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 outstanding_aio_calls++;
201 /****************************************************************************
202 Set up an aio request from a SMBwriteX call.
203 *****************************************************************************/
205 bool schedule_aio_write_and_X(connection_struct *conn,
206 struct smb_request *req,
207 files_struct *fsp, char *data,
211 struct aio_extra *aio_ex;
214 bool write_through = BITSETW(req->vwv+7,0);
215 size_t min_aio_write_size = lp_aio_write_size(SNUM(conn));
218 if (fsp->base_fsp != NULL) {
219 /* No AIO on streams yet */
220 DEBUG(10, ("AIO on streams not yet supported\n"));
224 if ((!min_aio_write_size || (numtowrite < min_aio_write_size))
225 && !SMB_VFS_AIO_FORCE(fsp)) {
226 /* Too small a write for aio request. */
227 DEBUG(10,("schedule_aio_write_and_X: write size (%u) too "
228 "small for minimum aio_write of %u\n",
229 (unsigned int)numtowrite,
230 (unsigned int)min_aio_write_size ));
234 /* Only do this on non-chained and non-chaining reads not using the
236 if (req_is_in_chain(req) || (lp_write_cache_size(SNUM(conn)) != 0)) {
240 if (outstanding_aio_calls >= aio_pending_size) {
241 DEBUG(3,("schedule_aio_write_and_X: Already have %d aio "
242 "activities outstanding.\n",
243 outstanding_aio_calls ));
244 DEBUG(10,("schedule_aio_write_and_X: failed to schedule "
245 "aio_write for file %s, offset %.0f, len = %u "
247 fsp->fsp_name, (double)startpos,
248 (unsigned int)numtowrite,
249 (unsigned int)req->mid ));
253 bufsize = smb_size + 6*2;
255 if (!(aio_ex = create_aio_extra(fsp, bufsize))) {
256 DEBUG(0,("schedule_aio_write_and_X: malloc fail.\n"));
259 aio_ex->handle_completion = handle_aio_write_complete;
261 construct_reply_common_req(req, aio_ex->outbuf);
262 srv_set_message(aio_ex->outbuf, 6, 0, True);
263 SCVAL(aio_ex->outbuf,smb_vwv0,0xFF); /* Never a chained reply. */
267 /* Now set up the aio record for the write call. */
269 a->aio_fildes = fsp->fh->fd;
271 a->aio_nbytes = numtowrite;
272 a->aio_offset = startpos;
273 a->aio_sigevent.sigev_notify = SIGEV_SIGNAL;
274 a->aio_sigevent.sigev_signo = RT_SIGNAL_AIO;
275 a->aio_sigevent.sigev_value.sival_int = req->mid;
277 ret = SMB_VFS_AIO_WRITE(fsp, a);
279 DEBUG(3,("schedule_aio_wrote_and_X: aio_write failed. "
280 "Error %s\n", strerror(errno) ));
285 aio_ex->req = talloc_move(aio_ex, &req);
287 /* This should actually be improved to span the write. */
288 contend_level2_oplocks_begin(fsp, LEVEL2_CONTEND_WRITE);
289 contend_level2_oplocks_end(fsp, LEVEL2_CONTEND_WRITE);
291 if (!write_through && !lp_syncalways(SNUM(fsp->conn))
292 && fsp->aio_write_behind) {
293 /* Lie to the client and immediately claim we finished the
295 SSVAL(aio_ex->outbuf,smb_vwv2,numtowrite);
296 SSVAL(aio_ex->outbuf,smb_vwv4,(numtowrite>>16)&1);
297 show_msg(aio_ex->outbuf);
298 if (!srv_send_smb(smbd_server_fd(),aio_ex->outbuf,
299 true, aio_ex->req->seqnum+1,
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 outstanding_aio_calls++;
310 DEBUG(10,("schedule_aio_write_and_X: scheduled aio_write for file "
311 "%s, offset %.0f, len = %u (mid = %u) "
312 "outstanding_aio_calls = %d\n",
313 fsp->fsp_name, (double)startpos, (unsigned int)numtowrite,
314 (unsigned int)aio_ex->req->mid, outstanding_aio_calls ));
320 /****************************************************************************
321 Complete the read and return the data or error back to the client.
322 Returns errno or zero if all ok.
323 *****************************************************************************/
325 static int handle_aio_read_complete(struct aio_extra *aio_ex)
329 char *outbuf = aio_ex->outbuf;
330 char *data = smb_buf(outbuf);
331 ssize_t nread = SMB_VFS_AIO_RETURN(aio_ex->fsp,&aio_ex->acb);
334 /* We're relying here on the fact that if the fd is
335 closed then the aio will complete and aio_return
336 will return an error. Hopefully this is
339 /* If errno is ECANCELED then don't return anything to the
341 if (errno == ECANCELED) {
345 DEBUG( 3,( "handle_aio_read_complete: file %s nread == -1. "
347 aio_ex->fsp->fsp_name, strerror(errno) ));
350 ERROR_NT(map_nt_error_from_unix(ret));
351 outsize = srv_set_message(outbuf,0,0,true);
353 outsize = srv_set_message(outbuf,12,nread,False);
354 SSVAL(outbuf,smb_vwv2,0xFFFF); /* Remaining - must be * -1. */
355 SSVAL(outbuf,smb_vwv5,nread);
356 SSVAL(outbuf,smb_vwv6,smb_offset(data,outbuf));
357 SSVAL(outbuf,smb_vwv7,((nread >> 16) & 1));
358 SSVAL(smb_buf(outbuf),-2,nread);
360 aio_ex->fsp->fh->pos = aio_ex->acb.aio_offset + nread;
361 aio_ex->fsp->fh->position_information = aio_ex->fsp->fh->pos;
363 DEBUG( 3, ( "handle_aio_read_complete file %s max=%d "
365 aio_ex->fsp->fsp_name,
366 (int)aio_ex->acb.aio_nbytes, (int)nread ) );
369 smb_setlen(outbuf,outsize - 4);
371 if (!srv_send_smb(smbd_server_fd(),outbuf,
372 true, aio_ex->req->seqnum+1,
373 IS_CONN_ENCRYPTED(aio_ex->fsp->conn), NULL)) {
374 exit_server_cleanly("handle_aio_read_complete: srv_send_smb "
378 DEBUG(10,("handle_aio_read_complete: scheduled aio_read completed "
379 "for file %s, offset %.0f, len = %u\n",
380 aio_ex->fsp->fsp_name, (double)aio_ex->acb.aio_offset,
381 (unsigned int)nread ));
386 /****************************************************************************
387 Complete the write and return the data or error back to the client.
388 Returns errno or zero if all ok.
389 *****************************************************************************/
391 static int handle_aio_write_complete(struct aio_extra *aio_ex)
394 files_struct *fsp = aio_ex->fsp;
395 char *outbuf = aio_ex->outbuf;
396 ssize_t numtowrite = aio_ex->acb.aio_nbytes;
397 ssize_t nwritten = SMB_VFS_AIO_RETURN(fsp,&aio_ex->acb);
399 if (fsp->aio_write_behind) {
400 if (nwritten != numtowrite) {
401 if (nwritten == -1) {
402 DEBUG(5,("handle_aio_write_complete: "
403 "aio_write_behind failed ! File %s "
404 "is corrupt ! Error %s\n",
405 fsp->fsp_name, strerror(errno) ));
408 DEBUG(0,("handle_aio_write_complete: "
409 "aio_write_behind failed ! File %s "
410 "is corrupt ! Wanted %u bytes but "
411 "only wrote %d\n", fsp->fsp_name,
412 (unsigned int)numtowrite,
417 DEBUG(10,("handle_aio_write_complete: "
418 "aio_write_behind completed for file %s\n",
424 /* We don't need outsize or set_message here as we've already set the
425 fixed size length when we set up the aio call. */
428 DEBUG( 3,( "handle_aio_write: file %s wanted %u bytes. "
429 "nwritten == %d. Error = %s\n",
430 fsp->fsp_name, (unsigned int)numtowrite,
431 (int)nwritten, strerror(errno) ));
433 /* If errno is ECANCELED then don't return anything to the
435 if (errno == ECANCELED) {
440 ERROR_BOTH(map_nt_error_from_unix(ret), ERRHRD, ERRdiskfull);
441 srv_set_message(outbuf,0,0,true);
443 bool write_through = BITSETW(aio_ex->req->vwv+7,0);
446 SSVAL(outbuf,smb_vwv2,nwritten);
447 SSVAL(outbuf,smb_vwv4,(nwritten>>16)&1);
448 if (nwritten < (ssize_t)numtowrite) {
449 SCVAL(outbuf,smb_rcls,ERRHRD);
450 SSVAL(outbuf,smb_err,ERRdiskfull);
453 DEBUG(3,("handle_aio_write: fnum=%d num=%d wrote=%d\n",
454 fsp->fnum, (int)numtowrite, (int)nwritten));
455 status = sync_file(fsp->conn,fsp, write_through);
456 if (!NT_STATUS_IS_OK(status)) {
458 ERROR_BOTH(map_nt_error_from_unix(ret),
459 ERRHRD, ERRdiskfull);
460 srv_set_message(outbuf,0,0,true);
461 DEBUG(5,("handle_aio_write: sync_file for %s returned %s\n",
462 fsp->fsp_name, nt_errstr(status) ));
465 aio_ex->fsp->fh->pos = aio_ex->acb.aio_offset + nwritten;
469 if (!srv_send_smb(smbd_server_fd(),outbuf,
470 true, aio_ex->req->seqnum+1,
471 IS_CONN_ENCRYPTED(fsp->conn),
473 exit_server_cleanly("handle_aio_write: srv_send_smb failed.");
476 DEBUG(10,("handle_aio_write_complete: scheduled aio_write completed "
477 "for file %s, offset %.0f, requested %u, written = %u\n",
478 fsp->fsp_name, (double)aio_ex->acb.aio_offset,
479 (unsigned int)numtowrite, (unsigned int)nwritten ));
484 /****************************************************************************
485 Handle any aio completion. Returns True if finished (and sets *perr if err
486 was non-zero), False if not.
487 *****************************************************************************/
489 static bool handle_aio_completed(struct aio_extra *aio_ex, int *perr)
494 DEBUG(3, ("handle_aio_completed: Non-existing aio_ex passed\n"));
498 /* Ensure the operation has really completed. */
499 if (SMB_VFS_AIO_ERROR(aio_ex->fsp, &aio_ex->acb) == EINPROGRESS) {
500 DEBUG(10,( "handle_aio_completed: operation mid %u still in "
501 "process for file %s\n",
502 aio_ex->req->mid, aio_ex->fsp->fsp_name ));
506 err = aio_ex->handle_completion(aio_ex);
508 *perr = err; /* Only save non-zero errors. */
514 /****************************************************************************
515 Handle any aio completion inline.
516 Returns non-zero errno if fail or zero if all ok.
517 *****************************************************************************/
519 void smbd_aio_complete_mid(unsigned int mid)
521 files_struct *fsp = NULL;
522 struct aio_extra *aio_ex = find_aio_ex(mid);
525 DEBUG(10,("smbd_aio_complete_mid: mid[%u]\n", mid));
528 DEBUG(3,("smbd_aio_complete_mid: Can't find record to "
529 "match mid %u.\n", mid));
535 /* file was closed whilst I/O was outstanding. Just
537 DEBUG( 3,( "smbd_aio_complete_mid: file closed whilst "
538 "aio outstanding (mid[%u]).\n", mid));
542 if (!handle_aio_completed(aio_ex, &ret)) {
549 static void smbd_aio_signal_handler(struct tevent_context *ev_ctx,
550 struct tevent_signal *se,
551 int signum, int count,
552 void *_info, void *private_data)
554 siginfo_t *info = (siginfo_t *)_info;
555 unsigned int mid = (unsigned int)info->si_value.sival_int;
557 smbd_aio_complete_mid(mid);
560 /****************************************************************************
561 We're doing write behind and the client closed the file. Wait up to 30
562 seconds (my arbitrary choice) for the aio to complete. Return 0 if all writes
563 completed, errno to return if not.
564 *****************************************************************************/
566 #define SMB_TIME_FOR_AIO_COMPLETE_WAIT 29
568 int wait_for_aio_completion(files_struct *fsp)
570 struct aio_extra *aio_ex;
571 const SMB_STRUCT_AIOCB **aiocb_list;
572 int aio_completion_count = 0;
573 time_t start_time = time(NULL);
576 for (seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT;
577 seconds_left >= 0;) {
582 aio_completion_count = 0;
583 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
584 if (aio_ex->fsp == fsp) {
585 aio_completion_count++;
589 if (!aio_completion_count) {
593 DEBUG(3,("wait_for_aio_completion: waiting for %d aio events "
594 "to complete.\n", aio_completion_count ));
596 aiocb_list = SMB_MALLOC_ARRAY(const SMB_STRUCT_AIOCB *,
597 aio_completion_count);
602 for( i = 0, aio_ex = aio_list_head;
604 aio_ex = aio_ex->next) {
605 if (aio_ex->fsp == fsp) {
606 aiocb_list[i++] = &aio_ex->acb;
610 /* Now wait up to seconds_left for completion. */
611 ts.tv_sec = seconds_left;
614 DEBUG(10,("wait_for_aio_completion: %d events, doing a wait "
616 aio_completion_count, seconds_left ));
618 err = SMB_VFS_AIO_SUSPEND(fsp, aiocb_list,
619 aio_completion_count, &ts);
621 DEBUG(10,("wait_for_aio_completion: returned err = %d, "
622 "errno = %s\n", err, strerror(errno) ));
624 if (err == -1 && errno == EAGAIN) {
625 DEBUG(0,("wait_for_aio_completion: aio_suspend timed "
626 "out waiting for %d events after a wait of "
627 "%d seconds\n", aio_completion_count,
630 cancel_aio_by_fsp(fsp);
631 SAFE_FREE(aiocb_list);
635 /* One or more events might have completed - process them if
637 for( i = 0; i < aio_completion_count; i++) {
638 uint16 mid = aiocb_list[i]->aio_sigevent.sigev_value.sival_int;
640 aio_ex = find_aio_ex(mid);
643 DEBUG(0, ("wait_for_aio_completion: mid %u "
644 "doesn't match an aio record\n",
645 (unsigned int)mid ));
649 if (!handle_aio_completed(aio_ex, &err)) {
655 SAFE_FREE(aiocb_list);
656 seconds_left = SMB_TIME_FOR_AIO_COMPLETE_WAIT
657 - (time(NULL) - start_time);
660 /* We timed out - we don't know why. Return ret if already an error,
662 DEBUG(10,("wait_for_aio_completion: aio_suspend timed out waiting "
664 aio_completion_count));
669 /****************************************************************************
670 Cancel any outstanding aio requests. The client doesn't care about the reply.
671 *****************************************************************************/
673 void cancel_aio_by_fsp(files_struct *fsp)
675 struct aio_extra *aio_ex;
677 for( aio_ex = aio_list_head; aio_ex; aio_ex = aio_ex->next) {
678 if (aio_ex->fsp == fsp) {
679 /* Don't delete the aio_extra record as we may have
680 completed and don't yet know it. Just do the
681 aio_cancel call and return. */
682 SMB_VFS_AIO_CANCEL(fsp, &aio_ex->acb);
683 aio_ex->fsp = NULL; /* fsp will be closed when we
689 /****************************************************************************
690 Initialize the signal handler for aio read/write.
691 *****************************************************************************/
693 void initialize_async_io_handler(void)
695 aio_signal_event = tevent_add_signal(smbd_event_context(),
696 smbd_event_context(),
697 RT_SIGNAL_AIO, SA_SIGINFO,
698 smbd_aio_signal_handler,
700 if (!aio_signal_event) {
701 exit_server("Failed to setup RT_SIGNAL_AIO handler");
704 /* tevent supports 100 signal with SA_SIGINFO */
705 aio_pending_size = 100;
709 void initialize_async_io_handler(void)
713 bool schedule_aio_read_and_X(connection_struct *conn,
714 struct smb_request *req,
715 files_struct *fsp, SMB_OFF_T startpos,
721 bool schedule_aio_write_and_X(connection_struct *conn,
722 struct smb_request *req,
723 files_struct *fsp, char *data,
730 void cancel_aio_by_fsp(files_struct *fsp)
734 int wait_for_aio_completion(files_struct *fsp)
739 void smbd_aio_complete_mid(unsigned int mid);