--- IO-AIO/AIO.xs 2006/10/24 00:26:32 1.65 +++ IO-AIO/AIO.xs 2007/05/09 06:45:12 1.98 @@ -1,8 +1,4 @@ -#if __linux -# define _GNU_SOURCE -#endif - -#define _REENTRANT 1 +#include "xthread.h" #include @@ -12,9 +8,8 @@ #include "autoconf/config.h" -#include - #include +#include #include #include #include @@ -41,28 +36,26 @@ # 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 -#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 #define dBUF \ - char *aio_buf = malloc (AIO_BUFSIZE); \ + char *aio_buf; \ + LOCK (wrklock); \ + self->dbuf = aio_buf = malloc (AIO_BUFSIZE); \ + UNLOCK (wrklock); \ if (!aio_buf) \ return -1; -#define fBUF free (aio_buf) +typedef SV SV8; /* byte-sv, used for argument-checking */ enum { REQ_QUIT, @@ -71,11 +64,11 @@ 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_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" @@ -85,17 +78,16 @@ { 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, *fh; + SV *sv1, *sv2; + void *ptr1, *ptr2; + off_t offs; + size_t size; ssize_t result; - STRLEN dataoffset; + STRLEN stroffset; int type; - int fd, fd2; + int int1, int2; int errorno; mode_t mode; /* open */ @@ -107,7 +99,9 @@ } aio_cb; enum { - FLAG_CANCELLED = 0x01, + FLAG_CANCELLED = 0x01, /* request was cancelled */ + FLAG_SV1_RO_OFF = 0x40, /* data was set readonly */ + FLAG_PTR2_FREE = 0x80, /* need to free(ptr2) */ }; typedef aio_cb *aio_req; @@ -119,30 +113,209 @@ 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 () +{ + 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); +} + static int next_pri = DEFAULT_PRI + PRI_BIAS; -static int started, wanted; -static volatile int nreqs; -static int max_outstanding = 1<<30; +static unsigned int started, idle, wanted; + +/* worker threads management */ +static mutex_t wrklock = 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 [2]; -#if __linux && defined (PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP) -# define AIO_MUTEX_INIT PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP +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 AIO_MUTEX_INIT PTHREAD_MUTEX_INITIALIZER + +# define get_nready() nready +# define get_npending() npending +# define get_nthreads() started + #endif -static pthread_mutex_t reslock = AIO_MUTEX_INIT; -static pthread_mutex_t reqlock = AIO_MUTEX_INIT; -static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER; +/* + * 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; + } + } -static volatile aio_req reqs, reqe; /* queue start, queue end */ -static volatile aio_req ress, rese; /* queue start, queue end */ + abort (); +} -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); /* must be called at most once */ static SV *req_sv (aio_req req, const char *klass) @@ -170,11 +343,13 @@ 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; @@ -183,163 +358,164 @@ 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) - { - req_invoke (grp); - req_free (grp); - } -} - -static void poll_wait () -{ - fd_set rfd; - - while (nreqs) + if (!grp->size && grp->int1) { - aio_req req; -#if !(__i386 || __x86_64) /* safe without sempahore on this archs */ - pthread_mutex_lock (&reslock); -#endif - req = ress; -#if !(__i386 || __x86_64) /* safe without sempahore on this archs */ - pthread_mutex_unlock (&reslock); -#endif - - if (req) - return; + block_sig (); - 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; - int errorno = errno; - - if (req->flags & FLAG_CANCELLED || !SvOK (req->callback)) - return; - - errno = req->errorno; - ENTER; - SAVETMPS; - PUSHMARK (SP); - EXTEND (SP, 1); + if (req->flags & FLAG_SV1_RO_OFF) + SvREADONLY_off (req->sv1); - 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_push (av, sv); - buf += SvCUR (sv) + 1; - req->result--; - } + av_extend (av, req->result - 1); - rv = sv_2mortal (newRV_noinc ((SV *)av)); - } + for (i = 0; i < req->result; ++i) + { + SV *sv = newSVpv (buf, 0); - PUSHs (rv); - } - break; + av_store (av, i, sv); + buf += SvCUR (sv) + 1; + } - case REQ_OPEN: - { - /* convert fd to fh */ - SV *fh; + rv = sv_2mortal (newRV_noinc ((SV *)av)); + } - PUSHs (sv_2mortal (newSViv (req->result))); - PUTBACK; - call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); - SPAGAIN; + PUSHs (rv); + } + break; - fh = SvREFCNT_inc (POPs); + case REQ_OPEN: + { + /* convert fd to fh */ + SV *fh; - PUSHMARK (SP); - XPUSHs (sv_2mortal (fh)); - } - break; + PUSHs (sv_2mortal (newSViv (req->result))); + PUTBACK; + call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); + SPAGAIN; + + fh = POPs; + PUSHMARK (SP); + XPUSHs (fh); + } + break; - case REQ_GROUP: - req->fd = 2; /* mark group as finished */ + case REQ_GROUP: + req->int1 = 2; /* mark group as finished */ - if (req->data) - { - int i; - AV *av = (AV *)req->data; + if (req->sv1) + { + int i; + AV *av = (AV *)req->sv1; - 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_SLEEP: - break; + case REQ_NOP: + case REQ_BUSY: + break; - default: - PUSHs (sv_2mortal (newSViv (req->result))); - break; - } + case REQ_READLINK: + if (req->result > 0) + { + SvCUR_set (req->sv1, req->result); + *SvEND (req->sv1) = 0; + PUSHs (req->sv1); + } + 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; - PUTBACK; - call_sv (req->callback, G_VOID | G_EVAL); - SPAGAIN; + case REQ_READ: + SvCUR_set (req->sv1, req->stroffset + (req->result > 0 ? req->result : 0)); + *SvEND (req->sv1) = 0; + PUSHs (sv_2mortal (newSViv (req->result))); + break; - FREETMPS; - LEAVE; + default: + PUSHs (sv_2mortal (newSViv (req->result))); + break; + } - errno = errorno; + errno = req->errorno; - if (SvTRUE (ERRSV)) - { - req_free (req); - croak (0); + PUTBACK; + call_sv (req->callback, G_VOID | G_EVAL); + SPAGAIN; + + FREETMPS; + LEAVE; } -} -static void req_free (aio_req req) -{ if (req->grp) { aio_req grp = req->grp; @@ -354,166 +530,127 @@ aio_grp_dec (grp); } + return !SvTRUE (ERRSV); +} + +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->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->flags |= FLAG_CANCELLED; + 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; - - for (;;) - { - pthread_mutex_lock (&reslock); - req = ress; - - 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) - ; - - rese = 0; - } - } - - pthread_mutex_unlock (&reslock); - - if (!req) - break; - - --nreqs; + req->flags |= FLAG_CANCELLED; - 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)); + req_cancel_subs (req); +} - if (req->data2ptr && (req->type == REQ_READ || req->type == REQ_WRITE)) - SvREADONLY_off (req->data); +static void *aio_proc(void *arg); - if (req->statdata) - { - PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; - PL_laststatval = req->result; - PL_statcache = *(req->statdata); - } +static void start_thread (void) +{ + worker *wrk = calloc (1, sizeof (worker)); - req_invoke (req); + if (!wrk) + croak ("unable to allocate worker thread data"); - count++; - } + 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; + UNLOCK (wrklock); } -static void *aio_proc(void *arg); - -static void start_thread (void) +static void maybe_start_thread () { - 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; - - pthread_cond_signal (&reqwait); - pthread_mutex_unlock (&reqlock); + LOCK (reqlock); + ++nready; + reqq_push (&req_queue, req); + COND_SIGNAL (reqwait); + UNLOCK (reqlock); - if (nreqs > max_outstanding) - for (;;) - { - poll_cb (); + unblock_sig (); - 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; + + LOCK (reqlock); + reqq_push (&req_queue, req); + COND_SIGNAL (reqwait); + UNLOCK (reqlock); - req_send (req); + 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) @@ -524,22 +661,120 @@ 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 (); + + ++maxreqs; } + + unblock_sig (); + return count; } static void create_pipe () @@ -566,19 +801,19 @@ * 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 = 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); + 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; } @@ -588,12 +823,12 @@ 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; } @@ -604,9 +839,9 @@ #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) { dBUF; @@ -619,23 +854,22 @@ count -= len; } - fBUF; - errno = 0; } + #endif #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) { struct dirent *e; int errorno; - pthread_mutex_lock (&readdirlock); + LOCK (readdirlock); e = readdir (dirp); errorno = errno; @@ -648,7 +882,7 @@ else *res = 0; - pthread_mutex_unlock (&readdirlock); + UNLOCK (readdirlock); errno = errorno; return e ? 0 : -1; @@ -656,7 +890,7 @@ #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; @@ -677,7 +911,7 @@ 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; } @@ -743,15 +977,13 @@ res += cnt; count -= cnt; } - - fBUF; } return res; } /* read a full directory */ -static int scandir_ (const char *path, void **namesp) +static void scandir_ (aio_req req, worker *self) { DIR *dirp; union @@ -766,14 +998,14 @@ int res = 0; int errorno; - dirp = opendir (path); - if (!dirp) - return -1; - - u = malloc (sizeof (*u)); - names = malloc (memlen); + LOCK (wrklock); + self->dirp = dirp = opendir (req->ptr1); + self->dbuf = u = malloc (sizeof (*u)); + req->flags |= FLAG_PTR2_FREE; + req->ptr2 = names = malloc (memlen); + UNLOCK (wrklock); - if (u && names) + if (dirp && u && names) for (;;) { errno = 0; @@ -793,7 +1025,10 @@ while (memofs + len > memlen) { memlen *= 2; - names = realloc (names, memlen); + LOCK (wrklock); + req->ptr2 = names = realloc (names, memlen); + UNLOCK (wrklock); + if (!names) break; } @@ -803,19 +1038,10 @@ } } - errorno = errno; - free (u); - closedir (dirp); - - if (errorno) - { - free (names); - errno = errorno; - res = -1; - } - - *namesp = (void *)names; - return res; + if (errno) + res = -1; + + req->result = res; } /*****************************************************************************/ @@ -823,73 +1049,100 @@ 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 (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; } - pthread_mutex_unlock (&reqlock); + --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); 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_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_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_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->int1); break; + case REQ_FSYNC: req->result = fsync (req->int1); break; + case REQ_READDIR: scandir_ (req, self); break; - case REQ_SLEEP: + case REQ_BUSY: { struct timeval tv; - tv.tv_sec = req->fd; - tv.tv_usec = req->fd2; + tv.tv_sec = req->int1; + tv.tv_usec = req->int2; req->result = select (0, 0, 0, 0, &tv); } case REQ_GROUP: case REQ_NOP: - case REQ_QUIT: break; + case REQ_QUIT: + goto quit; + default: req->result = ENOSYS; break; @@ -897,26 +1150,30 @@ req->errorno = errno; - pthread_mutex_lock (&reslock); + LOCK (reslock); - req->next = 0; + ++npending; - if (rese) + if (!reqq_push (&res_queue, req)) { - rese->next = req; - rese = req; - } - else - { - rese = ress = req; - /* 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); } - pthread_mutex_unlock (&reslock); + self->req = 0; + worker_clear (self); + + UNLOCK (reslock); } - while (type != REQ_QUIT); + +quit: + LOCK (wrklock); + worker_free (self); + UNLOCK (wrklock); return 0; } @@ -925,51 +1182,56 @@ 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]); @@ -1006,35 +1268,57 @@ 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 -max_parallel (nthreads) - int nthreads +min_parallel (int nthreads) PROTOTYPE: $ +void +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 + SV8 * pathname int flags int mode SV * callback @@ -1044,9 +1328,9 @@ 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; @@ -1066,8 +1350,8 @@ dREQ; req->type = ix; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); + req->fh = newSVsv (fh); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh))); REQ_SEND (req); } @@ -1077,7 +1361,7 @@ SV * fh UV offset UV length - SV * data + SV8 * data UV dataoffset SV * callback ALIAS: @@ -1086,7 +1370,6 @@ PROTOTYPE: $$$$$;$ PPCODE: { - aio_req req; STRLEN svlen; char *svptr = SvPVbyte (data, svlen); @@ -1108,7 +1391,7 @@ else { /* read: grow scalar as necessary */ - svptr = SvGROW (data, length + dataoffset); + svptr = SvGROW (data, length + dataoffset + 1); } if (length < 0) @@ -1118,18 +1401,19 @@ dREQ; req->type = ix; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) - : IoOFP (sv_2io (fh))); - req->offset = offset; - req->length = length; - req->data = SvREFCNT_inc (data); - req->dataptr = (char *)svptr + dataoffset; + req->fh = newSVsv (fh); + req->int1 = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) + : IoOFP (sv_2io (fh))); + req->offs = offset; + req->size = length; + req->sv1 = SvREFCNT_inc (data); + req->ptr1 = (char *)svptr + dataoffset; + req->stroffset = dataoffset; if (!SvREADONLY (data)) { SvREADONLY_on (data); - req->data2ptr = (void *)data; + req->flags |= FLAG_SV1_RO_OFF; } REQ_SEND; @@ -1137,6 +1421,28 @@ } void +aio_readlink (path,callback=&PL_sv_undef) + SV8 * path + SV * callback + PROTOTYPE: $$;$ + PPCODE: +{ + SV *data; + dREQ; + + data = newSV (NAME_MAX); + SvPOK_on (data); + + req->type = REQ_READLINK; + req->fh = newSVsv (path); + req->ptr2 = SvPVbyte_nolen (req->fh); + req->sv1 = data; + req->ptr1 = SvPVbyte_nolen (data); + + REQ_SEND; +} + +void aio_sendfile (out_fh,in_fh,in_offset,length,callback=&PL_sv_undef) SV * out_fh SV * in_fh @@ -1149,12 +1455,12 @@ 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->fh = 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; } @@ -1171,17 +1477,17 @@ dREQ; req->type = REQ_READAHEAD; - req->fh = newSVsv (fh); - req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); - req->offset = offset; - req->length = length; + req->fh = 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 + SV8 * fh_or_path SV * callback ALIAS: aio_stat = REQ_STAT @@ -1190,24 +1496,26 @@ { 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; + if (SvPOK (fh_or_path)) { req->type = ix; - req->data = newSVsv (fh_or_path); - req->dataptr = SvPVbyte_nolen (req->data); + req->sv1 = newSVsv (fh_or_path); + 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->fh = newSVsv (fh_or_path); + req->int1 = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); } REQ_SEND; @@ -1215,8 +1523,8 @@ void aio_unlink (pathname,callback=&PL_sv_undef) - SV * pathname - SV * callback + SV8 * pathname + SV * callback ALIAS: aio_unlink = REQ_UNLINK aio_rmdir = REQ_RMDIR @@ -1226,17 +1534,34 @@ 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 (pathname,mode,callback=&PL_sv_undef) + SV8 * pathname + UV mode + SV * callback + 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 + SV8 * oldpath + SV8 * newpath + SV * callback ALIAS: aio_link = REQ_LINK aio_symlink = REQ_SYMLINK @@ -1246,25 +1571,44 @@ 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->fh = newSVsv (oldpath); + req->ptr2 = SvPVbyte_nolen (req->fh); + req->sv1 = newSVsv (newpath); + req->ptr1 = SvPVbyte_nolen (req->sv1); + + REQ_SEND; +} + +void +aio_mknod (pathname,mode,dev,callback=&PL_sv_undef) + SV8 * pathname + UV mode + UV dev + SV * callback + 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 (delay,callback=&PL_sv_undef) + double delay + SV * callback 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->int1 = delay < 0. ? 0 : delay; + req->int2 = delay < 0. ? 0 : 1000. * (delay - req->int1); REQ_SEND; } @@ -1278,14 +1622,14 @@ 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 + SV * callback PPCODE: { dREQ; @@ -1295,16 +1639,27 @@ REQ_SEND; } -#if 0 +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 -aio_pri (int pri = DEFAULT_PRI) +aioreq_nice (int nice = 0) CODE: - if (pri < PRI_MIN) pri = PRI_MIN; - if (pri > PRI_MAX) pri = PRI_MAX; - next_pri = pri + PRI_BIAS; - -#endif + 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 () @@ -1316,15 +1671,14 @@ poll_cb (); } -void +int poll() PROTOTYPE: CODE: - if (nreqs) - { - poll_wait (); - poll_cb (); - } + poll_wait (); + RETVAL = poll_cb (); + OUTPUT: + RETVAL int poll_fileno() @@ -1346,8 +1700,44 @@ 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"); + + 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 nreqs() @@ -1357,13 +1747,38 @@ 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) 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_cancel (req); @@ -1382,7 +1797,10 @@ 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 ) @@ -1394,7 +1812,7 @@ if (req) { - ++grp->length; + ++grp->size; req->grp = grp; req->grp_prev = 0; @@ -1409,34 +1827,48 @@ } 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 -feed_limit (aio_req grp, int limit) +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); }