio_write_complete: scheduled aio_write completed "
478 "for file %s, offset %.0f, requested %u, written = %u\n",
479 fsp
->fsp_name
, (double)aio_ex
->acb
.aio_offset
,
480 (unsigned int)numtowrite
, (unsigned int)nwritten
));
485 /****************************************************************************
486 Handle any aio completion. Returns True if finished (and sets *perr if err
487 was non-zero), False if not.
488 *****************************************************************************/
490 static bool handle_aio_completed(struct aio_extra
*aio_ex
, int *perr
)
495 DEBUG(3, ("handle_aio_completed: Non-existing aio_ex passed\n"));
499 /* Ensure the operation has really completed. */
500 if (SMB_VFS_AIO_ERROR(aio_ex
->fsp
, &aio_ex
->acb
) == EINPROGRESS
) {
501 DEBUG(10,( "handle_aio_completed: operation mid %u still in "
502 "process for file %s\n",
503 aio_ex
->req
->mid
, aio_ex
->fsp
->fsp_name
));
507 err
= aio_ex
->handle_completion(aio_ex
);
509 *perr
= err
; /* Only save non-zero errors. */
515 /****************************************************************************
516 Handle any aio completion inline.
517 Returns non-zero errno if fail or zero if all ok.
518 *****************************************************************************/
520 void smbd_aio_complete_mid(unsigned int mid
)
522 files_struct
*fsp
= NULL
;
523 struct aio_extra
*aio_ex
= find_aio_ex(mid
);
526 DEBUG(10,("smbd_aio_complete_mid: mid[%u]\n", mid
));
529 DEBUG(3,("smbd_aio_complete_mid: Can't find record to "
530 "match mid %u.\n", mid
));
531 srv_cancel_sign_response(mid
, false);
537 /* file was closed whilst I/O was outstanding. Just
539 DEBUG( 3,( "smbd_aio_complete_mid: file closed whilst "
540 "aio outstanding (mid[%u]).\n", mid
));
541 srv_cancel_sign_response(mid
, false);
545 if (!handle_aio_completed(aio_ex
, &ret
)) {
552 static void smbd_aio_signal_handler(struct tevent_context
*ev_ctx
,
553 struct tevent_signal
*se
,
554 int signum
, int count
,
555 void *_info
, void *private_data
)
557 siginfo_t
*info
= (siginfo_t
*)_info
;
558 unsigned int mid
= (unsigned int)info
->si_value
.sival_int
;
560 smbd_aio_complete_mid(mid
);
563 /****************************************************************************
564 We're doing write behind and the client closed the file. Wait up to 30
565 seconds (my arbitrary choice) for the aio to complete. Return 0 if all writes
566 completed, errno to return if not.
567 *****************************************************************************/
569 #define SMB_TIME_FOR_AIO_COMPLETE_WAIT 29
571 int wait_for_aio_completion(files_struct
*fsp
)
573 struct aio_extra
*aio_ex
;
574 const SMB_STRUCT_AIOCB
**aiocb_list
;
575 int aio_completion_count
= 0;
576 time_t start_time
= time(NULL
);
579 for (seconds_left
= SMB_TIME_FOR_AIO_COMPLETE_WAIT
;
580 seconds_left
>= 0;) {
585 aio_completion_count
= 0;
586 for( aio_ex
= aio_list_head
; aio_ex
; aio_ex
= aio_ex
->next
) {
587 if (aio_ex
->fsp
== fsp
) {
588 aio_completion_count
++;
592 if (!aio_completion_count
) {
596 DEBUG(3,("wait_for_aio_completion: waiting for %d aio events "
597 "to complete.\n", aio_completion_count
));
599 aiocb_list
= SMB_MALLOC_ARRAY(const SMB_STRUCT_AIOCB
*,
600 aio_completion_count
);
605 for( i
= 0, aio_ex
= aio_list_head
;
607 aio_ex
= aio_ex
->next
) {
608 if (aio_ex
->fsp
== fsp
) {
609 aiocb_list
[i
++] = &aio_ex
->acb
;
613 /* Now wait up to seconds_left for completion. */
614 ts
.tv_sec
= seconds_left
;
617 DEBUG(10,("wait_for_aio_completion: %d events, doing a wait "
619 aio_completion_count
, seconds_left
));
621 err
= SMB_VFS_AIO_SUSPEND(fsp
, aiocb_list
,
622 aio_completion_count
, &ts
);
624 DEBUG(10,("wait_for_aio_completion: returned err = %d, "
625 "errno = %s\n", err
, strerror(errno
) ));
627 if (err
== -1 && errno
== EAGAIN
) {
628 DEBUG(0,("wait_for_aio_completion: aio_suspend timed "
629 "out waiting for %d events after a wait of "
630 "%d seconds\n", aio_completion_count
,
633 cancel_aio_by_fsp(fsp
);
634 SAFE_FREE(aiocb_list
);
638 /* One or more events might have completed - process them if
640 for( i
= 0; i
< aio_completion_count
; i
++) {
641 uint16 mid
= aiocb_list
[i
]->aio_sigevent
.sigev_value
.sival_int
;
643 aio_ex
= find_aio_ex(mid
);
646 DEBUG(0, ("wait_for_aio_completion: mid %u "
647 "doesn't match an aio record\n",
648 (unsigned int)mid
));
652 if (!handle_aio_completed(aio_ex
, &err
)) {
658 SAFE_FREE(aiocb_list
);
659 seconds_left
= SMB_TIME_FOR_AIO_COMPLETE_WAIT
660 - (time(NULL
) - start_time
);
663 /* We timed out - we don't know why. Return ret if already an error,
665 DEBUG(10,("wait_for_aio_completion: aio_suspend timed out waiting "
667 aio_completion_count
));
672 /****************************************************************************
673 Cancel any outstanding aio requests. The client doesn't care about the reply.
674 *****************************************************************************/
676 void cancel_aio_by_fsp(files_struct
*fsp
)
678 struct aio_extra
*aio_ex
;
680 for( aio_ex
= aio_list_head
; aio_ex
; aio_ex
= aio_ex
->next
) {
681 if (aio_ex
->fsp
== fsp
) {
682 /* Don't delete the aio_extra record as we may have
683 completed and don't yet know it. Just do the
684 aio_cancel call and return. */
685 SMB_VFS_AIO_CANCEL(fsp
, &aio_ex
->acb
);
686 aio_ex
->fsp
= NULL
; /* fsp will be closed when we
692 /****************************************************************************
693 Initialize the signal handler for aio read/write.
694 *****************************************************************************/
696 void initialize_async_io_handler(void)
698 aio_signal_event
= tevent_add_signal(smbd_event_context(),
699 smbd_event_context(),
700 RT_SIGNAL_AIO
, SA_SIGINFO
,
701 smbd_aio_signal_handler
,
703 if (!aio_signal_event
) {
704 exit_server("Failed to setup RT_SIGNAL_AIO handler");
707 /* tevent supports 100 signal with SA_SIGINFO */
708 aio_pending_size
= 100;
712 void initialize_async_io_handler(void)
716 bool schedule_aio_read_and_X(connection_struct
*conn
,
717 struct smb_request
*req
,
718 files_struct
*fsp
, SMB_OFF_T startpos
,
724 bool schedule_aio_write_and_X(connection_struct
*conn
,
725 struct smb_request
*req
,
726 files_struct
*fsp
, char *data
,
733 void cancel_aio_by_fsp(files_struct
*fsp
)
737 int wait_for_aio_completion(files_struct
*fsp
)
742 void smbd_aio_complete_mid(unsigned int mid
);