From d71b808a96f0fd1b6cd82c41698b518bf345692b Mon Sep 17 00:00:00 2001 From: Ulrich Drepper Date: Mon, 8 Dec 1997 03:06:47 +0000 Subject: Update. 1997-12-08 03:24 Ulrich Drepper * Makeconfig: Define $(rt) to rt if linuxthreads is available. * Makefile (subdirs): Replace rt by $(rt). * shlib-versions: Use version .2 for NSS modules. * grp/fgetgrent_r.c (LINE_PARSER): Call parse_line with extra argument. * pwd/fgetpwent_r.c: Likewise. * spwd/fgetspent_r.c: Likewise. * spwd/sgetspent_r.c: Likewise. * hesiod/nss_hesiod/hesiod-grp.c (lookup): Add new argument. Store error number in *ERRNOP. Little optimization. * hesiod/nss_hesiod/hesiod-pwd.c: Likewise. * hesiod/nss_hesiod/hesiod-service.c: Likewise. * inet/getnetgrent_r.c (__internal_setnetgrent_reuse): Likewise. (__internal_getnetgrent_r): New argument. Call __internal_setnetgrent_reuse with new argument. (__getnetgrent_r): Call __internal_getnetgrent_r with new argument. * inet/netgroup.h: Add argument to prototypes. * nis/nisplus-parser.h: Likewise. * nis/nss_compat/compat-grp.c: Change everything to store error code through provided pointer and not use errno. * nis/nss_compat/compat-pwd.c: Likewise. * nis/nss_compat/compat-spwd.c: Likewise. * nis/nss_nis/nis-alias.c: Likewise. * nis/nss_nis/nis-ethers.c: Likewise. * nis/nss_nis/nis-grp.c: Likewise. * nis/nss_nis/nis-hosts.c: Likewise. * nis/nss_nis/nis-netgrp.c: Likewise. * nis/nss_nis/nis-network.c: Likewise. * nis/nss_nis/nis-proto.c: Likewise. * nis/nss_nis/nis-publickey.c: Likewise. * nis/nss_nis/nis-pwd.c: Likewise. * nis/nss_nis/nis-rpc.c: Likewise. * nis/nss_nis/nis-service.c: Likewise. * nis/nss_nis/nis-spwd.c: Likewise. * nis/nss_nisplus/nisplus-alias.c: Likewise. * nis/nss_nisplus/nisplus-ethers.c: Likewise. * nis/nss_nisplus/nisplus-grp.c: Likewise. * nis/nss_nisplus/nisplus-hosts.c: Likewise. * nis/nss_nisplus/nisplus-netgrp.c: Likewise. * nis/nss_nisplus/nisplus-network.c: Likewise. * nis/nss_nisplus/nisplus-parser.c: Likewise. * nis/nss_nisplus/nisplus-proto.c: Likewise. * nis/nss_nisplus/nisplus-publickey.c: Likewise. * nis/nss_nisplus/nisplus-pwd.c: Likewise. * nis/nss_nisplus/nisplus-rpc.c: Likewise. * nis/nss_nisplus/nisplus-service.c: Likewise. * nis/nss_nisplus/nisplus-spwd.c: Likewise. * nss/nss_db/db-XXX.c: Likewise. * nss/nss_db/db-alias.c: Likewise. * nss/nss_db/db-netgrp.c: Likewise. * nss/nss_files/files-XXX.c: Likewise * nss/nss_files/files-alias.c: Likewise * nss/nss_files/files-netgrp.c: Likewise * nss/nss_files/files-parse.c: Likewise * nss/getXXbyYY_r.c: Call with __errno_location() as value for new parameter of get functions. * nss/getXXent_r.c: Likewise. * localedata/tst-fmon.c: Print debug info if test fails. * rt/Makefile (routines): Add aio_notify. * rt/aio.h: Remove non-public definitions. Add aioinit stuff. * rt/aio_cancel.c: Rewrite to support lio_listio and aio_suspend. * rt/aio_fsync.c: Likewise. * rt/aio_misc.c: Likewise. * rt/aio_misc.h: Likewise. * rt/aio_read.c: Likewise. * rt/aio_read64.c: Likewise. * rt/aio_suspend.c: Likewise. * rt/aio_write.c: Likewise. * rt/aio_write64.c: Likewise. * rt/lio_listio.c: Likewise. * rt/lio_listio64.c: Likewise. * rt/aio_notify.c: New file. * sysdeps/generic/bits/sigset.h: Pretty print. * sysdeps/posix/pwrite.c: Define weak alias only if function is really defined as __pwrite. * sysdeps/unix/sysv/linux/pread.c: Rewrite to use syscall directly instead of pread64. * sysdeps/unix/sysv/linux/pwrite.c: Likewise. * sysdeps/unix/sysv/linux/Makefile [$(subdir)=signal] (sysdep_routines): Add rt_sigpending and rt_sigreturn. * sysdeps/unix/sysv/linux/sigaction.c: Use correct value for sigset_t size in syscall. * sysdeps/unix/sysv/linux/sigpending.c: Likewise. * sysdeps/unix/sysv/linux/sigprocmask.c: Likewise. * sysdeps/unix/sysv/linux/sigreturn.c: Likewise. * sysdeps/unix/sysv/linux/sigsuspend.c: Likewise. * sysdeps/unix/sysv/linux/sigtimedwait.c: Likewise. * sysdeps/unix/sysv/linux/sigwaitinfo.c: Likewise. * sysdeps/unix/sysv/linux/i386/sigaction.c: Likewise. Take care for ACT being NULL. * sysdeps/unix/sysv/linux/syscalls.list: Don't mention user for RT signals. * sysdeps/unix/sysv/linux/i386/s_pread64.s: Return ENOSYS if __NR_pread is not defined. * sysdeps/unix/sysv/linux/i386/s_pwrite64.s: Return ENOSYS if __NR_pwrite is not defined. 1997-12-07 11:15 Thorsten Kukuk * nis/nss_compat/compat-grp.c: If buffer is to small, set the file handle for /etc/group back in some more cases. * nis/nss_compat/compat-pwd.c: If buffer is to small, set the file handle for /etc/passwd back in some more cases. * nis/nss_compat/compat-spwd.c: If buffer is to small, set the file handle for /etc/shadow back in some more cases. 1997-12-06 17:00 H.J. Lu * sysdeps/posix/sigblock.c (__sigblock): Don't check int mask beyound its size. * sysdeps/posix/sigsetmask.c (__sigsetmask): Ditto. * sysdeps/posix/sigvec.c (convert_mask, __sigvec): Ditto. 1997-12-06 04:38 Ulrich Drepper * sysdeps/wordsize-32/inttypes.h: Rewrite {str,wcs}to{i,u}max handling. Don't use macros, we need functions. Use inline functions for optimization. * sysdeps/wordsize-64/inttypes.h: Likewise. * sysdeps/wordsize-32/Makefile: Add strtoimax, stroumax, wcstoimax, and wcstoumax. * sysdeps/wordsize-64/Makefile: Likewise. * sysdeps/wordsize-32/Dist: Likewise. * sysdeps/wordsize-64/Dist: Likewise. * sysdeps/wordsize-32/strtoimax.c: New file. * sysdeps/wordsize-32/strtoumax.c: New file. * sysdeps/wordsize-32/wcstoimax.c: New file. * sysdeps/wordsize-32/wcstoumax.c: New file. * sysdeps/wordsize-64/strtoimax.c: New file. * sysdeps/wordsize-64/strtoumax.c: New file. * sysdeps/wordsize-64/wcstoimax.c: New file. * sysdeps/wordsize-64/wcstoumax.c: New file. 1997-12-04 10:40 Philip Blundell * sysdeps/generic/wait3.c (__wait3): Use ANSI-style definition so that transparent union works. * sysdeps/generic/wait4.c (__wait4): Likewise. * sysdeps/generic/wait.c (__wait): Likewise. * sysdeps/generic/getpeername.c (getpeername): Use socklen_t not size_t. * sysdeps/generic/syscall.c (syscall): Return value and arg are long, to match prototype. * malloc/malloc.c (HAVE_MREMAP): Set to 0 for ARM, as we don't have mremap() yet. * sysdeps/generic/usleep.c (usleep): Returns void, to match prototype. * sysdeps/unix/sysv/linux/arm/syscalls.list: New file. * sysdeps/generic/bits/time.h: Use __time_t not time_t. * sysdeps/generic/bits/resource.h: Include for __rlim_t. * sysdeps/standalone/arm/bits/errno.h (ENOTTY): Added. * sysdeps/generic/bits/dirent.h (struct dirent64): Added. * io/sys/stat.h (S_ISLNK): Always false if we don't have __S_IFLNK. * sysdeps/standalone/close.c (_STDIO_H): Define before including . * sysdeps/standalone/filedesc.h (_STDIO_H): Likewise. 1997-12-06 01:09 Ulrich Drepper * posix/sys/wait.h: Don't use transparent unions in C++. 1997-12-05 Andreas Jaeger * time/time.h: Add strfxtime. * libc.map: Add tcgetsid, strcasestr, wmemrtombs, wmemrtowcs, sysv_signal, strfxtime, pread64, pwrite64. 1997-11-30 21:57 Richard Henderson Alpha changes for EGCS: * config.h.in (ASM_ALPHA_NG_SYMBOL_PREFIX): New entry. * configure.in (libc_cv_gcc_alpha_ng_prefix): New check. * sysdeps/alpha/dl-machine.h (_dl_runtime_resolve): Care for prefix. (_start): Likewise. * sysdeps/unix/sysv/linux/alpha/init-first.h: Likewise. --- rt/Makefile | 3 +- rt/aio.h | 39 ++-- rt/aio_cancel.c | 195 ++++++++---------- rt/aio_fsync.c | 2 +- rt/aio_misc.c | 588 ++++++++++++++++++++++++++++++++++++++---------------- rt/aio_misc.h | 81 +++++++- rt/aio_notify.c | 97 +++++++++ rt/aio_read.c | 2 +- rt/aio_read64.c | 2 +- rt/aio_suspend.c | 79 +++++++- rt/aio_write.c | 2 +- rt/aio_write64.c | 2 +- rt/lio_listio.c | 112 +++++++++-- rt/lio_listio64.c | 113 +++++++++-- 14 files changed, 978 insertions(+), 339 deletions(-) create mode 100644 rt/aio_notify.c (limited to 'rt') diff --git a/rt/Makefile b/rt/Makefile index 4eb9dfcb7a..e0f03d78a0 100644 --- a/rt/Makefile +++ b/rt/Makefile @@ -28,7 +28,8 @@ extra-libs-others := $(extra-libs) librt-routines := aio_cancel aio_error aio_fsync aio_misc aio_read \ aio_read64 aio_return aio_suspend aio_write \ - aio_write64 lio_listio lio_listio64 aio_sigqueue + aio_write64 lio_listio lio_listio64 aio_sigqueue \ + aio_notify librt-map := librt.map diff --git a/rt/aio.h b/rt/aio.h index 561776c2d3..a49b8c7a67 100644 --- a/rt/aio.h +++ b/rt/aio.h @@ -43,8 +43,6 @@ struct aiocb struct sigevent aio_sigevent; /* Signal number and value. */ /* Internal members. */ - struct aiocb *__last_fd; - struct aiocb *__next_fd; struct aiocb *__next_prio; int __abs_prio; int __policy; @@ -53,10 +51,11 @@ struct aiocb #ifndef __USE_FILE_OFFSET64 __off_t aio_offset; /* File offset. */ + char __pad[sizeof (__off64_t) - sizeof (__off_t)]; #else __off64_t aio_offset; /* File offset. */ #endif - int __unused[32]; + char __unused[32]; }; /* The same for the 64bit offsets. */ @@ -71,20 +70,33 @@ struct aiocb64 struct sigevent aio_sigevent; /* Signal number and value. */ /* Internal members. */ - struct aiocb64 *__last_fd; - struct aiocb64 *__next_fd; - struct aiocb64 *__next_prio; int __abs_prio; int __policy; int __error_code; __ssize_t __return_value; __off64_t aio_offset; /* File offset. */ - int __unused[32]; + char __unused[32]; }; #endif +#ifdef __USE_GNU +/* To customize the implementation one can use the following struct. + This implementation follows the one in Irix. */ +struct aioinit + { + int aio_threads; /* Maximal number of threads. */ + int aio_num; /* Number of expected simultanious requests. */ + int aio_locks; /* Not used. */ + int aio_usedba; /* Not used. */ + int aio_debug; /* Not used. */ + int aio_numusers; /* Not used. */ + int aio_reserved[2]; + }; +#endif + + /* Return values of cancelation function. */ enum { @@ -104,12 +116,8 @@ enum #define LIO_READ LIO_READ LIO_WRITE, #define LIO_WRITE LIO_WRITE - LIO_NOP, + LIO_NOP #define LIO_NOP LIO_NOP - __LIO_DSYNC, - __LIO_SYNC, - __LIO_READ64 = LIO_READ | 128, - __LIO_WRITE64 = LIO_WRITE | 128 }; @@ -123,6 +131,13 @@ enum }; +/* Allow user to specify optimization. */ +#ifdef __USE_GNU +extern void __aio_init __P ((__const struct aioinit *__init)); +extern void aio_init __P ((__const struct aioinit *__init)); +#endif + + /* Enqueue read request for given number of bytes and the given priority. */ #ifndef __USE_FILE_OFFSET64 extern int aio_read __P ((struct aiocb *__aiocbp)); diff --git a/rt/aio_cancel.c b/rt/aio_cancel.c index f2d9389650..e802a8533f 100644 --- a/rt/aio_cancel.c +++ b/rt/aio_cancel.c @@ -30,7 +30,6 @@ #undef aio_cancel64 #include -#include #include "aio_misc.h" @@ -43,128 +42,106 @@ aio_cancel (fildes, aiocbp) int fildes; struct aiocb *aiocbp; { - struct aiocb *firstp; + struct requestlist *req = NULL; int result = AIO_ALLDONE; - /* Request the semaphore. */ - sem_wait (&__aio_requests_sema); + /* Request the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); - /* Search for the list of requests associated with the given file - descriptor. */ - for (firstp = (struct aiocb *) __aio_requests; firstp != NULL; - firstp = firstp->__next_fd) - if (firstp->aio_fildes == fildes) - break; - - /* If the file descriptor is not found all work seems to done - already. Otherwise try to cancel the request(s). */ - if (firstp != NULL) + /* We are asked to cancel a specific AIO request. */ + if (aiocbp != NULL) { - if (aiocbp != NULL) + /* If the AIO request is not for this descriptor it has no value + to look for the request block. */ + if (aiocbp->aio_fildes == fildes) { - /* Locate the entry corresponding to the AIOCBP parameter. */ - if (aiocbp == firstp) - /* The requests is currently handled, therefore don't - cancel it and signal this to the user. */ - result = AIO_NOTCANCELED; - else + struct requestlist *last = NULL; + + req = __aio_find_req_fd (fildes); + + while (req->aiocbp != (aiocb_union *) aiocbp) { - while (firstp->__next_prio != NULL - && aiocbp != firstp->__next_prio) - firstp = firstp->__next_prio; - - if (firstp->__next_prio != NULL) - { - /* The request the user wants to cancel is in the - queue. Simply remove it. */ - firstp->__next_prio = aiocbp->__next_prio; - - /* Mark as canceled. */ - aiocbp->__error_code = ECANCELED; - aiocbp->__return_value = -1; - - /* Send the signal to notify about canceled - processing of the request. */ - if (aiocbp->aio_sigevent.sigev_notify == SIGEV_THREAD) - { - /* We have to start a thread. */ - pthread_t tid; - pthread_attr_t attr, *pattr; - - pattr = (pthread_attr_t *) - aiocbp->aio_sigevent.sigev_notify_attributes; - if (pattr == NULL) - { - pthread_attr_init (&attr); - pthread_attr_setdetachstate (&attr, - PTHREAD_CREATE_DETACHED); - pattr = &attr; - } - - pthread_create (&tid, pattr, - (void *(*) (void *)) - aiocbp->aio_sigevent.sigev_notify_function, - aiocbp->aio_sigevent.sigev_value.sival_ptr); - } - else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_SIGNAL) - /* We have to send a signal. */ - __aio_sigqueue (aiocbp->aio_sigevent.sigev_signo, - aiocbp->aio_sigevent.sigev_value); - - result = AIO_CANCELED; - } + last = req; + req = req->next_prio; } - } - else - { - /* First dequeue all waiting requests. */ - aiocbp = firstp; - while ((firstp = firstp->__next_prio) != NULL) + /* Don't remove the entry if a thread is already working on it. */ + if (req->running == allocated) + result = AIO_NOTCANCELED; + else { - firstp->__error_code = ECANCELED; - firstp->__return_value = -1; - - - /* Send the signal to notify about canceled processing - of the request. */ - if (firstp->aio_sigevent.sigev_notify == SIGEV_THREAD) - { - /* We have to start a thread. */ - pthread_t tid; - pthread_attr_t attr, *pattr; - - pattr = (pthread_attr_t *) - aiocbp->aio_sigevent.sigev_notify_attributes; - if (pattr == NULL) - { - pthread_attr_init (&attr); - pthread_attr_setdetachstate (&attr, - PTHREAD_CREATE_DETACHED); - pattr = &attr; - } - - pthread_create (&tid, pattr, - (void *(*) (void *)) - firstp->aio_sigevent.sigev_notify_function, - firstp->aio_sigevent.sigev_value.sival_ptr); - } - else if (firstp->aio_sigevent.sigev_notify == SIGEV_SIGNAL) - /* We have to send a signal. */ - __aio_sigqueue (firstp->aio_sigevent.sigev_signo, - firstp->aio_sigevent.sigev_value); + /* We can remove the entry. */ + if (last != NULL) + last->next_prio = req->next_prio; + else + if (req->next_prio == NULL) + { + if (req->last_fd != NULL) + req->last_fd->next_fd = req->next_fd; + if (req->next_fd != NULL) + req->next_fd->last_fd = req->last_fd; + } + else + { + if (req->last_fd != NULL) + req->last_fd->next_fd = req->next_prio; + if (req->next_fd != NULL) + req->next_fd->last_fd = req->next_prio; + req->next_prio->last_fd = req->last_fd; + req->next_prio->next_fd = req->next_fd; + + /* Mark this entry as runnable. */ + req->next_prio->running = yes; + } + + result = AIO_CANCELED; } - /* We have to signal that not all requests could be canceled - since the first requests is currently processed. */ - result = AIO_NOTCANCELED; - - aiocbp->__next_prio = NULL; + req->next_prio = NULL; } } + else + { + /* Find the beginning of the list of all requests for this + desriptor. */ + req = __aio_find_req_fd (fildes); + + /* If any request is worked on by a thread it must be the first. + So either we can delete all requests or all but the first. */ + if (req != NULL) + if (req->running == allocated) + { + struct requestlist *old = req; + req = req->next_prio; + old->next_prio = NULL; + + result = AIO_NOTCANCELED; + } + else + { + /* Remove entry from the file descriptor list. */ + if (req->last_fd != NULL) + req->last_fd->next_fd = req->next_fd; + if (req->next_fd != NULL) + req->next_fd->last_fd = req->last_fd; + + result = AIO_CANCELED; + } + } + + /* Mark requests as canceled and send signal. */ + while (req != NULL) + { + struct requestlist *old = req; + req->aiocbp->aiocb.__error_code = ECANCELED; + req->aiocbp->aiocb.__return_value = -1; + __aio_notify (req); + req = req->next_prio; + __aio_free_request (old); + } - /* Release the semaphore. */ - sem_post (&__aio_requests_sema); + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); return result; } diff --git a/rt/aio_fsync.c b/rt/aio_fsync.c index 6daaca1e97..c813654b61 100644 --- a/rt/aio_fsync.c +++ b/rt/aio_fsync.c @@ -36,7 +36,7 @@ int aio_fsync (int op, struct aiocb *aiocbp) { return __aio_enqueue_request ((aiocb_union *) aiocbp, - op == O_SYNC ? __LIO_SYNC : __LIO_DSYNC, 1); + op == O_SYNC ? LIO_SYNC : LIO_DSYNC) != NULL; } weak_alias (aio_fsync, aio_fsync64) diff --git a/rt/aio_misc.c b/rt/aio_misc.c index e4bb12c500..6ea30c2158 100644 --- a/rt/aio_misc.c +++ b/rt/aio_misc.c @@ -21,7 +21,6 @@ #include #include #include -#include #include #include #include @@ -29,40 +28,199 @@ #include "aio_misc.h" -/* We need a list of pending operations. This is sorted according to - the priority given in the aio_reqprio member. */ -aiocb_union *__aio_requests; +/* Pool of request list entries. */ +static struct requestlist **pool; -/* Since the list is global we need a semaphore protecting it. */ -sem_t __aio_requests_sema; +/* Number of total and allocated pool entries. */ +static size_t pool_tab_size; +static size_t pool_size; +/* We implement a two dimensional array but allocate each row separately. + The macro below determines how many entries should be used per row. + It should better be a power of two. */ +#define ENTRIES_PER_ROW 16 -/* The initialization function. It gets automatically called if any - aio_* function is used in the program. */ -static void -__attribute__ ((unused)) -aio_initialize (void) +/* The row table is incremented in units of this. */ +#define ROW_STEP 8 + +/* List of available entries. */ +static struct requestlist *freelist; + +/* List of request waiting to be processed. */ +static struct requestlist *runlist; + +/* Structure list of all currently processed requests. */ +static struct requestlist *requests; + +/* Number of threads currently running. */ +static int nthreads; + + +/* These are the values used to optimize the use of AIO. The user can + overwrite them by using the `aio_init' function. */ +static struct aioinit optim = +{ + 20, /* int aio_threads; Maximal number of threads. */ + 256, /* int aio_num; Number of expected simultanious requests. */ + 0, + 0, + 0, + 0, + { 0, } +}; + + +/* Since the list is global we need a mutex protecting it. */ +pthread_mutex_t __aio_requests_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP; + + +/* Functions to handle request list pool. */ +static struct requestlist * +get_elem (void) { - /* Initialize the semaphore. We allow exactly one user at a time. */ - sem_init (&__aio_requests_sema, 0, 1); + struct requestlist *result; + + if (freelist == NULL) + { + struct requestlist *new_row; + size_t new_size; + + /* Compute new size. */ + new_size = pool_size ? pool_size + ENTRIES_PER_ROW : optim.aio_num; + + if ((new_size / ENTRIES_PER_ROW) >= pool_tab_size) + { + size_t new_tab_size = new_size / ENTRIES_PER_ROW; + struct requestlist **new_tab; + + new_tab = (struct requestlist **) + realloc (pool, (new_tab_size * sizeof (struct requestlist *))); + + if (new_tab == NULL) + return NULL; + + pool_tab_size = new_tab_size; + pool = new_tab; + } + + if (pool_size == 0) + { + size_t cnt; + + new_row = (struct requestlist *) + calloc (new_size, sizeof (struct requestlist)); + + if (new_row == NULL) + return NULL; + + for (cnt = 0; cnt < new_size / ENTRIES_PER_ROW; ++cnt) + pool[cnt] = &new_row[cnt * ENTRIES_PER_ROW]; + } + else + { + /* Allocat one new row. */ + new_row = (struct requestlist *) + calloc (ENTRIES_PER_ROW, sizeof (struct requestlist)); + if (new_row == NULL) + return NULL; + + pool[new_size / ENTRIES_PER_ROW] = new_row; + } + + /* Put all the new entries in the freelist. */ + do + { + new_row->next_prio = freelist; + freelist = new_row++; + } + while (++pool_size < new_size); + } + + result = freelist; + freelist = freelist->next_prio; + + return result; } -text_set_element (__libc_subinit, aio_initialize); + +void +__aio_free_req (struct requestlist *elem) +{ + elem->running = no; + elem->next_prio = freelist; + freelist = elem; +} + + +struct requestlist * +__aio_find_req (aiocb_union *elem) +{ + struct requestlist *runp = requests; + int fildes = elem->aiocb.aio_fildes; + + while (runp != NULL && runp->aiocbp->aiocb.aio_fildes < fildes) + runp = runp->next_fd; + + if (runp != NULL) + if (runp->aiocbp->aiocb.aio_fildes != fildes) + runp = NULL; + else + while (runp != NULL && runp->aiocbp != elem) + runp = runp->next_prio; + + return runp; +} + + +struct requestlist * +__aio_find_req_fd (int fildes) +{ + struct requestlist *runp = requests; + + while (runp != NULL && runp->aiocbp->aiocb.aio_fildes < fildes) + runp = runp->next_fd; + + return (runp != NULL && runp->aiocbp->aiocb.aio_fildes == fildes + ? runp : NULL); +} /* The thread handler. */ static void *handle_fildes_io (void *arg); +/* User optimization. */ +void +__aio_init (const struct aioinit *init) +{ + /* Get the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); + + /* Only allow writing new values if the table is not yet allocated. */ + if (pool == NULL) + { + optim.aio_threads = init->aio_threads < 1 ? 1 : init->aio_threads; + optim.aio_num = (init->aio_num < ENTRIES_PER_ROW + ? ENTRIES_PER_ROW + : init->aio_num & ~ENTRIES_PER_ROW); + } + + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); +} +weak_alias (__aio_init, aio_init) + + /* The main function of the async I/O handling. It enqueues requests and if necessary starts and handles threads. */ -int -__aio_enqueue_request (aiocb_union *aiocbp, int operation, int require_lock) +struct requestlist * +__aio_enqueue_request (aiocb_union *aiocbp, int operation) { - int result; + int result = 0; int policy, prio; struct sched_param param; - aiocb_union *runp; + struct requestlist *last, *runp, *newp; + int running = no; if (aiocbp->aiocb.aio_reqprio < 0 || aiocbp->aiocb.aio_reqprio > AIO_PRIO_DELTA_MAX) @@ -71,94 +229,160 @@ __aio_enqueue_request (aiocb_union *aiocbp, int operation, int require_lock) __set_errno (EINVAL); aiocbp->aiocb.__error_code = EINVAL; aiocbp->aiocb.__return_value = -1; - return -1; - } - - if (pthread_getschedparam (pthread_self (), &policy, ¶m) < 0) - { - /* Something went wrong. */ - aiocbp->aiocb.__error_code = errno; - aiocbp->aiocb.__return_value = -1; - return -1; + return NULL; } /* Compute priority for this request. */ + pthread_getschedparam (pthread_self (), &policy, ¶m); prio = param.sched_priority - aiocbp->aiocb.aio_reqprio; + /* Get the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); - /* Get the semaphore. */ - if (require_lock) - sem_wait (&__aio_requests_sema); - - runp = __aio_requests; + last = NULL; + runp = requests; /* First look whether the current file descriptor is currently worked with. */ - while (runp != NULL && runp->aiocb.aio_fildes < aiocbp->aiocb.aio_fildes) - runp = (aiocb_union *) runp->aiocb.__next_fd; + while (runp != NULL + && runp->aiocbp->aiocb.aio_fildes < aiocbp->aiocb.aio_fildes) + { + last = runp; + runp = runp->next_fd; + } - if (runp != NULL) + /* Get a new element for the waiting list. */ + newp = get_elem (); + if (newp == NULL) + { + __set_errno (EAGAIN); + pthread_mutex_unlock (&__aio_requests_mutex); + return NULL; + } + newp->aiocbp = aiocbp; + newp->waiting = NULL; + + aiocbp->aiocb.__abs_prio = prio; + aiocbp->aiocb.__policy = policy; + aiocbp->aiocb.aio_lio_opcode = operation; + aiocbp->aiocb.__error_code = EINPROGRESS; + aiocbp->aiocb.__return_value = 0; + + if (runp != NULL + && runp->aiocbp->aiocb.aio_fildes == aiocbp->aiocb.aio_fildes) { /* The current file descriptor is worked on. It makes no sense - to start another thread since this new thread would have to - wait for the previous one to terminate. Simply enqueue it - after the running one according to the priority. */ - while (runp->aiocb.__next_prio != NULL - && runp->aiocb.__next_prio->__abs_prio >= prio) - runp = (aiocb_union *) runp->aiocb.__next_prio; - - aiocbp->aiocb.__next_prio = runp->aiocb.__next_prio; - aiocbp->aiocb.__abs_prio = prio; - aiocbp->aiocb.__policy = policy; - aiocbp->aiocb.aio_lio_opcode = operation; - aiocbp->aiocb.__error_code = EINPROGRESS; - aiocbp->aiocb.__return_value = 0; - runp->aiocb.__next_prio = (struct aiocb *) aiocbp; - - result = 0; + to start another thread since this new thread would fight + with the running thread for the resources. But we also cannot + say that the thread processing this desriptor shall imeediately + after finishing the current job process this request if there + are other threads in the running queue which have a higher + priority. */ + + /* Simply enqueue it after the running one according to the + priority. */ + while (runp->next_prio != NULL + && runp->next_prio->aiocbp->aiocb.__abs_prio >= prio) + runp = runp->next_prio; + + newp->next_prio = runp->next_prio; + runp->next_prio = newp; + + running = queued; } else { - /* We create a new thread for this file descriptor. The + /* Enqueue this request for a new descriptor. */ + if (last == NULL) + { + newp->last_fd = NULL; + newp->next_fd = requests; + if (requests != NULL) + requests->last_fd = newp; + requests = newp; + } + else + { + newp->next_fd = last->next_fd; + newp->last_fd = last; + last->next_fd = newp; + if (newp->next_fd != NULL) + newp->next_fd->last_fd = newp; + } + + newp->next_prio = NULL; + } + + if (running == no) + { + /* We try to create a new thread for this file descriptor. The function which gets called will handle all available requests for this descriptor and when all are processed it will - terminate. */ - pthread_t thid; - pthread_attr_t attr; - - /* First enqueue the request (the list is empty). */ - aiocbp->aiocb.__next_fd = NULL; - aiocbp->aiocb.__last_fd = NULL; - - aiocbp->aiocb.__next_prio = NULL; - aiocbp->aiocb.__abs_prio = prio; - aiocbp->aiocb.__policy = policy; - aiocbp->aiocb.aio_lio_opcode = operation; - aiocbp->aiocb.__error_code = EINPROGRESS; - aiocbp->aiocb.__return_value = 0; - - /* Make sure the thread is created detached. */ - pthread_attr_init (&attr); - pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); - - /* Now try to start a thread. */ - if (pthread_create (&thid, &attr, handle_fildes_io, aiocbp) < 0) + terminate. + + If no new thread can be created or if the specified limit of + threads for AIO is reached we queue the request. */ + + /* See if we can create a thread. */ + if (nthreads < optim.aio_threads) { - result = -1; - aiocbp->aiocb.__error_code = errno; - aiocbp->aiocb.__return_value = -1; + pthread_t thid; + pthread_attr_t attr; + + /* Make sure the thread is created detached. */ + pthread_attr_init (&attr); + pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); + + /* Now try to start a thread. */ + if (pthread_create (&thid, &attr, handle_fildes_io, newp) == 0) + { + /* We managed to enqueue the request. All errors which can + happen now can be recognized by calls to `aio_return' and + `aio_error'. */ + running = allocated; + ++nthreads; + } + else if (nthreads == 0) + /* We cannot create a thread in the moment and there is + also no thread running. This is a problem. `errno' is + set to EAGAIN if this is only a temporary problem. */ + result = -1; + } + } + + /* Enqueue the request in the run queue if it is not yet running. */ + if (running < yes && result == 0) + { + if (runlist == NULL || runlist->aiocbp->aiocb.__abs_prio < prio) + { + newp->next_run = runlist; + runlist = newp; } else - /* We managed to enqueue the request. All errors which can - happen now can be recognized by calls to `aio_return' and - `aio_error'. */ - result = 0; + { + runp = runlist; + + while (runp->next_run != NULL + && runp->next_run->aiocbp->aiocb.__abs_prio >= prio) + runp = runp->next_run; + + newp->next_run = runp->next_run; + runp->next_run = newp; + } } - /* Release the semaphore. */ - if (require_lock) - sem_post (&__aio_requests_sema); + if (result == 0) + newp->running = running; + else + { + /* Something went wrong. */ + __aio_free_req (newp); + newp = NULL; + } - return result; + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); + + return newp; } @@ -167,140 +391,156 @@ handle_fildes_io (void *arg) { pthread_t self = pthread_self (); struct sched_param param; - aiocb_union *runp = (aiocb_union *) arg; + struct requestlist *runp = (struct requestlist *) arg; + aiocb_union *aiocbp = runp->aiocbp; int policy; - int fildes = runp->aiocb.aio_fildes; /* This is always the same. */ + int fildes = runp->aiocbp->aiocb.aio_fildes; pthread_getschedparam (self, &policy, ¶m); do { /* Change the priority to the requested value (if necessary). */ - if (runp->aiocb.__abs_prio != param.sched_priority - || runp->aiocb.__policy != policy) + if (aiocbp->aiocb.__abs_prio != param.sched_priority + || aiocbp->aiocb.__policy != policy) { - param.sched_priority = runp->aiocb.__abs_prio; - policy = runp->aiocb.__policy; + param.sched_priority = aiocbp->aiocb.__abs_prio; + policy = aiocbp->aiocb.__policy; pthread_setschedparam (self, policy, ¶m); } /* Process request pointed to by RUNP. We must not be disturbed by signals. */ - if ((runp->aiocb.aio_lio_opcode & 127) == LIO_READ) + if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_READ) { - if (runp->aiocb.aio_lio_opcode & 128) - runp->aiocb.__return_value = + if (aiocbp->aiocb.aio_lio_opcode & 128) + aiocbp->aiocb.__return_value = TEMP_FAILURE_RETRY (__pread64 (fildes, - (void *) runp->aiocb64.aio_buf, - runp->aiocb64.aio_nbytes, - runp->aiocb64.aio_offset)); + (void *) aiocbp->aiocb64.aio_buf, + aiocbp->aiocb64.aio_nbytes, + aiocbp->aiocb64.aio_offset)); else - runp->aiocb.__return_value = + aiocbp->aiocb.__return_value = TEMP_FAILURE_RETRY (__pread (fildes, - (void *) runp->aiocb.aio_buf, - runp->aiocb.aio_nbytes, - runp->aiocb.aio_offset)); + (void *) aiocbp->aiocb.aio_buf, + aiocbp->aiocb.aio_nbytes, + aiocbp->aiocb.aio_offset)); } - else if ((runp->aiocb.aio_lio_opcode & 127) == LIO_WRITE) + else if ((aiocbp->aiocb.aio_lio_opcode & 127) == LIO_WRITE) { - if (runp->aiocb.aio_lio_opcode & 128) - runp->aiocb.__return_value = + if (aiocbp->aiocb.aio_lio_opcode & 128) + aiocbp->aiocb.__return_value = TEMP_FAILURE_RETRY (__pwrite64 (fildes, - (const void *) runp->aiocb64.aio_buf, - runp->aiocb64.aio_nbytes, - runp->aiocb64.aio_offset)); + (const void *) aiocbp->aiocb64.aio_buf, + aiocbp->aiocb64.aio_nbytes, + aiocbp->aiocb64.aio_offset)); else - runp->aiocb.__return_value = + aiocbp->aiocb.__return_value = TEMP_FAILURE_RETRY (__pwrite (fildes, - (const void *) runp->aiocb.aio_buf, - runp->aiocb.aio_nbytes, - runp->aiocb.aio_offset)); + (const void *) aiocbp->aiocb.aio_buf, + aiocbp->aiocb.aio_nbytes, + aiocbp->aiocb.aio_offset)); } - else if (runp->aiocb.aio_lio_opcode == __LIO_DSYNC) - runp->aiocb.__return_value = TEMP_FAILURE_RETRY (fdatasync (fildes)); - else if (runp->aiocb.aio_lio_opcode == __LIO_SYNC) - runp->aiocb.__return_value = TEMP_FAILURE_RETRY (fsync (fildes)); + else if (aiocbp->aiocb.aio_lio_opcode == LIO_DSYNC) + aiocbp->aiocb.__return_value = TEMP_FAILURE_RETRY (fdatasync (fildes)); + else if (aiocbp->aiocb.aio_lio_opcode == LIO_SYNC) + aiocbp->aiocb.__return_value = TEMP_FAILURE_RETRY (fsync (fildes)); else { /* This is an invalid opcode. */ - runp->aiocb.__return_value = -1; + aiocbp->aiocb.__return_value = -1; __set_errno (EINVAL); } - if (runp->aiocb.__return_value == -1) - runp->aiocb.__error_code = errno; + /* Get the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); + + if (aiocbp->aiocb.__return_value == -1) + aiocbp->aiocb.__error_code = errno; else - runp->aiocb.__error_code = 0; + aiocbp->aiocb.__error_code = 0; /* Send the signal to notify about finished processing of the request. */ - if (runp->aiocb.aio_sigevent.sigev_notify == SIGEV_THREAD) + __aio_notify (runp); + + /* Now dequeue the current request. */ + if (runp->next_prio == NULL) { - /* We have to start a thread. */ - pthread_t tid; - pthread_attr_t attr, *pattr; + /* No outstanding request for this descriptor. Process the + runlist if necessary. */ + if (runp->next_fd != NULL) + runp->next_fd->last_fd = runp->last_fd; + if (runp->last_fd != NULL) + runp->last_fd->next_fd = runp->next_fd; + } + else + { + runp->next_prio->last_fd = runp->last_fd; + runp->next_prio->next_fd = runp->next_fd; + runp->next_prio->running = yes; + if (runp->next_fd != NULL) + runp->next_fd->last_fd = runp->next_prio; + if (runp->last_fd != NULL) + runp->last_fd->next_fd = runp->next_prio; + } + + /* Free the old element. */ + __aio_free_req (runp); - pattr = (pthread_attr_t *) - runp->aiocb.aio_sigevent.sigev_notify_attributes; - if (pattr == NULL) + runp = freelist; + if (runp != NULL) + { + /* We must not run requests which are not marked `running'. */ + if (runp->running == yes) { - pthread_attr_init (&attr); - pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); - pattr = &attr; + freelist = runp->next_run; + runp->running = allocated; } - - if (pthread_create (&tid, - (pthread_attr_t *) - runp->aiocb.aio_sigevent.sigev_notify_attributes, - (void *(*) (void *)) - runp->aiocb.aio_sigevent.sigev_notify_function, - runp->aiocb.aio_sigevent.sigev_value.sival_ptr) - < 0) + else { - /* XXX What shall we do if already an error is set by - read/write/fsync? */ - runp->aiocb.__error_code = errno; - runp->aiocb.__return_value = -1; + struct requestlist *old; + + do + { + old = runp; + runp = runp->next_run; + } + while (runp != NULL && runp->running != yes); + + if (runp != NULL) + old->next_run = runp->next_run; } } - else if (runp->aiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) - /* We have to send a signal. */ - if (__aio_sigqueue (runp->aiocb.aio_sigevent.sigev_signo, - runp->aiocb.aio_sigevent.sigev_value) < 0) - { - /* XXX What shall we do if already an error is set by - read/write/fsync? */ - runp->aiocb.__error_code = errno; - runp->aiocb.__return_value = -1; - } - - /* Get the semaphore. */ - sem_wait (&__aio_requests_sema); - /* Now dequeue the current request. */ - if (runp->aiocb.__next_prio == NULL) - { - if (runp->aiocb.__next_fd != NULL) - runp->aiocb.__next_fd->__last_fd = runp->aiocb.__last_fd; - if (runp->aiocb.__last_fd != NULL) - runp->aiocb.__last_fd->__next_fd = runp->aiocb.__next_fd; - runp = NULL; - } - else - { - runp->aiocb.__next_prio->__last_fd = runp->aiocb.__last_fd; - runp->aiocb.__next_prio->__next_fd = runp->aiocb.__next_fd; - if (runp->aiocb.__next_fd != NULL) - runp->aiocb.__next_fd->__last_fd = runp->aiocb.__next_prio; - if (runp->aiocb.__last_fd != NULL) - runp->aiocb.__last_fd->__next_fd = runp->aiocb.__next_prio; - runp = (aiocb_union *) runp->aiocb.__next_prio; - } + /* If no request to work on we will stop the thread. */ + if (runp == NULL) + --nthreads; - /* Release the semaphore. */ - sem_post (&__aio_requests_sema); + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); } while (runp != NULL); pthread_exit (NULL); } + + +/* Free allocated resources. */ +static void +__attribute__ ((unused)) +free_res (void) +{ + size_t row; + + /* The first block of rows as specified in OPTIM is allocated in + one chunk. */ + free (pool[0]); + + for (row = optim.aio_num / ENTRIES_PER_ROW; row < pool_tab_size; ++row) + free (pool[row]); + + free (pool); +} + +text_set_element (__libc_subfreeres, free_res); diff --git a/rt/aio_misc.h b/rt/aio_misc.h index c2eb9fcea1..af8e2479ba 100644 --- a/rt/aio_misc.h +++ b/rt/aio_misc.h @@ -19,7 +19,19 @@ #ifndef _AIO_MISC_H #define _AIO_MISC_H 1 -#include +#include +#include + + +/* Extend the operation enum. */ +enum +{ + LIO_DSYNC = LIO_READ + 1, + LIO_SYNC, + LIO_READ64 = LIO_READ | 128, + LIO_WRITE64 = LIO_WRITE | 128 +}; + /* Union of the two request types. */ typedef union @@ -28,18 +40,73 @@ typedef union struct aiocb64 aiocb64; } aiocb_union; -/* List of enqueued requests. */ -extern aiocb_union *__aio_requests; + +/* Used to synchronize. */ +struct waitlist + { + struct waitlist *next; + + pthread_cond_t *cond; + volatile int *counterp; + /* The next field is used in asynchronous `lio_listio' operations. */ + struct sigevent *sigevp; + }; + + +/* Status of a request. */ +enum +{ + no, + queued, + yes, + allocated +}; + + +/* Used to queue requests.. */ +struct requestlist + { + int running; + + struct requestlist *last_fd; + struct requestlist *next_fd; + struct requestlist *next_prio; + struct requestlist *next_run; + + /* Pointer to the actual data. */ + aiocb_union *aiocbp; + + /* List of waiting processes. */ + struct waitlist *waiting; + }; + /* Lock for global I/O list of requests. */ -extern sem_t __aio_requests_sema; +extern pthread_mutex_t __aio_requests_mutex; /* Enqueue request. */ -extern int __aio_enqueue_request (aiocb_union *aiocbp, int operation, - int require_lock); +extern struct requestlist *__aio_enqueue_request (aiocb_union *aiocbp, + int operation) + internal_function; + +/* Find request entry for given AIO control block. */ +extern struct requestlist *__aio_find_req (aiocb_union *elem) + internal_function; + +/* Find request entry for given file descriptor. */ +extern struct requestlist *__aio_find_req_fd (int fildes) internal_function; + +/* Release the entry for the request. */ +extern void __aio_free_request (struct requestlist *req) internal_function; + +/* Notify initiator of request and tell this everybody listening. */ +extern void __aio_notify (struct requestlist *req) internal_function; + +/* Notify initiator of request. */ +extern int __aio_notify_only (struct sigevent *sigev) internal_function; /* Send the signal. */ -extern int __aio_sigqueue (int sig, const union sigval val); +extern int __aio_sigqueue (int sig, const union sigval val) internal_function; #endif /* aio_misc.h */ diff --git a/rt/aio_notify.c b/rt/aio_notify.c new file mode 100644 index 0000000000..be2e1e246b --- /dev/null +++ b/rt/aio_notify.c @@ -0,0 +1,97 @@ +/* Notify initiator of AIO request. + Copyright (C) 1997 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Ulrich Drepper , 1997. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Library General Public License as + published by the Free Software Foundation; either version 2 of the + License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Library General Public License for more details. + + You should have received a copy of the GNU Library General Public + License along with the GNU C Library; see the file COPYING.LIB. If not, + write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330, + Boston, MA 02111-1307, USA. */ + +#include +#include +#include "aio_misc.h" + +int +__aio_notify_only (struct sigevent *sigev) +{ + int result = 0; + + /* Send the signal to notify about finished processing of the request. */ + if (sigev->sigev_notify == SIGEV_THREAD) + { + /* We have to start a thread. */ + pthread_t tid; + pthread_attr_t attr, *pattr; + + pattr = (pthread_attr_t *) sigev->sigev_notify_attributes; + if (pattr == NULL) + { + pthread_attr_init (&attr); + pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); + pattr = &attr; + } + + if (pthread_create (&tid, pattr, + (void *(*) (void *)) sigev->sigev_notify_function, + sigev->sigev_value.sival_ptr) < 0) + result = -1; + } + else if (sigev->sigev_notify == SIGEV_SIGNAL) + /* We have to send a signal. */ + if (__aio_sigqueue (sigev->sigev_signo, sigev->sigev_value) < 0) + result = -1; + + return result; +} + + +void +__aio_notify (struct requestlist *req) +{ + struct waitlist *waitlist; + struct aiocb *aiocbp = &req->aiocbp->aiocb; + + if (__aio_notify_only (&aiocbp->aio_sigevent) != 0) + { + /* XXX What shall we do if already an error is set by + read/write/fsync? */ + aiocbp->__error_code = errno; + aiocbp->__return_value = -1; + } + + /* Now also notify possibly waiting threads. */ + waitlist = req->waiting; + while (waitlist != NULL) + { + struct waitlist *next = waitlist->next; + + /* Decrement the counter. This is used in both cases. */ + --*waitlist->counterp; + + if (waitlist->sigevp == NULL) + pthread_cond_signal (waitlist->cond); + else + /* This is part of a asynchronous `lio_listio' operation. If + this request is the last one, send the signal. */ + if (*waitlist->counterp == 0) + { + __aio_notify_only (waitlist->sigevp); + /* This is tricky. See lio_listio.c for the reason why + this works. */ + free ((void *) waitlist->counterp); + } + + waitlist = next; + } +} diff --git a/rt/aio_read.c b/rt/aio_read.c index 8286ba9334..871ce99767 100644 --- a/rt/aio_read.c +++ b/rt/aio_read.c @@ -27,5 +27,5 @@ int aio_read (aiocbp) struct aiocb *aiocbp; { - return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ, 1); + return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ) != NULL; } diff --git a/rt/aio_read64.c b/rt/aio_read64.c index bf808fb3da..b194dda6db 100644 --- a/rt/aio_read64.c +++ b/rt/aio_read64.c @@ -27,5 +27,5 @@ int aio_read64 (aiocbp) struct aiocb64 *aiocbp; { - return __aio_enqueue_request ((aiocb_union *) aiocbp, __LIO_READ64, 1); + return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_READ64) != NULL; } diff --git a/rt/aio_suspend.c b/rt/aio_suspend.c index 75bf9ba94b..6123b7baa0 100644 --- a/rt/aio_suspend.c +++ b/rt/aio_suspend.c @@ -40,16 +40,89 @@ aio_suspend (list, nent, timeout) int nent; const struct timespec *timeout; { + pthread_cond_t cond; + struct waitlist waitlist[nent]; + struct requestlist *requestlist[nent]; int cnt; + int result = 0; + + /* Request the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); /* First look whether there is already a terminated request. */ for (cnt = 0; cnt < nent; ++cnt) if (list[cnt] != NULL && list[cnt]->__error_code != EINPROGRESS) - return 0; + break; + + if (cnt == nent) + { + int oldstate; + + /* There is not yet a finished request. Signal the request that + we are working for it. */ + for (cnt = 0; cnt < nent; ++cnt) + if (list[cnt] != NULL && list[cnt]->__error_code == EINPROGRESS) + { + requestlist[cnt] = __aio_find_req ((aiocb_union *) list[cnt]); + + if (requestlist[cnt] != NULL) + { + waitlist[cnt].cond = &cond; + waitlist[cnt].next = requestlist[cnt]->waiting; + waitlist[cnt].counterp = NULL; + waitlist[cnt].sigevp = NULL; + requestlist[cnt]->waiting = &waitlist[cnt]; + } + } + + /* Since `pthread_cond_wait'/`pthread_cond_timedwait' are cancelation + points we must be careful. We added entries to the waiting lists + which we must remove. So defer cancelation for now. */ + pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &oldstate); + + if (timeout == NULL) + result = pthread_cond_wait (&cond, &__aio_requests_mutex); + else + result = pthread_cond_timedwait (&cond, &__aio_requests_mutex, + timeout); + + /* Now remove the entry in the waiting list for all requests + which didn't terminate */ + for (cnt = 0; cnt < nent; ++cnt) + if (list[cnt] != NULL && list[cnt]->__error_code == EINPROGRESS + && requestlist[cnt] != NULL) + { + struct waitlist **listp = &requestlist[cnt]->waiting; + + /* There is the chance that we cannot find our entry anymore. + This could happen if the request terminated and restarted + again. */ + while (*listp != NULL && *listp != &waitlist[cnt]) + listp = &(*listp)->next; + + if (*listp != NULL) + *listp = (*listp)->next; + } + + /* Now it's time to restore the cancelation state. */ + pthread_setcancelstate (oldstate, NULL); + + if (result != 0) + { + /* An error occurred. Possibly it's EINTR. We have to translate + the timeout error report of `pthread_cond_timedwait' to the + form expected from `aio_suspend'. */ + if (result == ETIMEDOUT) + __set_errno (EAGAIN); + + result = -1; + } + } - /* XXX We have to write code which waits. */ + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); - return -1; + return result; } weak_alias (aio_suspend, aio_suspend64) diff --git a/rt/aio_write.c b/rt/aio_write.c index 80b827c1aa..8801d89110 100644 --- a/rt/aio_write.c +++ b/rt/aio_write.c @@ -27,5 +27,5 @@ int aio_write (aiocbp) struct aiocb *aiocbp; { - return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE, 1); + return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE) != NULL; } diff --git a/rt/aio_write64.c b/rt/aio_write64.c index 477ce2cbe4..4d599b1062 100644 --- a/rt/aio_write64.c +++ b/rt/aio_write64.c @@ -27,5 +27,5 @@ int aio_write64 (aiocbp) struct aiocb64 *aiocbp; { - return __aio_enqueue_request ((aiocb_union *) aiocbp, __LIO_WRITE64, 1); + return __aio_enqueue_request ((aiocb_union *) aiocbp, LIO_WRITE64) != NULL; } diff --git a/rt/lio_listio.c b/rt/lio_listio.c index 73df5c2f5c..b389c6a5f6 100644 --- a/rt/lio_listio.c +++ b/rt/lio_listio.c @@ -20,11 +20,20 @@ #include #include -#include +#include #include "aio_misc.h" +/* We need this special structure to handle asynchronous I/O. */ +struct async_waitlist + { + int counter; + struct sigevent sigev; + struct waitlist list[0]; + }; + + int lio_listio (mode, list, nent, sig) int mode; @@ -32,8 +41,9 @@ lio_listio (mode, list, nent, sig) int nent; struct sigevent *sig; { + struct requestlist *requests[nent]; int cnt; - int total = 0; + volatile int total = 0; int result = 0; /* Check arguments. */ @@ -43,26 +53,100 @@ lio_listio (mode, list, nent, sig) return -1; } - /* Request the semaphore. */ - sem_wait (&__aio_requests_sema); + /* Request the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); /* Now we can enqueue all requests. Since we already acquired the - semaphore the enqueue function need not do this. */ + mutex the enqueue function need not do this. */ for (cnt = 0; cnt < nent; ++cnt) if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP) - if (__aio_enqueue_request ((aiocb_union *) list[cnt], - list[cnt]->aio_lio_opcode, 0) >= 0) - /* Successfully enqueued. */ - ++total; + { + requests[cnt] = __aio_enqueue_request ((aiocb_union *) list[cnt], + list[cnt]->aio_lio_opcode); + + if (requests[cnt] != NULL) + /* Successfully enqueued. */ + ++total; + else + /* Signal that we've seen an error. `errno' and the error code + of the aiocb will tell more. */ + result = -1; + } + + if (total == 0) + { + /* We don't have anything to do except signalling if we work + asynchronously. */ + if (mode == LIO_NOWAIT) + __aio_notify_only (sig); + } + else if (mode == LIO_WAIT) + { + pthread_cond_t cond; + struct waitlist waitlist[nent]; + int oldstate; + + total = 0; + for (cnt = 0; cnt < nent; ++cnt) + if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP + && requests[cnt] != NULL) + { + waitlist[cnt].cond = &cond; + waitlist[cnt].next = requests[cnt]->waiting; + waitlist[cnt].counterp = NULL; + waitlist[cnt].sigevp = NULL; + requests[cnt]->waiting = &waitlist[cnt]; + ++total; + } + + /* Since `pthread_cond_wait'/`pthread_cond_timedwait' are cancelation + points we must be careful. We added entries to the waiting lists + which we must remove. So defer cancelation for now. */ + pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &oldstate); + + while (total > 0) + if (pthread_cond_wait (&cond, &__aio_requests_mutex) == 0) + --total; + + /* Now it's time to restore the cancelation state. */ + pthread_setcancelstate (oldstate, NULL); + } + else + { + struct async_waitlist *waitlist; + + waitlist = (struct async_waitlist *) + malloc (sizeof (struct async_waitlist) + + (nent * sizeof (struct waitlist))); + + if (waitlist == NULL) + { + __set_errno (EAGAIN); + result = -1; + } else - /* Signal that we've seen an error. `errno' and the error code - of the aiocb will tell more. */ - result = -1; + { + total = 0; + for (cnt = 0; cnt < nent; ++cnt) + if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP + && requests[cnt] != NULL) + { + waitlist->list[cnt].cond = NULL; + waitlist->list[cnt].next = requests[cnt]->waiting; + waitlist->list[cnt].counterp = &waitlist->counter; + waitlist->list[cnt].sigevp = &waitlist->sigev; + requests[cnt]->waiting = &waitlist->list[cnt]; + ++total; + } + waitlist->counter = total; + waitlist->sigev = *sig; + } + } - /* Release the semaphore. */ - sem_post (&__aio_requests_sema); + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); return result; } diff --git a/rt/lio_listio64.c b/rt/lio_listio64.c index b17e9edae2..19f56adab1 100644 --- a/rt/lio_listio64.c +++ b/rt/lio_listio64.c @@ -20,11 +20,20 @@ #include #include -#include +#include #include "aio_misc.h" +/* We need this special structure to handle asynchronous I/O. */ +struct async_waitlist + { + int counter; + struct sigevent sigev; + struct waitlist list[0]; + }; + + int lio_listio64 (mode, list, nent, sig) int mode; @@ -32,8 +41,9 @@ lio_listio64 (mode, list, nent, sig) int nent; struct sigevent *sig; { + struct requestlist *requests[nent]; int cnt; - int total = 0; + volatile int total = 0; int result = 0; /* Check arguments. */ @@ -43,24 +53,99 @@ lio_listio64 (mode, list, nent, sig) return -1; } - /* Request the semaphore. */ - sem_wait (&__aio_requests_sema); + /* Request the mutex. */ + pthread_mutex_lock (&__aio_requests_mutex); /* Now we can enqueue all requests. Since we already acquired the - semaphore the enqueue function need not do this. */ + mutex the enqueue function need not do this. */ for (cnt = 0; cnt < nent; ++cnt) if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP) - if (__aio_enqueue_request ((aiocb_union *) list[cnt], - list[cnt]->aio_lio_opcode | 128, 0) >= 0) - /* Successfully enqueued. */ - ++total; + { + requests[cnt] = __aio_enqueue_request ((aiocb_union *) list[cnt], + (list[cnt]->aio_lio_opcode + | 128)); + if (requests[cnt] != NULL) + /* Successfully enqueued. */ + ++total; + else + /* Signal that we've seen an error. `errno' and the error code + of the aiocb will tell more. */ + result = -1; + } + + if (total == 0) + { + /* We don't have anything to do except signalling if we work + asynchronously. */ + if (mode == LIO_NOWAIT) + __aio_notify_only (sig); + } + else if (mode == LIO_WAIT) + { + pthread_cond_t cond; + struct waitlist waitlist[nent]; + int oldstate; + + total = 0; + for (cnt = 0; cnt < nent; ++cnt) + if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP + && requests[cnt] != NULL) + { + waitlist[cnt].cond = &cond; + waitlist[cnt].next = requests[cnt]->waiting; + waitlist[cnt].counterp = &total; + waitlist[cnt].sigevp = NULL; + requests[cnt]->waiting = &waitlist[cnt]; + ++total; + } + + /* Since `pthread_cond_wait'/`pthread_cond_timedwait' are cancelation + points we must be careful. We added entries to the waiting lists + which we must remove. So defer cancelation for now. */ + pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &oldstate); + + while (total > 0) + pthread_cond_wait (&cond, &__aio_requests_mutex); + + /* Now it's time to restore the cancelation state. */ + pthread_setcancelstate (oldstate, NULL); + } + else + { + struct async_waitlist *waitlist; + + waitlist = (struct async_waitlist *) + malloc (sizeof (struct async_waitlist) + + (nent * sizeof (struct waitlist))); + + if (waitlist == NULL) + { + __set_errno (EAGAIN); + result = -1; + } else - /* Signal that we've seen an error. `errno' and the error code - of the aiocb will tell more. */ - result = -1; + { + total = 0; + + for (cnt = 0; cnt < nent; ++cnt) + if (list[cnt] != NULL && list[cnt]->aio_lio_opcode != LIO_NOP + && requests[cnt] != NULL) + { + waitlist->list[cnt].cond = NULL; + waitlist->list[cnt].next = requests[cnt]->waiting; + waitlist->list[cnt].counterp = &waitlist->counter; + waitlist->list[cnt].sigevp = &waitlist->sigev; + requests[cnt]->waiting = &waitlist->list[cnt]; + ++total; + } + + waitlist->counter = total; + waitlist->sigev = *sig; + } + } - /* Release the semaphore. */ - sem_post (&__aio_requests_sema); + /* Release the mutex. */ + pthread_mutex_unlock (&__aio_requests_mutex); return result; } -- cgit v1.2.3