--- IO-AIO/AIO.xs 2005/07/10 18:16:49 1.2 +++ IO-AIO/AIO.xs 2005/08/16 23:33:34 1.23 @@ -1,42 +1,44 @@ -#define PERL_NO_GET_CONTEXT +#define _REENTRANT 1 +#include #include "EXTERN.h" #include "perl.h" #include "XSUB.h" +#include "autoconf/config.h" + #include #include + #include #include #include #include -#include #include -#include typedef void *InputStream; /* hack, but 5.6.1 is simply toooo old ;) */ typedef void *OutputStream; /* hack, but 5.6.1 is simply toooo old ;) */ typedef void *InOutStream; /* hack, but 5.6.1 is simply toooo old ;) */ -#if __i386 || __amd64 -# define STACKSIZE ( 256 * sizeof (long)) -#elif __ia64 -# define STACKSIZE (8192 * sizeof (long)) +#if __ia64 +# define STACKSIZE 65536 #else -# define STACKSIZE ( 512 * sizeof (long)) +# define STACKSIZE 4096 #endif enum { REQ_QUIT, REQ_OPEN, REQ_CLOSE, REQ_READ, REQ_WRITE, REQ_READAHEAD, - REQ_STAT, REQ_LSTAT, REQ_FSTAT, REQ_UNLINK, + REQ_STAT, REQ_LSTAT, REQ_FSTAT, REQ_FSYNC, REQ_FDATASYNC, + REQ_UNLINK, REQ_RMDIR, + REQ_SYMLINK, }; typedef struct aio_cb { - struct aio_cb *next; + struct aio_cb *volatile next; int type; @@ -46,8 +48,8 @@ ssize_t result; mode_t mode; /* open */ int errorno; - SV *data, *callback; - void *dataptr; + SV *data, *callback, *fh; + void *dataptr, *data2ptr; STRLEN dataoffset; Stat_t *statdata; @@ -56,10 +58,126 @@ typedef aio_cb *aio_req; static int started; -static int nreqs; -static int reqpipe[2], respipe[2]; +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_mutex_t frklock = PTHREAD_MUTEX_INITIALIZER; +static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER; + +static volatile aio_req reqs, reqe; /* queue start, queue end */ +static volatile aio_req ress, rese; /* queue start, queue end */ + +static void +poll_wait () +{ + if (nreqs && !ress) + { + fd_set rfd; + FD_ZERO(&rfd); + FD_SET(respipe [0], &rfd); + + select (respipe [0] + 1, &rfd, 0, 0, 0); + } +} + +static int +poll_cb () +{ + dSP; + int count = 0; + aio_req req, prv; + + pthread_mutex_lock (&reslock); + + { + /* read any signals sent by the worker threads */ + char buf [32]; + while (read (respipe [0], buf, 32) == 32) + ; + } + + req = ress; + ress = rese = 0; + + pthread_mutex_unlock (&reslock); + + while (req) + { + nreqs--; + + if (req->type == REQ_QUIT) + started--; + else + { + int errorno = errno; + errno = req->errorno; + + if (req->type == REQ_READ) + SvCUR_set (req->data, req->dataoffset + + req->result > 0 ? req->result : 0); + + if (req->data) + SvREFCNT_dec (req->data); + + if (req->fh) + SvREFCNT_dec (req->fh); + + if (req->type == REQ_STAT || req->type == REQ_LSTAT || req->type == REQ_FSTAT) + { + PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; + PL_laststatval = req->result; + PL_statcache = *(req->statdata); + + Safefree (req->statdata); + } + + ENTER; + PUSHMARK (SP); + XPUSHs (sv_2mortal (newSViv (req->result))); + + if (req->type == REQ_OPEN) + { + /* convert fd to fh */ + SV *fh; + + PUTBACK; + call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); + SPAGAIN; + + fh = SvREFCNT_inc (POPs); + + PUSHMARK (SP); + XPUSHs (sv_2mortal (fh)); + } + + if (SvOK (req->callback)) + { + PUTBACK; + call_sv (req->callback, G_VOID | G_EVAL); + SPAGAIN; + } + + LEAVE; + + if (req->callback) + SvREFCNT_dec (req->callback); + + errno = errorno; + count++; + } + + prv = req; + req = req->next; + Safefree (prv); + + /* TODO: croak on errors? */ + } -static aio_req qs, qe; /* queue start, queue end */ + return count; +} static void *aio_proc(void *arg); @@ -84,31 +202,30 @@ } static void -send_reqs (void) -{ - /* this write is atomic */ - while (qs && write (reqpipe[1], &qs, sizeof qs) == sizeof qs) - { - qs = qs->next; - if (!qs) qe = 0; - } -} - -static void send_req (aio_req req) { nreqs++; + + pthread_mutex_lock (&reqlock); + req->next = 0; - if (qe) + if (reqe) { - qe->next = req; - qe = req; + reqe->next = req; + reqe = req; } else - qe = qs = req; + reqe = reqs = req; + + pthread_cond_signal (&reqwait); + pthread_mutex_unlock (&reqlock); - send_reqs (); + while (nreqs > max_outstanding) + { + poll_wait (); + poll_cb (); + } } static void @@ -121,156 +238,191 @@ send_req (req); } -static void -read_write (pTHX_ - int dowrite, int fd, off_t offset, size_t length, - SV *data, STRLEN dataoffset, SV *callback) -{ - aio_req req; - STRLEN svlen; - char *svptr = SvPV (data, svlen); - SvUPGRADE (data, SVt_PV); - SvPOK_on (data); - - if (dataoffset < 0) - dataoffset += svlen; - - if (dataoffset < 0 || dataoffset > svlen) - croak ("data offset outside of string"); +static void min_parallel (int nthreads) +{ + while (nthreads > started) + start_thread (); +} - if (dowrite) - { - /* write: check length and adjust. */ - if (length < 0 || length + dataoffset > svlen) - length = svlen - dataoffset; +static void max_parallel (int nthreads) +{ + int cur = started; + while (cur > nthreads) + { + end_thread (); + cur--; } - else + + while (started > nthreads) { - /* read: grow scalar as necessary */ - svptr = SvGROW (data, length + dataoffset); + poll_wait (); + poll_cb (); } +} - if (length < 0) - croak ("length must not be negative"); +static int fork_started; - Newz (0, req, 1, aio_cb); +static void atfork_prepare (void) +{ + pthread_mutex_lock (&frklock); - if (!req) - croak ("out of memory during aio_req allocation"); + fork_started = started; - req->type = dowrite ? REQ_WRITE : REQ_READ; - req->fd = fd; - req->offset = offset; - req->length = length; - req->data = SvREFCNT_inc (data); - req->dataptr = (char *)svptr + dataoffset; - req->callback = SvREFCNT_inc (callback); + for (;;) { + while (nreqs) + { + poll_wait (); + poll_cb (); + } - send_req (req); + max_parallel (0); + + pthread_mutex_lock (&reqlock); + + if (!nreqs && !started) + break; + + pthread_mutex_unlock (&reqlock); + + min_parallel (fork_started); + } + + pthread_mutex_lock (&reslock); + + assert (!started); + assert (!nreqs); + assert (!reqs && !reqe); + assert (!ress && !rese); } -static void -poll_wait () +static void atfork_parent (void) { - fd_set rfd; - FD_ZERO(&rfd); - FD_SET(respipe[0], &rfd); - - select (respipe[0] + 1, &rfd, 0, 0, 0); + pthread_mutex_unlock (&reslock); + min_parallel (fork_started); + pthread_mutex_unlock (&reqlock); + pthread_mutex_unlock (&frklock); } -static int -poll_cb (pTHX) +static void atfork_child (void) { - dSP; - int count = 0; - aio_req req; + reqs = reqe = 0; - while (read (respipe[0], (void *)&req, sizeof (req)) == sizeof (req)) - { - nreqs--; + atfork_parent (); +} - if (req->type == REQ_QUIT) - started--; - else - { - int errorno = errno; - errno = req->errorno; +/*****************************************************************************/ +/* work around various missing functions */ - if (req->type == REQ_READ) - SvCUR_set (req->data, req->dataoffset - + req->result > 0 ? req->result : 0); +#if !HAVE_PREADWRITE +# define pread aio_pread +# define pwrite aio_pwrite - if (req->data) - SvREFCNT_dec (req->data); +/* + * make our pread/pwrite safe against themselves, but not against + * normal read/write by using a mutex. slows down execution a lot, + * but that's your problem, not mine. + */ +static pthread_mutex_t iolock = PTHREAD_MUTEX_INITIALIZER; - if (req->type == REQ_STAT || req->type == REQ_LSTAT || req->type == REQ_FSTAT) - { - PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT; - PL_laststatval = req->result; - PL_statcache = *(req->statdata); +static ssize_t +pread (int fd, void *buf, size_t count, off_t offset) +{ + ssize_t res; + off_t ooffset; - Safefree (req->statdata); - } + pthread_mutex_lock (&iolock); + ooffset = lseek (fd, 0, SEEK_CUR); + lseek (fd, offset, SEEK_SET); + res = read (fd, buf, count); + lseek (fd, ooffset, SEEK_SET); + pthread_mutex_unlock (&iolock); - PUSHMARK (SP); - XPUSHs (sv_2mortal (newSViv (req->result))); + return res; +} - if (req->type == REQ_OPEN) - { - /* convert fd to fh */ - SV *fh; +static ssize_t +pwrite (int fd, void *buf, size_t count, off_t offset) +{ + ssize_t res; + off_t ooffset; - PUTBACK; - call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); - SPAGAIN; + pthread_mutex_lock (&iolock); + ooffset = lseek (fd, 0, SEEK_CUR); + lseek (fd, offset, SEEK_SET); + res = write (fd, buf, count); + lseek (fd, offset, SEEK_SET); + pthread_mutex_unlock (&iolock); - fh = POPs; + return res; +} +#endif - PUSHMARK (SP); - XPUSHs (fh); - } +#if !HAVE_FDATASYNC +# define fdatasync fsync +#endif - PUTBACK; - call_sv (req->callback, G_VOID | G_EVAL); - SPAGAIN; - - if (req->callback) - SvREFCNT_dec (req->callback); +#if !HAVE_READAHEAD +# define readahead aio_readahead - errno = errorno; - count++; - } +static char readahead_buf[4096]; - Safefree (req); - } +static ssize_t +readahead (int fd, off_t offset, size_t count) +{ + while (count > 0) + { + size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf); - if (qs) - send_reqs (); + pread (fd, readahead_buf, len, offset); + offset += len; + count -= len; + } - return count; + errno = 0; } +#endif + +/*****************************************************************************/ static void * aio_proc (void *thr_arg) { aio_req req; + int type; - /* then loop */ - while (read (reqpipe[0], (void *)&req, sizeof (req)) == sizeof (req)) + do { + pthread_mutex_lock (&reqlock); + + for (;;) + { + req = reqs; + + if (reqs) + { + reqs = reqs->next; + if (!reqs) reqe = 0; + } + + if (req) + break; + + pthread_cond_wait (&reqwait, &reqlock); + } + + pthread_mutex_unlock (&reqlock); + errno = 0; /* strictly unnecessary */ - switch (req->type) + type = req->type; + + switch (type) { - case REQ_READ: req->result = pread64 (req->fd, req->dataptr, req->length, req->offset); break; - case REQ_WRITE: req->result = pwrite64 (req->fd, req->dataptr, req->length, req->offset); break; -#if SYS_readahead + 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; -#else - case REQ_READAHEAD: req->result = -1; errno = ENOSYS; break; -#endif case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break; case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break; @@ -279,13 +431,14 @@ 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_SYMLINK: req->result = symlink (req->data2ptr, req->dataptr); break; - case REQ_FSYNC: req->result = fsync (req->fd); break; case REQ_FDATASYNC: req->result = fdatasync (req->fd); break; + case REQ_FSYNC: req->result = fsync (req->fd); break; case REQ_QUIT: - write (respipe[1], (void *)&req, sizeof (req)); - return 0; + break; default: req->result = ENOSYS; @@ -293,222 +446,290 @@ } req->errorno = errno; - write (respipe[1], (void *)&req, sizeof (req)); + + pthread_mutex_lock (&reslock); + + req->next = 0; + + if (rese) + { + 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); + } + + pthread_mutex_unlock (&reslock); } + while (type != REQ_QUIT); return 0; } +#define dREQ \ + aio_req req; \ + \ + if (SvOK (callback) && !SvROK (callback)) \ + croak ("clalback must be undef or of reference type"); \ + \ + Newz (0, req, 1, aio_cb); \ + if (!req) \ + croak ("out of memory during aio_req allocation"); \ + \ + req->callback = SvREFCNT_inc (callback); + MODULE = IO::AIO PACKAGE = IO::AIO +PROTOTYPES: ENABLE + BOOT: { - if (pipe (reqpipe) || pipe (respipe)) - croak ("unable to initialize request or result pipe"); + if (pipe (respipe)) + croak ("unable to initialize result pipe"); - if (fcntl (reqpipe[1], F_SETFL, O_NONBLOCK)) + if (fcntl (respipe [0], F_SETFL, O_NONBLOCK)) croak ("cannot set result pipe to nonblocking mode"); - if (fcntl (respipe[0], F_SETFL, O_NONBLOCK)) + if (fcntl (respipe [1], F_SETFL, O_NONBLOCK)) croak ("cannot set result pipe to nonblocking mode"); + + pthread_atfork (atfork_prepare, atfork_parent, atfork_child); } void min_parallel(nthreads) int nthreads PROTOTYPE: $ - CODE: - while (nthreads > started) - start_thread (); void max_parallel(nthreads) int nthreads PROTOTYPE: $ - CODE: -{ - int cur = started; - while (cur > nthreads) - { - end_thread (); - cur--; - } - while (started > nthreads) - { - poll_wait (); - poll_cb (aTHX); - } -} +int +max_outstanding(nreqs) + int nreqs + PROTOTYPE: $ + CODE: + RETVAL = max_outstanding; + max_outstanding = nreqs; void -aio_open(pathname,flags,mode,callback) +aio_open(pathname,flags,mode,callback=&PL_sv_undef) SV * pathname int flags int mode SV * callback - PROTOTYPE: $$$$ + PROTOTYPE: $$$;$ CODE: { - aio_req req; - - Newz (0, req, 1, aio_cb); - - if (!req) - croak ("out of memory during aio_req allocation"); + dREQ; req->type = REQ_OPEN; req->data = newSVsv (pathname); - req->dataptr = SvPV_nolen (req->data); + req->dataptr = SvPVbyte_nolen (req->data); req->fd = flags; req->mode = mode; - req->callback = SvREFCNT_inc (callback); send_req (req); } void -aio_close(fh,callback) - InputStream fh - SV * callback - PROTOTYPE: $$ +aio_close(fh,callback=&PL_sv_undef) + SV * fh + SV * callback + PROTOTYPE: $;$ ALIAS: aio_close = REQ_CLOSE aio_fsync = REQ_FSYNC aio_fdatasync = REQ_FDATASYNC CODE: { - aio_req req; - - Newz (0, req, 1, aio_cb); - - if (!req) - croak ("out of memory during aio_req allocation"); + dREQ; req->type = ix; - req->fd = PerlIO_fileno (fh); - req->callback = SvREFCNT_inc (callback); + req->fh = newSVsv (fh); + req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); send_req (req); } void -aio_read(fh,offset,length,data,dataoffset,callback) - InputStream fh - UV offset - IV length - SV * data - IV dataoffset - SV * callback - PROTOTYPE: $$$$$$ - CODE: - read_write (aTHX_ 0, PerlIO_fileno (fh), offset, length, data, dataoffset, callback); - -void -aio_write(fh,offset,length,data,dataoffset,callback) - OutputStream fh - UV offset - IV length - SV * data - IV dataoffset - SV * callback - PROTOTYPE: $$$$$$ - CODE: - read_write (aTHX_ 1, PerlIO_fileno (fh), offset, length, data, dataoffset, callback); - -void -aio_readahead(fh,offset,length,callback) - InputStream fh - UV offset - IV length - SV * callback - PROTOTYPE: $$$$ +aio_read(fh,offset,length,data,dataoffset,callback=&PL_sv_undef) + SV * fh + UV offset + IV length + SV * data + IV dataoffset + SV * callback + ALIAS: + aio_read = REQ_READ + aio_write = REQ_WRITE + PROTOTYPE: $$$$$;$ CODE: { aio_req req; + STRLEN svlen; + char *svptr = SvPVbyte (data, svlen); + + SvUPGRADE (data, SVt_PV); + SvPOK_on (data); + + if (dataoffset < 0) + dataoffset += svlen; + + if (dataoffset < 0 || dataoffset > svlen) + croak ("data offset outside of string"); + + if (ix == REQ_WRITE) + { + /* write: check length and adjust. */ + if (length < 0 || length + dataoffset > svlen) + length = svlen - dataoffset; + } + else + { + /* read: grow scalar as necessary */ + svptr = SvGROW (data, length + dataoffset); + } if (length < 0) croak ("length must not be negative"); - Newz (0, req, 1, aio_cb); + { + dREQ; + + req->type = ix; + req->fh = newSVsv (fh); + req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh)) + : IoOFP (sv_2io (fh))); + req->offset = offset; + req->length = length; + req->data = SvREFCNT_inc (data); + req->dataptr = (char *)svptr + dataoffset; + req->callback = SvREFCNT_inc (callback); - if (!req) - croak ("out of memory during aio_req allocation"); + send_req (req); + } +} + +void +aio_readahead(fh,offset,length,callback=&PL_sv_undef) + SV * fh + UV offset + IV length + SV * callback + PROTOTYPE: $$$;$ + CODE: +{ + dREQ; req->type = REQ_READAHEAD; - req->fd = PerlIO_fileno (fh); + req->fh = newSVsv (fh); + req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); req->offset = offset; req->length = length; - req->callback = SvREFCNT_inc (callback); send_req (req); } void -aio_stat(fh_or_path,callback) +aio_stat(fh_or_path,callback=&PL_sv_undef) SV * fh_or_path SV * callback - PROTOTYPE: $$ ALIAS: - aio_lstat = 1 + aio_stat = REQ_STAT + aio_lstat = REQ_LSTAT CODE: { - aio_req req; - - Newz (0, req, 1, aio_cb); - - if (!req) - croak ("out of memory during aio_req allocation"); + dREQ; New (0, req->statdata, 1, Stat_t); - if (!req->statdata) - croak ("out of memory during aio_req->statdata allocation"); + croak ("out of memory during aio_req->statdata allocation (sorry, i just leaked memory, too)"); if (SvPOK (fh_or_path)) { - req->type = ix ? REQ_LSTAT : REQ_STAT; + req->type = ix; req->data = newSVsv (fh_or_path); - req->dataptr = SvPV_nolen (req->data); + req->dataptr = SvPVbyte_nolen (req->data); } else { req->type = REQ_FSTAT; + req->fh = newSVsv (fh_or_path); req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); } - req->callback = SvREFCNT_inc (callback); - send_req (req); } void -aio_unlink(pathname,callback) +aio_unlink(pathname,callback=&PL_sv_undef) SV * pathname SV * callback - PROTOTYPE: $$ + ALIAS: + aio_unlink = REQ_UNLINK + aio_rmdir = REQ_RMDIR CODE: { - aio_req req; + dREQ; - Newz (0, req, 1, aio_cb); + req->type = ix; + req->data = newSVsv (pathname); + req->dataptr = SvPVbyte_nolen (req->data); - if (!req) - croak ("out of memory during aio_req allocation"); + send_req (req); +} + +void +aio_symlink(oldpath,newpath,callback=&PL_sv_undef) + SV * oldpath + SV * newpath + SV * callback + CODE: +{ + dREQ; - req->type = REQ_UNLINK; - req->data = newSVsv (pathname); - req->dataptr = SvPV_nolen (req->data); - req->callback = SvREFCNT_inc (callback); + req->type = REQ_SYMLINK; + req->fh = newSVsv (oldpath); + req->data2ptr = SvPVbyte_nolen (req->fh); + req->data = newSVsv (newpath); + req->dataptr = SvPVbyte_nolen (req->data); send_req (req); } +void +flush() + PROTOTYPE: + CODE: + while (nreqs) + { + poll_wait (); + poll_cb (); + } + +void +poll() + PROTOTYPE: + CODE: + if (nreqs) + { + poll_wait (); + poll_cb (); + } + int poll_fileno() PROTOTYPE: CODE: - RETVAL = respipe[0]; + RETVAL = respipe [0]; OUTPUT: RETVAL @@ -516,7 +737,7 @@ poll_cb(...) PROTOTYPE: CODE: - RETVAL = poll_cb (aTHX); + RETVAL = poll_cb (); OUTPUT: RETVAL @@ -524,7 +745,8 @@ poll_wait() PROTOTYPE: CODE: - poll_wait (); + if (nreqs) + poll_wait (); int nreqs()