--- libeio/eio.c 2008/05/13 19:34:11 1.13 +++ libeio/eio.c 2009/06/06 19:44:17 1.33 @@ -1,4 +1,47 @@ +/* + * libeio implementation + * + * Copyright (c) 2007,2008,2009 Marc Alexander Lehmann + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without modifica- + * tion, are permitted provided that the following conditions are met: + * + * 1. Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED + * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER- + * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO + * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE- + * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; + * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, + * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH- + * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED + * OF THE POSSIBILITY OF SUCH DAMAGE. + * + * Alternatively, the contents of this file may be used under the terms of + * the GNU General Public License ("GPL") version 2 or any later version, + * in which case the provisions of the GPL are applicable instead of + * the above. If you wish to allow the use of your version of this file + * only under the terms of the GPL and not to allow others to use your + * version of this file under the BSD license, indicate your decision + * by deleting the provisions above and replace them with the notice + * and other provisions required by the GPL. If you do not delete the + * provisions above, a recipient may use your version of this file under + * either the BSD or the GPL. + */ + #include "eio.h" + +#ifdef EIO_STACKSIZE +# define XTHREAD_STACKSIZE EIO_STACKSIZE +#endif #include "xthread.h" #include @@ -27,7 +70,6 @@ #ifdef _WIN32 /*doh*/ - #else # include "config.h" @@ -38,6 +80,24 @@ # include # include +/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */ +# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) +# define _DIRENT_HAVE_D_TYPE /* sigh */ +# define D_INO(de) (de)->d_fileno +# define D_NAMLEN(de) (de)->d_namlen +# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 +# define D_INO(de) (de)->d_ino +# endif + +#ifdef _D_EXACT_NAMLEN +# undef D_NAMLEN +# define D_NAMLEN(de) _D_EXACT_NAMLEN (de) +#endif + +# ifdef _DIRENT_HAVE_D_TYPE +# define D_TYPE(de) (de)->d_type +# endif + # ifndef EIO_STRUCT_DIRENT # define EIO_STRUCT_DIRENT struct dirent # endif @@ -59,6 +119,16 @@ # endif #endif +#ifndef D_TYPE +# define D_TYPE(de) 0 +#endif +#ifndef D_INO +# define D_INO(de) 0 +#endif +#ifndef D_NAMLEN +# define D_NAMLEN(de) strlen ((de)->d_name) +#endif + /* number of seconds after which an idle threads exit */ #define IDLE_TIMEOUT 10 @@ -72,9 +142,9 @@ #define dBUF \ char *eio_buf; \ - X_LOCK (etplock); \ + ETP_WORKER_LOCK (self); \ self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \ - X_UNLOCK (etplock); \ + ETP_WORKER_UNLOCK (self); \ errno = ENOMEM; \ if (!eio_buf) \ return -1; @@ -83,6 +153,50 @@ /*****************************************************************************/ +#if __GNUC__ >= 3 +# define expect(expr,value) __builtin_expect ((expr),(value)) +#else +# define expect(expr,value) (expr) +#endif + +#define expect_false(expr) expect ((expr) != 0, 0) +#define expect_true(expr) expect ((expr) != 0, 1) + +/*****************************************************************************/ + +#define ETP_PRI_MIN EIO_PRI_MIN +#define ETP_PRI_MAX EIO_PRI_MAX + +struct etp_worker; + +#define ETP_REQ eio_req +#define ETP_DESTROY(req) eio_destroy (req) +static int eio_finish (eio_req *req); +#define ETP_FINISH(req) eio_finish (req) +static void eio_execute (struct etp_worker *self, eio_req *req); +#define ETP_EXECUTE(wrk,req) eio_execute (wrk,req) + +#define ETP_WORKER_CLEAR(req) \ + if (wrk->dbuf) \ + { \ + free (wrk->dbuf); \ + wrk->dbuf = 0; \ + } \ + \ + if (wrk->dirp) \ + { \ + closedir (wrk->dirp); \ + wrk->dirp = 0; \ + } + +#define ETP_WORKER_COMMON \ + void *dbuf; \ + DIR *dirp; + +/*****************************************************************************/ + +#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) + /* calculcate time difference in ~1/EIO_TICKS of a second */ static int tvdiff (struct timeval *tv1, struct timeval *tv2) { @@ -92,56 +206,57 @@ static unsigned int started, idle, wanted = 4; -typedef struct etp_pool -{ - void (*want_poll_cb) (void); - void (*done_poll_cb) (void); +static void (*want_poll_cb) (void); +static void (*done_poll_cb) (void); - unsigned int max_poll_time; - unsigned int max_poll_reqs; -} etp_pool; +static unsigned int max_poll_time; /* reslock */ +static unsigned int max_poll_reqs; /* reslock */ -static volatile unsigned int nreqs, nready, npending; +static volatile unsigned int nreqs; /* reqlock */ +static volatile unsigned int nready; /* reqlock */ +static volatile unsigned int npending; /* reqlock */ static volatile unsigned int max_idle = 4; -static mutex_t etplock = X_MUTEX_INIT; +static mutex_t wrklock = X_MUTEX_INIT; static mutex_t reslock = X_MUTEX_INIT; static mutex_t reqlock = X_MUTEX_INIT; static cond_t reqwait = X_COND_INIT; -typedef struct worker +#if !HAVE_PREADWRITE +/* + * make our pread/pwrite emulation 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 mutex_t preadwritelock = X_MUTEX_INIT; +#endif + +typedef struct etp_worker { - /* locked by etplock */ - struct worker *prev, *next; + /* locked by wrklock */ + struct etp_worker *prev, *next; thread_t tid; - /* locked by reslock, reqlock or etplock */ - eio_req *req; /* currently processed request */ - void *dbuf; - DIR *dirp; -} worker; + /* locked by reslock, reqlock or wrklock */ + ETP_REQ *req; /* currently processed request */ -static worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */ + ETP_WORKER_COMMON +} etp_worker; + +static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */ + +#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock) +#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock) /* worker threads management */ -static void worker_clear (worker *wrk) +static void etp_worker_clear (etp_worker *wrk) { - if (wrk->dirp) - { - closedir (wrk->dirp); - wrk->dirp = 0; - } - - if (wrk->dbuf) - { - free (wrk->dbuf); - wrk->dbuf = 0; - } + ETP_WORKER_CLEAR (wrk); } -static void worker_free (worker *wrk) +static void etp_worker_free (etp_worker *wrk) { wrk->next->prev = wrk->prev; wrk->prev->next = wrk->next; @@ -149,12 +264,16 @@ free (wrk); } -unsigned int eio_nreqs (void) +static unsigned int etp_nreqs (void) { - return nreqs; + int retval; + if (WORDACCESS_UNSAFE) X_LOCK (reqlock); + retval = nreqs; + if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); + return retval; } -unsigned int eio_nready (void) +static unsigned int etp_nready (void) { unsigned int retval; @@ -165,7 +284,7 @@ return retval; } -unsigned int eio_npending (void) +static unsigned int etp_npending (void) { unsigned int retval; @@ -176,7 +295,7 @@ return retval; } -unsigned int eio_nthreads (void) +static unsigned int etp_nthreads (void) { unsigned int retval; @@ -193,14 +312,14 @@ * per shift, the most expensive operation. */ typedef struct { - eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */ + ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */ int size; -} reqq; +} etp_reqq; -static reqq req_queue; -static reqq res_queue; +static etp_reqq req_queue; +static etp_reqq res_queue; -static int reqq_push (reqq *q, eio_req *req) +static int reqq_push (etp_reqq *q, ETP_REQ *req) { int pri = req->pri; req->next = 0; @@ -216,7 +335,7 @@ return q->size++; } -static eio_req *reqq_shift (reqq *q) +static ETP_REQ *reqq_shift (etp_reqq *q) { int pri; @@ -225,7 +344,7 @@ --q->size; - for (pri = EIO_NUM_PRI; pri--; ) + for (pri = ETP_NUM_PRI; pri--; ) { eio_req *req = q->qs[pri]; @@ -243,49 +362,43 @@ static void etp_atfork_prepare (void) { - X_LOCK (etplock); + X_LOCK (wrklock); X_LOCK (reqlock); X_LOCK (reslock); #if !HAVE_PREADWRITE X_LOCK (preadwritelock); #endif -#if !HAVE_READDIR_R - X_LOCK (readdirlock); -#endif } static void etp_atfork_parent (void) { -#if !HAVE_READDIR_R - X_UNLOCK (readdirlock); -#endif #if !HAVE_PREADWRITE X_UNLOCK (preadwritelock); #endif X_UNLOCK (reslock); X_UNLOCK (reqlock); - X_UNLOCK (etplock); + X_UNLOCK (wrklock); } static void etp_atfork_child (void) { - eio_req *prv; + ETP_REQ *prv; - while (prv = reqq_shift (&req_queue)) - eio_destroy (prv); + while ((prv = reqq_shift (&req_queue))) + ETP_DESTROY (prv); - while (prv = reqq_shift (&res_queue)) - eio_destroy (prv); + while ((prv = reqq_shift (&res_queue))) + ETP_DESTROY (prv); while (wrk_first.next != &wrk_first) { - worker *wrk = wrk_first.next; + etp_worker *wrk = wrk_first.next; if (wrk->req) - eio_destroy (wrk->req); + ETP_DESTROY (wrk->req); - worker_clear (wrk); - worker_free (wrk); + etp_worker_clear (wrk); + etp_worker_free (wrk); } started = 0; @@ -304,19 +417,220 @@ } static int -etp_init (etp_pool *etp, void (*want_poll)(void), void (*done_poll)(void)) +etp_init (void (*want_poll)(void), void (*done_poll)(void)) { static pthread_once_t doinit = PTHREAD_ONCE_INIT; pthread_once (&doinit, etp_once_init); - memset (etp, 0, sizeof *etp); + want_poll_cb = want_poll; + done_poll_cb = done_poll; + + return 0; +} + +X_THREAD_PROC (etp_proc); + +static void etp_start_thread (void) +{ + etp_worker *wrk = calloc (1, sizeof (etp_worker)); + + /*TODO*/ + assert (("unable to allocate worker thread data", wrk)); + + X_LOCK (wrklock); + + if (thread_create (&wrk->tid, etp_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); + + X_UNLOCK (wrklock); +} + +static void etp_maybe_start_thread (void) +{ + if (expect_true (etp_nthreads () >= wanted)) + return; + + /* todo: maybe use idle here, but might be less exact */ + if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ())) + return; + + etp_start_thread (); +} + +static void etp_end_thread (void) +{ + eio_req *req = calloc (1, sizeof (eio_req)); + + req->type = -1; + req->pri = ETP_PRI_MAX - ETP_PRI_MIN; + + X_LOCK (reqlock); + reqq_push (&req_queue, req); + X_COND_SIGNAL (reqwait); + X_UNLOCK (reqlock); + + X_LOCK (wrklock); + --started; + X_UNLOCK (wrklock); +} + +static int etp_poll (void) +{ + unsigned int maxreqs; + unsigned int maxtime; + struct timeval tv_start, tv_now; + + X_LOCK (reslock); + maxreqs = max_poll_reqs; + maxtime = max_poll_time; + X_UNLOCK (reslock); + + if (maxtime) + gettimeofday (&tv_start, 0); + + for (;;) + { + ETP_REQ *req; + + etp_maybe_start_thread (); + + X_LOCK (reslock); + req = reqq_shift (&res_queue); + + if (req) + { + --npending; + + if (!res_queue.size && done_poll_cb) + done_poll_cb (); + } + + X_UNLOCK (reslock); + + if (!req) + return 0; + + X_LOCK (reqlock); + --nreqs; + X_UNLOCK (reqlock); + + if (expect_false (req->type == EIO_GROUP && req->size)) + { + req->int1 = 1; /* mark request as delayed */ + continue; + } + else + { + int res = ETP_FINISH (req); + if (expect_false (res)) + return res; + } + + if (expect_false (maxreqs && !--maxreqs)) + break; + + if (maxtime) + { + gettimeofday (&tv_now, 0); + + if (tvdiff (&tv_start, &tv_now) >= maxtime) + break; + } + } + + errno = EAGAIN; + return -1; +} + +static void etp_cancel (ETP_REQ *req) +{ + X_LOCK (wrklock); + req->flags |= EIO_FLAG_CANCELLED; + X_UNLOCK (wrklock); - etp->want_poll_cb = want_poll; - etp->done_poll_cb = done_poll; + eio_grp_cancel (req); } -static etp_pool etp; +static void etp_submit (ETP_REQ *req) +{ + req->pri -= ETP_PRI_MIN; + + if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN; + if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN; + + if (expect_false (req->type == EIO_GROUP)) + { + /* I hope this is worth it :/ */ + X_LOCK (reqlock); + ++nreqs; + X_UNLOCK (reqlock); + + X_LOCK (reslock); + + ++npending; + + if (!reqq_push (&res_queue, req) && want_poll_cb) + want_poll_cb (); + + X_UNLOCK (reslock); + } + else + { + X_LOCK (reqlock); + ++nreqs; + ++nready; + reqq_push (&req_queue, req); + X_COND_SIGNAL (reqwait); + X_UNLOCK (reqlock); + + etp_maybe_start_thread (); + } +} + +static void etp_set_max_poll_time (double nseconds) +{ + if (WORDACCESS_UNSAFE) X_LOCK (reslock); + max_poll_time = nseconds; + if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); +} + +static void etp_set_max_poll_reqs (unsigned int maxreqs) +{ + if (WORDACCESS_UNSAFE) X_LOCK (reslock); + max_poll_reqs = maxreqs; + if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); +} + +static void etp_set_max_idle (unsigned int nthreads) +{ + if (WORDACCESS_UNSAFE) X_LOCK (reqlock); + max_idle = nthreads <= 0 ? 1 : nthreads; + if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); +} + +static void etp_set_min_parallel (unsigned int nthreads) +{ + if (wanted < nthreads) + wanted = nthreads; +} + +static void etp_set_max_parallel (unsigned int nthreads) +{ + if (wanted > nthreads) + wanted = nthreads; + + while (started > wanted) + etp_end_thread (); +} /*****************************************************************************/ @@ -324,12 +638,12 @@ { while (grp->size < grp->int2 && !EIO_CANCELLED (grp)) { - int old_len = grp->size; + grp->flags &= ~EIO_FLAG_GROUPADD; EIO_FEED (grp); /* stop if no progress has been made */ - if (old_len == grp->size) + if (!(grp->flags & EIO_FLAG_GROUPADD)) { grp->feed = 0; break; @@ -337,8 +651,6 @@ } } -static int eio_finish (eio_req *req); - static int grp_dec (eio_req *grp) { --grp->size; @@ -396,194 +708,73 @@ void eio_cancel (eio_req *req) { - X_LOCK (etplock); - req->flags |= EIO_FLAG_CANCELLED; - X_UNLOCK (etplock); - - eio_grp_cancel (req); + etp_cancel (req); } -X_THREAD_PROC (eio_proc); - -static void start_thread (void) +void eio_submit (eio_req *req) { - worker *wrk = calloc (1, sizeof (worker)); - - /*TODO*/ - assert (("unable to allocate worker thread data", wrk)); - - X_LOCK (etplock); - - if (thread_create (&wrk->tid, eio_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); - - X_UNLOCK (etplock); + etp_submit (req); } -static void maybe_start_thread (void) +unsigned int eio_nreqs (void) { - if (eio_nthreads () >= wanted) - return; - - /* todo: maybe use idle here, but might be less exact */ - if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ()) - return; - - start_thread (); + return etp_nreqs (); } -void eio_submit (eio_req *req) +unsigned int eio_nready (void) { - req->pri += EIO_PRI_BIAS; - - if (req->pri < EIO_PRI_MIN + EIO_PRI_BIAS) req->pri = EIO_PRI_MIN + EIO_PRI_BIAS; - if (req->pri > EIO_PRI_MAX + EIO_PRI_BIAS) req->pri = EIO_PRI_MAX + EIO_PRI_BIAS; - - ++nreqs; - - X_LOCK (reqlock); - ++nready; - reqq_push (&req_queue, req); - X_COND_SIGNAL (reqwait); - X_UNLOCK (reqlock); - - maybe_start_thread (); + return etp_nready (); } -static void end_thread (void) +unsigned int eio_npending (void) { - eio_req *req = calloc (1, sizeof (eio_req)); - - req->type = EIO_QUIT; - req->pri = EIO_PRI_MAX + EIO_PRI_BIAS; - - X_LOCK (reqlock); - reqq_push (&req_queue, req); - X_COND_SIGNAL (reqwait); - X_UNLOCK (reqlock); + return etp_npending (); +} - X_LOCK (etplock); - --started; - X_UNLOCK (etplock); +unsigned int eio_nthreads (void) +{ + return etp_nthreads (); } void eio_set_max_poll_time (double nseconds) { - if (WORDACCESS_UNSAFE) X_LOCK (reqlock); - etp.max_poll_time = nseconds; - if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); + etp_set_max_poll_time (nseconds); } void eio_set_max_poll_reqs (unsigned int maxreqs) { - if (WORDACCESS_UNSAFE) X_LOCK (reqlock); - etp.max_poll_reqs = maxreqs; - if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); + etp_set_max_poll_reqs (maxreqs); } void eio_set_max_idle (unsigned int nthreads) { - if (WORDACCESS_UNSAFE) X_LOCK (reqlock); - max_idle = nthreads <= 0 ? 1 : nthreads; - if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); + etp_set_max_idle (nthreads); } void eio_set_min_parallel (unsigned int nthreads) { - if (wanted < nthreads) - wanted = nthreads; + etp_set_min_parallel (nthreads); } void eio_set_max_parallel (unsigned int nthreads) { - if (wanted > nthreads) - wanted = nthreads; - - while (started > wanted) - end_thread (); + etp_set_max_parallel (nthreads); } int eio_poll (void) { - int maxreqs = etp.max_poll_reqs; - struct timeval tv_start, tv_now; - eio_req *req; - - if (etp.max_poll_time) - gettimeofday (&tv_start, 0); - - for (;;) - { - maybe_start_thread (); - - X_LOCK (reslock); - req = reqq_shift (&res_queue); - - if (req) - { - --npending; - - if (!res_queue.size && etp.done_poll_cb) - etp.done_poll_cb (); - } - - X_UNLOCK (reslock); - - if (!req) - return 0; - - --nreqs; - - if (req->type == EIO_GROUP && req->size) - { - req->int1 = 1; /* mark request as delayed */ - continue; - } - else - { - int res = eio_finish (req); - if (res) - return res; - } - - if (maxreqs && !--maxreqs) - break; - - if (etp.max_poll_time) - { - gettimeofday (&tv_now, 0); - - if (tvdiff (&tv_start, &tv_now) >= etp.max_poll_time) - break; - } - } - - errno = EAGAIN; - return -1; + return etp_poll (); } /*****************************************************************************/ /* work around various missing functions */ #if !HAVE_PREADWRITE +# undef pread +# undef pwrite # define pread eio__pread # define pwrite eio__pwrite -/* - * 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 mutex_t preadwritelock = X_MUTEX_INIT; - static ssize_t eio__pread (int fd, void *buf, size_t count, off_t offset) { @@ -610,7 +801,7 @@ ooffset = lseek (fd, 0, SEEK_CUR); lseek (fd, offset, SEEK_SET); res = write (fd, buf, count); - lseek (fd, offset, SEEK_SET); + lseek (fd, ooffset, SEEK_SET); X_UNLOCK (preadwritelock); return res; @@ -619,6 +810,8 @@ #ifndef HAVE_FUTIMES +# undef utimes +# undef futimes # define utimes(path,times) eio__utimes (path, times) # define futimes(fd,times) eio__futimes (fd, times) @@ -647,14 +840,44 @@ #endif #if !HAVE_FDATASYNC -# define fdatasync fsync +# undef fdatasync +# define fdatasync(fd) fsync (fd) +#endif + +/* sync_file_range always needs emulation */ +int +eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags) +{ +#if HAVE_SYNC_FILE_RANGE + int res; + + if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE + || EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE + || EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER) + { + flags = 0 + | (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0) + | (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0) + | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0); + } + + res = sync_file_range (fd, offset, nbytes, flags); + + if (!res || errno != ENOSYS) + return res; #endif + /* even though we could play tricks with the flags, it's better to always + * call fdatasync, as thta matches the expectation of it's users best */ + return fdatasync (fd); +} + #if !HAVE_READAHEAD +# undef readahead # define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) static ssize_t -eio__readahead (int fd, off_t offset, size_t count, worker *self) +eio__readahead (int fd, off_t offset, size_t count, etp_worker *self) { size_t todo = count; dBUF; @@ -674,40 +897,9 @@ #endif -#if !HAVE_READDIR_R -# define readdir_r eio__readdir_r - -static mutex_t readdirlock = X_MUTEX_INIT; - -static int -eio__readdir_r (DIR *dirp, EIO_STRUCT_DIRENT *ent, EIO_STRUCT_DIRENT **res) -{ - EIO_STRUCT_DIRENT *e; - int errorno; - - X_LOCK (readdirlock); - - e = readdir (dirp); - errorno = errno; - - if (e) - { - *res = ent; - strcpy (ent->d_name, e->d_name); - } - else - *res = 0; - - X_UNLOCK (readdirlock); - - errno = errorno; - return e ? 0 : -1; -} -#endif - /* sendfile always needs emulation */ static ssize_t -eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self) +eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self) { ssize_t res; @@ -799,66 +991,259 @@ return res; } +static int +eio_dent_cmp (const void *a_, const void *b_) +{ + const eio_dirent *a = (const eio_dirent *)a_; + const eio_dirent *b = (const eio_dirent *)b_; + + return (int)b->score - (int)a->score ? (int)b->score - (int)a->score + : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; /* int might be < ino_t */ +} + /* read a full directory */ static void -eio__scandir (eio_req *req, worker *self) +eio__scandir (eio_req *req, etp_worker *self) { DIR *dirp; - union - { - EIO_STRUCT_DIRENT d; - char b [offsetof (EIO_STRUCT_DIRENT, d_name) + NAME_MAX + 1]; - } *u; EIO_STRUCT_DIRENT *entp; - char *name, *names; - int memlen = 4096; - int memofs = 0; - int res = 0; + unsigned char *name, *names; + int namesalloc = 4096; + int namesoffs = 0; + int flags = req->int1; + eio_dirent *dents = 0; + int dentalloc = 128; + int dentoffs = 0; - X_LOCK (etplock); + req->result = -1; + + if (!(flags & EIO_READDIR_DENTS)) + flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); + + X_LOCK (wrklock); + /* the corresponding closedir is in ETP_WORKER_CLEAR */ self->dirp = dirp = opendir (req->ptr1); - self->dbuf = u = malloc (sizeof (*u)); - req->flags |= EIO_FLAG_PTR2_FREE; - req->ptr2 = names = malloc (memlen); - X_UNLOCK (etplock); + req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; + req->ptr1 = names = malloc (namesalloc); + req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; + X_UNLOCK (wrklock); - if (dirp && u && names) + if (dirp && names && (!flags || dents)) for (;;) { errno = 0; - readdir_r (dirp, &u->d, &entp); + entp = readdir (dirp); if (!entp) - break; + { + if (errno) + break; + + /* sort etc. */ + req->int1 = flags; + req->result = dentoffs; + + if (dents) + { + eio_dirent *ent = dents + dentoffs; + + while (ent > dents) + (--ent)->name = names + (size_t)ent->name; + } + + if (flags & EIO_READDIR_STAT_ORDER + || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) + { + /* pray your qsort doesn't use quicksort */ + qsort (dents, dentoffs, sizeof (*dents), eio_dent_cmp); /* score depends of DIRS_FIRST */ + } + else if (flags & EIO_READDIR_DIRS_FIRST) + { + /* in this case, all is known, and we just put dirs first and sort them */ + eio_dirent *ent = dents + dentoffs; + eio_dirent *dir = dents; + + while (ent > dir) + { + if (dir->type == DT_DIR) + ++dir; + else + { + --ent; + + if (ent->type == DT_DIR) + { + eio_dirent tmp = *dir; + *dir = *ent; + *ent = tmp; + + ++dir; + } + } + } + /* now sort the dirs only */ + qsort (dents, dir - dents, sizeof (*dents), eio_dent_cmp); + } + + /* only provide the names array unless DENTS is specified */ + if (!(flags & EIO_READDIR_DENTS)) + { + X_LOCK (wrklock); + assert (!dents); + req->ptr1 = 0; + req->ptr2 = names; + X_UNLOCK (wrklock); + } + + break; + } + + /* now add the entry to our list(s) */ name = entp->d_name; + /* skip . and .. entries */ if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) { - int len = strlen (name) + 1; + int len = D_NAMLEN (entp) + 1; - res++; - - while (memofs + len > memlen) + while (expect_false (namesoffs + len > namesalloc)) { - memlen *= 2; - X_LOCK (etplock); - req->ptr2 = names = realloc (names, memlen); - X_UNLOCK (etplock); + namesalloc *= 2; + X_LOCK (wrklock); + req->ptr1 = names = realloc (names, namesalloc); + X_UNLOCK (wrklock); if (!names) break; } - memcpy (names + memofs, name, len); - memofs += len; + memcpy (names + namesoffs, name, len); + + if (dents) + { + struct eio_dirent *ent; + + if (expect_false (dentoffs == dentalloc)) + { + dentalloc *= 2; + X_LOCK (wrklock); + req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); + X_UNLOCK (wrklock); + + if (!dents) + break; + } + + ent = dents + dentoffs; + + ent->name = (char *)(size_t)namesoffs; /* rather dirtily we store the offset in the pointer */ + ent->namelen = len - 1; + ent->inode = D_INO (entp); + + switch (D_TYPE (entp)) + { + default: + ent->type = EIO_DT_UNKNOWN; + flags |= EIO_READDIR_FOUND_UNKNOWN; + break; + + #ifdef DT_FIFO + case DT_FIFO: ent->type = EIO_DT_FIFO; break; + #endif + #ifdef DT_CHR + case DT_CHR: ent->type = EIO_DT_CHR; break; + #endif + #ifdef DT_MPC + case DT_MPC: ent->type = EIO_DT_MPC; break; + #endif + #ifdef DT_DIR + case DT_DIR: ent->type = EIO_DT_DIR; break; + #endif + #ifdef DT_NAM + case DT_NAM: ent->type = EIO_DT_NAM; break; + #endif + #ifdef DT_BLK + case DT_BLK: ent->type = EIO_DT_BLK; break; + #endif + #ifdef DT_MPB + case DT_MPB: ent->type = EIO_DT_MPB; break; + #endif + #ifdef DT_REG + case DT_REG: ent->type = EIO_DT_REG; break; + #endif + #ifdef DT_NWK + case DT_NWK: ent->type = EIO_DT_NWK; break; + #endif + #ifdef DT_CMP + case DT_CMP: ent->type = EIO_DT_CMP; break; + #endif + #ifdef DT_LNK + case DT_LNK: ent->type = EIO_DT_LNK; break; + #endif + #ifdef DT_SOCK + case DT_SOCK: ent->type = EIO_DT_SOCK; break; + #endif + #ifdef DT_DOOR + case DT_DOOR: ent->type = EIO_DT_DOOR; break; + #endif + #ifdef DT_WHT + case DT_WHT: ent->type = EIO_DT_WHT; break; + #endif + } + + ent->score = 0; + + if (flags & EIO_READDIR_DIRS_FIRST) + { + if (ent->type == EIO_DT_UNKNOWN) + { + if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ + ent->score = 98; + else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ + ent->score = len <= 2 ? len + 60 : len <= 4 ? 50 : len <= 7 ? 40 : 10; /* shorter == more likely dir, but avoid too many classes */ + } + else if (ent->type == EIO_DT_DIR) + ent->score = 100; + } + } + + namesoffs += len; + ++dentoffs; } } + else + req->result = -1; +} - if (errno) - res = -1; - - req->result = res; +#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) +# undef msync +# define msync(a,b,c) ((errno = ENOSYS), -1) +#endif + +int +eio__mtouch (void *mem, size_t len, int flags) +{ + intptr_t addr = (intptr_t)mem; + intptr_t end = addr + len; +#ifdef PAGESIZE + const intptr_t page = PAGESIZE; +#else + static intptr_t page; + + if (!page) + page = sysconf (_SC_PAGESIZE); +#endif + + addr &= ~(page - 1); /* assume page size is always a power of two */ + + if (addr < end) + if (flags) /* modify */ + do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); + else + do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); + + return 0; } /*****************************************************************************/ @@ -866,9 +1251,9 @@ #define ALLOC(len) \ if (!req->ptr2) \ { \ - X_LOCK (etplock); \ + X_LOCK (wrklock); \ req->flags |= EIO_FLAG_PTR2_FREE; \ - X_UNLOCK (etplock); \ + X_UNLOCK (wrklock); \ req->ptr2 = malloc (len); \ if (!req->ptr2) \ { \ @@ -878,19 +1263,17 @@ } \ } -X_THREAD_PROC (eio_proc) +X_THREAD_PROC (etp_proc) { - eio_req *req; + ETP_REQ *req; struct timespec ts; - worker *self = (worker *)thr_arg; + etp_worker *self = (etp_worker *)thr_arg; /* try to distribute timeouts somewhat randomly */ ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL); for (;;) { - ts.tv_sec = time (0) + IDLE_TIMEOUT; - X_LOCK (reqlock); for (;;) @@ -902,21 +1285,21 @@ ++idle; + ts.tv_sec = time (0) + IDLE_TIMEOUT; if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT) { if (idle > max_idle) { --idle; X_UNLOCK (reqlock); - X_LOCK (etplock); + X_LOCK (wrklock); --started; - X_UNLOCK (etplock); + X_UNLOCK (wrklock); goto quit; } /* we are allowed to idle, so do so without any timeout */ X_COND_WAIT (reqwait, reqlock); - ts.tv_sec = time (0) + IDLE_TIMEOUT; } --idle; @@ -926,127 +1309,29 @@ X_UNLOCK (reqlock); - errno = 0; /* strictly unnecessary */ + if (req->type < 0) + goto quit; if (!EIO_CANCELLED (req)) - switch (req->type) - { - case EIO_READ: ALLOC (req->size); - req->result = req->offs >= 0 - ? pread (req->int1, req->ptr2, req->size, req->offs) - : read (req->int1, req->ptr2, req->size); break; - case EIO_WRITE: req->result = req->offs >= 0 - ? pwrite (req->int1, req->ptr2, req->size, req->offs) - : write (req->int1, req->ptr2, req->size); break; - - case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break; - case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break; - - case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT)); - req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; - case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); - req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; - case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); - req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; - - case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; - case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; - case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; - case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; - case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; - case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break; - - case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break; - case EIO_CLOSE: req->result = close (req->int1); break; - case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break; - case EIO_UNLINK: req->result = unlink (req->ptr1); break; - case EIO_RMDIR: req->result = rmdir (req->ptr1); break; - case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break; - case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; - case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; - case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; - case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break; - - case EIO_READLINK: ALLOC (NAME_MAX); - req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; - - case EIO_SYNC: req->result = 0; sync (); break; - case EIO_FSYNC: req->result = fsync (req->int1); break; - case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; - - case EIO_READDIR: eio__scandir (req, self); break; - - case EIO_BUSY: -#ifdef _WIN32 - Sleep (req->nv1 * 1000.); -#else - { - struct timeval tv; - - tv.tv_sec = req->nv1; - tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.; - - req->result = select (0, 0, 0, 0, &tv); - } -#endif - break; - - case EIO_UTIME: - case EIO_FUTIME: - { - struct timeval tv[2]; - struct timeval *times; - - if (req->nv1 != -1. || req->nv2 != -1.) - { - tv[0].tv_sec = req->nv1; - tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.; - tv[1].tv_sec = req->nv2; - tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.; - - times = tv; - } - else - times = 0; - - - req->result = req->type == EIO_FUTIME - ? futimes (req->int1, times) - : utimes (req->ptr1, times); - } - - case EIO_GROUP: - case EIO_NOP: - req->result = 0; - break; - - case EIO_QUIT: - goto quit; - - default: - req->result = -1; - break; - } - - req->errorno = errno; + ETP_EXECUTE (self, req); X_LOCK (reslock); ++npending; - if (!reqq_push (&res_queue, req) && etp.want_poll_cb) - etp.want_poll_cb (); + if (!reqq_push (&res_queue, req) && want_poll_cb) + want_poll_cb (); self->req = 0; - worker_clear (self); + etp_worker_clear (self); X_UNLOCK (reslock); } quit: - X_LOCK (etplock); - worker_free (self); - X_UNLOCK (etplock); + X_LOCK (wrklock); + etp_worker_free (self); + X_UNLOCK (wrklock); return 0; } @@ -1055,7 +1340,7 @@ int eio_init (void (*want_poll)(void), void (*done_poll)(void)) { - etp_init (&etp, want_poll, done_poll); + return etp_init (want_poll, done_poll); } static void eio_api_destroy (eio_req *req) @@ -1087,6 +1372,119 @@ return 0; \ } +static void eio_execute (etp_worker *self, eio_req *req) +{ + errno = 0; + + switch (req->type) + { + case EIO_READ: ALLOC (req->size); + req->result = req->offs >= 0 + ? pread (req->int1, req->ptr2, req->size, req->offs) + : read (req->int1, req->ptr2, req->size); break; + case EIO_WRITE: req->result = req->offs >= 0 + ? pwrite (req->int1, req->ptr2, req->size, req->offs) + : write (req->int1, req->ptr2, req->size); break; + + case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break; + case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break; + + case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT)); + req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; + case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); + req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; + case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); + req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; + + case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; + case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; + case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; + case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; + case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; + case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break; + + case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break; + case EIO_CLOSE: req->result = close (req->int1); break; + case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break; + case EIO_UNLINK: req->result = unlink (req->ptr1); break; + case EIO_RMDIR: req->result = rmdir (req->ptr1); break; + case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break; + case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; + case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; + case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; + case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break; + + case EIO_READLINK: ALLOC (NAME_MAX); + req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; + + case EIO_SYNC: req->result = 0; sync (); break; + case EIO_FSYNC: req->result = fsync (req->int1); break; + case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; + case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break; + case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; + case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break; + + case EIO_READDIR: eio__scandir (req, self); break; + + case EIO_BUSY: +#ifdef _WIN32 + Sleep (req->nv1 * 1000.); +#else + { + struct timeval tv; + + tv.tv_sec = req->nv1; + tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.; + + req->result = select (0, 0, 0, 0, &tv); + } +#endif + break; + + case EIO_UTIME: + case EIO_FUTIME: + { + struct timeval tv[2]; + struct timeval *times; + + if (req->nv1 != -1. || req->nv2 != -1.) + { + tv[0].tv_sec = req->nv1; + tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.; + tv[1].tv_sec = req->nv2; + tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.; + + times = tv; + } + else + times = 0; + + + req->result = req->type == EIO_FUTIME + ? futimes (req->int1, times) + : utimes (req->ptr1, times); + } + break; + + case EIO_GROUP: + abort (); /* handled in eio_request */ + + case EIO_NOP: + req->result = 0; + break; + + case EIO_CUSTOM: + ((void (*)(eio_req *))req->feed) (req); + break; + + default: + req->result = -1; + break; + } + + req->errorno = errno; +} + #ifndef EIO_NO_WRAPPERS eio_req *eio_nop (int pri, eio_cb cb, void *data) @@ -1109,6 +1507,21 @@ REQ (EIO_FSYNC); req->int1 = fd; SEND; } +eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) +{ + REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; +} + +eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) +{ + REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; +} + +eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data) +{ + REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND; +} + eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data) { REQ (EIO_FDATASYNC); req->int1 = fd; SEND; @@ -1230,14 +1643,14 @@ return eio__1path (EIO_RMDIR, path, pri, cb, data); } -eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data) +eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data) { - return eio__1path (EIO_READDIR, path, pri, cb, data); + REQ (EIO_READDIR); PATH; req->int1 = flags; SEND; } eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) { - REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND; + REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND; } static eio_req * @@ -1271,6 +1684,11 @@ return eio__2path (EIO_RENAME, path, new_path, pri, cb, data); } +eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data) +{ + REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND; +} + #endif eio_req *eio_grp (eio_cb cb, void *data) @@ -1306,6 +1724,8 @@ { assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2)); + grp->flags |= EIO_FLAG_GROUPADD; + ++grp->size; req->grp = grp; @@ -1323,7 +1743,7 @@ ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) { - worker wrk; + etp_worker wrk; wrk.dbuf = 0;