ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libeio/eio.c
(Generate patch)

Comparing libeio/eio.c (file contents):
Revision 1.16 by root, Fri May 30 04:50:21 2008 UTC vs.
Revision 1.30 by root, Wed Jun 3 12:24:49 2009 UTC

… …
36 * provisions above, a recipient may use your version of this file under 36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL. 37 * either the BSD or the GPL.
38 */ 38 */
39 39
40#include "eio.h" 40#include "eio.h"
41
42#ifdef EIO_STACKSIZE
43# define XTHREAD_STACKSIZE EIO_STACKSIZE
44#endif
41#include "xthread.h" 45#include "xthread.h"
42 46
43#include <errno.h> 47#include <errno.h>
44#include <stddef.h> 48#include <stddef.h>
45#include <stdlib.h> 49#include <stdlib.h>
… …
120 124
121#define EIO_TICKS ((1000000 + 1023) >> 10) 125#define EIO_TICKS ((1000000 + 1023) >> 10)
122 126
123/*****************************************************************************/ 127/*****************************************************************************/
124 128
129#if __GNUC__ >= 3
130# define expect(expr,value) __builtin_expect ((expr),(value))
131#else
132# define expect(expr,value) (expr)
133#endif
134
135#define expect_false(expr) expect ((expr) != 0, 0)
136#define expect_true(expr) expect ((expr) != 0, 1)
137
138/*****************************************************************************/
139
125#define ETP_PRI_MIN EIO_PRI_MIN 140#define ETP_PRI_MIN EIO_PRI_MIN
126#define ETP_PRI_MAX EIO_PRI_MAX 141#define ETP_PRI_MAX EIO_PRI_MAX
127 142
128struct etp_worker; 143struct etp_worker;
129 144
… …
177static mutex_t wrklock = X_MUTEX_INIT; 192static mutex_t wrklock = X_MUTEX_INIT;
178static mutex_t reslock = X_MUTEX_INIT; 193static mutex_t reslock = X_MUTEX_INIT;
179static mutex_t reqlock = X_MUTEX_INIT; 194static mutex_t reqlock = X_MUTEX_INIT;
180static cond_t reqwait = X_COND_INIT; 195static cond_t reqwait = X_COND_INIT;
181 196
197#if !HAVE_PREADWRITE
198/*
199 * make our pread/pwrite emulation safe against themselves, but not against
200 * normal read/write by using a mutex. slows down execution a lot,
201 * but that's your problem, not mine.
202 */
203static mutex_t preadwritelock = X_MUTEX_INIT;
204#endif
205
182typedef struct etp_worker 206typedef struct etp_worker
183{ 207{
184 /* locked by wrklock */ 208 /* locked by wrklock */
185 struct etp_worker *prev, *next; 209 struct etp_worker *prev, *next;
186 210
… …
330 354
331static void etp_atfork_child (void) 355static void etp_atfork_child (void)
332{ 356{
333 ETP_REQ *prv; 357 ETP_REQ *prv;
334 358
335 while (prv = reqq_shift (&req_queue)) 359 while ((prv = reqq_shift (&req_queue)))
336 ETP_DESTROY (prv); 360 ETP_DESTROY (prv);
337 361
338 while (prv = reqq_shift (&res_queue)) 362 while ((prv = reqq_shift (&res_queue)))
339 ETP_DESTROY (prv); 363 ETP_DESTROY (prv);
340 364
341 while (wrk_first.next != &wrk_first) 365 while (wrk_first.next != &wrk_first)
342 { 366 {
343 etp_worker *wrk = wrk_first.next; 367 etp_worker *wrk = wrk_first.next;
… …
371 395
372 pthread_once (&doinit, etp_once_init); 396 pthread_once (&doinit, etp_once_init);
373 397
374 want_poll_cb = want_poll; 398 want_poll_cb = want_poll;
375 done_poll_cb = done_poll; 399 done_poll_cb = done_poll;
400
401 return 0;
376} 402}
377 403
378X_THREAD_PROC (etp_proc); 404X_THREAD_PROC (etp_proc);
379 405
380static void etp_start_thread (void) 406static void etp_start_thread (void)
… …
400 X_UNLOCK (wrklock); 426 X_UNLOCK (wrklock);
401} 427}
402 428
403static void etp_maybe_start_thread (void) 429static void etp_maybe_start_thread (void)
404{ 430{
405 if (etp_nthreads () >= wanted) 431 if (expect_true (etp_nthreads () >= wanted))
406 return; 432 return;
407 433
408 /* todo: maybe use idle here, but might be less exact */ 434 /* todo: maybe use idle here, but might be less exact */
409 if (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()) 435 if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
410 return; 436 return;
411 437
412 etp_start_thread (); 438 etp_start_thread ();
413} 439}
414 440
… …
467 493
468 X_LOCK (reqlock); 494 X_LOCK (reqlock);
469 --nreqs; 495 --nreqs;
470 X_UNLOCK (reqlock); 496 X_UNLOCK (reqlock);
471 497
472 if (req->type == EIO_GROUP && req->size) 498 if (expect_false (req->type == EIO_GROUP && req->size))
473 { 499 {
474 req->int1 = 1; /* mark request as delayed */ 500 req->int1 = 1; /* mark request as delayed */
475 continue; 501 continue;
476 } 502 }
477 else 503 else
478 { 504 {
479 int res = ETP_FINISH (req); 505 int res = ETP_FINISH (req);
480 if (res) 506 if (expect_false (res))
481 return res; 507 return res;
482 } 508 }
483 509
484 if (maxreqs && !--maxreqs) 510 if (expect_false (maxreqs && !--maxreqs))
485 break; 511 break;
486 512
487 if (maxtime) 513 if (maxtime)
488 { 514 {
489 gettimeofday (&tv_now, 0); 515 gettimeofday (&tv_now, 0);
… …
508 534
509static void etp_submit (ETP_REQ *req) 535static void etp_submit (ETP_REQ *req)
510{ 536{
511 req->pri -= ETP_PRI_MIN; 537 req->pri -= ETP_PRI_MIN;
512 538
513 if (req->pri < ETP_PRI_MIN - ETP_PRI_MIN) req->pri = ETP_PRI_MIN - ETP_PRI_MIN; 539 if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
514 if (req->pri > ETP_PRI_MAX - ETP_PRI_MIN) req->pri = ETP_PRI_MAX - ETP_PRI_MIN; 540 if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
515 541
542 if (expect_false (req->type == EIO_GROUP))
543 {
544 /* I hope this is worth it :/ */
516 X_LOCK (reqlock); 545 X_LOCK (reqlock);
517 ++nreqs; 546 ++nreqs;
547 X_UNLOCK (reqlock);
548
549 X_LOCK (reslock);
550
551 ++npending;
552
553 if (!reqq_push (&res_queue, req) && want_poll_cb)
554 want_poll_cb ();
555
556 X_UNLOCK (reslock);
557 }
558 else
559 {
560 X_LOCK (reqlock);
561 ++nreqs;
518 ++nready; 562 ++nready;
519 reqq_push (&req_queue, req); 563 reqq_push (&req_queue, req);
520 X_COND_SIGNAL (reqwait); 564 X_COND_SIGNAL (reqwait);
521 X_UNLOCK (reqlock); 565 X_UNLOCK (reqlock);
522 566
523 etp_maybe_start_thread (); 567 etp_maybe_start_thread ();
568 }
524} 569}
525 570
526static void etp_set_max_poll_time (double nseconds) 571static void etp_set_max_poll_time (double nseconds)
527{ 572{
528 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 573 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
… …
563 608
564static void grp_try_feed (eio_req *grp) 609static void grp_try_feed (eio_req *grp)
565{ 610{
566 while (grp->size < grp->int2 && !EIO_CANCELLED (grp)) 611 while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
567 { 612 {
568 int old_len = grp->size; 613 grp->flags &= ~EIO_FLAG_GROUPADD;
569 614
570 EIO_FEED (grp); 615 EIO_FEED (grp);
571 616
572 /* stop if no progress has been made */ 617 /* stop if no progress has been made */
573 if (old_len == grp->size) 618 if (!(grp->flags & EIO_FLAG_GROUPADD))
574 { 619 {
575 grp->feed = 0; 620 grp->feed = 0;
576 break; 621 break;
577 } 622 }
578 } 623 }
… …
695 740
696/*****************************************************************************/ 741/*****************************************************************************/
697/* work around various missing functions */ 742/* work around various missing functions */
698 743
699#if !HAVE_PREADWRITE 744#if !HAVE_PREADWRITE
745# undef pread
746# undef pwrite
700# define pread eio__pread 747# define pread eio__pread
701# define pwrite eio__pwrite 748# define pwrite eio__pwrite
702
703/*
704 * make our pread/pwrite safe against themselves, but not against
705 * normal read/write by using a mutex. slows down execution a lot,
706 * but that's your problem, not mine.
707 */
708static mutex_t preadwritelock = X_MUTEX_INIT;
709 749
710static ssize_t 750static ssize_t
711eio__pread (int fd, void *buf, size_t count, off_t offset) 751eio__pread (int fd, void *buf, size_t count, off_t offset)
712{ 752{
713 ssize_t res; 753 ssize_t res;
… …
731 771
732 X_LOCK (preadwritelock); 772 X_LOCK (preadwritelock);
733 ooffset = lseek (fd, 0, SEEK_CUR); 773 ooffset = lseek (fd, 0, SEEK_CUR);
734 lseek (fd, offset, SEEK_SET); 774 lseek (fd, offset, SEEK_SET);
735 res = write (fd, buf, count); 775 res = write (fd, buf, count);
736 lseek (fd, offset, SEEK_SET); 776 lseek (fd, ooffset, SEEK_SET);
737 X_UNLOCK (preadwritelock); 777 X_UNLOCK (preadwritelock);
738 778
739 return res; 779 return res;
740} 780}
741#endif 781#endif
742 782
743#ifndef HAVE_FUTIMES 783#ifndef HAVE_FUTIMES
744 784
785# undef utimes
786# undef futimes
745# define utimes(path,times) eio__utimes (path, times) 787# define utimes(path,times) eio__utimes (path, times)
746# define futimes(fd,times) eio__futimes (fd, times) 788# define futimes(fd,times) eio__futimes (fd, times)
747 789
748static int 790static int
749eio__utimes (const char *filename, const struct timeval times[2]) 791eio__utimes (const char *filename, const struct timeval times[2])
… …
768} 810}
769 811
770#endif 812#endif
771 813
772#if !HAVE_FDATASYNC 814#if !HAVE_FDATASYNC
815# undef fdatasync
773# define fdatasync fsync 816# define fdatasync(fd) fsync (fd)
774#endif 817#endif
818
819/* sync_file_range always needs emulation */
820int
821eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
822{
823#if HAVE_SYNC_FILE_RANGE
824 int res;
825
826 if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
827 || EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
828 || EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
829 {
830 flags = 0
831 | (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
832 | (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
833 | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
834 }
835
836 res = sync_file_range (fd, offset, nbytes, flags);
837
838 if (!res || errno != ENOSYS)
839 return res;
840#endif
841
842 /* even though we could play tricks with the flags, it's better to always
843 * call fdatasync, as thta matches the expectation of it's users best */
844 return fdatasync (fd);
845}
775 846
776#if !HAVE_READAHEAD 847#if !HAVE_READAHEAD
848# undef readahead
777# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) 849# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
778 850
779static ssize_t 851static ssize_t
780eio__readahead (int fd, off_t offset, size_t count, worker *self) 852eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
781{ 853{
782 size_t todo = count; 854 size_t todo = count;
783 dBUF; 855 dBUF;
784 856
785 while (todo > 0) 857 while (todo > 0)
… …
901 int memlen = 4096; 973 int memlen = 4096;
902 int memofs = 0; 974 int memofs = 0;
903 int res = 0; 975 int res = 0;
904 976
905 X_LOCK (wrklock); 977 X_LOCK (wrklock);
978 /* the corresponding closedir is in ETP_WORKER_CLEAR */
906 self->dirp = dirp = opendir (req->ptr1); 979 self->dirp = dirp = opendir (req->ptr1);
907 req->flags |= EIO_FLAG_PTR2_FREE; 980 req->flags |= EIO_FLAG_PTR2_FREE;
908 req->ptr2 = names = malloc (memlen); 981 req->ptr2 = names = malloc (memlen);
909 X_UNLOCK (wrklock); 982 X_UNLOCK (wrklock);
910 983
… …
943 1016
944 if (errno) 1017 if (errno)
945 res = -1; 1018 res = -1;
946 1019
947 req->result = res; 1020 req->result = res;
1021}
1022
1023#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1024# undef msync
1025# define msync(a,b,c) ((errno = ENOSYS), -1)
1026#endif
1027
1028int
1029eio__mtouch (void *mem, size_t len, int flags)
1030{
1031 intptr_t addr = (intptr_t)mem;
1032 intptr_t end = addr + len;
1033#ifdef PAGESIZE
1034 const intptr_t page = PAGESIZE;
1035#else
1036 static intptr_t page;
1037
1038 if (!page)
1039 page = sysconf (_SC_PAGESIZE);
1040#endif
1041
1042 addr &= ~(page - 1); /* assume page size is always a power of two */
1043
1044 if (addr < end)
1045 if (flags) /* modify */
1046 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1047 else
1048 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1049
1050 return 0;
948} 1051}
949 1052
950/*****************************************************************************/ 1053/*****************************************************************************/
951 1054
952#define ALLOC(len) \ 1055#define ALLOC(len) \
… …
1039 1142
1040/*****************************************************************************/ 1143/*****************************************************************************/
1041 1144
1042int eio_init (void (*want_poll)(void), void (*done_poll)(void)) 1145int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1043{ 1146{
1044 etp_init (want_poll, done_poll); 1147 return etp_init (want_poll, done_poll);
1045} 1148}
1046 1149
1047static void eio_api_destroy (eio_req *req) 1150static void eio_api_destroy (eio_req *req)
1048{ 1151{
1049 free (req); 1152 free (req);
… …
1111 case EIO_RMDIR: req->result = rmdir (req->ptr1); break; 1214 case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
1112 case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break; 1215 case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
1113 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1216 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1114 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1217 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1115 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1218 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1116 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break; 1219 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1117 1220
1118 case EIO_READLINK: ALLOC (NAME_MAX); 1221 case EIO_READLINK: ALLOC (NAME_MAX);
1119 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1222 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1120 1223
1121 case EIO_SYNC: req->result = 0; sync (); break; 1224 case EIO_SYNC: req->result = 0; sync (); break;
1122 case EIO_FSYNC: req->result = fsync (req->int1); break; 1225 case EIO_FSYNC: req->result = fsync (req->int1); break;
1123 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1226 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1227 case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break;
1228 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1229 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1124 1230
1125 case EIO_READDIR: eio__scandir (req, self); break; 1231 case EIO_READDIR: eio__scandir (req, self); break;
1126 1232
1127 case EIO_BUSY: 1233 case EIO_BUSY:
1128#ifdef _WIN32 1234#ifdef _WIN32
… …
1160 1266
1161 req->result = req->type == EIO_FUTIME 1267 req->result = req->type == EIO_FUTIME
1162 ? futimes (req->int1, times) 1268 ? futimes (req->int1, times)
1163 : utimes (req->ptr1, times); 1269 : utimes (req->ptr1, times);
1164 } 1270 }
1271 break;
1165 1272
1166 case EIO_GROUP: 1273 case EIO_GROUP:
1274 abort (); /* handled in eio_request */
1275
1167 case EIO_NOP: 1276 case EIO_NOP:
1168 req->result = 0; 1277 req->result = 0;
1169 break; 1278 break;
1170 1279
1171 case EIO_CUSTOM: 1280 case EIO_CUSTOM:
1172 req->feed (req); 1281 ((void (*)(eio_req *))req->feed) (req);
1173 break; 1282 break;
1174 1283
1175 default: 1284 default:
1176 req->result = -1; 1285 req->result = -1;
1177 break; 1286 break;
… …
1200eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data) 1309eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1201{ 1310{
1202 REQ (EIO_FSYNC); req->int1 = fd; SEND; 1311 REQ (EIO_FSYNC); req->int1 = fd; SEND;
1203} 1312}
1204 1313
1314eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1315{
1316 REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1317}
1318
1319eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1320{
1321 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1322}
1323
1324eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1325{
1326 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1327}
1328
1205eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data) 1329eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1206{ 1330{
1207 REQ (EIO_FDATASYNC); req->int1 = fd; SEND; 1331 REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1208} 1332}
1209 1333
… …
1328 return eio__1path (EIO_READDIR, path, pri, cb, data); 1452 return eio__1path (EIO_READDIR, path, pri, cb, data);
1329} 1453}
1330 1454
1331eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) 1455eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1332{ 1456{
1333 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND; 1457 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
1334} 1458}
1335 1459
1336static eio_req * 1460static eio_req *
1337eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data) 1461eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1338{ 1462{
… …
1364 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data); 1488 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1365} 1489}
1366 1490
1367eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data) 1491eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
1368{ 1492{
1369 REQ (EIO_CUSTOM); req->feed = execute; SEND; 1493 REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
1370} 1494}
1371 1495
1372#endif 1496#endif
1373 1497
1374eio_req *eio_grp (eio_cb cb, void *data) 1498eio_req *eio_grp (eio_cb cb, void *data)
… …
1402 1526
1403void eio_grp_add (eio_req *grp, eio_req *req) 1527void eio_grp_add (eio_req *grp, eio_req *req)
1404{ 1528{
1405 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2)); 1529 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1406 1530
1531 grp->flags |= EIO_FLAG_GROUPADD;
1532
1407 ++grp->size; 1533 ++grp->size;
1408 req->grp = grp; 1534 req->grp = grp;
1409 1535
1410 req->grp_prev = 0; 1536 req->grp_prev = 0;
1411 req->grp_next = grp->grp_first; 1537 req->grp_next = grp->grp_first;

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines