--- IO-AIO/AIO.xs 2006/10/23 00:34:36 1.54 +++ IO-AIO/AIO.xs 2008/04/16 16:45:18 1.113 @@ -1,26 +1,71 @@ -#define _REENTRANT 1 +#include "xthread.h" + #include #include "EXTERN.h" #include "perl.h" #include "XSUB.h" -#include "autoconf/config.h" - -#include - #include +#include #include -#include -#include #include #include #include -#include #include -#include #include +#ifdef _WIN32 + +# define SIGIO 0 + typedef Direntry_t X_DIRENT; +#undef malloc +#undef free + +// perl overrides all those nice win32 functions +# undef open +# undef read +# undef write +# undef send +# undef recv +# undef stat +# undef fstat +# define lstat stat +# undef truncate +# undef ftruncate +# undef open +# undef close +# undef unlink +# undef rmdir +# undef rename +# undef lseek + +# define chown(a,b,c) (errno = ENOSYS, -1) +# define fchown(a,b,c) (errno = ENOSYS, -1) +# define fchmod(a,b) (errno = ENOSYS, -1) +# define symlink(a,b) (errno = ENOSYS, -1) +# define readlink(a,b,c) (errno = ENOSYS, -1) +# define mknod(a,b,c) (errno = ENOSYS, -1) +# define truncate(a,b) (errno = ENOSYS, -1) +# define ftruncate(fd,o) chsize ((fd), (o)) +# define fsync(fd) _commit (fd) +# define opendir(fd) (errno = ENOSYS, 0) +# define readdir(fd) (errno = ENOSYS, -1) +# define closedir(fd) (errno = ENOSYS, -1) +# define mkdir(a,b) mkdir (a) + +#else + +# include "autoconf/config.h" +# include +# include +# include +# include +# include + typedef struct dirent X_DIRENT; + +#endif + #if HAVE_SENDFILE # if __linux # include @@ -36,29 +81,52 @@ # 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 -#if __ia64 -# define STACKSIZE 65536 +/* buffer size for various temporary buffers */ +#define AIO_BUFSIZE 65536 + +/* use NV for 32 bit perls as it allows larger offsets */ +#if IVSIZE >= 8 +# define SvVAL64 SvIV #else -# define STACKSIZE 8192 +# define SvVAL64 SvNV #endif +static HV *stash; + +#define dBUF \ + char *aio_buf; \ + X_LOCK (wrklock); \ + self->dbuf = aio_buf = malloc (AIO_BUFSIZE); \ + X_UNLOCK (wrklock); \ + 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_READ, REQ_WRITE, + REQ_READAHEAD, REQ_SENDFILE, REQ_STAT, REQ_LSTAT, REQ_FSTAT, - REQ_FSYNC, REQ_FDATASYNC, - REQ_UNLINK, REQ_RMDIR, REQ_RENAME, - REQ_READDIR, - REQ_LINK, REQ_SYMLINK, + REQ_TRUNCATE, REQ_FTRUNCATE, + REQ_UTIME, REQ_FUTIME, + REQ_CHMOD, REQ_FCHMOD, + REQ_CHOWN, REQ_FCHOWN, + REQ_SYNC, REQ_FSYNC, REQ_FDATASYNC, + REQ_UNLINK, REQ_RMDIR, REQ_MKDIR, REQ_RENAME, + REQ_MKNOD, REQ_READDIR, + REQ_LINK, REQ_SYMLINK, REQ_READLINK, REQ_GROUP, REQ_NOP, - REQ_SLEEP, + REQ_BUSY, }; #define AIO_REQ_KLASS "IO::AIO::REQ" @@ -68,43 +136,245 @@ { struct aio_cb *volatile next; - struct aio_cb *grp, *grp_prev, *grp_next, *grp_first; - - SV *self; /* the perl counterpart of this request, if any */ - - 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 stroffset; int type; - int fd, fd2; + int int1, int2, int3; int errorno; - STRLEN dataoffset; mode_t mode; /* open */ - unsigned char cancelled; + + unsigned char flags; + unsigned char pri; + + SV *self; /* the perl counterpart of this request, if any */ + struct aio_cb *grp, *grp_prev, *grp_next, *grp_first; } aio_cb; +enum { + 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; typedef aio_cb *aio_req_ornot; -static int started, wanted; -static volatile int nreqs; -static int max_outstanding = 1<<30; -static int respipe [2]; - -static pthread_mutex_t reslock = PTHREAD_MUTEX_INITIALIZER; -static pthread_mutex_t reqlock = PTHREAD_MUTEX_INITIALIZER; -static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER; +enum { + PRI_MIN = -4, + PRI_MAX = 4, + + DEFAULT_PRI = 0, + PRI_BIAS = -PRI_MIN, + NUM_PRI = PRI_MAX + PRI_BIAS + 1, +}; + +#define AIO_TICKS ((1000000 + 1023) >> 10) + +static unsigned int max_poll_time = 0; +static unsigned int max_poll_reqs = 0; + +/* 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 (void) +{ + sigset_t ss; -static volatile aio_req reqs, reqe; /* queue start, queue end */ -static volatile aio_req ress, rese; /* queue start, queue end */ + if (block_sig_level++) + return; + + if (!main_sig) + return; + + sigemptyset (&ss); + sigaddset (&ss, main_sig); + pthread_sigmask (SIG_BLOCK, &ss, 0); +} + +void unblock_sig (void) +{ + sigset_t ss; + + if (--block_sig_level) + return; + + if (!main_sig) + return; + + sigemptyset (&ss); + sigaddset (&ss, main_sig); + pthread_sigmask (SIG_UNBLOCK, &ss, 0); +} + +static int next_pri = DEFAULT_PRI + PRI_BIAS; + +static unsigned int started, idle, wanted; + +/* worker threads management */ +static mutex_t wrklock = X_MUTEX_INIT; + +typedef struct worker { + /* locked by wrklock */ + struct worker *prev, *next; + + thread_t tid; + + /* locked by reslock, reqlock or wrklock */ + aio_req req; /* currently processed request */ + void *dbuf; + DIR *dirp; +} worker; + +static worker wrk_first = { &wrk_first, &wrk_first, 0 }; + +static void worker_clear (worker *wrk) +{ + if (wrk->dirp) + { + closedir (wrk->dirp); + wrk->dirp = 0; + } + + if (wrk->dbuf) + { + free (wrk->dbuf); + wrk->dbuf = 0; + } +} + +static void worker_free (worker *wrk) +{ + wrk->next->prev = wrk->prev; + wrk->prev->next = wrk->next; + + free (wrk); +} + +static volatile unsigned int nreqs, nready, npending; +static volatile unsigned int max_idle = 4; +static volatile unsigned int max_outstanding = 0xffffffff; +static int respipe_osf [2], respipe [2] = { -1, -1 }; + +static mutex_t reslock = X_MUTEX_INIT; +static mutex_t reqlock = X_MUTEX_INIT; +static cond_t reqwait = X_COND_INIT; + +#if WORDACCESS_UNSAFE + +static unsigned int get_nready (void) +{ + unsigned int retval; -static void req_invoke (aio_req req); -static void req_free (aio_req req); + X_LOCK (reqlock); + retval = nready; + X_UNLOCK (reqlock); + + return retval; +} + +static unsigned int get_npending (void) +{ + unsigned int retval; + + X_LOCK (reslock); + retval = npending; + X_UNLOCK (reslock); + + return retval; +} + +static unsigned int get_nthreads (void) +{ + unsigned int retval; + + X_LOCK (wrklock); + retval = started; + X_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 + * with 8 priorities this actually needs <20 insns + * per shift, the most expensive operation. + */ +typedef struct { + aio_req qs[NUM_PRI], qe[NUM_PRI]; /* qstart, qend */ + int size; +} reqq; + +static reqq req_queue; +static reqq res_queue; + +int reqq_push (reqq *q, aio_req req) +{ + int pri = req->pri; + req->next = 0; + + if (q->qe[pri]) + { + q->qe[pri]->next = req; + q->qe[pri] = req; + } + else + q->qe[pri] = q->qs[pri] = req; + + return q->size++; +} + +aio_req reqq_shift (reqq *q) +{ + int pri; + + if (!q->size) + return 0; + + --q->size; + + for (pri = NUM_PRI; pri--; ) + { + aio_req req = q->qs[pri]; + + if (req) + { + if (!(q->qs[pri] = req->next)) + q->qe[pri] = 0; + + return req; + } + } + + abort (); +} + +static int poll_cb (void); +static int req_invoke (aio_req req); +static void req_destroy (aio_req req); +static void req_cancel (aio_req req); /* must be called at most once */ static SV *req_sv (aio_req req, const char *klass) @@ -132,11 +402,13 @@ static void aio_grp_feed (aio_req grp) { - while (grp->length < grp->fd2) + 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; @@ -145,150 +417,181 @@ PUSHMARK (SP); XPUSHs (req_sv (grp, AIO_GRP_KLASS)); PUTBACK; - call_sv (grp->fh2, G_VOID | G_EVAL); + 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 () -{ - if (nreqs && !ress) - { - fd_set rfd; - FD_ZERO(&rfd); - FD_SET(respipe [0], &rfd); + if (!req_invoke (grp)) + { + req_destroy (grp); + unblock_sig (); + croak (0); + } - select (respipe [0] + 1, &rfd, 0, 0, 0); + req_destroy (grp); + unblock_sig (); } } -static void req_invoke (aio_req req) +static int req_invoke (aio_req req) { dSP; - int errorno = errno; - - if (req->cancelled || !SvOK (req->callback)) - return; - errno = req->errorno; + if (req->flags & FLAG_SV2_RO_OFF) + SvREADONLY_off (req->sv2); - ENTER; - SAVETMPS; - PUSHMARK (SP); - EXTEND (SP, 1); - - switch (req->type) + if (!(req->flags & FLAG_CANCELLED) && SvOK (req->callback)) { - case REQ_READDIR: - { - SV *rv = &PL_sv_undef; + ENTER; + SAVETMPS; + PUSHMARK (SP); + EXTEND (SP, 1); - if (req->result >= 0) + switch (req->type) + { + case REQ_READDIR: { - char *buf = req->data2ptr; - AV *av = newAV (); + SV *rv = &PL_sv_undef; - while (req->result) + if (req->result >= 0) { - SV *sv = newSVpv (buf, 0); + int i; + char *buf = req->ptr2; + AV *av = newAV (); + + av_extend (av, req->result - 1); + + for (i = 0; i < req->result; ++i) + { + SV *sv = newSVpv (buf, 0); - av_push (av, sv); - buf += SvCUR (sv) + 1; - req->result--; + av_store (av, i, sv); + buf += SvCUR (sv) + 1; + } + + rv = sv_2mortal (newRV_noinc ((SV *)av)); } - rv = sv_2mortal (newRV_noinc ((SV *)av)); + PUSHs (rv); } + break; - PUSHs (rv); - } - break; + case REQ_OPEN: + { + /* convert fd to fh */ + SV *fh = &PL_sv_undef; - case REQ_OPEN: - { - /* convert fd to fh */ - SV *fh; + if (req->result >= 0) + { + GV *gv = (GV *)sv_newmortal (); + int flags = req->int1 & (O_RDONLY | O_WRONLY | O_RDWR); + char sym [64]; + int symlen; + + symlen = snprintf (sym, sizeof (sym), "fd#%d", req->result); + gv_init (gv, stash, sym, symlen, 0); + + symlen = snprintf ( + sym, + sizeof (sym), + "%s&=%d", + flags == O_RDONLY ? "<" : flags == O_WRONLY ? ">" : "+<", + req->result + ); - PUSHs (sv_2mortal (newSViv (req->result))); - PUTBACK; - call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); - SPAGAIN; + if (do_open (gv, sym, symlen, 0, 0, 0, 0)) + fh = (SV *)gv; + } - fh = SvREFCNT_inc (POPs); + PUSHs (fh); + } + break; - PUSHMARK (SP); - XPUSHs (sv_2mortal (fh)); - } - break; + case REQ_GROUP: + req->int1 = 2; /* mark group as finished */ - case REQ_GROUP: - req->fd = 2; /* mark group as finished */ + if (req->sv1) + { + int i; + AV *av = (AV *)req->sv1; - if (req->data) - { - int i; - AV *av = (AV *)req->data; + EXTEND (SP, AvFILL (av) + 1); + for (i = 0; i <= AvFILL (av); ++i) + PUSHs (*av_fetch (av, i, 0)); + } + break; - EXTEND (SP, AvFILL (av) + 1); - for (i = 0; i <= AvFILL (av); ++i) - PUSHs (*av_fetch (av, i, 0)); - } - break; + case REQ_NOP: + case REQ_BUSY: + break; - case REQ_NOP: - case REQ_SLEEP: - break; + case REQ_READLINK: + if (req->result > 0) + { + SvCUR_set (req->sv2, req->result); + *SvEND (req->sv2) = 0; + PUSHs (req->sv2); + } + break; - default: - PUSHs (sv_2mortal (newSViv (req->result))); - 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; - PUTBACK; - call_sv (req->callback, G_VOID | G_EVAL); - SPAGAIN; + default: + PUSHs (sv_2mortal (newSViv (req->result))); + break; + } - FREETMPS; - LEAVE; + errno = req->errorno; - errno = errorno; + PUTBACK; + call_sv (req->callback, G_VOID | G_EVAL | G_DISCARD); + SPAGAIN; - if (SvTRUE (ERRSV)) - { - req_free (req); - croak (0); + FREETMPS; + LEAVE; + + PUTBACK; } -} -static void req_free (aio_req req) -{ if (req->grp) { aio_req grp = req->grp; @@ -303,166 +606,170 @@ aio_grp_dec (grp); } + return !SvTRUE (ERRSV); +} + +static void req_destroy (aio_req req) +{ if (req->self) { sv_unmagic (req->self, PERL_MAGIC_ext); 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 && req->result >= 0) - free (req->data2ptr); + if (req->flags & FLAG_PTR2_FREE) + free (req->ptr2); Safefree (req); } -static void req_cancel (aio_req req) +static void req_cancel_subs (aio_req grp) { - req->cancelled = 1; + aio_req sub; - if (req->type == REQ_GROUP) - { - aio_req sub; + if (grp->type != REQ_GROUP) + return; - for (sub = req->grp_first; sub; sub = sub->grp_next) - req_cancel (sub); - } + SvREFCNT_dec (grp->sv2); + grp->sv2 = 0; + + for (sub = grp->grp_first; sub; sub = sub->grp_next) + req_cancel (sub); } -static int poll_cb () +static void req_cancel (aio_req req) { - dSP; - int count = 0; - int do_croak = 0; - aio_req req; + req->flags |= FLAG_CANCELLED; - for (;;) - { - pthread_mutex_lock (&reslock); - req = ress; + req_cancel_subs (req); +} - if (req) - { - ress = req->next; +#ifdef USE_SOCKETS_AS_HANDLES +# define TO_SOCKET(x) (win32_get_osfhandle (x)) +#else +# define TO_SOCKET(x) (x) +#endif - if (!ress) - { - /* read any signals sent by the worker threads */ - char buf [32]; - while (read (respipe [0], buf, 32) == 32) - ; +static void +create_respipe (void) +{ + int old_readfd = respipe [0]; - rese = 0; - } - } + if (respipe [1] >= 0) + respipe_close (TO_SOCKET (respipe [1])); - pthread_mutex_unlock (&reslock); +#ifdef _WIN32 + if (PerlSock_socketpair (AF_UNIX, SOCK_STREAM, 0, respipe)) +#else + if (pipe (respipe)) +#endif + croak ("unable to initialize result pipe"); - if (!req) - break; + if (old_readfd >= 0) + { + if (dup2 (TO_SOCKET (respipe [0]), TO_SOCKET (old_readfd)) < 0) + croak ("unable to initialize result pipe(2)"); - --nreqs; + respipe_close (respipe [0]); + respipe [0] = old_readfd; + } - 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)); +#ifdef _WIN32 + int arg = 1; + if (ioctlsocket (TO_SOCKET (respipe [0]), FIONBIO, &arg) + || ioctlsocket (TO_SOCKET (respipe [1]), FIONBIO, &arg)) +#else + if (fcntl (respipe [0], F_SETFL, O_NONBLOCK) + || fcntl (respipe [1], F_SETFL, O_NONBLOCK)) +#endif + croak ("unable to initialize result pipe(3)"); - if (req->data2ptr && (req->type == REQ_READ || req->type == REQ_WRITE)) - SvREADONLY_off (req->data); + respipe_osf [0] = TO_SOCKET (respipe [0]); + respipe_osf [1] = TO_SOCKET (respipe [1]); +} - if (req->statdata) - { - PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; - PL_laststatval = req->result; - PL_statcache = *(req->statdata); - } +X_THREAD_PROC (aio_proc); + +static void start_thread (void) +{ + worker *wrk = calloc (1, sizeof (worker)); - req_invoke (req); + if (!wrk) + croak ("unable to allocate worker thread data"); - count++; - } + X_LOCK (wrklock); - req_free (req); + 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; } + else + free (wrk); - return count; + X_UNLOCK (wrklock); } -static void *aio_proc(void *arg); - -static void start_thread (void) +static void maybe_start_thread (void) { - sigset_t fullsigset, oldsigset; - pthread_t tid; - pthread_attr_t attr; - - pthread_attr_init (&attr); - pthread_attr_setstacksize (&attr, STACKSIZE); - pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); - - sigfillset (&fullsigset); - sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset); - - if (pthread_create (&tid, &attr, aio_proc, 0) == 0) - started++; + 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; - sigprocmask (SIG_SETMASK, &oldsigset, 0); + start_thread (); } static void req_send (aio_req req) { - while (started < wanted && nreqs >= started) - start_thread (); + block_sig (); ++nreqs; - pthread_mutex_lock (&reqlock); - - req->next = 0; - - if (reqe) - { - reqe->next = req; - reqe = req; - } - else - reqe = reqs = req; + X_LOCK (reqlock); + ++nready; + reqq_push (&req_queue, req); + X_COND_SIGNAL (reqwait); + X_UNLOCK (reqlock); - pthread_cond_signal (&reqwait); - pthread_mutex_unlock (&reqlock); + unblock_sig (); - if (nreqs > max_outstanding) - for (;;) - { - poll_cb (); - - if (nreqs <= max_outstanding) - break; - - poll_wait (); - } + maybe_start_thread (); } static void end_thread (void) { aio_req req; + Newz (0, req, 1, aio_cb); + req->type = REQ_QUIT; + req->pri = PRI_MAX + PRI_BIAS; - req_send (req); + X_LOCK (reqlock); + reqq_push (&req_queue, req); + X_COND_SIGNAL (reqwait); + X_UNLOCK (reqlock); + + X_LOCK (wrklock); + --started; + X_UNLOCK (wrklock); +} + +static void set_max_idle (int nthreads) +{ + if (WORDACCESS_UNSAFE) X_LOCK (reqlock); + max_idle = nthreads <= 0 ? 1 : nthreads; + if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); } static void min_parallel (int nthreads) @@ -473,34 +780,120 @@ static void max_parallel (int nthreads) { - int cur = started; - if (wanted > nthreads) wanted = nthreads; - while (cur > wanted) - { - end_thread (); - cur--; - } - while (started > wanted) + end_thread (); +} + +static void poll_wait (void) +{ + fd_set rfd; + + while (nreqs) { - poll_wait (); - poll_cb (); + int size; + if (WORDACCESS_UNSAFE) X_LOCK (reslock); + size = res_queue.size; + if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); + + if (size) + return; + + maybe_start_thread (); + + FD_ZERO (&rfd); + FD_SET (respipe [0], &rfd); + + PerlSock_select (respipe [0] + 1, &rfd, 0, 0, 0); } } -static void create_pipe () +static int poll_cb (void) { - if (pipe (respipe)) - croak ("unable to initialize result pipe"); + 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 (); + + X_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 (respipe_read (respipe [0], buf, 4) == 4) + ; + } + } + + X_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_destroy (req); + unblock_sig (); + croak (0); + } + + count++; + } + + req_destroy (req); + + if (maxreqs && !--maxreqs) + break; + + if (max_poll_time) + { + gettimeofday (&tv_now, 0); - if (fcntl (respipe [0], F_SETFL, O_NONBLOCK)) - croak ("cannot set result pipe to nonblocking mode"); + if (tvdiff (&tv_start, &tv_now) >= max_poll_time) + break; + } + } + + if (nreqs <= max_outstanding) + break; - if (fcntl (respipe [1], F_SETFL, O_NONBLOCK)) - croak ("cannot set result pipe to nonblocking mode"); + poll_wait (); + + ++maxreqs; + } + + unblock_sig (); + return count; } /*****************************************************************************/ @@ -515,19 +908,19 @@ * 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 = X_MUTEX_INIT; static ssize_t pread (int fd, void *buf, size_t count, off_t offset) { ssize_t res; off_t ooffset; - pthread_mutex_lock (&preadwritelock); + X_LOCK (preadwritelock); ooffset = lseek (fd, 0, SEEK_CUR); lseek (fd, offset, SEEK_SET); res = read (fd, buf, count); lseek (fd, ooffset, SEEK_SET); - pthread_mutex_unlock (&preadwritelock); + X_UNLOCK (preadwritelock); return res; } @@ -537,52 +930,83 @@ ssize_t res; off_t ooffset; - pthread_mutex_lock (&preadwritelock); + X_LOCK (preadwritelock); ooffset = lseek (fd, 0, SEEK_CUR); lseek (fd, offset, SEEK_SET); res = write (fd, buf, count); lseek (fd, offset, SEEK_SET); - pthread_mutex_unlock (&preadwritelock); + X_UNLOCK (preadwritelock); return res; } #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 #if !HAVE_READAHEAD -# define readahead aio_readahead +# define readahead(fd,offset,count) aio_readahead (fd, offset, count, self) -static ssize_t readahead (int fd, off_t offset, size_t count) +static ssize_t aio_readahead (int fd, off_t offset, size_t count, worker *self) { - char readahead_buf[4096]; + size_t todo = count; + dBUF; - while (count > 0) + while (todo > 0) { - size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf); + size_t len = todo < AIO_BUFSIZE ? todo : AIO_BUFSIZE; - pread (fd, readahead_buf, len, offset); + pread (fd, aio_buf, len, offset); offset += len; - count -= len; + todo -= len; } errno = 0; + return count; } + #endif #if !HAVE_READDIR_R # define readdir_r aio_readdir_r -static pthread_mutex_t readdirlock = PTHREAD_MUTEX_INITIALIZER; +static mutex_t readdirlock = X_MUTEX_INIT; -static int readdir_r (DIR *dirp, struct dirent *ent, struct dirent **res) +static int readdir_r (DIR *dirp, X_DIRENT *ent, X_DIRENT **res) { - struct dirent *e; + X_DIRENT *e; int errorno; - pthread_mutex_lock (&readdirlock); + X_LOCK (readdirlock); e = readdir (dirp); errorno = errno; @@ -595,7 +1019,7 @@ else *res = 0; - pthread_mutex_unlock (&readdirlock); + X_UNLOCK (readdirlock); errno = errorno; return e ? 0 : -1; @@ -603,7 +1027,7 @@ #endif /* sendfile always needs emulation */ -static ssize_t sendfile_ (int ofd, int ifd, off_t offset, size_t count) +static ssize_t sendfile_ (int ofd, int ifd, off_t offset, size_t count, worker *self) { ssize_t res; @@ -624,7 +1048,7 @@ res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); if (res < 0 && sbytes) - /* maybe only on EAGAIN only: as usual, the manpage leaves you guessing */ + /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ res = sbytes; } @@ -662,14 +1086,15 @@ ) { /* emulate sendfile. this is a major pain in the ass */ - char buf[4096]; + dBUF; + res = 0; while (count) { ssize_t cnt; - cnt = pread (ifd, buf, count > 4096 ? 4096 : count, offset); + cnt = pread (ifd, aio_buf, count > AIO_BUFSIZE ? AIO_BUFSIZE : count, offset); if (cnt <= 0) { @@ -677,7 +1102,7 @@ break; } - cnt = write (ofd, buf, cnt); + cnt = write (ofd, aio_buf, cnt); if (cnt <= 0) { @@ -695,166 +1120,264 @@ } /* read a full directory */ -static int scandir_ (const char *path, void **namesp) +static void scandir_ (aio_req req, worker *self) { - DIR *dirp = opendir (path); + DIR *dirp; union { - struct dirent d; - char b [offsetof (struct dirent, d_name) + NAME_MAX + 1]; - } u; - struct dirent *entp; + X_DIRENT d; + char b [offsetof (X_DIRENT, d_name) + NAME_MAX + 1]; + } *u; + X_DIRENT *entp; char *name, *names; int memlen = 4096; int memofs = 0; int res = 0; - int errorno; - if (!dirp) - return -1; + X_LOCK (wrklock); + self->dirp = dirp = opendir (req->ptr1); + self->dbuf = u = malloc (sizeof (*u)); + req->flags |= FLAG_PTR2_FREE; + req->ptr2 = names = malloc (memlen); + X_UNLOCK (wrklock); - names = malloc (memlen); + if (dirp && u && names) + for (;;) + { + errno = 0; + readdir_r (dirp, &u->d, &entp); - for (;;) - { - errno = 0, readdir_r (dirp, &u.d, &entp); + if (!entp) + break; - if (!entp) - break; + name = entp->d_name; - name = entp->d_name; + if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) + { + int len = strlen (name) + 1; - if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) - { - int len = strlen (name) + 1; + res++; - res++; + while (memofs + len > memlen) + { + memlen *= 2; + X_LOCK (wrklock); + req->ptr2 = names = realloc (names, memlen); + X_UNLOCK (wrklock); - while (memofs + len > memlen) - { - memlen *= 2; - names = realloc (names, memlen); - if (!names) - break; - } + if (!names) + break; + } - memcpy (names + memofs, name, len); - memofs += len; - } - } + memcpy (names + memofs, name, len); + memofs += len; + } + } - errorno = errno; - closedir (dirp); + if (errno) + res = -1; + + req->result = res; +} + +static int +aio_close (int fd) +{ + static int close_pipe = -1; /* dummy fd to close fds via dup2 */ - if (errorno) + X_LOCK (wrklock); + + if (close_pipe < 0) { - free (names); - errno = errorno; - res = -1; + int pipefd [2]; + + if (pipe (pipefd) < 0 + || close (pipefd [1]) < 0 + || fcntl (pipefd [0], F_SETFD, FD_CLOEXEC) < 0) + { + X_UNLOCK (wrklock); + return -1; + } + + close_pipe = pipefd [0]; } - *namesp = (void *)names; - return res; + X_UNLOCK (wrklock); + + return dup2 (close_pipe, fd) < 0 ? -1 : 0; } /*****************************************************************************/ -static void *aio_proc (void *thr_arg) +X_THREAD_PROC (aio_proc) { aio_req req; - int type; + struct timespec ts; + worker *self = (worker *)thr_arg; + + /* try to distribute timeouts somewhat randomly */ + ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL); - do + for (;;) { - pthread_mutex_lock (&reqlock); + ts.tv_sec = time (0) + IDLE_TIMEOUT; + + X_LOCK (reqlock); for (;;) { - req = reqs; - - if (reqs) - { - reqs = reqs->next; - if (!reqs) reqe = 0; - } + self->req = req = reqq_shift (&req_queue); if (req) break; - pthread_cond_wait (&reqwait, &reqlock); + ++idle; + + if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) + == ETIMEDOUT) + { + if (idle > max_idle) + { + --idle; + X_UNLOCK (reqlock); + X_LOCK (wrklock); + --started; + X_UNLOCK (wrklock); + goto quit; + } + + /* we are allowed to idle, so do so without any timeout */ + X_COND_WAIT (reqwait, reqlock); + ts.tv_sec = time (0) + IDLE_TIMEOUT; + } + + --idle; } - pthread_mutex_unlock (&reqlock); + --nready; + + X_UNLOCK (reqlock); errno = 0; /* strictly unnecessary */ - if (!req->cancelled) - switch (type = req->type) /* remember type for QUIT check */ + if (!(req->flags & FLAG_CANCELLED)) + 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_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); 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_FDATASYNC: req->result = fdatasync (req->fd); break; - case REQ_FSYNC: req->result = fsync (req->fd); break; - case REQ_READDIR: req->result = scandir_ (req->dataptr, &req->data2ptr); break; - - case REQ_SLEEP: + case REQ_READ: req->result = req->offs >= 0 + ? pread (req->int1, req->ptr1, req->size, req->offs) + : read (req->int1, req->ptr1, req->size); break; + case REQ_WRITE: req->result = req->offs >= 0 + ? pwrite (req->int1, req->ptr1, req->size, req->offs) + : write (req->int1, req->ptr1, req->size); 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->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_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; + case REQ_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break; + + case REQ_OPEN: req->result = open (req->ptr1, req->int1, req->mode); break; + case REQ_CLOSE: req->result = aio_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_SYNC: req->result = 0; sync (); break; + case REQ_FSYNC: req->result = fsync (req->int1); break; + case REQ_FDATASYNC: req->result = fdatasync (req->int1); break; + + case REQ_READDIR: scandir_ (req, self); break; + + case REQ_BUSY: +#ifdef _WIN32 + Sleep (req->nv1 * 1000.); +#else { 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_sec) * 1000000.; req->result = select (0, 0, 0, 0, &tv); } +#endif + break; - case REQ_GROUP: /*TODO: should not be handled here */ - case REQ_NOP: /*TODO: should not be handled here */ - case REQ_QUIT: + 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: break; + case REQ_QUIT: + goto quit; + default: - req->result = ENOSYS; + req->result = -1; break; } req->errorno = errno; - pthread_mutex_lock (&reslock); + X_LOCK (reslock); - req->next = 0; + ++npending; - if (rese) - { - rese->next = req; - rese = req; - } - else + if (!reqq_push (&res_queue, req)) { - rese = ress = req; - /* write a dummy byte to the pipe so fh becomes ready */ - write (respipe [1], &respipe, 1); + respipe_write (respipe_osf [1], (const void *)&respipe_osf, 1); + + /* optionally signal the main thread asynchronously */ + if (main_sig) + pthread_kill (main_tid, main_sig); } - pthread_mutex_unlock (&reslock); + self->req = 0; + worker_clear (self); + + X_UNLOCK (reslock); } - while (type != REQ_QUIT); + +quit: + X_LOCK (wrklock); + worker_free (self); + X_UNLOCK (wrklock); return 0; } @@ -863,61 +1386,66 @@ static void atfork_prepare (void) { - pthread_mutex_lock (&reqlock); - pthread_mutex_lock (&reslock); + X_LOCK (wrklock); + X_LOCK (reqlock); + X_LOCK (reslock); #if !HAVE_PREADWRITE - pthread_mutex_lock (&preadwritelock); + X_LOCK (preadwritelock); #endif #if !HAVE_READDIR_R - pthread_mutex_lock (&readdirlock); + X_LOCK (readdirlock); #endif } static void atfork_parent (void) { #if !HAVE_READDIR_R - pthread_mutex_unlock (&readdirlock); + X_UNLOCK (readdirlock); #endif #if !HAVE_PREADWRITE - pthread_mutex_unlock (&preadwritelock); + X_UNLOCK (preadwritelock); #endif - pthread_mutex_unlock (&reslock); - pthread_mutex_unlock (&reqlock); + X_UNLOCK (reslock); + X_UNLOCK (reqlock); + X_UNLOCK (wrklock); } static void atfork_child (void) { aio_req prv; - started = 0; + while (prv = reqq_shift (&req_queue)) + req_destroy (prv); - while (reqs) - { - prv = reqs; - reqs = prv->next; - req_free (prv); - } + while (prv = reqq_shift (&res_queue)) + req_destroy (prv); - reqs = reqe = 0; - - while (ress) + while (wrk_first.next != &wrk_first) { - prv = ress; - ress = prv->next; - req_free (prv); + worker *wrk = wrk_first.next; + + if (wrk->req) + req_destroy (wrk->req); + + worker_clear (wrk); + worker_free (wrk); } - - ress = rese = 0; - close (respipe [0]); - close (respipe [1]); - create_pipe (); + started = 0; + idle = 0; + nreqs = 0; + nready = 0; + npending = 0; + + create_respipe (); atfork_parent (); } #define dREQ \ aio_req req; \ + int req_pri = next_pri; \ + next_pri = DEFAULT_PRI + PRI_BIAS; \ \ if (SvOK (callback) && !SvROK (callback)) \ croak ("callback must be undef or of reference type"); \ @@ -926,7 +1454,8 @@ if (!req) \ croak ("out of memory during aio_req allocation"); \ \ - req->callback = newSVsv (callback) + req->callback = newSVsv (callback); \ + req->pri = req_pri #define REQ_SEND \ req_send (req); \ @@ -940,60 +1469,87 @@ BOOT: { - HV *stash = gv_stashpv ("IO::AIO", 1); + 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)); +#ifdef _WIN32 + X_MUTEX_CHECK (wrklock); + X_MUTEX_CHECK (reslock); + X_MUTEX_CHECK (reqlock); + X_MUTEX_CHECK (reqwait); + X_MUTEX_CHECK (preadwritelock); + X_MUTEX_CHECK (readdirlock); + + X_COND_CHECK (reqwait); +#else + newCONSTSUB (stash, "S_IFIFO", newSViv (S_IFIFO)); + newCONSTSUB (stash, "SIGIO", newSViv (SIGIO)); +#endif - create_pipe (); - pthread_atfork (atfork_prepare, atfork_parent, atfork_child); + create_respipe (); + + X_THREAD_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 +max_idle (int nthreads) + PROTOTYPE: $ + CODE: + set_max_idle (nthreads); + int -max_outstanding (nreqs) - int nreqs - PROTOTYPE: $ +max_outstanding (int maxreqs) + PROTOTYPE: $ CODE: RETVAL = max_outstanding; - max_outstanding = nreqs; + max_outstanding = maxreqs; + OUTPUT: + RETVAL void -aio_open (pathname,flags,mode,callback=&PL_sv_undef) - SV * pathname - int flags - int mode - SV * callback +aio_open (SV8 *pathname, int flags, int mode, SV *callback=&PL_sv_undef) PROTOTYPE: $$$;$ PPCODE: { dREQ; req->type = REQ_OPEN; - req->data = newSVsv (pathname); - req->dataptr = SvPVbyte_nolen (req->data); - req->fd = flags; + req->sv1 = newSVsv (pathname); + req->ptr1 = SvPVbyte_nolen (req->sv1); + req->int1 = flags; req->mode = mode; REQ_SEND; } void -aio_close (fh,callback=&PL_sv_undef) - SV * fh - SV * callback +aio_fsync (SV *fh, SV *callback=&PL_sv_undef) PROTOTYPE: $;$ ALIAS: - aio_close = REQ_CLOSE aio_fsync = REQ_FSYNC aio_fdatasync = REQ_FDATASYNC PPCODE: @@ -1001,29 +1557,36 @@ dREQ; req->type = ix; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); + req->sv1 = newSVsv (fh); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh))); REQ_SEND (req); } void -aio_read (fh,offset,length,data,dataoffset,callback=&PL_sv_undef) - SV * fh - UV offset - UV length - SV * data - UV dataoffset - SV * callback +aio_close (SV *fh, SV *callback=&PL_sv_undef) + PROTOTYPE: $;$ + PPCODE: +{ + dREQ; + + req->type = REQ_CLOSE; + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh))); + + REQ_SEND (req); +} + +void +aio_read (SV *fh, SV *offset, SV *length, SV8 *data, IV dataoffset, SV *callback=&PL_sv_undef) ALIAS: aio_read = REQ_READ aio_write = REQ_WRITE PROTOTYPE: $$$$$;$ PPCODE: { - aio_req req; STRLEN svlen; char *svptr = SvPVbyte (data, svlen); + UV len = SvUV (length); SvUPGRADE (data, SVt_PV); SvPOK_on (data); @@ -1032,39 +1595,40 @@ dataoffset += svlen; if (dataoffset < 0 || dataoffset > svlen) - croak ("data offset outside of string"); + croak ("dataoffset outside of data scalar"); if (ix == REQ_WRITE) { /* write: check length and adjust. */ - if (length < 0 || length + dataoffset > svlen) - length = svlen - dataoffset; + if (!SvOK (length) || len + dataoffset > svlen) + len = svlen - dataoffset; } else { /* read: grow scalar as necessary */ - svptr = SvGROW (data, length + dataoffset); + svptr = SvGROW (data, len + dataoffset + 1); } - if (length < 0) + if (len < 0) croak ("length must not be negative"); { dREQ; req->type = ix; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) - : IoOFP (sv_2io (fh))); - req->offset = offset; - req->length = length; - req->data = SvREFCNT_inc (data); - req->dataptr = (char *)svptr + dataoffset; + req->sv1 = newSVsv (fh); + req->int1 = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) + : IoOFP (sv_2io (fh))); + req->offs = SvOK (offset) ? SvVAL64 (offset) : -1; + req->size = len; + req->sv2 = SvREFCNT_inc (data); + req->ptr1 = (char *)svptr + dataoffset; + req->stroffset = dataoffset; if (!SvREADONLY (data)) { SvREADONLY_on (data); - req->data2ptr = (void *)data; + req->flags |= FLAG_SV2_RO_OFF; } REQ_SEND; @@ -1072,52 +1636,61 @@ } void -aio_sendfile (out_fh,in_fh,in_offset,length,callback=&PL_sv_undef) - SV * out_fh - SV * in_fh - UV in_offset - UV length - SV * callback +aio_readlink (SV8 *path, SV *callback=&PL_sv_undef) + PROTOTYPE: $$;$ + 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, SV *in_offset, UV length, SV *callback=&PL_sv_undef) PROTOTYPE: $$$$;$ PPCODE: { dREQ; req->type = REQ_SENDFILE; - req->fh = newSVsv (out_fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (out_fh))); - req->fh2 = newSVsv (in_fh); - req->fd2 = PerlIO_fileno (IoIFP (sv_2io (in_fh))); - req->offset = in_offset; - req->length = length; + req->sv1 = newSVsv (out_fh); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (out_fh))); + req->sv2 = newSVsv (in_fh); + req->int2 = PerlIO_fileno (IoIFP (sv_2io (in_fh))); + req->offs = SvVAL64 (in_offset); + req->size = length; REQ_SEND; } void -aio_readahead (fh,offset,length,callback=&PL_sv_undef) - SV * fh - UV offset - IV length - SV * callback +aio_readahead (SV *fh, SV *offset, IV length, SV *callback=&PL_sv_undef) PROTOTYPE: $$$;$ 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 = SvVAL64 (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 @@ -1125,33 +1698,126 @@ { 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"); + req_destroy (req); + 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_unlink (pathname,callback=&PL_sv_undef) - SV * pathname - SV * callback +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_truncate (SV8 *fh_or_path, SV *offset, SV *callback=&PL_sv_undef) + PPCODE: +{ + dREQ; + + req->sv1 = newSVsv (fh_or_path); + req->offs = SvOK (offset) ? SvVAL64 (offset) : -1; + + if (SvPOK (fh_or_path)) + { + req->type = REQ_TRUNCATE; + req->ptr1 = SvPVbyte_nolen (req->sv1); + } + else + { + req->type = REQ_FTRUNCATE; + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); + } + + REQ_SEND; +} + +void +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 @@ -1161,17 +1827,28 @@ 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 @@ -1181,53 +1858,90 @@ 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_sleep (delay,callback=&PL_sv_undef) - double delay - SV * callback +aio_busy (double delay, SV *callback=&PL_sv_undef) PPCODE: { dREQ; - req->type = REQ_SLEEP; - req->fd = delay < 0. ? 0 : delay; - req->fd2 = delay < 0. ? 0 : 1000. * (delay - req->fd); + req->type = REQ_BUSY; + 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); XPUSHs (req_sv (req, AIO_GRP_KLASS)); } void -aio_nop (callback=&PL_sv_undef) - SV * callback +aio_nop (SV *callback=&PL_sv_undef) + ALIAS: + aio_nop = REQ_NOP + aio_sync = REQ_SYNC PPCODE: { dREQ; - req->type = REQ_NOP; + req->type = ix; REQ_SEND; } +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; + void flush () PROTOTYPE: @@ -1238,15 +1952,14 @@ poll_cb (); } -void +int poll() PROTOTYPE: CODE: - if (nreqs) - { - poll_wait (); - poll_cb (); - } + poll_wait (); + RETVAL = poll_cb (); + OUTPUT: + RETVAL int poll_fileno() @@ -1268,8 +1981,43 @@ poll_wait() PROTOTYPE: CODE: - if (nreqs) - poll_wait (); + poll_wait (); + +void +setsig (int signum = SIGIO) + PROTOTYPE: ;$ + CODE: +{ + if (block_sig_level) + croak ("cannot call IO::AIO::setsig from within aio_block/callback"); + + X_LOCK (reslock); + main_tid = pthread_self (); + main_sig = signum; + X_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); + unblock_sig (); + + if (SvTRUE (ERRSV)) + croak (0); + + XSRETURN (count); +} int nreqs() @@ -1279,16 +2027,47 @@ OUTPUT: RETVAL +int +nready() + PROTOTYPE: + CODE: + RETVAL = get_nready (); + OUTPUT: + RETVAL + +int +npending() + PROTOTYPE: + CODE: + RETVAL = get_npending (); + OUTPUT: + RETVAL + +int +nthreads() + PROTOTYPE: + CODE: + if (WORDACCESS_UNSAFE) X_LOCK (wrklock); + RETVAL = started; + if (WORDACCESS_UNSAFE) X_UNLOCK (wrklock); + OUTPUT: + RETVAL + PROTOTYPES: DISABLE MODULE = IO::AIO PACKAGE = IO::AIO::REQ void cancel (aio_req_ornot req) - PROTOTYPE: CODE: req_cancel (req); +void +cb (aio_req_ornot req, SV *callback=&PL_sv_undef) + CODE: + SvREFCNT_dec (req->callback); + req->callback = newSVsv (callback); + MODULE = IO::AIO PACKAGE = IO::AIO::GRP void @@ -1298,7 +2077,10 @@ 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 ) @@ -1310,7 +2092,7 @@ if (req) { - ++grp->length; + ++grp->size; req->grp = grp; req->grp_prev = 0; @@ -1325,44 +2107,48 @@ } void +cancel_subs (aio_req_ornot req) + CODE: + req_cancel_subs (req); + +void result (aio_req grp, ...) 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 -lock (aio_req grp) - CODE: - ++grp->length; - -void -unlock (aio_req grp) - CODE: - aio_grp_dec (grp); +errno (aio_req grp, int errorno = errno) + CODE: + grp->errorno = errorno; void -feeder_limit (aio_req grp, int limit) +limit (aio_req grp, int limit) CODE: - grp->fd2 = limit; + grp->int2 = limit; aio_grp_feed (grp); void -set_feeder (aio_req grp, SV *callback=&PL_sv_undef) +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); }