--- IO-AIO/AIO.xs 2005/08/22 23:21:57 1.33 +++ IO-AIO/AIO.xs 2006/10/21 23:06:04 1.43 @@ -9,9 +9,11 @@ #include +#include +#include #include #include - +#include #include #include #include @@ -25,13 +27,22 @@ # include # elif __hpux # include +# elif __solaris /* not yet */ +# include +# else +# error sendfile support requested but not available # endif #endif +/* used for struct dirent, AIX doesn't provide it */ +#ifndef NAME_MAX +# define NAME_MAX 4096 +#endif + #if __ia64 # define STACKSIZE 65536 #else -# define STACKSIZE 4096 +# define STACKSIZE 8192 #endif enum { @@ -41,30 +52,40 @@ REQ_SENDFILE, REQ_STAT, REQ_LSTAT, REQ_FSTAT, REQ_FSYNC, REQ_FDATASYNC, - REQ_UNLINK, REQ_RMDIR, - REQ_SYMLINK, + REQ_UNLINK, REQ_RMDIR, REQ_RENAME, + REQ_READDIR, + REQ_LINK, REQ_SYMLINK, }; -typedef struct aio_cb { +#define AIO_CB_KLASS "IO::AIO::CB" + +typedef struct aio_cb +{ + struct aio_cb *grp_prev, *grp_next; + struct aio_grp *grp; + struct aio_cb *volatile next; - int type; + SV *self; /* the perl counterpart of this request, if any */ - int fd, fd2; + SV *data, *callback; + SV *fh, *fh2; + void *dataptr, *data2ptr; + Stat_t *statdata; off_t offset; size_t length; ssize_t result; - mode_t mode; /* open */ + + int type; + int fd, fd2; int errorno; - SV *data, *callback; - SV *fh, *fh2; - void *dataptr, *data2ptr; STRLEN dataoffset; - - Stat_t *statdata; + mode_t mode; /* open */ + unsigned char cancelled; } aio_cb; typedef aio_cb *aio_req; +typedef aio_cb *aio_req_ornot; static int started, wanted; static volatile int nreqs; @@ -78,8 +99,66 @@ static volatile aio_req reqs, reqe; /* queue start, queue end */ static volatile aio_req ress, rese; /* queue start, queue end */ -static void free_req (aio_req req) +typedef struct aio_grp +{ + struct aio_cb *first, *last; + SV *callback; + int busycount; +} aio_grp; + +static void aio_grp_begin (aio_grp *grp) +{ + ++grp->busycount; +} + +static void aio_grp_end (aio_grp *grp) +{ + --grp->busycount; + + if (grp->busycount) + return; + + SvREFCNT_dec (grp->callback); + grp->callback = 0; +} + +static aio_grp *aio_grp_new () { + aio_grp *grp; + + Newz (0, grp, 1, aio_grp); + aio_grp_begin (grp); + + return grp; +} + +/* must be called at most once */ +static SV *req_sv (aio_req req) +{ + req->self = (SV *)newHV (); + sv_magic (req->self, 0, PERL_MAGIC_ext, (char *)req, 0); + + return sv_bless (newRV_noinc (req->self), gv_stashpv (AIO_CB_KLASS, 1)); +} + +static aio_req SvAIO_REQ (SV *sv) +{ + if (!sv_derived_from (sv, AIO_CB_KLASS) || !SvROK (sv)) + croak ("object of class " AIO_CB_KLASS " expected"); + + MAGIC *mg = mg_find (SvRV (sv), PERL_MAGIC_ext); + + return mg ? (aio_req)mg->mg_ptr : 0; +} + +static void req_free (aio_req req) +{ + if (req->self) + { + sv_unmagic (req->self, PERL_MAGIC_ext); + SvREFCNT_dec (req->self); + } + if (req->data) SvREFCNT_dec (req->data); @@ -95,6 +174,9 @@ if (req->callback) SvREFCNT_dec (req->callback); + if (req->type == REQ_READDIR && req->result >= 0) + free (req->data2ptr); + Safefree (req); } @@ -168,24 +250,51 @@ ENTER; PUSHMARK (SP); - XPUSHs (sv_2mortal (newSViv (req->result))); - if (req->type == REQ_OPEN) + if (req->type == REQ_READDIR) { - /* convert fd to fh */ - SV *fh; + SV *rv = &PL_sv_undef; - PUTBACK; - call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL); - SPAGAIN; + if (req->result >= 0) + { + char *buf = req->data2ptr; + AV *av = newAV (); + + while (req->result) + { + SV *sv = newSVpv (buf, 0); + + av_push (av, sv); + buf += SvCUR (sv) + 1; + req->result--; + } + + rv = sv_2mortal (newRV_noinc ((SV *)av)); + } - fh = SvREFCNT_inc (POPs); + XPUSHs (rv); + } + else + { + XPUSHs (sv_2mortal (newSViv (req->result))); + + if (req->type == REQ_OPEN) + { + /* convert fd to fh */ + SV *fh; - PUSHMARK (SP); - XPUSHs (sv_2mortal (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)) + if (SvOK (req->callback) && !req->cancelled) { PUTBACK; call_sv (req->callback, G_VOID | G_EVAL); @@ -193,7 +302,7 @@ if (SvTRUE (ERRSV)) { - free_req (req); + req_free (req); croak (0); } } @@ -204,7 +313,7 @@ count++; } - free_req (req); + req_free (req); } return count; @@ -233,7 +342,7 @@ } static void -send_req (aio_req req) +req_send (aio_req req) { while (started < wanted && nreqs >= started) start_thread (); @@ -274,7 +383,7 @@ Newz (0, req, 1, aio_cb); req->type = REQ_QUIT; - send_req (req); + req_send (req); } static void min_parallel (int nthreads) @@ -315,49 +424,6 @@ croak ("cannot set result pipe to nonblocking mode"); } -static void atfork_prepare (void) -{ - pthread_mutex_lock (&reqlock); - pthread_mutex_lock (&reslock); -} - -static void atfork_parent (void) -{ - pthread_mutex_unlock (&reslock); - pthread_mutex_unlock (&reqlock); -} - -static void atfork_child (void) -{ - aio_req prv; - - started = 0; - - while (reqs) - { - prv = reqs; - reqs = prv->next; - free_req (prv); - } - - reqs = reqe = 0; - - while (ress) - { - prv = ress; - ress = prv->next; - free_req (prv); - } - - ress = rese = 0; - - close (respipe [0]); - close (respipe [1]); - create_pipe (); - - atfork_parent (); -} - /*****************************************************************************/ /* work around various missing functions */ @@ -370,7 +436,7 @@ * 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; +static pthread_mutex_t preadwritelock = PTHREAD_MUTEX_INITIALIZER; static ssize_t pread (int fd, void *buf, size_t count, off_t offset) @@ -378,12 +444,12 @@ ssize_t res; off_t ooffset; - pthread_mutex_lock (&iolock); + pthread_mutex_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 (&iolock); + pthread_mutex_unlock (&preadwritelock); return res; } @@ -394,12 +460,12 @@ ssize_t res; off_t ooffset; - pthread_mutex_lock (&iolock); + pthread_mutex_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 (&iolock); + pthread_mutex_unlock (&preadwritelock); return res; } @@ -412,11 +478,11 @@ #if !HAVE_READAHEAD # define readahead aio_readahead -static char readahead_buf[4096]; - static ssize_t readahead (int fd, off_t offset, size_t count) { + char readahead_buf[4096]; + while (count > 0) { size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf); @@ -430,78 +496,192 @@ } #endif +#if !HAVE_READDIR_R +# define readdir_r aio_readdir_r + +static pthread_mutex_t readdirlock = PTHREAD_MUTEX_INITIALIZER; + +static int +readdir_r (DIR *dirp, struct dirent *ent, struct dirent **res) +{ + struct dirent *e; + int errorno; + + pthread_mutex_lock (&readdirlock); + + e = readdir (dirp); + errorno = errno; + + if (e) + { + *res = ent; + strcpy (ent->d_name, e->d_name); + } + else + *res = 0; + + pthread_mutex_unlock (&readdirlock); + + errno = errorno; + return e ? 0 : -1; +} +#endif + /* sendfile always needs emulation */ static ssize_t sendfile_ (int ofd, int ifd, off_t offset, size_t count) { - ssize_t res; + ssize_t res; + + if (!count) + return 0; - if (!count) - return 0; +#if HAVE_SENDFILE +# if __linux + res = sendfile (ofd, ifd, &offset, count); -#if __linux - res = sendfile (ofd, ifd, &offset, count); +# elif __freebsd + /* + * Of course, the freebsd sendfile is a dire hack with no thoughts + * wasted on making it similar to other I/O functions. + */ + { + off_t sbytes; + 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 */ + res = sbytes; + } -#elif __freebsd - /* - * Of course, the freebsd sendfile is a dire hack with no thoughts - * wasted on making it similar to other i/o functions. - */ - { - off_t sbytes; - res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); - - if (!res && errno == EAGAIN) - /* maybe on others, too, as usual, the manpage leaves you guessing */ - res = sbytes; - } +# elif __hpux + res = sendfile (ofd, ifd, offset, count, 0, 0); -#elif __hpux - res = sendfile (ofd, ifd, offset, count, 0, 0); +# elif __solaris + { + struct sendfilevec vec; + size_t sbytes; + + vec.sfv_fd = ifd; + vec.sfv_flag = 0; + vec.sfv_off = offset; + vec.sfv_len = count; + + res = sendfilev (ofd, &vec, 1, &sbytes); + + if (res < 0 && sbytes) + res = sbytes; + } +# endif #else - res = -1; - errno = ENOSYS; + res = -1; + errno = ENOSYS; #endif - if (res < 0 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK)) - { - /* emulate sendfile. this is a major pain in the ass */ - char *buf = malloc (4096); - res = 0; - - for (;;) - { - ssize_t cnt; - - cnt = pread (ifd, buf, 4096, offset); - - if (cnt <= 0) - { - if (cnt && !res) res = -1; - break; - } - - cnt = write (ofd, buf, cnt); - - if (cnt <= 0) - { - if (cnt && !res) res = -1; - break; - } - - offset += cnt; - res += cnt; - } - - { - int errorno = errno; - free (buf); - errno = errorno; - } - } + if (res < 0 + && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK +#if __solaris + || errno == EAFNOSUPPORT || errno == EPROTOTYPE +#endif + ) + ) + { + /* emulate sendfile. this is a major pain in the ass */ + char buf[4096]; + res = 0; + + while (count) + { + ssize_t cnt; + + cnt = pread (ifd, buf, count > 4096 ? 4096 : count, offset); - return res; + if (cnt <= 0) + { + if (cnt && !res) res = -1; + break; + } + + cnt = write (ofd, buf, cnt); + + if (cnt <= 0) + { + if (cnt && !res) res = -1; + break; + } + + offset += cnt; + res += cnt; + count -= cnt; + } + } + + return res; +} + +/* read a full directory */ +static int +scandir_ (const char *path, void **namesp) +{ + DIR *dirp = opendir (path); + union + { + struct dirent d; + char b [offsetof (struct dirent, d_name) + NAME_MAX + 1]; + } u; + struct dirent *entp; + char *name, *names; + int memlen = 4096; + int memofs = 0; + int res = 0; + int errorno; + + if (!dirp) + return -1; + + names = malloc (memlen); + + for (;;) + { + errno = 0, readdir_r (dirp, &u.d, &entp); + + if (!entp) + break; + + name = entp->d_name; + + if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) + { + int len = strlen (name) + 1; + + res++; + + while (memofs + len > memlen) + { + memlen *= 2; + names = realloc (names, memlen); + if (!names) + break; + } + + memcpy (names + memofs, name, len); + memofs += len; + } + } + + errorno = errno; + closedir (dirp); + + if (errorno) + { + free (names); + errno = errorno; + res = -1; + } + + *namesp = (void *)names; + return res; } /*****************************************************************************/ @@ -536,36 +716,38 @@ errno = 0; /* strictly unnecessary */ - type = req->type; - - switch (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_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; + if (!req->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_QUIT: - 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; - default: - req->result = ENOSYS; - 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_QUIT: + break; + + default: + req->result = ENOSYS; + break; + } req->errorno = errno; @@ -593,17 +775,80 @@ return 0; } +/*****************************************************************************/ + +static void atfork_prepare (void) +{ + pthread_mutex_lock (&reqlock); + pthread_mutex_lock (&reslock); +#if !HAVE_PREADWRITE + pthread_mutex_lock (&preadwritelock); +#endif +#if !HAVE_READDIR_R + pthread_mutex_lock (&readdirlock); +#endif +} + +static void atfork_parent (void) +{ +#if !HAVE_READDIR_R + pthread_mutex_unlock (&readdirlock); +#endif +#if !HAVE_PREADWRITE + pthread_mutex_unlock (&preadwritelock); +#endif + pthread_mutex_unlock (&reslock); + pthread_mutex_unlock (&reqlock); +} + +static void atfork_child (void) +{ + aio_req prv; + + started = 0; + + while (reqs) + { + prv = reqs; + reqs = prv->next; + req_free (prv); + } + + reqs = reqe = 0; + + while (ress) + { + prv = ress; + ress = prv->next; + req_free (prv); + } + + ress = rese = 0; + + close (respipe [0]); + close (respipe [1]); + create_pipe (); + + atfork_parent (); +} + #define dREQ \ aio_req req; \ \ if (SvOK (callback) && !SvROK (callback)) \ - croak ("clalback must be undef or of reference type"); \ + croak ("callback 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 = newSVsv (callback); + req->callback = newSVsv (callback) + +#define REQ_SEND \ + req_send (req); \ + \ + if (GIMME_V != G_VOID) \ + XPUSHs (req_sv (req)); MODULE = IO::AIO PACKAGE = IO::AIO @@ -611,22 +856,27 @@ 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)); + create_pipe (); pthread_atfork (atfork_prepare, atfork_parent, atfork_child); } void -min_parallel(nthreads) +min_parallel (nthreads) int nthreads PROTOTYPE: $ void -max_parallel(nthreads) +max_parallel (nthreads) int nthreads PROTOTYPE: $ int -max_outstanding(nreqs) +max_outstanding (nreqs) int nreqs PROTOTYPE: $ CODE: @@ -634,13 +884,13 @@ max_outstanding = nreqs; void -aio_open(pathname,flags,mode,callback=&PL_sv_undef) +aio_open (pathname,flags,mode,callback=&PL_sv_undef) SV * pathname int flags int mode SV * callback PROTOTYPE: $$$;$ - CODE: + PPCODE: { dREQ; @@ -650,11 +900,11 @@ req->fd = flags; req->mode = mode; - send_req (req); + REQ_SEND; } void -aio_close(fh,callback=&PL_sv_undef) +aio_close (fh,callback=&PL_sv_undef) SV * fh SV * callback PROTOTYPE: $;$ @@ -662,7 +912,7 @@ aio_close = REQ_CLOSE aio_fsync = REQ_FSYNC aio_fdatasync = REQ_FDATASYNC - CODE: + PPCODE: { dREQ; @@ -670,11 +920,11 @@ req->fh = newSVsv (fh); req->fd = PerlIO_fileno (IoIFP (sv_2io (fh))); - send_req (req); + REQ_SEND (req); } void -aio_read(fh,offset,length,data,dataoffset,callback=&PL_sv_undef) +aio_read (fh,offset,length,data,dataoffset,callback=&PL_sv_undef) SV * fh UV offset UV length @@ -685,7 +935,7 @@ aio_read = REQ_READ aio_write = REQ_WRITE PROTOTYPE: $$$$$;$ - CODE: + PPCODE: { aio_req req; STRLEN svlen; @@ -733,19 +983,19 @@ req->data2ptr = (void *)data; } - send_req (req); + REQ_SEND; } } void -aio_sendfile(out_fh,in_fh,in_offset,length,callback=&PL_sv_undef) +aio_sendfile (out_fh,in_fh,in_offset,length,callback=&PL_sv_undef) SV * out_fh SV * in_fh UV in_offset UV length SV * callback PROTOTYPE: $$$$;$ - CODE: + PPCODE: { dREQ; @@ -757,17 +1007,17 @@ req->offset = in_offset; req->length = length; - send_req (req); + REQ_SEND; } void -aio_readahead(fh,offset,length,callback=&PL_sv_undef) +aio_readahead (fh,offset,length,callback=&PL_sv_undef) SV * fh UV offset IV length SV * callback PROTOTYPE: $$$;$ - CODE: + PPCODE: { dREQ; @@ -777,24 +1027,24 @@ req->offset = offset; req->length = length; - send_req (req); + REQ_SEND; } void -aio_stat(fh_or_path,callback=&PL_sv_undef) +aio_stat (fh_or_path,callback=&PL_sv_undef) SV * fh_or_path SV * callback ALIAS: aio_stat = REQ_STAT aio_lstat = REQ_LSTAT - CODE: + PPCODE: { dREQ; New (0, req->statdata, 1, Stat_t); if (!req->statdata) { - free_req (req); + req_free (req); croak ("out of memory during aio_req->statdata allocation"); } @@ -811,17 +1061,18 @@ req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path))); } - send_req (req); + REQ_SEND; } void -aio_unlink(pathname,callback=&PL_sv_undef) +aio_unlink (pathname,callback=&PL_sv_undef) SV * pathname SV * callback ALIAS: - aio_unlink = REQ_UNLINK - aio_rmdir = REQ_RMDIR - CODE: + aio_unlink = REQ_UNLINK + aio_rmdir = REQ_RMDIR + aio_readdir = REQ_READDIR + PPCODE: { dREQ; @@ -829,29 +1080,56 @@ req->data = newSVsv (pathname); req->dataptr = SvPVbyte_nolen (req->data); - send_req (req); + REQ_SEND; } void -aio_symlink(oldpath,newpath,callback=&PL_sv_undef) +aio_link (oldpath,newpath,callback=&PL_sv_undef) SV * oldpath SV * newpath SV * callback - CODE: + ALIAS: + aio_link = REQ_LINK + aio_symlink = REQ_SYMLINK + aio_rename = REQ_RENAME + PPCODE: { dREQ; - req->type = REQ_SYMLINK; + req->type = ix; req->fh = newSVsv (oldpath); req->data2ptr = SvPVbyte_nolen (req->fh); req->data = newSVsv (newpath); req->dataptr = SvPVbyte_nolen (req->data); - send_req (req); + REQ_SEND; +} + +#if 0 + +# undocumented, because it does not cancel active requests +void +cancel_most_requests () + PROTOTYPE: + CODE: +{ + aio_req *req; + + 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); } +#endif + void -flush() +flush () PROTOTYPE: CODE: while (nreqs) @@ -901,3 +1179,11 @@ OUTPUT: RETVAL +MODULE = IO::AIO PACKAGE = IO::AIO::CB + +void +cancel (aio_req_ornot req) + PROTOTYPE: + CODE: + req->cancelled = 1; +