--- libeio/eio.c 2008/06/19 09:05:43 1.20 +++ libeio/eio.c 2011/06/20 07:28:15 1.74 @@ -1,7 +1,7 @@ /* * libeio implementation * - * Copyright (c) 2007,2008 Marc Alexander Lehmann + * Copyright (c) 2007,2008,2009,2010,2011 Marc Alexander Lehmann * All rights reserved. * * Redistribution and use in source and binary forms, with or without modifica- @@ -37,7 +37,16 @@ * either the BSD or the GPL. */ +#ifndef _WIN32 +# include "config.h" +#endif + #include "eio.h" +#include "ecb.h" + +#ifdef EIO_STACKSIZE +# define XTHREAD_STACKSIZE EIO_STACKSIZE +#endif #include "xthread.h" #include @@ -47,10 +56,17 @@ #include #include #include +#include #include #include #include +/* intptr_t comes from unistd.h, says POSIX/UNIX/tradition */ +/* intptr_t only comes form stdint.h, says idiot openbsd coder */ +#if HAVE_STDINT_H +# include +#endif + #ifndef EIO_FINISH # define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0 #endif @@ -66,10 +82,8 @@ #ifdef _WIN32 /*doh*/ - #else -# include "config.h" # include # include # include @@ -77,6 +91,28 @@ # include # include +#if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES +# include +#endif + +/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */ +# if __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 __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 @@ -86,26 +122,38 @@ #if HAVE_SENDFILE # if __linux # include -# elif __freebsd +# elif __FreeBSD__ || defined __APPLE__ # include # include # elif __hpux # include -# elif __solaris /* not yet */ +# elif __solaris # include # else # error sendfile support requested but not available # endif #endif -/* number of seconds after which an idle threads exit */ -#define IDLE_TIMEOUT 10 +#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 /* used for struct dirent, AIX doesn't provide it */ #ifndef NAME_MAX # define NAME_MAX 4096 #endif +/* used for readlink etc. */ +#ifndef PATH_MAX +# define PATH_MAX 4096 +#endif + /* buffer size for various temporary buffers */ #define EIO_BUFSIZE 65536 @@ -120,8 +168,6 @@ #define EIO_TICKS ((1000000 + 1023) >> 10) -/*****************************************************************************/ - #define ETP_PRI_MIN EIO_PRI_MIN #define ETP_PRI_MAX EIO_PRI_MAX @@ -146,6 +192,7 @@ closedir (wrk->dirp); \ wrk->dirp = 0; \ } + #define ETP_WORKER_COMMON \ void *dbuf; \ DIR *dirp; @@ -154,8 +201,9 @@ #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) +/* calculate time difference in ~1/EIO_TICKS of a second */ +ecb_inline int +tvdiff (struct timeval *tv1, struct timeval *tv2) { return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS + ((tv2->tv_usec - tv1->tv_usec) >> 10); @@ -172,19 +220,29 @@ 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 volatile unsigned int max_idle = 4; /* maximum number of threads that can idle indefinitely */ +static volatile unsigned int idle_timeout = 10; /* number of seconds after which an idle threads exit */ -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; +static xmutex_t wrklock; +static xmutex_t reslock; +static xmutex_t reqlock; +static xcond_t reqwait; + +#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 xmutex_t preadwritelock = X_MUTEX_INIT; +#endif typedef struct etp_worker { /* locked by wrklock */ struct etp_worker *prev, *next; - thread_t tid; + xthread_t tid; /* locked by reslock, reqlock or wrklock */ ETP_REQ *req; /* currently processed request */ @@ -199,12 +257,14 @@ /* worker threads management */ -static void etp_worker_clear (etp_worker *wrk) +static void ecb_cold +etp_worker_clear (etp_worker *wrk) { ETP_WORKER_CLEAR (wrk); } -static void etp_worker_free (etp_worker *wrk) +static void ecb_cold +etp_worker_free (etp_worker *wrk) { wrk->next->prev = wrk->prev; wrk->prev->next = wrk->next; @@ -212,7 +272,8 @@ free (wrk); } -static unsigned int etp_nreqs (void) +static unsigned int +etp_nreqs (void) { int retval; if (WORDACCESS_UNSAFE) X_LOCK (reqlock); @@ -221,7 +282,8 @@ return retval; } -static unsigned int etp_nready (void) +static unsigned int +etp_nready (void) { unsigned int retval; @@ -232,7 +294,8 @@ return retval; } -static unsigned int etp_npending (void) +static unsigned int +etp_npending (void) { unsigned int retval; @@ -243,7 +306,8 @@ return retval; } -static unsigned int etp_nthreads (void) +static unsigned int +etp_nthreads (void) { unsigned int retval; @@ -267,7 +331,8 @@ static etp_reqq req_queue; static etp_reqq res_queue; -static int reqq_push (etp_reqq *q, ETP_REQ *req) +static int ecb_noinline +reqq_push (etp_reqq *q, ETP_REQ *req) { int pri = req->pri; req->next = 0; @@ -283,7 +348,8 @@ return q->size++; } -static ETP_REQ *reqq_shift (etp_reqq *q) +static ETP_REQ * ecb_noinline +reqq_shift (etp_reqq *q) { int pri; @@ -308,7 +374,17 @@ abort (); } -static void etp_atfork_prepare (void) +static void ecb_cold +etp_thread_init (void) +{ + X_MUTEX_CREATE (wrklock); + X_MUTEX_CREATE (reslock); + X_MUTEX_CREATE (reqlock); + X_COND_CREATE (reqwait); +} + +static void ecb_cold +etp_atfork_prepare (void) { X_LOCK (wrklock); X_LOCK (reqlock); @@ -318,7 +394,8 @@ #endif } -static void etp_atfork_parent (void) +static void ecb_cold +etp_atfork_parent (void) { #if !HAVE_PREADWRITE X_UNLOCK (preadwritelock); @@ -328,14 +405,15 @@ X_UNLOCK (wrklock); } -static void etp_atfork_child (void) +static void ecb_cold +etp_atfork_child (void) { ETP_REQ *prv; - while (prv = reqq_shift (&req_queue)) + while ((prv = reqq_shift (&req_queue))) ETP_DESTROY (prv); - while (prv = reqq_shift (&res_queue)) + while ((prv = reqq_shift (&res_queue))) ETP_DESTROY (prv); while (wrk_first.next != &wrk_first) @@ -355,16 +433,17 @@ nready = 0; npending = 0; - etp_atfork_parent (); + etp_thread_init (); } -static void +static void ecb_cold etp_once_init (void) -{ +{ + etp_thread_init (); X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child); } -static int +static int ecb_cold etp_init (void (*want_poll)(void), void (*done_poll)(void)) { static pthread_once_t doinit = PTHREAD_ONCE_INIT; @@ -379,7 +458,8 @@ X_THREAD_PROC (etp_proc); -static void etp_start_thread (void) +static void ecb_cold +etp_start_thread (void) { etp_worker *wrk = calloc (1, sizeof (etp_worker)); @@ -402,19 +482,21 @@ X_UNLOCK (wrklock); } -static void etp_maybe_start_thread (void) +static void +etp_maybe_start_thread (void) { - if (etp_nthreads () >= wanted) + if (ecb_expect_true (etp_nthreads () >= wanted)) return; /* todo: maybe use idle here, but might be less exact */ - if (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()) + if (ecb_expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ())) return; etp_start_thread (); } -static void etp_end_thread (void) +static void ecb_cold +etp_end_thread (void) { eio_req *req = calloc (1, sizeof (eio_req)); @@ -431,7 +513,8 @@ X_UNLOCK (wrklock); } -static int etp_poll (void) +static int +etp_poll (void) { unsigned int maxreqs; unsigned int maxtime; @@ -471,7 +554,7 @@ --nreqs; X_UNLOCK (reqlock); - if (req->type == EIO_GROUP && req->size) + if (ecb_expect_false (req->type == EIO_GROUP && req->size)) { req->int1 = 1; /* mark request as delayed */ continue; @@ -479,11 +562,11 @@ else { int res = ETP_FINISH (req); - if (res) + if (ecb_expect_false (res)) return res; } - if (maxreqs && !--maxreqs) + if (ecb_expect_false (maxreqs && !--maxreqs)) break; if (maxtime) @@ -499,7 +582,8 @@ return -1; } -static void etp_cancel (ETP_REQ *req) +static void +etp_cancel (ETP_REQ *req) { X_LOCK (wrklock); req->flags |= EIO_FLAG_CANCELLED; @@ -508,51 +592,84 @@ eio_grp_cancel (req); } -static void etp_submit (ETP_REQ *req) +static void +etp_submit (ETP_REQ *req) { req->pri -= ETP_PRI_MIN; - if (req->pri < ETP_PRI_MIN - ETP_PRI_MIN) req->pri = ETP_PRI_MIN - ETP_PRI_MIN; - if (req->pri > ETP_PRI_MAX - ETP_PRI_MIN) req->pri = ETP_PRI_MAX - ETP_PRI_MIN; + if (ecb_expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN; + if (ecb_expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN; - X_LOCK (reqlock); - ++nreqs; - ++nready; - reqq_push (&req_queue, req); - X_COND_SIGNAL (reqwait); - X_UNLOCK (reqlock); + if (ecb_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 (); - etp_maybe_start_thread (); + 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) +static void ecb_cold +etp_set_max_poll_time (double nseconds) { if (WORDACCESS_UNSAFE) X_LOCK (reslock); - max_poll_time = nseconds; + max_poll_time = nseconds * EIO_TICKS; if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); } -static void etp_set_max_poll_reqs (unsigned int maxreqs) +static void ecb_cold +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) +static void ecb_cold +etp_set_max_idle (unsigned int nthreads) { if (WORDACCESS_UNSAFE) X_LOCK (reqlock); - max_idle = nthreads <= 0 ? 1 : nthreads; + max_idle = nthreads; if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); } -static void etp_set_min_parallel (unsigned int nthreads) +static void ecb_cold +etp_set_idle_timeout (unsigned int seconds) +{ + if (WORDACCESS_UNSAFE) X_LOCK (reqlock); + idle_timeout = seconds; + if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock); +} + +static void ecb_cold +etp_set_min_parallel (unsigned int nthreads) { if (wanted < nthreads) wanted = nthreads; } -static void etp_set_max_parallel (unsigned int nthreads) +static void ecb_cold +etp_set_max_parallel (unsigned int nthreads) { if (wanted > nthreads) wanted = nthreads; @@ -563,16 +680,17 @@ /*****************************************************************************/ -static void grp_try_feed (eio_req *grp) +static void +grp_try_feed (eio_req *grp) { 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; @@ -580,7 +698,8 @@ } } -static int grp_dec (eio_req *grp) +static int +grp_dec (eio_req *grp) { --grp->size; @@ -594,7 +713,8 @@ return 0; } -void eio_destroy (eio_req *req) +void +eio_destroy (eio_req *req) { if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1); if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2); @@ -602,7 +722,8 @@ EIO_DESTROY (req); } -static int eio_finish (eio_req *req) +static int +eio_finish (eio_req *req) { int res = EIO_FINISH (req); @@ -629,63 +750,81 @@ return res; } -void eio_grp_cancel (eio_req *grp) +void +eio_grp_cancel (eio_req *grp) { for (grp = grp->grp_first; grp; grp = grp->grp_next) eio_cancel (grp); } -void eio_cancel (eio_req *req) +void +eio_cancel (eio_req *req) { etp_cancel (req); } -void eio_submit (eio_req *req) +void +eio_submit (eio_req *req) { etp_submit (req); } -unsigned int eio_nreqs (void) +unsigned int +eio_nreqs (void) { return etp_nreqs (); } -unsigned int eio_nready (void) +unsigned int +eio_nready (void) { return etp_nready (); } -unsigned int eio_npending (void) +unsigned int +eio_npending (void) { return etp_npending (); } -unsigned int eio_nthreads (void) +unsigned int ecb_cold +eio_nthreads (void) { return etp_nthreads (); } -void eio_set_max_poll_time (double nseconds) +void ecb_cold +eio_set_max_poll_time (double nseconds) { etp_set_max_poll_time (nseconds); } -void eio_set_max_poll_reqs (unsigned int maxreqs) +void ecb_cold +eio_set_max_poll_reqs (unsigned int maxreqs) { etp_set_max_poll_reqs (maxreqs); } -void eio_set_max_idle (unsigned int nthreads) +void ecb_cold +eio_set_max_idle (unsigned int nthreads) { etp_set_max_idle (nthreads); } -void eio_set_min_parallel (unsigned int nthreads) +void ecb_cold +eio_set_idle_timeout (unsigned int seconds) +{ + etp_set_idle_timeout (seconds); +} + +void ecb_cold +eio_set_min_parallel (unsigned int nthreads) { etp_set_min_parallel (nthreads); } -void eio_set_max_parallel (unsigned int nthreads) +void ecb_cold +eio_set_max_parallel (unsigned int nthreads) { etp_set_max_parallel (nthreads); } @@ -699,16 +838,11 @@ /* 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) { @@ -735,17 +869,17 @@ 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; } #endif -#ifndef HAVE_FUTIMES +#ifndef HAVE_UTIMES -# define utimes(path,times) eio__utimes (path, times) -# define futimes(fd,times) eio__futimes (fd, times) +# undef utimes +# define utimes(path,times) eio__utimes (path, times) static int eio__utimes (const char *filename, const struct timeval times[2]) @@ -763,7 +897,15 @@ return utime (filename, 0); } -static int eio__futimes (int fd, const struct timeval tv[2]) +#endif + +#ifndef HAVE_FUTIMES + +# undef futimes +# define futimes(fd,times) eio__futimes (fd, times) + +static int +eio__futimes (int fd, const struct timeval tv[2]) { errno = ENOSYS; return -1; @@ -772,10 +914,40 @@ #endif #if !HAVE_FDATASYNC -# define fdatasync fsync +# undef fdatasync +# define fdatasync(fd) fsync (fd) #endif +/* sync_file_range always needs emulation */ +static 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 that matches the expectation of its users best */ + return fdatasync (fd); +} + #if !HAVE_READAHEAD +# undef readahead # define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) static ssize_t @@ -803,56 +975,110 @@ static ssize_t eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self) { + ssize_t written = 0; ssize_t res; if (!count) return 0; + for (;;) + { #if HAVE_SENDFILE # if __linux - res = sendfile (ofd, ifd, &offset, count); + off_t soffset = offset; + res = sendfile (ofd, ifd, &soffset, 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: 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 0 /* according to the manpage, this is correct, but broken behaviour */ + /* freebsd' sendfile will return 0 on success */ + /* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */ + /* not on e.g. EIO or EPIPE - sounds broken */ + if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0) + res = sbytes; + #endif + + /* according to source inspection, this is correct, and useful behaviour */ + if (sbytes) + res = sbytes; + +# elif defined (__APPLE__) + off_t sbytes = count; + res = sendfile (ifd, ofd, offset, &sbytes, 0, 0); + + /* according to the manpage, sbytes is always valid */ + if (sbytes) + res = sbytes; # elif __hpux - res = sendfile (ofd, ifd, offset, count, 0, 0); + res = sendfile (ofd, ifd, offset, count, 0, 0); # elif __solaris - { - struct sendfilevec vec; - size_t sbytes; + struct sendfilevec vec; + size_t sbytes; - vec.sfv_fd = ifd; - vec.sfv_flag = 0; - vec.sfv_off = offset; - vec.sfv_len = count; + vec.sfv_fd = ifd; + vec.sfv_flag = 0; + vec.sfv_off = offset; + vec.sfv_len = count; - res = sendfilev (ofd, &vec, 1, &sbytes); + res = sendfilev (ofd, &vec, 1, &sbytes); - if (res < 0 && sbytes) - res = sbytes; - } + if (res < 0 && sbytes) + res = sbytes; # endif + +#elif defined (_WIN32) + /* does not work, just for documentation of what would need to be done */ + /* actually, cannot be done like this, as TransmitFile changes the file offset, */ + /* libeio guarantees that the file offset does not change, and windows */ + /* has no way to get an independent handle to the same file description */ + HANDLE h = TO_SOCKET (ifd); + SetFilePointer (h, offset, 0, FILE_BEGIN); + res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0); + #else - res = -1; - errno = ENOSYS; + res = -1; + errno = ENOSYS; #endif - if (res < 0 + /* we assume sendfile can copy at least 128mb in one go */ + if (res <= 128 * 1024 * 1024) + { + if (res > 0) + written += res; + + if (written) + return written; + + break; + } + else + { + /* if we requested more, then probably the kernel was lazy */ + written += res; + offset += res; + count -= res; + + if (!count) + return written; + } + } + + if (res < 0 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK + /* BSDs */ +#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */ + || errno == ENOTSUP +#endif + || errno == EOPNOTSUPP /* BSDs */ #if __solaris || errno == EAFNOSUPPORT || errno == EPROTOTYPE #endif @@ -893,6 +1119,171 @@ return res; } +static signed char +eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) +{ + return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */ + : a->inode < b->inode ? -1 + : a->inode > b->inode ? 1 + : 0; +} + +#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 + +#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ +#define EIO_SORT_FAST 60 /* when to only use insertion sort */ + +static void +eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) +{ + unsigned char bits [9 + sizeof (ino_t) * 8]; + unsigned char *bit = bits; + + assert (CHAR_BIT == 8); + assert (sizeof (eio_dirent) * 8 < 256); + assert (offsetof (eio_dirent, inode)); /* we use bit #0 as sentinel */ + assert (offsetof (eio_dirent, score)); /* we use bit #0 as sentinel */ + + if (size <= EIO_SORT_FAST) + return; + + /* first prepare an array of bits to test in our radix sort */ + /* try to take endianness into account, as well as differences in ino_t sizes */ + /* inode_bits must contain all inodes ORed together */ + /* which is used to skip bits that are 0 everywhere, which is very common */ + { + ino_t endianness; + int i, j; + + /* we store the byte offset of byte n into byte n of "endianness" */ + for (i = 0; i < sizeof (ino_t); ++i) + ((unsigned char *)&endianness)[i] = i; + + *bit++ = 0; + + for (i = 0; i < sizeof (ino_t); ++i) + { + /* shifting off the byte offsets out of "endianness" */ + int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8; + endianness >>= 8; + + for (j = 0; j < 8; ++j) + if (inode_bits & (((ino_t)1) << (i * 8 + j))) + *bit++ = offs + j; + } + + for (j = 0; j < 8; ++j) + if (score_bits & (1 << j)) + *bit++ = offsetof (eio_dirent, score) * 8 + j; + } + + /* now actually do the sorting (a variant of MSD radix sort) */ + { + eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base; + eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end; + unsigned char *bit_stk [9 + sizeof (ino_t) * 8]; + int stk_idx = 0; + + base_stk [stk_idx] = dents; + end_stk [stk_idx] = dents + size; + bit_stk [stk_idx] = bit - 1; + + do + { + base = base_stk [stk_idx]; + end = end_stk [stk_idx]; + bit = bit_stk [stk_idx]; + + for (;;) + { + unsigned char O = *bit >> 3; + unsigned char M = 1 << (*bit & 7); + + eio_dirent *a = base; + eio_dirent *b = end; + + if (b - a < EIO_SORT_CUTOFF) + break; + + /* now bit-partition the array on the bit */ + /* this ugly asymmetric loop seems to perform much better than typical */ + /* partition algos found in the literature */ + do + if (!(((unsigned char *)a)[O] & M)) + ++a; + else if (!(((unsigned char *)--b)[O] & M)) + { + eio_dirent tmp = *a; *a = *b; *b = tmp; + ++a; + } + while (b > a); + + /* next bit, or stop, if no bits left in this path */ + if (!*--bit) + break; + + base_stk [stk_idx] = a; + end_stk [stk_idx] = end; + bit_stk [stk_idx] = bit; + ++stk_idx; + + end = a; + } + } + while (stk_idx--); + } +} + +static void +eio_dent_insertion_sort (eio_dirent *dents, int size) +{ + /* first move the smallest element to the front, to act as a sentinel */ + { + int i; + eio_dirent *min = dents; + + /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */ + for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; ) + if (EIO_DENT_CMP (dents [i], <, *min)) + min = &dents [i]; + + /* swap elements 0 and j (minimum) */ + { + eio_dirent tmp = *dents; *dents = *min; *min = tmp; + } + } + + /* then do standard insertion sort, assuming that all elements are >= dents [0] */ + { + eio_dirent *i, *j; + + for (i = dents + 1; i < dents + size; ++i) + { + eio_dirent value = *i; + + for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j) + j [1] = j [0]; + + j [1] = value; + } + } +} + +static void +eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) +{ + if (size <= 1) + return; /* our insertion sort relies on size > 0 */ + + /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */ + /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */ + eio_dent_radix_sort (dents, size, score_bits, inode_bits); + + /* use an insertion sort at the end, or for small arrays, */ + /* as insertion sort is more efficient for small partitions */ + eio_dent_insertion_sort (dents, size); +} + /* read a full directory */ static void eio__scandir (eio_req *req, etp_worker *self) @@ -900,54 +1291,315 @@ DIR *dirp; EIO_STRUCT_DIRENT *entp; char *name, *names; - int memlen = 4096; - int memofs = 0; - int res = 0; + int namesalloc = 4096; + int namesoffs = 0; + int flags = req->int1; + eio_dirent *dents = 0; + int dentalloc = 128; + int dentoffs = 0; + ino_t inode_bits = 0; + + 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); - req->flags |= EIO_FLAG_PTR2_FREE; - req->ptr2 = names = malloc (memlen); + + if (req->flags & EIO_FLAG_PTR1_FREE) + free (req->ptr1); + + req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; + req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; + req->ptr2 = names = malloc (namesalloc); X_UNLOCK (wrklock); - if (dirp && names) + if (dirp && names && (!flags || dents)) for (;;) { errno = 0; entp = readdir (dirp); if (!entp) - break; + { + if (errno) + break; + /* sort etc. */ + req->int1 = flags; + req->result = dentoffs; + + if (flags & EIO_READDIR_STAT_ORDER) + eio_dent_sort (dents, dentoffs, flags & EIO_READDIR_DIRS_FIRST ? 7 : 0, inode_bits); + else if (flags & EIO_READDIR_DIRS_FIRST) + if (flags & EIO_READDIR_FOUND_UNKNOWN) + eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */ + else + { + /* in this case, all is known, and we just put dirs first and sort them */ + eio_dirent *oth = dents + dentoffs; + eio_dirent *dir = dents; + + /* now partition dirs to the front, and non-dirs to the back */ + /* by walking from both sides and swapping if necessary */ + while (oth > dir) + { + if (dir->type == EIO_DT_DIR) + ++dir; + else if ((--oth)->type == EIO_DT_DIR) + { + eio_dirent tmp = *dir; *dir = *oth; *oth = tmp; + + ++dir; + } + } + + /* now sort the dirs only (dirs all have the same score) */ + eio_dent_sort (dents, dir - dents, 0, inode_bits); + } + + 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; - - res++; + int len = D_NAMLEN (entp) + 1; - while (memofs + len > memlen) + while (ecb_expect_false (namesoffs + len > namesalloc)) { - memlen *= 2; + namesalloc *= 2; X_LOCK (wrklock); - req->ptr2 = names = realloc (names, memlen); + req->ptr2 = 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 (ecb_expect_false (dentoffs == dentalloc)) + { + dentalloc *= 2; + X_LOCK (wrklock); + req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); + X_UNLOCK (wrklock); + + if (!dents) + break; + } + + ent = dents + dentoffs; + + ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */ + ent->namelen = len - 1; + ent->inode = D_INO (entp); + + inode_bits |= ent->inode; + + 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 = 7; + + 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 = 1; + else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ + ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */ + } + else if (ent->type == EIO_DT_DIR) + ent->score = 0; + } + } + + namesoffs += len; + ++dentoffs; + } + + if (EIO_CANCELLED (req)) + { + errno = ECANCELED; + break; } } +} - if (errno) - res = -1; - - req->result = res; +#ifdef PAGESIZE +# define eio_pagesize() PAGESIZE +#else +static intptr_t +eio_pagesize (void) +{ + static intptr_t page; + + if (!page) + page = sysconf (_SC_PAGESIZE); + + return page; +} +#endif + +static void +eio_page_align (void **addr, size_t *length) +{ + intptr_t mask = eio_pagesize () - 1; + + /* round down addr */ + intptr_t adj = mask & (intptr_t)*addr; + + *addr = (void *)((intptr_t)*addr - adj); + *length += adj; + + /* round up length */ + *length = (*length + mask) & ~mask; +} + +#if !_POSIX_MEMLOCK +# define eio__mlockall(a) ((errno = ENOSYS), -1) +#else + +static int +eio__mlockall (int flags) +{ + #if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7 + extern int mallopt (int, int); + mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */ + #endif + + if (EIO_MCL_CURRENT != MCL_CURRENT + || EIO_MCL_FUTURE != MCL_FUTURE) + { + flags = 0 + | (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0) + | (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0); + } + + return mlockall (flags); +} +#endif + +#if !_POSIX_MEMLOCK_RANGE +# define eio__mlock(a,b) ((errno = ENOSYS), -1) +#else + +static int +eio__mlock (void *addr, size_t length) +{ + eio_page_align (&addr, &length); + + return mlock (addr, length); +} + +#endif + +#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) +# define eio__msync(a,b,c) ((errno = ENOSYS), -1) +#else + +static int +eio__msync (void *mem, size_t len, int flags) +{ + eio_page_align (&mem, &len); + + if (EIO_MS_ASYNC != MS_SYNC + || EIO_MS_INVALIDATE != MS_INVALIDATE + || EIO_MS_SYNC != MS_SYNC) + { + flags = 0 + | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0) + | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0) + | (flags & EIO_MS_SYNC ? MS_SYNC : 0); + } + + return msync (mem, len, flags); +} + +#endif + +static int +eio__mtouch (eio_req *req) +{ + void *mem = req->ptr2; + size_t len = req->size; + int flags = req->int1; + + eio_page_align (&mem, &len); + + { + intptr_t addr = (intptr_t)mem; + intptr_t end = addr + len; + intptr_t page = eio_pagesize (); + + if (addr < end) + if (flags & EIO_MT_MODIFY) /* modify */ + do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len && !EIO_CANCELLED (req)); + else + do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len && !EIO_CANCELLED (req)); + } + + return 0; } /*****************************************************************************/ @@ -989,7 +1641,7 @@ ++idle; - ts.tv_sec = time (0) + IDLE_TIMEOUT; + ts.tv_sec = time (0) + idle_timeout; if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT) { if (idle > max_idle) @@ -1042,12 +1694,14 @@ /*****************************************************************************/ -int eio_init (void (*want_poll)(void), void (*done_poll)(void)) +int ecb_cold +eio_init (void (*want_poll)(void), void (*done_poll)(void)) { return etp_init (want_poll, done_poll); } -static void eio_api_destroy (eio_req *req) +ecb_inline void +eio_api_destroy (eio_req *req) { free (req); } @@ -1076,10 +1730,9 @@ return 0; \ } -static void eio_execute (etp_worker *self, eio_req *req) +static void +eio_execute (etp_worker *self, eio_req *req) { - errno = 0; - switch (req->type) { case EIO_READ: ALLOC (req->size); @@ -1100,6 +1753,11 @@ case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; + case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS)); + req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break; + case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS)); + req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)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; @@ -1116,26 +1774,31 @@ 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_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_READLINK: ALLOC (PATH_MAX); + req->result = readlink (req->ptr1, req->ptr2, PATH_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 = eio__msync (req->ptr2, req->size, req->int1); break; + case EIO_MTOUCH: req->result = eio__mtouch (req); break; + case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break; + case EIO_MLOCKALL: req->result = eio__mlockall (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.); + Sleep (req->nv1 * 1e3); #else { struct timeval tv; tv.tv_sec = req->nv1; - tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.; + tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6; req->result = select (0, 0, 0, 0, &tv); } @@ -1160,13 +1823,15 @@ 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; @@ -1176,6 +1841,7 @@ break; default: + errno = ENOSYS; req->result = -1; break; } @@ -1205,6 +1871,31 @@ 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_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data) +{ + REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND; +} + +eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data) +{ + REQ (EIO_MLOCKALL); 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; @@ -1235,6 +1926,11 @@ REQ (EIO_FSTAT); req->int1 = fd; SEND; } +eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data) +{ + REQ (EIO_FSTATVFS); req->int1 = fd; SEND; +} + eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) { REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND; @@ -1316,6 +2012,11 @@ return eio__1path (EIO_LSTAT, path, pri, cb, data); } +eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data) +{ + return eio__1path (EIO_STATVFS, path, pri, cb, data); +} + eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) { return eio__1path (EIO_UNLINK, path, pri, cb, data); @@ -1326,14 +2027,14 @@ 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->int3 = (long)dev; SEND; + REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->offs = (off_t)dev; SEND; } static eio_req * @@ -1367,7 +2068,7 @@ 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) +eio_req *eio_custom (void (*execute)(eio_req *), int pri, eio_cb cb, void *data) { REQ (EIO_CUSTOM); req->feed = execute; SEND; } @@ -1388,7 +2089,8 @@ /*****************************************************************************/ /* grp functions */ -void eio_grp_feed (eio_req *grp, void (*feed)(eio_req *req), int limit) +void +eio_grp_feed (eio_req *grp, void (*feed)(eio_req *req), int limit) { grp->int2 = limit; grp->feed = feed; @@ -1396,17 +2098,21 @@ grp_try_feed (grp); } -void eio_grp_limit (eio_req *grp, int limit) +void +eio_grp_limit (eio_req *grp, int limit) { grp->int2 = limit; grp_try_feed (grp); } -void eio_grp_add (eio_req *grp, eio_req *req) +void +eio_grp_add (eio_req *grp, eio_req *req) { 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; @@ -1422,15 +2128,19 @@ /*****************************************************************************/ /* misc garbage */ -ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) +ssize_t +eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) { etp_worker wrk; + ssize_t ret; wrk.dbuf = 0; - eio__sendfile (ofd, ifd, offset, count, &wrk); + ret = eio__sendfile (ofd, ifd, offset, count, &wrk); if (wrk.dbuf) free (wrk.dbuf); + + return ret; }