--- IO-AIO/AIO.xs 2012/04/01 17:46:02 1.209 +++ IO-AIO/AIO.xs 2012/10/11 03:20:52 1.224 @@ -21,6 +21,15 @@ # include #endif +#if __linux__ +# include +# ifdef FS_IOC_FIEMAP +# include +# include +# define HAVE_FIEMAP 1 +# endif +#endif + /* perl namespace pollution */ #undef VERSION @@ -79,17 +88,6 @@ #endif -/* use NV for 32 bit perls as it allows larger offsets */ -#if IVSIZE >= 8 -# define VAL64 IV -# define SvVAL64 SvIV -# define newSVval64 newSViv -#else -# define VAL64 NV -# define SvVAL64 SvNV -# define newSVval64 newSVnv -#endif - /*****************************************************************************/ #if __GNUC__ >= 3 @@ -128,170 +126,14 @@ #include "libeio/eio.c" -/* Linux/others */ -#ifndef O_ASYNC -# define O_ASYNC 0 -#endif -#ifndef O_DIRECT -# define O_DIRECT 0 -#endif -#ifndef O_NOATIME -# define O_NOATIME 0 -#endif - -/* POSIX */ -#ifndef O_CLOEXEC -# define O_CLOEXEC 0 -#endif -#ifndef O_NOFOLLOW -# define O_NOFOLLOW 0 -#endif -#ifndef O_NOCTTY -# define O_NOCTTY 0 -#endif -#ifndef O_NONBLOCK -# define O_NONBLOCK 0 -#endif -#ifndef O_EXEC -# define O_EXEC 0 -#endif -#ifndef O_SEARCH -# define O_SEARCH 0 -#endif -#ifndef O_DIRECTORY -# define O_DIRECTORY 0 -#endif -#ifndef O_DSYNC -# define O_DSYNC 0 -#endif -#ifndef O_RSYNC -# define O_RSYNC 0 -#endif -#ifndef O_SYNC -# define O_SYNC 0 -#endif -#ifndef O_TTY_INIT -# define O_TTY_INIT 0 -#endif - -#ifndef POSIX_FADV_NORMAL -# define POSIX_FADV_NORMAL 0 -#endif -#ifndef POSIX_FADV_SEQUENTIAL -# define POSIX_FADV_SEQUENTIAL 0 -#endif -#ifndef POSIX_FADV_RANDOM -# define POSIX_FADV_RANDOM 0 -#endif -#ifndef POSIX_FADV_NOREUSE -# define POSIX_FADV_NOREUSE 0 -#endif -#ifndef POSIX_FADV_WILLNEED -# define POSIX_FADV_WILLNEED 0 -#endif -#ifndef POSIX_FADV_DONTNEED -# define POSIX_FADV_DONTNEED 0 -#endif - #if !HAVE_POSIX_FADVISE # define posix_fadvise(a,b,c,d) errno = ENOSYS /* also return ENOSYS */ #endif -#ifndef POSIX_MADV_NORMAL -# define POSIX_MADV_NORMAL 0 -#endif -#ifndef POSIX_MADV_SEQUENTIAL -# define POSIX_MADV_SEQUENTIAL 0 -#endif -#ifndef POSIX_MADV_RANDOM -# define POSIX_MADV_RANDOM 0 -#endif -#ifndef POSIX_MADV_WILLNEED -# define POSIX_MADV_WILLNEED 0 -#endif -#ifndef POSIX_MADV_DONTNEED -# define POSIX_MADV_DONTNEED 0 -#endif - #if !HAVE_POSIX_MADVISE # define posix_madvise(a,b,c) errno = ENOSYS /* also return ENOSYS */ #endif -#ifndef PROT_NONE -# define PROT_NONE 0 -#endif -#ifndef PROT_READ -# define PROT_READ 0 -#endif -#ifndef PROT_WRITE -# define PROT_READ 0 -#endif -#ifndef PROT_EXEC -# define PROT_EXEC 0 -#endif - -#ifndef ST_RDONLY -# define ST_RDONLY 0 -#endif -#ifndef ST_NOSUID -# define ST_NOSUID 0 -#endif -#ifndef ST_NODEV -# define ST_NODEV 0 -#endif -#ifndef ST_NOEXEC -# define ST_NOEXEC 0 -#endif -#ifndef ST_SYNCHRONOUS -# define ST_SYNCHRONOUS 0 -#endif -#ifndef ST_MANDLOCK -# define ST_MANDLOCK 0 -#endif -#ifndef ST_WRITE -# define ST_WRITE 0 -#endif -#ifndef ST_APPEND -# define ST_APPEND 0 -#endif -#ifndef ST_IMMUTABLE -# define ST_IMMUTABLE 0 -#endif -#ifndef ST_NOATIME -# define ST_NOATIME 0 -#endif -#ifndef ST_NODIRATIME -# define ST_NODIRATIME 0 -#endif -#ifndef ST_RELATIME -# define ST_RELATIME 0 -#endif - -#ifndef S_IFIFO -# define S_IFIFO 0 -#endif -#ifndef S_IFCHR -# define S_IFCHR 0 -#endif -#ifndef S_IFBLK -# define S_IFBLK 0 -#endif -#ifndef S_IFLNK -# define S_IFLNK 0 -#endif -#ifndef S_IFREG -# define S_IFREG 0 -#endif -#ifndef S_IFDIR -# define S_IFDIR 0 -#endif -#ifndef S_IFWHT -# define S_IFWHT 0 -#endif -#ifndef S_IFSOCK -# define S_IFSOCK 0 -#endif - #ifndef MAP_ANONYMOUS # ifdef MAP_ANON # define MAP_ANONYMOUS MAP_ANON @@ -299,21 +141,9 @@ # define MAP_ANONYMOUS MAP_FIXED /* and hope this fails */ # endif #endif -#ifndef MAP_HUGETLB -# define MAP_HUGETLB 0 -#endif -#ifndef MAP_LOCKED -# define MAP_LOCKED 0 -#endif -#ifndef MAP_NORESERVE -# define MAP_NORESERVE 0 -#endif -#ifndef MAP_POPULATE -# define MAP_POPULATE 0 -#endif -#ifndef MAP_NONBLOCK -# define MAP_NONBLOCK 0 -#endif + +/* defines all sorts of constants to 0 unless they are already defined */ +#include "def0.h" #ifndef makedev # define makedev(maj,min) (((maj) << 8) | (min)) @@ -329,6 +159,103 @@ # define PAGESIZE sysconf (_SC_PAGESIZE) #endif +/*****************************************************************************/ + +static void +fiemap (eio_req *req) +{ + req->result = -1; + +#if HAVE_FIEMAP + /* assume some c99 */ + struct fiemap *fiemap = 0; + size_t end_offset; + int count = req->int3; + + req->flags |= EIO_FLAG_PTR1_FREE; + + /* heuristic: start with 512 bytes (8 extents), and if that isn't enough, */ + /* increase in 3.5kb steps */ + if (count < 0) + count = 8; + + fiemap = malloc (sizeof (*fiemap) + sizeof (struct fiemap_extent) * count); + errno = ENOMEM; + if (!fiemap) + return; + + req->ptr1 = fiemap; + + fiemap->fm_start = req->offs; + fiemap->fm_length = req->size; + fiemap->fm_flags = req->int2; + fiemap->fm_extent_count = count; + + if (ioctl (req->int1, FS_IOC_FIEMAP, fiemap) < 0) + return; + + if (req->int3 >= 0 /* not autosizing */ + || !fiemap->fm_mapped_extents /* no more extents */ + || fiemap->fm_extents [fiemap->fm_mapped_extents - 1].fe_flags & FIEMAP_EXTENT_LAST /* hit eof */) + goto done; + + /* else we have to loop - + * it would be tempting (atcually I tried that first) to just query the + * number of extents needed, but linux often feels like not returning all + * extents, without telling us it left any out. this complicates + * this quite a bit. + */ + + end_offset = fiemap->fm_length + (fiemap->fm_length == FIEMAP_MAX_OFFSET ? 0 : fiemap->fm_start); + + for (;;) + { + /* we go in 54 extent steps - 3kb, in the hope that this fits nicely on the eio stack (normally 16+ kb) */ + char scratch[3072]; + struct fiemap *incmap = (struct fiemap *)scratch; + + incmap->fm_start = fiemap->fm_extents [fiemap->fm_mapped_extents - 1].fe_logical + + fiemap->fm_extents [fiemap->fm_mapped_extents - 1].fe_length; + incmap->fm_length = fiemap->fm_length - (incmap->fm_start - fiemap->fm_start); + incmap->fm_flags = fiemap->fm_flags; + incmap->fm_extent_count = (sizeof (scratch) - sizeof (struct fiemap)) / sizeof (struct fiemap_extent); + + if (ioctl (req->int1, FS_IOC_FIEMAP, incmap) < 0) + return; + + count = fiemap->fm_mapped_extents + incmap->fm_mapped_extents; + fiemap = realloc (fiemap, sizeof (*fiemap) + sizeof (struct fiemap_extent) * count); + errno = ENOMEM; + if (!fiemap) + return; + + req->ptr1 = fiemap; + + for (count = 0; count < incmap->fm_mapped_extents; ++count) + { + struct fiemap_extent *e = incmap->fm_extents + count; + + if (e->fe_logical + e->fe_length >= end_offset) + goto done; + + fiemap->fm_extents [fiemap->fm_mapped_extents++] = *e; + + if (e->fe_flags & FIEMAP_EXTENT_LAST) + goto done; + + } + } + +done: + req->result = 0; + +#else + errno = ENOSYS; +#endif +} + +/*****************************************************************************/ + enum { FLAG_SV2_RO_OFF = 0x40, /* data was set readonly */ }; @@ -582,6 +509,11 @@ { EIO_STRUCT_STATVFS *f = EIO_STATVFS_BUF (req); HV *hv = newHV (); + /* POSIX requires fsid to be unsigned long, but AIX in its infinite wisdom + * chooses to make it a struct. + */ + unsigned long fsid = 0; + memcpy (&fsid, &f->f_fsid, sizeof (unsigned long) < sizeof (f->f_fsid) ? sizeof (unsigned long) : sizeof (f->f_fsid)); rv = sv_2mortal (newRV_noinc ((SV *)hv)); @@ -593,7 +525,7 @@ hv_store (hv, "files" , sizeof ("files" ) - 1, newSVval64 (f->f_files ), 0); hv_store (hv, "ffree" , sizeof ("ffree" ) - 1, newSVval64 (f->f_ffree ), 0); hv_store (hv, "favail" , sizeof ("favail" ) - 1, newSVval64 (f->f_favail ), 0); - hv_store (hv, "fsid" , sizeof ("fsid" ) - 1, newSVval64 (f->f_fsid ), 0); + hv_store (hv, "fsid" , sizeof ("fsid" ) - 1, newSVval64 (fsid ), 0); hv_store (hv, "flag" , sizeof ("flag" ) - 1, newSVval64 (f->f_flag ), 0); hv_store (hv, "namemax", sizeof ("namemax") - 1, newSVval64 (f->f_namemax), 0); } @@ -654,6 +586,46 @@ } break; + case EIO_CUSTOM: + if (req->feed == fiemap) + { +#if HAVE_FIEMAP + if (!req->result) + { + struct fiemap *fiemap = (struct fiemap *)req->ptr1; + + if (fiemap->fm_extent_count) + { + AV *av = newAV (); + int i; + + while (fiemap->fm_mapped_extents) + { + struct fiemap_extent *extent = &fiemap->fm_extents [--fiemap->fm_mapped_extents]; + AV *ext_av = newAV (); + + av_store (ext_av, 3, newSVuv (extent->fe_flags)); + av_store (ext_av, 2, newSVval64 (extent->fe_length)); + av_store (ext_av, 1, newSVval64 (extent->fe_physical)); + av_store (ext_av, 0, newSVval64 (extent->fe_logical)); + + av_store (av, fiemap->fm_mapped_extents, newRV_noinc ((SV *)ext_av)); + } + + PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); + } + else + { + SvIV_set (sv_result, fiemap->fm_mapped_extents); + PUSHs (sv_result); + } + } +#endif + } + else + PUSHs (sv_result); + break; + case EIO_DUP2: /* EIO_DUP2 actually means aio_close(), so fudge result value */ if (req->result > 0) SvIV_set (sv_result, 0); @@ -770,12 +742,12 @@ /*****************************************************************************/ #if !_POSIX_MAPPED_FILES -# define mmap(addr,length,prot,flags,fd,offs) (errno = ENOSYS, -1) -# define munmap(addr,length) (errno = ENOSYS, -1) +# define mmap(addr,length,prot,flags,fd,offs) EIO_ENOSYS () +# define munmap(addr,length) EIO_ENOSYS () #endif #if !_POSIX_MEMORY_PROTECTION -# define mprotect(addr,len,prot) (errno = ENOSYS, -1) +# define mprotect(addr,len,prot) EIO_ENOSYS () # define PROT_NONE 0 # define PROT_WRITE 0 # define MAP_PRIVATE 0 @@ -929,8 +901,13 @@ # define const_niv(name, value) { # name, (IV) value }, # define const_iv(name) { # name, (IV) name }, # define const_eio(name) { # name, (IV) EIO_ ## name }, - const_iv (EXDEV) + + /* you have to re-run ./gendef0 after adding/Removing any constants here */ + const_iv (ENOSYS) + const_iv (EXDEV) + const_iv (EBADR) + const_iv (O_RDONLY) const_iv (O_WRONLY) const_iv (O_RDWR) @@ -1008,6 +985,30 @@ const_iv (MAP_POPULATE) const_iv (MAP_NONBLOCK) + const_iv (FIEMAP_FLAG_SYNC) + const_iv (FIEMAP_FLAG_XATTR) + const_iv (FIEMAP_FLAGS_COMPAT) + const_iv (FIEMAP_EXTENT_LAST) + const_iv (FIEMAP_EXTENT_UNKNOWN) + const_iv (FIEMAP_EXTENT_DELALLOC) + const_iv (FIEMAP_EXTENT_ENCODED) + const_iv (FIEMAP_EXTENT_DATA_ENCRYPTED) + const_iv (FIEMAP_EXTENT_NOT_ALIGNED) + const_iv (FIEMAP_EXTENT_DATA_INLINE) + const_iv (FIEMAP_EXTENT_DATA_TAIL) + const_iv (FIEMAP_EXTENT_UNWRITTEN) + const_iv (FIEMAP_EXTENT_MERGED) + const_iv (FIEMAP_EXTENT_SHARED) + + const_iv (SPLICE_F_MOVE) + const_iv (SPLICE_F_NONBLOCK) + const_iv (SPLICE_F_MORE) + const_iv (SPLICE_F_GIFT) + + const_iv (SEEK_DATA) + const_iv (SEEK_HOLE) + + /* libeio constants */ const_eio (SEEK_SET) const_eio (SEEK_CUR) const_eio (SEEK_END) @@ -1026,6 +1027,7 @@ const_eio (SYNC_FILE_RANGE_WAIT_AFTER) const_eio (FALLOC_FL_KEEP_SIZE) + const_eio (FALLOC_FL_PUNCH_HOLE) const_eio (READDIR_DENTS) const_eio (READDIR_DIRS_FIRST) @@ -1136,7 +1138,7 @@ aio_syncfs = EIO_SYNCFS PPCODE: { - int fd = s_fileno_croak (fh, 0); + int fd = s_fileno_croak (fh, 0); dREQ; req->type = ix; @@ -1150,7 +1152,7 @@ aio_sync_file_range (SV *fh, off_t offset, size_t nbytes, UV flags, SV *callback=&PL_sv_undef) PPCODE: { - int fd = s_fileno_croak (fh, 0); + int fd = s_fileno_croak (fh, 0); dREQ; req->type = EIO_SYNC_FILE_RANGE; @@ -1164,10 +1166,10 @@ } void -aio_fallocate (SV *fh, int mode, off_t offset, size_t len, SV *callback=&PL_sv_undef) +aio_allocate (SV *fh, int mode, off_t offset, size_t len, SV *callback=&PL_sv_undef) PPCODE: { - int fd = s_fileno_croak (fh, 0); + int fd = s_fileno_croak (fh, 0); dREQ; req->type = EIO_FALLOCATE; @@ -1185,7 +1187,7 @@ PPCODE: { static int close_fd = -1; /* dummy fd to close fds via dup2 */ - int fd = s_fileno_croak (fh, 0); + int fd = s_fileno_croak (fh, 0); dREQ; if (expect_false (close_fd < 0)) @@ -1218,7 +1220,6 @@ aio_seek (SV *fh, SV *offset, int whence, SV *callback=&PL_sv_undef) PPCODE: { - STRLEN svlen; int fd = s_fileno_croak (fh, 0); dREQ; @@ -1259,7 +1260,10 @@ { /* read: check type and grow scalar as necessary */ SvUPGRADE (data, SVt_PV); - svptr = SvGROW (data, len + dataoffset + 1); + if (SvLEN (data) >= SvCUR (data)) + svptr = SvGROW (data, len + dataoffset + 1); + else if (SvCUR (data) < len + dataoffset) + croak ("length + dataoffset outside of scalar, and cannot grow"); } { @@ -1303,8 +1307,8 @@ aio_sendfile (SV *out_fh, SV *in_fh, off_t in_offset, size_t length, SV *callback=&PL_sv_undef) PPCODE: { - int ifd = s_fileno_croak (in_fh , 0); - int ofd = s_fileno_croak (out_fh, 1); + int ifd = s_fileno_croak (in_fh , 0); + int ofd = s_fileno_croak (out_fh, 1); dREQ; req->type = EIO_SENDFILE; @@ -1322,7 +1326,7 @@ aio_readahead (SV *fh, off_t offset, size_t length, SV *callback=&PL_sv_undef) PPCODE: { - int fd = s_fileno_croak (fh, 0); + int fd = s_fileno_croak (fh, 0); dREQ; req->type = EIO_READAHEAD; @@ -1569,6 +1573,29 @@ } void +aio_fiemap (SV *fh, off_t start, SV *length, U32 flags, SV *count, SV *callback=&PL_sv_undef) + PPCODE: +{ + int fd = s_fileno_croak (fh, 0); + dREQ; + + req->type = EIO_CUSTOM; + req->sv1 = newSVsv (fh); + req->int1 = fd; + + req->feed = fiemap; +#if HAVE_FIEMAP + /* keep our fingers crossed that the next two types are 64 bit */ + req->offs = start; + req->size = SvOK (length) ? SvVAL64 (length) : ~0ULL; + req->int2 = flags; + req->int3 = SvOK (count) ? SvIV (count) : -1; +#endif + + REQ_SEND; +} + +void aio_busy (double delay, SV *callback=&PL_sv_undef) PPCODE: { @@ -1707,7 +1734,7 @@ RETVAL void -mmap (SV *scalar, size_t length, int prot, int flags, SV *fh, off_t offset = 0) +mmap (SV *scalar, size_t length, int prot, int flags, SV *fh = &PL_sv_undef, off_t offset = 0) PPCODE: sv_unmagic (scalar, MMAP_MAGIC); { @@ -1750,7 +1777,7 @@ CODE: { STRLEN svlen; - void *addr = SvPVbyte (scalar, svlen); + void *addr = SvPVbyte (scalar, svlen); size_t len = SvUV (length); if (offset < 0) @@ -1779,7 +1806,7 @@ CODE: { STRLEN svlen; - void *addr = SvPVbyte (scalar, svlen); + void *addr = SvPVbyte (scalar, svlen); size_t len = SvUV (length); if (offset < 0) @@ -1796,7 +1823,7 @@ #if _POSIX_MEMLOCK_RANGE RETVAL = munlock (addr, len); #else - RETVAL = ((errno = ENOSYS), -1); + RETVAL = EIO_ENOSYS (); #endif } OUTPUT: @@ -1808,12 +1835,40 @@ #if _POSIX_MEMLOCK munlockall (); #else - RETVAL = -1; - errno = ENOSYS; + RETVAL = EIO_ENOSYS (); #endif OUTPUT: RETVAL +int +splice (aio_rfd rfh, SV *off_in, aio_wfd wfh, SV *off_out, size_t length, unsigned int flags) + CODE: +{ +#if HAVE_LINUX_SPLICE + loff_t off_in_, off_out_; + RETVAL = splice ( + rfh, SvOK (off_in ) ? (off_in_ = SvVAL64 (off_in )), &off_in_ : 0, + wfh, SvOK (off_out) ? (off_out_ = SvVAL64 (off_out)), &off_out_ : 0, + length, flags + ); +#else + RETVAL = EIO_ENOSYS (); +#endif +} + OUTPUT: + RETVAL + +int +tee (aio_rfd rfh, aio_wfd wfh, size_t length, unsigned int flags) + CODE: +#if HAVE_LINUX_SPLICE + RETVAL = tee (rfh, wfh, length, flags); +#else + RETVAL = EIO_ENOSYS (); +#endif + OUTPUT: + RETVAL + void _on_next_submit (SV *cb) CODE: SvREFCNT_dec (on_next_submit);