--- IO-AIO/AIO.xs 2006/10/21 23:06:04 1.43 +++ IO-AIO/AIO.xs 2006/10/28 23:32:29 1.85 @@ -1,4 +1,13 @@ +/* solaris */ +#define _POSIX_PTHREAD_SEMANTICS 1 + +#if __linux && !defined(_GNU_SOURCE) +# define _GNU_SOURCE +#endif + +/* just in case */ #define _REENTRANT 1 + #include #include "EXTERN.h" @@ -11,6 +20,8 @@ #include #include +#include +#include #include #include #include @@ -34,17 +45,50 @@ # 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 8192 +# define STACKSIZE 16384 #endif +/* wether word reads are potentially non-atomic. + * this is conservatice, likely most arches this runs + * on have atomic word read/writes. + */ +#ifndef WORDACCESS_UNSAFE +# if __i386 || __x86_64 +# define WORDACCESS_UNSAFE 0 +# else +# define WORDACCESS_UNSAFE 1 +# endif +#endif + +/* buffer size for various temporary buffers */ +#define AIO_BUFSIZE 65536 + +#define dBUF \ + char *aio_buf; \ + LOCK (wrklock); \ + self->dbuf = aio_buf = malloc (AIO_BUFSIZE); \ + UNLOCK (wrklock); \ + if (!aio_buf) \ + return -1; + enum { REQ_QUIT, REQ_OPEN, REQ_CLOSE, @@ -53,21 +97,19 @@ REQ_STAT, REQ_LSTAT, REQ_FSTAT, REQ_FSYNC, REQ_FDATASYNC, REQ_UNLINK, REQ_RMDIR, REQ_RENAME, - REQ_READDIR, + REQ_MKNOD, REQ_READDIR, REQ_LINK, REQ_SYMLINK, + REQ_GROUP, REQ_NOP, + REQ_BUSY, }; -#define AIO_CB_KLASS "IO::AIO::CB" +#define AIO_REQ_KLASS "IO::AIO::REQ" +#define AIO_GRP_KLASS "IO::AIO::GRP" typedef struct aio_cb { - struct aio_cb *grp_prev, *grp_next; - struct aio_grp *grp; - struct aio_cb *volatile next; - SV *self; /* the perl counterpart of this request, if any */ - SV *data, *callback; SV *fh, *fh2; void *dataptr, *data2ptr; @@ -76,314 +118,521 @@ size_t length; ssize_t result; + STRLEN dataoffset; int type; int fd, fd2; 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, +}; + 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]; +enum { + PRI_MIN = -4, + PRI_MAX = 4, -static pthread_mutex_t reslock = PTHREAD_MUTEX_INITIALIZER; -static pthread_mutex_t reqlock = PTHREAD_MUTEX_INITIALIZER; -static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER; + DEFAULT_PRI = 0, + PRI_BIAS = -PRI_MIN, + NUM_PRI = PRI_MAX + PRI_BIAS + 1, +}; -static volatile aio_req reqs, reqe; /* queue start, queue end */ -static volatile aio_req ress, rese; /* queue start, queue end */ +#define AIO_TICKS ((1000000 + 1023) >> 10) -typedef struct aio_grp -{ - struct aio_cb *first, *last; - SV *callback; - int busycount; -} aio_grp; +static unsigned int max_poll_time = 0; +static unsigned int max_poll_reqs = 0; -static void aio_grp_begin (aio_grp *grp) +/* calculcate time difference in ~1/AIO_TICKS of a second */ +static int tvdiff (struct timeval *tv1, struct timeval *tv2) { - ++grp->busycount; + return (tv2->tv_sec - tv1->tv_sec ) * AIO_TICKS + + ((tv2->tv_usec - tv1->tv_usec) >> 10); } -static void aio_grp_end (aio_grp *grp) -{ - --grp->busycount; +static int next_pri = DEFAULT_PRI + PRI_BIAS; - if (grp->busycount) - return; +static unsigned int started, idle, wanted; - SvREFCNT_dec (grp->callback); - grp->callback = 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 + +#define LOCK(mutex) pthread_mutex_lock (&(mutex)) +#define UNLOCK(mutex) pthread_mutex_unlock (&(mutex)) + +/* worker threads management */ +static pthread_mutex_t wrklock = AIO_MUTEX_INIT; -static aio_grp *aio_grp_new () +typedef struct worker { + /* locked by wrklock */ + struct worker *prev, *next; + + pthread_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) { - aio_grp *grp; + if (wrk->dirp) + { + closedir (wrk->dirp); + wrk->dirp = 0; + } - Newz (0, grp, 1, aio_grp); - aio_grp_begin (grp); + if (wrk->dbuf) + { + free (wrk->dbuf); + wrk->dbuf = 0; + } +} - return grp; +static void worker_free (worker *wrk) +{ + wrk->next->prev = wrk->prev; + wrk->prev->next = wrk->next; + + free (wrk); } -/* must be called at most once */ -static SV *req_sv (aio_req req) +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 pthread_mutex_t reslock = AIO_MUTEX_INIT; +static pthread_mutex_t reqlock = AIO_MUTEX_INIT; +static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER; + +#if WORDACCESS_UNSAFE + +static unsigned int get_nready () { - req->self = (SV *)newHV (); - sv_magic (req->self, 0, PERL_MAGIC_ext, (char *)req, 0); + unsigned int retval; - return sv_bless (newRV_noinc (req->self), gv_stashpv (AIO_CB_KLASS, 1)); + LOCK (reqlock); + retval = nready; + UNLOCK (reqlock); + + return retval; } -static aio_req SvAIO_REQ (SV *sv) +static unsigned int get_npending () { - if (!sv_derived_from (sv, AIO_CB_KLASS) || !SvROK (sv)) - croak ("object of class " AIO_CB_KLASS " expected"); + unsigned int retval; - MAGIC *mg = mg_find (SvRV (sv), PERL_MAGIC_ext); + LOCK (reslock); + retval = npending; + UNLOCK (reslock); - return mg ? (aio_req)mg->mg_ptr : 0; + return retval; } -static void req_free (aio_req req) +static unsigned int get_nthreads () { - if (req->self) - { - sv_unmagic (req->self, PERL_MAGIC_ext); - SvREFCNT_dec (req->self); - } + unsigned int retval; - if (req->data) - SvREFCNT_dec (req->data); + LOCK (wrklock); + retval = started; + UNLOCK (wrklock); - if (req->fh) - SvREFCNT_dec (req->fh); + return retval; +} - if (req->fh2) - SvREFCNT_dec (req->fh2); +#else - if (req->statdata) - Safefree (req->statdata); +# define get_nready() nready +# define get_npending() npending +# define get_nthreads() started - if (req->callback) - SvREFCNT_dec (req->callback); +#endif - if (req->type == REQ_READDIR && req->result >= 0) - free (req->data2ptr); +/* + * 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; - Safefree (req); -} +static reqq req_queue; +static reqq res_queue; -static void -poll_wait () +int reqq_push (reqq *q, aio_req req) { - if (nreqs && !ress) - { - fd_set rfd; - FD_ZERO(&rfd); - FD_SET(respipe [0], &rfd); + int pri = req->pri; + req->next = 0; - select (respipe [0] + 1, &rfd, 0, 0, 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++; } -static int -poll_cb () +aio_req reqq_shift (reqq *q) { - dSP; - int count = 0; - int do_croak = 0; - aio_req req; + int pri; - for (;;) + if (!q->size) + return 0; + + --q->size; + + for (pri = NUM_PRI; pri--; ) { - pthread_mutex_lock (&reslock); - req = ress; + aio_req req = q->qs[pri]; if (req) { - ress = req->next; - - if (!ress) - { - /* read any signals sent by the worker threads */ - char buf [32]; - while (read (respipe [0], buf, 32) == 32) - ; + if (!(q->qs[pri] = req->next)) + q->qe[pri] = 0; - rese = 0; - } + return req; } + } - pthread_mutex_unlock (&reslock); + abort (); +} - if (!req) - break; +static int poll_cb (); +static void req_invoke (aio_req req); +static void req_free (aio_req req); +static void req_cancel (aio_req req); - nreqs--; +/* must be called at most once */ +static SV *req_sv (aio_req req, const char *klass) +{ + if (!req->self) + { + req->self = (SV *)newHV (); + sv_magic (req->self, 0, PERL_MAGIC_ext, (char *)req, 0); + } - if (req->type == REQ_QUIT) - started--; - else - { - int errorno = errno; - errno = req->errorno; + return sv_2mortal (sv_bless (newRV_inc (req->self), gv_stashpv (klass, 1))); +} - if (req->type == REQ_READ) - SvCUR_set (req->data, req->dataoffset + (req->result > 0 ? req->result : 0)); +static aio_req SvAIO_REQ (SV *sv) +{ + MAGIC *mg; - if (req->data2ptr && (req->type == REQ_READ || req->type == REQ_WRITE)) - SvREADONLY_off (req->data); + if (!sv_derived_from (sv, AIO_REQ_KLASS) || !SvROK (sv)) + croak ("object of class " AIO_REQ_KLASS " expected"); - if (req->statdata) - { - PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; - PL_laststatval = req->result; - PL_statcache = *(req->statdata); - } + mg = mg_find (SvRV (sv), PERL_MAGIC_ext); + + return mg ? (aio_req)mg->mg_ptr : 0; +} + +static void aio_grp_feed (aio_req grp) +{ + while (grp->length < grp->fd2 && !(grp->flags & FLAG_CANCELLED)) + { + int old_len = grp->length; + + if (grp->fh2 && SvOK (grp->fh2)) + { + dSP; ENTER; + SAVETMPS; PUSHMARK (SP); + XPUSHs (req_sv (grp, AIO_GRP_KLASS)); + PUTBACK; + call_sv (grp->fh2, G_VOID | G_EVAL | G_KEEPERR); + SPAGAIN; + FREETMPS; + LEAVE; + } + + /* stop if no progress has been made */ + if (old_len == grp->length) + { + SvREFCNT_dec (grp->fh2); + grp->fh2 = 0; + break; + } + } +} + +static void aio_grp_dec (aio_req grp) +{ + --grp->length; + + /* call feeder, if applicable */ + aio_grp_feed (grp); + + /* finish, if done */ + if (!grp->length && grp->fd) + { + req_invoke (grp); + req_free (grp); + } +} + +static void req_invoke (aio_req req) +{ + dSP; + + if (!(req->flags & FLAG_CANCELLED) && SvOK (req->callback)) + { + ENTER; + SAVETMPS; + PUSHMARK (SP); + EXTEND (SP, 1); - if (req->type == REQ_READDIR) + switch (req->type) + { + case REQ_READDIR: { SV *rv = &PL_sv_undef; if (req->result >= 0) { + int i; char *buf = req->data2ptr; AV *av = newAV (); - while (req->result) + av_extend (av, req->result - 1); + + for (i = 0; i < req->result; ++i) { SV *sv = newSVpv (buf, 0); - av_push (av, sv); + av_store (av, i, sv); buf += SvCUR (sv) + 1; - req->result--; } rv = sv_2mortal (newRV_noinc ((SV *)av)); } - XPUSHs (rv); + PUSHs (rv); } - else - { - XPUSHs (sv_2mortal (newSViv (req->result))); + break; - if (req->type == REQ_OPEN) - { - /* convert fd to fh */ - SV *fh; + case REQ_OPEN: + { + /* convert fd to fh */ + SV *fh; - PUTBACK; - call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); - SPAGAIN; + PUSHs (sv_2mortal (newSViv (req->result))); + PUTBACK; + call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); + SPAGAIN; - fh = SvREFCNT_inc (POPs); + fh = SvREFCNT_inc (POPs); - PUSHMARK (SP); - XPUSHs (sv_2mortal (fh)); - } + PUSHMARK (SP); + XPUSHs (sv_2mortal (fh)); } + break; - if (SvOK (req->callback) && !req->cancelled) - { - PUTBACK; - call_sv (req->callback, G_VOID | G_EVAL); - SPAGAIN; + case REQ_GROUP: + req->fd = 2; /* mark group as finished */ - if (SvTRUE (ERRSV)) - { - req_free (req); - croak (0); - } - } + 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; - LEAVE; + case REQ_NOP: + case REQ_BUSY: + break; - errno = errorno; - count++; + default: + PUSHs (sv_2mortal (newSViv (req->result))); + break; } + errno = req->errorno; + + PUTBACK; + call_sv (req->callback, G_VOID | G_EVAL); + SPAGAIN; + + FREETMPS; + LEAVE; + } + + if (req->grp) + { + aio_req grp = req->grp; + + /* unlink request */ + if (req->grp_next) req->grp_next->grp_prev = req->grp_prev; + if (req->grp_prev) req->grp_prev->grp_next = req->grp_next; + + if (grp->grp_first == req) + grp->grp_first = req->grp_next; + + aio_grp_dec (grp); + } + + if (SvTRUE (ERRSV)) + { req_free (req); + croak (0); } +} - return count; +static void req_free (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->callback); + Safefree (req->statdata); + + if (req->type == REQ_READDIR) + free (req->data2ptr); + + Safefree (req); +} + +static void req_cancel_subs (aio_req grp) +{ + aio_req sub; + + if (grp->type != REQ_GROUP) + return; + + SvREFCNT_dec (grp->fh2); + grp->fh2 = 0; + + for (sub = grp->grp_first; sub; sub = sub->grp_next) + req_cancel (sub); +} + +static void req_cancel (aio_req req) +{ + req->flags |= FLAG_CANCELLED; + + req_cancel_subs (req); } static void *aio_proc(void *arg); -static void -start_thread (void) +static void start_thread (void) { sigset_t fullsigset, oldsigset; - pthread_t tid; 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); +#ifdef PTHREAD_SCOPE_PROCESS + pthread_attr_setscope (&attr, PTHREAD_SCOPE_PROCESS); +#endif sigfillset (&fullsigset); + + LOCK (wrklock); sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset); - if (pthread_create (&tid, &attr, aio_proc, 0) == 0) - started++; + if (pthread_create (&wrk->tid, &attr, aio_proc, (void *)wrk) == 0) + { + wrk->prev = &wrk_first; + wrk->next = wrk_first.next; + wrk_first.next->prev = wrk; + wrk_first.next = wrk; + ++started; + } + else + free (wrk); sigprocmask (SIG_SETMASK, &oldsigset, 0); + UNLOCK (wrklock); } -static void -req_send (aio_req req) +static void maybe_start_thread () { - while (started < wanted && nreqs >= started) - start_thread (); - - nreqs++; - - pthread_mutex_lock (&reqlock); + 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; - req->next = 0; + start_thread (); +} - if (reqe) - { - reqe->next = req; - reqe = req; - } - else - reqe = reqs = req; +static void req_send (aio_req req) +{ + ++nreqs; + LOCK (reqlock); + ++nready; + reqq_push (&req_queue, req); pthread_cond_signal (&reqwait); - pthread_mutex_unlock (&reqlock); + UNLOCK (reqlock); - if (nreqs > max_outstanding) - for (;;) - { - poll_cb (); - - if (nreqs <= max_outstanding) - break; - - poll_wait (); - } + maybe_start_thread (); } -static void -end_thread (void) +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); + LOCK (reqlock); + reqq_push (&req_queue, req); + pthread_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) @@ -394,22 +643,125 @@ 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); + + 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 [32]; + while (read (respipe [0], buf, 32) == 32) + ; + } + } + + UNLOCK (reslock); + + if (!req) + break; + + --nreqs; + + 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); + + 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 (); + + ++maxreqs; } + + return count; } static void create_pipe () @@ -438,34 +790,32 @@ */ static pthread_mutex_t preadwritelock = PTHREAD_MUTEX_INITIALIZER; -static ssize_t -pread (int fd, void *buf, size_t count, off_t offset) +static ssize_t pread (int fd, void *buf, size_t count, off_t offset) { ssize_t res; off_t ooffset; - pthread_mutex_lock (&preadwritelock); + 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); + UNLOCK (preadwritelock); return res; } -static ssize_t -pwrite (int fd, void *buf, size_t count, off_t offset) +static ssize_t pwrite (int fd, void *buf, size_t count, off_t offset) { ssize_t res; off_t ooffset; - pthread_mutex_lock (&preadwritelock); + 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); + UNLOCK (preadwritelock); return res; } @@ -476,24 +826,24 @@ #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]; + dBUF; while (count > 0) { - size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf); + size_t len = count < AIO_BUFSIZE ? count : AIO_BUFSIZE; - pread (fd, readahead_buf, len, offset); + pread (fd, aio_buf, len, offset); offset += len; count -= len; } errno = 0; } + #endif #if !HAVE_READDIR_R @@ -501,13 +851,12 @@ static pthread_mutex_t readdirlock = PTHREAD_MUTEX_INITIALIZER; -static int -readdir_r (DIR *dirp, struct dirent *ent, struct dirent **res) +static int readdir_r (DIR *dirp, struct dirent *ent, struct dirent **res) { struct dirent *e; int errorno; - pthread_mutex_lock (&readdirlock); + LOCK (readdirlock); e = readdir (dirp); errorno = errno; @@ -520,7 +869,7 @@ else *res = 0; - pthread_mutex_unlock (&readdirlock); + UNLOCK (readdirlock); errno = errorno; return e ? 0 : -1; @@ -528,8 +877,7 @@ #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; @@ -550,7 +898,7 @@ 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; } @@ -588,14 +936,15 @@ ) { /* 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) { @@ -603,7 +952,7 @@ break; } - cnt = write (ofd, buf, cnt); + cnt = write (ofd, aio_buf, cnt); if (cnt <= 0) { @@ -621,15 +970,14 @@ } /* 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; + } *u; struct dirent *entp; char *name, *names; int memlen = 4096; @@ -637,93 +985,113 @@ int res = 0; int errorno; - if (!dirp) - return -1; - - names = malloc (memlen); - - for (;;) - { - errno = 0, readdir_r (dirp, &u.d, &entp); + LOCK (wrklock); + self->dirp = dirp = opendir (req->dataptr); + self->dbuf = u = malloc (sizeof (*u)); + req->data2ptr = names = malloc (memlen); + UNLOCK (wrklock); - if (!entp) - break; - - name = entp->d_name; + if (dirp && u && names) + for (;;) + { + errno = 0; + readdir_r (dirp, &u->d, &entp); - if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) - { - int len = strlen (name) + 1; + if (!entp) + break; - res++; + name = entp->d_name; - while (memofs + len > memlen) - { - memlen *= 2; - names = realloc (names, memlen); - if (!names) - break; - } + if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) + { + int len = strlen (name) + 1; - memcpy (names + memofs, name, len); - memofs += len; - } - } + res++; - errorno = errno; - closedir (dirp); + while (memofs + len > memlen) + { + memlen *= 2; + LOCK (wrklock); + req->data2ptr = names = realloc (names, memlen); + UNLOCK (wrklock); + + if (!names) + break; + } - if (errorno) - { - free (names); - errno = errorno; - res = -1; - } + memcpy (names + memofs, name, len); + memofs += len; + } + } - *namesp = (void *)names; - return res; + if (errno) + res = -1; + + req->result = res; } /*****************************************************************************/ -static void * -aio_proc (void *thr_arg) +static void *aio_proc (void *thr_arg) { aio_req req; - int type; + struct timespec ts; + worker *self = (worker *)thr_arg; + + /* try to distribute timeouts somewhat evenly */ + ts.tv_nsec = (((unsigned long)self + (unsigned long)ts.tv_sec) & 1023UL) + * (1000000000UL / 1024UL); - do + for (;;) { - pthread_mutex_lock (&reqlock); + ts.tv_sec = time (0) + IDLE_TIMEOUT; + + 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 (pthread_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 */ + pthread_cond_wait (&reqwait, &reqlock); + ts.tv_sec = time (0) + IDLE_TIMEOUT; + } + + --idle; } - pthread_mutex_unlock (&reqlock); + --nready; + + UNLOCK (reqlock); errno = 0; /* strictly unnecessary */ - if (!req->cancelled) + 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_SENDFILE: req->result = sendfile_ (req->fd, req->fd2, req->offset, req->length, self); break; case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break; case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break; @@ -736,14 +1104,29 @@ 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_MKNOD: req->result = mknod (req->data2ptr, req->mode, (dev_t)req->offset); 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_READDIR: scandir_ (req, self); break; - case REQ_QUIT: + case REQ_BUSY: + { + struct timeval tv; + + tv.tv_sec = req->fd; + tv.tv_usec = req->fd2; + + req->result = select (0, 0, 0, 0, &tv); + } + + case REQ_GROUP: + case REQ_NOP: break; + case REQ_QUIT: + goto quit; + default: req->result = ENOSYS; break; @@ -751,26 +1134,24 @@ req->errorno = errno; - pthread_mutex_lock (&reslock); + LOCK (reslock); - req->next = 0; + ++npending; - if (rese) - { - rese->next = req; - rese = req; - } - else - { - rese = ress = req; + 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); - } + self->req = 0; + worker_clear (self); - pthread_mutex_unlock (&reslock); + UNLOCK (reslock); } - while (type != REQ_QUIT); + +quit: + LOCK (wrklock); + worker_free (self); + UNLOCK (wrklock); return 0; } @@ -779,51 +1160,56 @@ static void atfork_prepare (void) { - pthread_mutex_lock (&reqlock); - pthread_mutex_lock (&reslock); + LOCK (wrklock); + LOCK (reqlock); + LOCK (reslock); #if !HAVE_PREADWRITE - pthread_mutex_lock (&preadwritelock); + LOCK (preadwritelock); #endif #if !HAVE_READDIR_R - pthread_mutex_lock (&readdirlock); + LOCK (readdirlock); #endif } static void atfork_parent (void) { #if !HAVE_READDIR_R - pthread_mutex_unlock (&readdirlock); + UNLOCK (readdirlock); #endif #if !HAVE_PREADWRITE - pthread_mutex_unlock (&preadwritelock); + UNLOCK (preadwritelock); #endif - pthread_mutex_unlock (&reslock); - pthread_mutex_unlock (&reqlock); + UNLOCK (reslock); + UNLOCK (reqlock); + UNLOCK (wrklock); } static void atfork_child (void) { aio_req prv; - started = 0; + while (prv = reqq_shift (&req_queue)) + req_free (prv); - while (reqs) - { - prv = reqs; - reqs = prv->next; - req_free (prv); - } + while (prv = reqq_shift (&res_queue)) + req_free (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_free (wrk->req); + + worker_clear (wrk); + worker_free (wrk); } - - ress = rese = 0; + + started = 0; + idle = 0; + nreqs = 0; + nready = 0; + npending = 0; close (respipe [0]); close (respipe [1]); @@ -834,6 +1220,8 @@ #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"); \ @@ -842,13 +1230,14 @@ 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); \ \ if (GIMME_V != G_VOID) \ - XPUSHs (req_sv (req)); + XPUSHs (req_sv (req, AIO_REQ_KLASS)); MODULE = IO::AIO PACKAGE = IO::AIO @@ -857,31 +1246,52 @@ 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)); create_pipe (); pthread_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 (int nthreads) PROTOTYPE: $ void -max_parallel (nthreads) - int nthreads +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) @@ -1105,28 +1515,87 @@ REQ_SEND; } -#if 0 +void +aio_mknod (pathname,mode,dev,callback=&PL_sv_undef) + SV * pathname + SV * callback + UV mode + UV dev + PPCODE: +{ + dREQ; + + req->type = REQ_MKNOD; + req->data = newSVsv (pathname); + req->dataptr = SvPVbyte_nolen (req->data); + req->mode = (mode_t)mode; + req->offset = dev; + + REQ_SEND; +} -# undocumented, because it does not cancel active requests void -cancel_most_requests () - PROTOTYPE: - CODE: +aio_busy (delay,callback=&PL_sv_undef) + double delay + SV * callback + PPCODE: { - aio_req *req; + dREQ; + + req->type = REQ_BUSY; + req->fd = delay < 0. ? 0 : delay; + req->fd2 = delay < 0. ? 0 : 1000. * (delay - req->fd); - pthread_mutex_lock (&reqlock); - for (req = reqs; req; req = req->next) - req->flags |= 1; - pthread_mutex_unlock (&reqlock); - - pthread_mutex_lock (&reslock); - for (req = ress; req; req = req->next) - req->flags |= 1; - pthread_mutex_unlock (&reslock); + REQ_SEND; } -#endif +void +aio_group (callback=&PL_sv_undef) + SV * callback + 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 + PPCODE: +{ + dREQ; + + req->type = REQ_NOP; + + 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 () @@ -1135,7 +1604,7 @@ while (nreqs) { poll_wait (); - poll_cb (); + poll_cb (0); } void @@ -1145,7 +1614,7 @@ if (nreqs) { poll_wait (); - poll_cb (); + poll_cb (0); } int @@ -1179,11 +1648,126 @@ OUTPUT: RETVAL -MODULE = IO::AIO PACKAGE = IO::AIO::CB +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) LOCK (wrklock); + RETVAL = started; + if (WORDACCESS_UNSAFE) UNLOCK (wrklock); + OUTPUT: + RETVAL + +PROTOTYPES: DISABLE + +MODULE = IO::AIO PACKAGE = IO::AIO::REQ void cancel (aio_req_ornot req) - PROTOTYPE: CODE: - req->cancelled = 1; + 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 +add (aio_req grp, ...) + PPCODE: +{ + int i; + aio_req req; + + if (grp->fd == 2) + croak ("cannot add requests to IO::AIO::GRP after the group finished"); + + for (i = 1; i < items; ++i ) + { + if (GIMME_V != G_VOID) + XPUSHs (sv_2mortal (newSVsv (ST (i)))); + + req = SvAIO_REQ (ST (i)); + + if (req) + { + ++grp->length; + req->grp = grp; + + req->grp_prev = 0; + req->grp_next = grp->grp_first; + + if (grp->grp_first) + grp->grp_first->grp_prev = req; + + grp->grp_first = req; + } + } +} + +void +cancel_subs (aio_req_ornot req) + CODE: + req_cancel_subs (req); + +void +result (aio_req grp, ...) + CODE: +{ + int i; + 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; +} + +void +errno (aio_req grp, int errorno = errno) + CODE: + grp->errorno = errorno; + +void +limit (aio_req grp, int limit) + CODE: + grp->fd2 = limit; + aio_grp_feed (grp); + +void +feed (aio_req grp, SV *callback=&PL_sv_undef) + CODE: +{ + SvREFCNT_dec (grp->fh2); + grp->fh2 = newSVsv (callback); + + if (grp->fd2 <= 0) + grp->fd2 = 2; + + aio_grp_feed (grp); +}