--- IO-AIO/AIO.xs 2006/10/26 13:25:40 1.77 +++ IO-AIO/AIO.xs 2007/06/01 06:00:40 1.100 @@ -1,12 +1,4 @@ -/* solaris */ -#define _POSIX_PTHREAD_SEMANTICS 1 - -#if __linux && !defined(_GNU_SOURCE) -# define _GNU_SOURCE -#endif - -/* just in case */ -#define _REENTRANT 1 +#include "xthread.h" #include @@ -16,9 +8,8 @@ #include "autoconf/config.h" -#include - #include +#include #include #include #include @@ -29,6 +20,7 @@ #include #include #include +#include #if HAVE_SENDFILE # if __linux @@ -45,24 +37,14 @@ # endif #endif +/* number of seconds after which idle threads exit */ +#define IDLE_TIMEOUT 10 + /* used for struct dirent, AIX doesn't provide it */ #ifndef NAME_MAX # define NAME_MAX 4096 #endif -#ifndef PTHREAD_STACK_MIN -/* care for broken platforms, e.g. windows */ -# define PTHREAD_STACK_MIN 16384 -#endif - -#if __ia64 -# define STACKSIZE 65536 -#elif __i386 || __x86_64 /* 16k is unreasonably high :( */ -# define STACKSIZE PTHREAD_STACK_MIN -#else -# define STACKSIZE 16384 -#endif - /* buffer size for various temporary buffers */ #define AIO_BUFSIZE 65536 @@ -74,16 +56,21 @@ if (!aio_buf) \ return -1; +typedef SV SV8; /* byte-sv, used for argument-checking */ + enum { REQ_QUIT, REQ_OPEN, REQ_CLOSE, REQ_READ, REQ_WRITE, REQ_READAHEAD, REQ_SENDFILE, REQ_STAT, REQ_LSTAT, REQ_FSTAT, + REQ_UTIME, REQ_FUTIME, /* must be consecutive */ + REQ_CHMOD, REQ_FCHMOD, /* must be consecutive */ + REQ_CHOWN, REQ_FCHOWN, /* must be consecutive */ REQ_FSYNC, REQ_FDATASYNC, - REQ_UNLINK, REQ_RMDIR, REQ_RENAME, - REQ_READDIR, - REQ_LINK, REQ_SYMLINK, + REQ_UNLINK, REQ_RMDIR, REQ_MKDIR, REQ_RENAME, + REQ_MKNOD, REQ_READDIR, + REQ_LINK, REQ_SYMLINK, REQ_READLINK, REQ_GROUP, REQ_NOP, REQ_BUSY, }; @@ -95,17 +82,17 @@ { struct aio_cb *volatile next; - SV *data, *callback; - SV *fh, *fh2; - void *dataptr, *data2ptr; - Stat_t *statdata; - off_t offset; - size_t length; + SV *callback; + SV *sv1, *sv2; + void *ptr1, *ptr2; + off_t offs; + size_t size; ssize_t result; + double nv1, nv2; - STRLEN dataoffset; + STRLEN stroffset; int type; - int fd, fd2; + int int1, int2, int3; int errorno; mode_t mode; /* open */ @@ -117,7 +104,9 @@ } aio_cb; enum { - FLAG_CANCELLED = 0x01, + FLAG_CANCELLED = 0x01, /* request was cancelled */ + FLAG_SV2_RO_OFF = 0x40, /* data was set readonly */ + FLAG_PTR2_FREE = 0x80, /* need to free(ptr2) */ }; typedef aio_cb *aio_req; @@ -132,29 +121,64 @@ NUM_PRI = PRI_MAX + PRI_BIAS + 1, }; -static int next_pri = DEFAULT_PRI + PRI_BIAS; +#define AIO_TICKS ((1000000 + 1023) >> 10) -static int started, wanted; -static volatile int nreqs; -static int respipe [2]; +static unsigned int max_poll_time = 0; +static unsigned int max_poll_reqs = 0; -#if __linux && defined (PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP) -# define AIO_MUTEX_INIT PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP -#else -# define AIO_MUTEX_INIT PTHREAD_MUTEX_INITIALIZER -#endif +/* calculcate time difference in ~1/AIO_TICKS of a second */ +static int tvdiff (struct timeval *tv1, struct timeval *tv2) +{ + return (tv2->tv_sec - tv1->tv_sec ) * AIO_TICKS + + ((tv2->tv_usec - tv1->tv_usec) >> 10); +} + +static thread_t main_tid; +static int main_sig; +static int block_sig_level; + +void block_sig () +{ + sigset_t ss; + + if (block_sig_level++) + return; + + if (!main_sig) + return; + + sigemptyset (&ss); + sigaddset (&ss, main_sig); + pthread_sigmask (SIG_BLOCK, &ss, 0); +} + +void unblock_sig () +{ + sigset_t ss; + + if (--block_sig_level) + return; + + if (!main_sig) + return; + + sigemptyset (&ss); + sigaddset (&ss, main_sig); + pthread_sigmask (SIG_UNBLOCK, &ss, 0); +} -#define LOCK(mutex) pthread_mutex_lock (&(mutex)) -#define UNLOCK(mutex) pthread_mutex_unlock (&(mutex)) +static int next_pri = DEFAULT_PRI + PRI_BIAS; + +static unsigned int started, idle, wanted; /* worker threads management */ -static pthread_mutex_t wrklock = AIO_MUTEX_INIT; +static mutex_t wrklock = MUTEX_INIT; typedef struct worker { /* locked by wrklock */ struct worker *prev, *next; - pthread_t tid; + thread_t tid; /* locked by reslock, reqlock or wrklock */ aio_req req; /* currently processed request */ @@ -187,9 +211,57 @@ free (wrk); } -static pthread_mutex_t reslock = AIO_MUTEX_INIT; -static pthread_mutex_t reqlock = AIO_MUTEX_INIT; -static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER; +static volatile unsigned int nreqs, nready, npending; +static volatile unsigned int max_idle = 4; +static volatile unsigned int max_outstanding = 0xffffffff; +static int respipe [2]; + +static mutex_t reslock = MUTEX_INIT; +static mutex_t reqlock = MUTEX_INIT; +static cond_t reqwait = COND_INIT; + +#if WORDACCESS_UNSAFE + +static unsigned int get_nready () +{ + unsigned int retval; + + LOCK (reqlock); + retval = nready; + UNLOCK (reqlock); + + return retval; +} + +static unsigned int get_npending () +{ + unsigned int retval; + + LOCK (reslock); + retval = npending; + UNLOCK (reslock); + + return retval; +} + +static unsigned int get_nthreads () +{ + unsigned int retval; + + LOCK (wrklock); + retval = started; + UNLOCK (wrklock); + + return retval; +} + +#else + +# define get_nready() nready +# define get_npending() npending +# define get_nthreads() started + +#endif /* * a somewhat faster data structure might be nice, but @@ -245,8 +317,8 @@ abort (); } -static int poll_cb (int max); -static void req_invoke (aio_req req); +static int poll_cb (); +static int req_invoke (aio_req req); static void req_free (aio_req req); static void req_cancel (aio_req req); @@ -276,11 +348,13 @@ static void aio_grp_feed (aio_req grp) { - while (grp->length < grp->fd2 && !(grp->flags & FLAG_CANCELLED)) + block_sig (); + + while (grp->size < grp->int2 && !(grp->flags & FLAG_CANCELLED)) { - int old_len = grp->length; + int old_len = grp->size; - if (grp->fh2 && SvOK (grp->fh2)) + if (grp->sv2 && SvOK (grp->sv2)) { dSP; @@ -289,70 +363,57 @@ PUSHMARK (SP); XPUSHs (req_sv (grp, AIO_GRP_KLASS)); PUTBACK; - call_sv (grp->fh2, G_VOID | G_EVAL | G_KEEPERR); + call_sv (grp->sv2, G_VOID | G_EVAL | G_KEEPERR); SPAGAIN; FREETMPS; LEAVE; } /* stop if no progress has been made */ - if (old_len == grp->length) + if (old_len == grp->size) { - SvREFCNT_dec (grp->fh2); - grp->fh2 = 0; + SvREFCNT_dec (grp->sv2); + grp->sv2 = 0; break; } } + + unblock_sig (); } static void aio_grp_dec (aio_req grp) { - --grp->length; + --grp->size; /* call feeder, if applicable */ aio_grp_feed (grp); /* finish, if done */ - if (!grp->length && grp->fd) + if (!grp->size && grp->int1) { - req_invoke (grp); - req_free (grp); - } -} + block_sig (); -static void poll_wait () -{ - fd_set rfd; - - while (nreqs) - { - int size; -#if !(__i386 || __x86_64) /* safe without sempahore on these archs */ - LOCK (reslock); -#endif - size = res_queue.size; -#if !(__i386 || __x86_64) /* safe without sempahore on these archs */ - UNLOCK (reslock); -#endif - - if (size) - return; - - FD_ZERO(&rfd); - FD_SET(respipe [0], &rfd); + if (!req_invoke (grp)) + { + req_free (grp); + unblock_sig (); + croak (0); + } - select (respipe [0] + 1, &rfd, 0, 0, 0); + req_free (grp); + unblock_sig (); } } -static void req_invoke (aio_req req) +static int req_invoke (aio_req req) { dSP; + if (req->flags & FLAG_SV2_RO_OFF) + SvREADONLY_off (req->sv2); + if (!(req->flags & FLAG_CANCELLED) && SvOK (req->callback)) { - errno = req->errorno; - ENTER; SAVETMPS; PUSHMARK (SP); @@ -367,7 +428,7 @@ if (req->result >= 0) { int i; - char *buf = req->data2ptr; + char *buf = req->ptr2; AV *av = newAV (); av_extend (av, req->result - 1); @@ -397,20 +458,19 @@ call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); SPAGAIN; - fh = SvREFCNT_inc (POPs); - + fh = POPs; PUSHMARK (SP); - XPUSHs (sv_2mortal (fh)); + XPUSHs (fh); } break; case REQ_GROUP: - req->fd = 2; /* mark group as finished */ + req->int1 = 2; /* mark group as finished */ - if (req->data) + if (req->sv1) { int i; - AV *av = (AV *)req->data; + AV *av = (AV *)req->sv1; EXTEND (SP, AvFILL (av) + 1); for (i = 0; i <= AvFILL (av); ++i) @@ -422,11 +482,36 @@ case REQ_BUSY: break; + case REQ_READLINK: + if (req->result > 0) + { + SvCUR_set (req->sv2, req->result); + *SvEND (req->sv2) = 0; + PUSHs (req->sv2); + } + break; + + case REQ_STAT: + case REQ_LSTAT: + case REQ_FSTAT: + PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; + PL_laststatval = req->result; + PL_statcache = *(Stat_t *)(req->ptr2); + PUSHs (sv_2mortal (newSViv (req->result))); + break; + + case REQ_READ: + SvCUR_set (req->sv2, req->stroffset + (req->result > 0 ? req->result : 0)); + *SvEND (req->sv2) = 0; + PUSHs (sv_2mortal (newSViv (req->result))); + break; + default: PUSHs (sv_2mortal (newSViv (req->result))); break; } + errno = req->errorno; PUTBACK; call_sv (req->callback, G_VOID | G_EVAL); @@ -450,11 +535,7 @@ aio_grp_dec (grp); } - if (SvTRUE (ERRSV)) - { - req_free (req); - croak (0); - } + return !SvTRUE (ERRSV); } static void req_free (aio_req req) @@ -465,14 +546,12 @@ SvREFCNT_dec (req->self); } - SvREFCNT_dec (req->data); - SvREFCNT_dec (req->fh); - SvREFCNT_dec (req->fh2); + SvREFCNT_dec (req->sv1); + SvREFCNT_dec (req->sv2); SvREFCNT_dec (req->callback); - Safefree (req->statdata); - if (req->type == REQ_READDIR) - free (req->data2ptr); + if (req->flags & FLAG_PTR2_FREE) + free (req->ptr2); Safefree (req); } @@ -484,8 +563,8 @@ if (grp->type != REQ_GROUP) return; - SvREFCNT_dec (grp->fh2); - grp->fh2 = 0; + SvREFCNT_dec (grp->sv2); + grp->sv2 = 0; for (sub = grp->grp_first; sub; sub = sub->grp_next) req_cancel (sub); @@ -498,116 +577,58 @@ req_cancel_subs (req); } -static int poll_cb (int max) -{ - dSP; - int count = 0; - int do_croak = 0; - aio_req req; - - while (max <= 0 || count < max) - { - LOCK (reslock); - req = reqq_shift (&res_queue); - - if (req) - { - if (!res_queue.size) - { - /* read any signals sent by the worker threads */ - char buf [32]; - while (read (respipe [0], buf, 32) == 32) - ; - } - } - - UNLOCK (reslock); - - if (!req) - break; - - --nreqs; - - if (req->type == REQ_QUIT) - started--; - else if (req->type == REQ_GROUP && req->length) - { - req->fd = 1; /* mark request as delayed */ - continue; - } - else - { - if (req->type == REQ_READ) - SvCUR_set (req->data, req->dataoffset + (req->result > 0 ? req->result : 0)); - - if (req->data2ptr && (req->type == REQ_READ || req->type == REQ_WRITE)) - SvREADONLY_off (req->data); - - if (req->statdata) - { - PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; - PL_laststatval = req->result; - PL_statcache = *(req->statdata); - } - - req_invoke (req); - - count++; - } - - req_free (req); - } - - return count; -} - static void *aio_proc(void *arg); static void start_thread (void) { - sigset_t fullsigset, oldsigset; - pthread_attr_t attr; - worker *wrk = calloc (1, sizeof (worker)); if (!wrk) croak ("unable to allocate worker thread data"); - pthread_attr_init (&attr); - pthread_attr_setstacksize (&attr, STACKSIZE); - pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); - - sigfillset (&fullsigset); - LOCK (wrklock); - sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset); - if (pthread_create (&wrk->tid, &attr, aio_proc, (void *)wrk) == 0) + if (thread_create (&wrk->tid, aio_proc, (void *)wrk)) { wrk->prev = &wrk_first; wrk->next = wrk_first.next; wrk_first.next->prev = wrk; wrk_first.next = wrk; - started++; + ++started; } else free (wrk); - sigprocmask (SIG_SETMASK, &oldsigset, 0); UNLOCK (wrklock); } +static void maybe_start_thread () +{ + if (get_nthreads () >= wanted) + return; + + /* todo: maybe use idle here, but might be less exact */ + if (0 <= (int)get_nthreads () + (int)get_npending () - (int)nreqs) + return; + + start_thread (); +} + static void req_send (aio_req req) { - while (started < wanted && nreqs >= started) - start_thread (); + block_sig (); ++nreqs; LOCK (reqlock); + ++nready; reqq_push (&req_queue, req); - pthread_cond_signal (&reqwait); + COND_SIGNAL (reqwait); UNLOCK (reqlock); + + unblock_sig (); + + maybe_start_thread (); } static void end_thread (void) @@ -619,7 +640,21 @@ req->type = REQ_QUIT; req->pri = PRI_MAX + PRI_BIAS; - req_send (req); + LOCK (reqlock); + reqq_push (&req_queue, req); + COND_SIGNAL (reqwait); + UNLOCK (reqlock); + + LOCK (wrklock); + --started; + UNLOCK (wrklock); +} + +static void set_max_idle (int nthreads) +{ + if (WORDACCESS_UNSAFE) LOCK (reqlock); + max_idle = nthreads <= 0 ? 1 : nthreads; + if (WORDACCESS_UNSAFE) UNLOCK (reqlock); } static void min_parallel (int nthreads) @@ -630,22 +665,120 @@ static void max_parallel (int nthreads) { - int cur = started; - if (wanted > nthreads) wanted = nthreads; - while (cur > wanted) + while (started > wanted) + end_thread (); +} + +static void poll_wait () +{ + fd_set rfd; + + while (nreqs) { - end_thread (); - cur--; + int size; + if (WORDACCESS_UNSAFE) LOCK (reslock); + size = res_queue.size; + if (WORDACCESS_UNSAFE) UNLOCK (reslock); + + if (size) + return; + + maybe_start_thread (); + + FD_ZERO(&rfd); + FD_SET(respipe [0], &rfd); + + select (respipe [0] + 1, &rfd, 0, 0, 0); } +} - while (started > wanted) +static int poll_cb () +{ + dSP; + int count = 0; + int maxreqs = max_poll_reqs; + int do_croak = 0; + struct timeval tv_start, tv_now; + aio_req req; + + if (max_poll_time) + gettimeofday (&tv_start, 0); + + block_sig (); + + for (;;) { + for (;;) + { + maybe_start_thread (); + + LOCK (reslock); + req = reqq_shift (&res_queue); + + if (req) + { + --npending; + + if (!res_queue.size) + { + /* read any signals sent by the worker threads */ + char buf [4]; + while (read (respipe [0], buf, 4) == 4) + ; + } + } + + UNLOCK (reslock); + + if (!req) + break; + + --nreqs; + + if (req->type == REQ_GROUP && req->size) + { + req->int1 = 1; /* mark request as delayed */ + continue; + } + else + { + if (!req_invoke (req)) + { + req_free (req); + unblock_sig (); + croak (0); + } + + count++; + } + + req_free (req); + + if (maxreqs && !--maxreqs) + break; + + if (max_poll_time) + { + gettimeofday (&tv_now, 0); + + if (tvdiff (&tv_start, &tv_now) >= max_poll_time) + break; + } + } + + if (nreqs <= max_outstanding) + break; + poll_wait (); - poll_cb (0); + + ++maxreqs; } + + unblock_sig (); + return count; } static void create_pipe () @@ -672,7 +805,7 @@ * normal read/write by using a mutex. slows down execution a lot, * but that's your problem, not mine. */ -static pthread_mutex_t preadwritelock = PTHREAD_MUTEX_INITIALIZER; +static mutex_t preadwritelock = MUTEX_INIT; static ssize_t pread (int fd, void *buf, size_t count, off_t offset) { @@ -705,6 +838,34 @@ } #endif +#ifndef HAVE_FUTIMES + +# define utimes(path,times) aio_utimes (path, times) +# define futimes(fd,times) aio_futimes (fd, times) + +int aio_utimes (const char *filename, const struct timeval times[2]) +{ + if (times) + { + struct utimbuf buf; + + buf.actime = times[0].tv_sec; + buf.modtime = times[1].tv_sec; + + return utime (filename, &buf); + } + else + return utime (filename, 0); +} + +int aio_futimes (int fd, const struct timeval tv[2]) +{ + errno = ENOSYS; + return -1; +} + +#endif + #if !HAVE_FDATASYNC # define fdatasync fsync #endif @@ -733,7 +894,7 @@ #if !HAVE_READDIR_R # define readdir_r aio_readdir_r -static pthread_mutex_t readdirlock = PTHREAD_MUTEX_INITIALIZER; +static mutex_t readdirlock = MUTEX_INIT; static int readdir_r (DIR *dirp, struct dirent *ent, struct dirent **res) { @@ -870,9 +1031,10 @@ int errorno; LOCK (wrklock); - self->dirp = dirp = opendir (req->dataptr); + self->dirp = dirp = opendir (req->ptr1); self->dbuf = u = malloc (sizeof (*u)); - req->data2ptr = names = malloc (memlen); + req->flags |= FLAG_PTR2_FREE; + req->ptr2 = names = malloc (memlen); UNLOCK (wrklock); if (dirp && u && names) @@ -896,7 +1058,7 @@ { memlen *= 2; LOCK (wrklock); - req->data2ptr = names = realloc (names, memlen); + req->ptr2 = names = realloc (names, memlen); UNLOCK (wrklock); if (!names) @@ -919,11 +1081,17 @@ static void *aio_proc (void *thr_arg) { aio_req req; - int type; + struct timespec ts; worker *self = (worker *)thr_arg; - do + /* try to distribute timeouts somewhat evenly */ + ts.tv_nsec = (((unsigned long)self + (unsigned long)ts.tv_sec) & 1023UL) + * (1000000000UL / 1024UL); + + for (;;) { + ts.tv_sec = time (0) + IDLE_TIMEOUT; + LOCK (reqlock); for (;;) @@ -933,56 +1101,111 @@ if (req) break; - pthread_cond_wait (&reqwait, &reqlock); + ++idle; + + if (COND_TIMEDWAIT (reqwait, reqlock, ts) + == ETIMEDOUT) + { + if (idle > max_idle) + { + --idle; + UNLOCK (reqlock); + LOCK (wrklock); + --started; + UNLOCK (wrklock); + goto quit; + } + + /* we are allowed to idle, so do so without any timeout */ + COND_WAIT (reqwait, reqlock); + ts.tv_sec = time (0) + IDLE_TIMEOUT; + } + + --idle; } + --nready; + UNLOCK (reqlock); errno = 0; /* strictly unnecessary */ - type = req->type; /* remember type for QUIT check */ if (!(req->flags & FLAG_CANCELLED)) - switch (type) + switch (req->type) { - case REQ_READ: req->result = pread (req->fd, req->dataptr, req->length, req->offset); break; - case REQ_WRITE: req->result = pwrite (req->fd, req->dataptr, req->length, req->offset); break; + case REQ_READ: req->result = pread (req->int1, req->ptr1, req->size, req->offs); break; + case REQ_WRITE: req->result = pwrite (req->int1, req->ptr1, req->size, req->offs); break; - case REQ_READAHEAD: req->result = readahead (req->fd, req->offset, req->length); break; - case REQ_SENDFILE: req->result = sendfile_ (req->fd, req->fd2, req->offset, req->length, self); break; + case REQ_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break; + case REQ_SENDFILE: req->result = sendfile_ (req->int1, req->int2, req->offs, req->size, self); break; - case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break; - case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break; - case REQ_FSTAT: req->result = fstat (req->fd , req->statdata); break; - - case REQ_OPEN: req->result = open (req->dataptr, req->fd, req->mode); break; - case REQ_CLOSE: req->result = close (req->fd); break; - case REQ_UNLINK: req->result = unlink (req->dataptr); break; - case REQ_RMDIR: req->result = rmdir (req->dataptr); break; - case REQ_RENAME: req->result = rename (req->data2ptr, req->dataptr); break; - case REQ_LINK: req->result = link (req->data2ptr, req->dataptr); break; - case REQ_SYMLINK: req->result = symlink (req->data2ptr, req->dataptr); break; + case REQ_STAT: req->result = stat (req->ptr1, (Stat_t *)req->ptr2); break; + case REQ_LSTAT: req->result = lstat (req->ptr1, (Stat_t *)req->ptr2); break; + case REQ_FSTAT: req->result = fstat (req->int1, (Stat_t *)req->ptr2); break; + + case REQ_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; + case REQ_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; + case REQ_CHMOD: req->result = chmod (req->ptr1, req->mode); break; + case REQ_FCHMOD: req->result = fchmod (req->int1, req->mode); break; + + case REQ_OPEN: req->result = open (req->ptr1, req->int1, req->mode); break; + case REQ_CLOSE: req->result = close (req->int1); break; + case REQ_UNLINK: req->result = unlink (req->ptr1); break; + case REQ_RMDIR: req->result = rmdir (req->ptr1); break; + case REQ_MKDIR: req->result = mkdir (req->ptr1, req->mode); break; + case REQ_RENAME: req->result = rename (req->ptr2, req->ptr1); break; + case REQ_LINK: req->result = link (req->ptr2, req->ptr1); break; + case REQ_SYMLINK: req->result = symlink (req->ptr2, req->ptr1); break; + case REQ_MKNOD: req->result = mknod (req->ptr2, req->mode, (dev_t)req->offs); break; + case REQ_READLINK: req->result = readlink (req->ptr2, req->ptr1, NAME_MAX); break; - case REQ_FDATASYNC: req->result = fdatasync (req->fd); break; - case REQ_FSYNC: req->result = fsync (req->fd); break; + case REQ_FDATASYNC: req->result = fdatasync (req->int1); break; + case REQ_FSYNC: req->result = fsync (req->int1); break; case REQ_READDIR: scandir_ (req, self); break; case REQ_BUSY: { struct timeval tv; - tv.tv_sec = req->fd; - tv.tv_usec = req->fd2; + tv.tv_sec = req->nv1; + tv.tv_usec = (req->nv1 - tv.tv_usec) * 1000000.; req->result = select (0, 0, 0, 0, &tv); } + case REQ_UTIME: + case REQ_FUTIME: + { + struct timeval tv[2]; + struct timeval *times; + + if (req->nv1 != -1. || req->nv2 != -1.) + { + tv[0].tv_sec = req->nv1; + tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.; + tv[1].tv_sec = req->nv2; + tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.; + + times = tv; + } + else + times = 0; + + + req->result = req->type == REQ_FUTIME + ? futimes (req->int1, times) + : utimes (req->ptr1, times); + } + case REQ_GROUP: case REQ_NOP: - case REQ_QUIT: break; + case REQ_QUIT: + goto quit; + default: - req->result = ENOSYS; + req->result = -1; break; } @@ -990,17 +1213,25 @@ LOCK (reslock); + ++npending; + if (!reqq_push (&res_queue, req)) - /* write a dummy byte to the pipe so fh becomes ready */ - write (respipe [1], &respipe, 1); + { + /* write a dummy byte to the pipe so fh becomes ready */ + write (respipe [1], &respipe, 1); + + /* optionally signal the main thread asynchronously */ + if (main_sig) + pthread_kill (main_tid, main_sig); + } self->req = 0; worker_clear (self); UNLOCK (reslock); } - while (type != REQ_QUIT); +quit: LOCK (wrklock); worker_free (self); UNLOCK (wrklock); @@ -1057,8 +1288,11 @@ worker_free (wrk); } - started = 0; - nreqs = 0; + started = 0; + idle = 0; + nreqs = 0; + nready = 0; + npending = 0; close (respipe [0]); close (respipe [1]); @@ -1095,48 +1329,72 @@ BOOT: { HV *stash = gv_stashpv ("IO::AIO", 1); + newCONSTSUB (stash, "EXDEV", newSViv (EXDEV)); newCONSTSUB (stash, "O_RDONLY", newSViv (O_RDONLY)); newCONSTSUB (stash, "O_WRONLY", newSViv (O_WRONLY)); + newCONSTSUB (stash, "O_CREAT", newSViv (O_CREAT)); + newCONSTSUB (stash, "O_TRUNC", newSViv (O_TRUNC)); + newCONSTSUB (stash, "S_IFIFO", newSViv (S_IFIFO)); + newCONSTSUB (stash, "SIGIO", newSViv (SIGIO)); create_pipe (); - pthread_atfork (atfork_prepare, atfork_parent, atfork_child); + ATFORK (atfork_prepare, atfork_parent, atfork_child); } void -min_parallel (nthreads) - int nthreads +max_poll_reqs (int nreqs) + PROTOTYPE: $ + CODE: + max_poll_reqs = nreqs; + +void +max_poll_time (double nseconds) + PROTOTYPE: $ + CODE: + max_poll_time = nseconds * AIO_TICKS; + +void +min_parallel (int nthreads) PROTOTYPE: $ void -max_parallel (nthreads) - int nthreads +max_parallel (int nthreads) PROTOTYPE: $ void -aio_open (pathname,flags,mode,callback=&PL_sv_undef) - SV * pathname - int flags - int mode - SV * callback +max_idle (int nthreads) + PROTOTYPE: $ + CODE: + set_max_idle (nthreads); + +int +max_outstanding (int maxreqs) + PROTOTYPE: $ + CODE: + RETVAL = max_outstanding; + max_outstanding = maxreqs; + OUTPUT: + RETVAL + +void +aio_open (SV8 *pathname, int flags, int mode, SV *callback=&PL_sv_undef) PROTOTYPE: $$$;$ PPCODE: { dREQ; req->type = REQ_OPEN; - req->data = newSVsv (pathname); - req->dataptr = SvPVbyte_nolen (req->data); - req->fd = flags; + req->sv1 = newSVsv (pathname); + req->ptr1 = SvPVbyte_nolen (req->sv1); + req->int1 = flags; req->mode = mode; REQ_SEND; } void -aio_close (fh,callback=&PL_sv_undef) - SV * fh - SV * callback +aio_close (SV *fh, SV *callback=&PL_sv_undef) PROTOTYPE: $;$ ALIAS: aio_close = REQ_CLOSE @@ -1147,27 +1405,20 @@ dREQ; req->type = ix; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); + req->sv1 = newSVsv (fh); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh))); REQ_SEND (req); } void -aio_read (fh,offset,length,data,dataoffset,callback=&PL_sv_undef) - SV * fh - UV offset - UV length - SV * data - UV dataoffset - SV * callback +aio_read (SV *fh, UV offset, UV length, SV8 *data, UV dataoffset, SV *callback=&PL_sv_undef) ALIAS: aio_read = REQ_READ aio_write = REQ_WRITE PROTOTYPE: $$$$$;$ PPCODE: { - aio_req req; STRLEN svlen; char *svptr = SvPVbyte (data, svlen); @@ -1189,7 +1440,7 @@ else { /* read: grow scalar as necessary */ - svptr = SvGROW (data, length + dataoffset); + svptr = SvGROW (data, length + dataoffset + 1); } if (length < 0) @@ -1199,18 +1450,19 @@ dREQ; req->type = ix; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) - : IoOFP (sv_2io (fh))); - req->offset = offset; - req->length = length; - req->data = SvREFCNT_inc (data); - req->dataptr = (char *)svptr + dataoffset; + req->sv1 = newSVsv (fh); + req->int1 = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) + : IoOFP (sv_2io (fh))); + req->offs = offset; + req->size = length; + req->sv2 = SvREFCNT_inc (data); + req->ptr1 = (char *)svptr + dataoffset; + req->stroffset = dataoffset; if (!SvREADONLY (data)) { SvREADONLY_on (data); - req->data2ptr = (void *)data; + req->flags |= FLAG_SV2_RO_OFF; } REQ_SEND; @@ -1218,52 +1470,61 @@ } void -aio_sendfile (out_fh,in_fh,in_offset,length,callback=&PL_sv_undef) - SV * out_fh - SV * in_fh - UV in_offset - UV length - SV * callback +aio_readlink (SV8 *path, SV *callback=&PL_sv_undef) + PROTOTYPE: $$;$ + PPCODE: +{ + SV *data; + dREQ; + + data = newSV (NAME_MAX); + SvPOK_on (data); + + req->type = REQ_READLINK; + req->sv1 = newSVsv (path); + req->ptr2 = SvPVbyte_nolen (req->sv1); + req->sv2 = data; + req->ptr1 = SvPVbyte_nolen (data); + + REQ_SEND; +} + +void +aio_sendfile (SV *out_fh, SV *in_fh, UV in_offset, UV length, SV *callback=&PL_sv_undef) PROTOTYPE: $$$$;$ PPCODE: { dREQ; req->type = REQ_SENDFILE; - req->fh = newSVsv (out_fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (out_fh))); - req->fh2 = newSVsv (in_fh); - req->fd2 = PerlIO_fileno (IoIFP (sv_2io (in_fh))); - req->offset = in_offset; - req->length = length; + req->sv1 = newSVsv (out_fh); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (out_fh))); + req->sv2 = newSVsv (in_fh); + req->int2 = PerlIO_fileno (IoIFP (sv_2io (in_fh))); + req->offs = in_offset; + req->size = length; REQ_SEND; } void -aio_readahead (fh,offset,length,callback=&PL_sv_undef) - SV * fh - UV offset - IV length - SV * callback +aio_readahead (SV *fh, UV offset, IV length, SV *callback=&PL_sv_undef) PROTOTYPE: $$$;$ PPCODE: { dREQ; req->type = REQ_READAHEAD; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); - req->offset = offset; - req->length = length; + req->sv1 = newSVsv (fh); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh))); + req->offs = offset; + req->size = length; REQ_SEND; } void -aio_stat (fh_or_path,callback=&PL_sv_undef) - SV * fh_or_path - SV * callback +aio_stat (SV8 *fh_or_path, SV *callback=&PL_sv_undef) ALIAS: aio_stat = REQ_STAT aio_lstat = REQ_LSTAT @@ -1271,33 +1532,103 @@ { dREQ; - New (0, req->statdata, 1, Stat_t); - if (!req->statdata) + req->ptr2 = malloc (sizeof (Stat_t)); + if (!req->ptr2) { req_free (req); - croak ("out of memory during aio_req->statdata allocation"); + croak ("out of memory during aio_stat statdata allocation"); } + req->flags |= FLAG_PTR2_FREE; + req->sv1 = newSVsv (fh_or_path); + if (SvPOK (fh_or_path)) { req->type = ix; - req->data = newSVsv (fh_or_path); - req->dataptr = SvPVbyte_nolen (req->data); + req->ptr1 = SvPVbyte_nolen (req->sv1); } else { req->type = REQ_FSTAT; - req->fh = newSVsv (fh_or_path); - req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); + } + + REQ_SEND; +} + +void +aio_utime (SV8 *fh_or_path, SV *atime, SV *mtime, SV *callback=&PL_sv_undef) + PPCODE: +{ + dREQ; + + req->nv1 = SvOK (atime) ? SvNV (atime) : -1.; + req->nv2 = SvOK (mtime) ? SvNV (mtime) : -1.; + req->sv1 = newSVsv (fh_or_path); + + if (SvPOK (fh_or_path)) + { + req->type = REQ_UTIME; + req->ptr1 = SvPVbyte_nolen (req->sv1); + } + else + { + req->type = REQ_FUTIME; + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); } REQ_SEND; } void -aio_unlink (pathname,callback=&PL_sv_undef) - SV * pathname - SV * callback +aio_chmod (SV8 *fh_or_path, int mode, SV *callback=&PL_sv_undef) + PPCODE: +{ + dREQ; + + req->mode = mode; + req->sv1 = newSVsv (fh_or_path); + + if (SvPOK (fh_or_path)) + { + req->type = REQ_CHMOD; + req->ptr1 = SvPVbyte_nolen (req->sv1); + } + else + { + req->type = REQ_FCHMOD; + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); + } + + REQ_SEND; +} + +void +aio_chown (SV8 *fh_or_path, SV *uid, SV *gid, SV *callback=&PL_sv_undef) + PPCODE: +{ + dREQ; + + req->int2 = SvOK (uid) ? SvIV (uid) : -1; + req->int3 = SvOK (gid) ? SvIV (gid) : -1; + req->sv1 = newSVsv (fh_or_path); + + if (SvPOK (fh_or_path)) + { + req->type = REQ_CHOWN; + req->ptr1 = SvPVbyte_nolen (req->sv1); + } + else + { + req->type = REQ_FCHOWN; + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); + } + + REQ_SEND; +} + +void +aio_unlink (SV8 *pathname, SV *callback=&PL_sv_undef) ALIAS: aio_unlink = REQ_UNLINK aio_rmdir = REQ_RMDIR @@ -1307,17 +1638,28 @@ dREQ; req->type = ix; - req->data = newSVsv (pathname); - req->dataptr = SvPVbyte_nolen (req->data); + req->sv1 = newSVsv (pathname); + req->ptr1 = SvPVbyte_nolen (req->sv1); + + REQ_SEND; +} + +void +aio_mkdir (SV8 *pathname, int mode, SV *callback=&PL_sv_undef) + PPCODE: +{ + dREQ; + req->type = REQ_MKDIR; + req->sv1 = newSVsv (pathname); + req->ptr1 = SvPVbyte_nolen (req->sv1); + req->mode = mode; + REQ_SEND; } void -aio_link (oldpath,newpath,callback=&PL_sv_undef) - SV * oldpath - SV * newpath - SV * callback +aio_link (SV8 *oldpath, SV8 *newpath, SV *callback=&PL_sv_undef) ALIAS: aio_link = REQ_LINK aio_symlink = REQ_SYMLINK @@ -1327,46 +1669,56 @@ dREQ; req->type = ix; - req->fh = newSVsv (oldpath); - req->data2ptr = SvPVbyte_nolen (req->fh); - req->data = newSVsv (newpath); - req->dataptr = SvPVbyte_nolen (req->data); + req->sv2 = newSVsv (oldpath); + req->ptr2 = SvPVbyte_nolen (req->sv2); + req->sv1 = newSVsv (newpath); + req->ptr1 = SvPVbyte_nolen (req->sv1); + + REQ_SEND; +} + +void +aio_mknod (SV8 *pathname, int mode, UV dev, SV *callback=&PL_sv_undef) + PPCODE: +{ + dREQ; + + req->type = REQ_MKNOD; + req->sv1 = newSVsv (pathname); + req->ptr1 = SvPVbyte_nolen (req->sv1); + req->mode = (mode_t)mode; + req->offs = dev; REQ_SEND; } void -aio_busy (delay,callback=&PL_sv_undef) - double delay - SV * callback +aio_busy (double delay, SV *callback=&PL_sv_undef) PPCODE: { dREQ; req->type = REQ_BUSY; - req->fd = delay < 0. ? 0 : delay; - req->fd2 = delay < 0. ? 0 : 1000. * (delay - req->fd); + req->nv1 = delay < 0. ? 0. : delay; REQ_SEND; } void -aio_group (callback=&PL_sv_undef) - SV * callback +aio_group (SV *callback=&PL_sv_undef) PROTOTYPE: ;$ PPCODE: { dREQ; req->type = REQ_GROUP; - req_send (req); + req_send (req); XPUSHs (req_sv (req, AIO_GRP_KLASS)); } void -aio_nop (callback=&PL_sv_undef) - SV * callback +aio_nop (SV *callback=&PL_sv_undef) PPCODE: { dREQ; @@ -1376,20 +1728,27 @@ REQ_SEND; } -void -aioreq_pri (int pri = DEFAULT_PRI) - CODE: - if (pri < PRI_MIN) pri = PRI_MIN; - if (pri > PRI_MAX) pri = PRI_MAX; - next_pri = pri + PRI_BIAS; +int +aioreq_pri (int pri = 0) + PROTOTYPE: ;$ + CODE: + RETVAL = next_pri - PRI_BIAS; + if (items > 0) + { + if (pri < PRI_MIN) pri = PRI_MIN; + if (pri > PRI_MAX) pri = PRI_MAX; + next_pri = pri + PRI_BIAS; + } + OUTPUT: + RETVAL void aioreq_nice (int nice = 0) - CODE: - nice = next_pri - nice; - if (nice < PRI_MIN) nice = PRI_MIN; - if (nice > PRI_MAX) nice = PRI_MAX; - next_pri = nice + PRI_BIAS; + CODE: + nice = next_pri - nice; + if (nice < PRI_MIN) nice = PRI_MIN; + if (nice > PRI_MAX) nice = PRI_MAX; + next_pri = nice + PRI_BIAS; void flush () @@ -1398,18 +1757,17 @@ while (nreqs) { poll_wait (); - poll_cb (0); + poll_cb (); } -void +int poll() PROTOTYPE: CODE: - if (nreqs) - { - poll_wait (); - poll_cb (0); - } + poll_wait (); + RETVAL = poll_cb (); + OUTPUT: + RETVAL int poll_fileno() @@ -1423,30 +1781,84 @@ poll_cb(...) PROTOTYPE: CODE: - RETVAL = poll_cb (0); + RETVAL = poll_cb (); OUTPUT: RETVAL +void +poll_wait() + PROTOTYPE: + CODE: + poll_wait (); + +void +setsig (int signum = SIGIO) + PROTOTYPE: ;$ + CODE: +{ + if (block_sig_level) + croak ("cannot call IO::AIO::setsig from within aio_block/callback"); + + LOCK (reslock); + main_tid = pthread_self (); + main_sig = signum; + UNLOCK (reslock); + + if (main_sig && npending) + pthread_kill (main_tid, main_sig); +} + +void +aio_block (SV *cb) + PROTOTYPE: & + PPCODE: +{ + int count; + + block_sig (); + PUSHMARK (SP); + PUTBACK; + count = call_sv (cb, GIMME_V | G_NOARGS | G_EVAL); + SPAGAIN; + unblock_sig (); + + if (SvTRUE (ERRSV)) + croak (0); + + XSRETURN (count); +} + int -poll_some(int max = 0) - PROTOTYPE: $ +nreqs() + PROTOTYPE: CODE: - RETVAL = poll_cb (max); + RETVAL = nreqs; OUTPUT: RETVAL -void -poll_wait() +int +nready() PROTOTYPE: CODE: - if (nreqs) - poll_wait (); + RETVAL = get_nready (); + OUTPUT: + RETVAL int -nreqs() +npending() PROTOTYPE: CODE: - RETVAL = nreqs; + RETVAL = get_npending (); + OUTPUT: + RETVAL + +int +nthreads() + PROTOTYPE: + CODE: + if (WORDACCESS_UNSAFE) LOCK (wrklock); + RETVAL = started; + if (WORDACCESS_UNSAFE) UNLOCK (wrklock); OUTPUT: RETVAL @@ -1474,7 +1886,10 @@ int i; aio_req req; - if (grp->fd == 2) + if (main_sig && !block_sig_level) + croak ("aio_group->add called outside aio_block/callback context while IO::AIO::setsig is in use"); + + if (grp->int1 == 2) croak ("cannot add requests to IO::AIO::GRP after the group finished"); for (i = 1; i < items; ++i ) @@ -1486,7 +1901,7 @@ if (req) { - ++grp->length; + ++grp->size; req->grp = grp; req->grp_prev = 0; @@ -1510,30 +1925,39 @@ CODE: { int i; - AV *av = newAV (); + AV *av; + + grp->errorno = errno; + + av = newAV (); for (i = 1; i < items; ++i ) av_push (av, newSVsv (ST (i))); - SvREFCNT_dec (grp->data); - grp->data = (SV *)av; + SvREFCNT_dec (grp->sv1); + grp->sv1 = (SV *)av; } void +errno (aio_req grp, int errorno = errno) + CODE: + grp->errorno = errorno; + +void limit (aio_req grp, int limit) CODE: - grp->fd2 = limit; + grp->int2 = limit; aio_grp_feed (grp); void feed (aio_req grp, SV *callback=&PL_sv_undef) CODE: { - SvREFCNT_dec (grp->fh2); - grp->fh2 = newSVsv (callback); + SvREFCNT_dec (grp->sv2); + grp->sv2 = newSVsv (callback); - if (grp->fd2 <= 0) - grp->fd2 = 2; + if (grp->int2 <= 0) + grp->int2 = 2; aio_grp_feed (grp); }