--- IO-AIO/Changes 2008/09/30 13:17:32 1.166 +++ IO-AIO/Changes 2009/06/14 20:36:59 1.194 @@ -3,13 +3,76 @@ TODO: better autoconf.pm that can return LIBS etc. TODO: aio_cptree/mvtree TODO: reduce condvar fairness: schedule hot-cache-threads first? -TODO: splice/tee/vmsplice/sync_file_range? (http://kerneltrap.org/node/6505 http://lwn.net/Articles/178199/) +TODO: splice/tee/vmsplice? (http://kerneltrap.org/node/6505 http://lwn.net/Articles/178199/) TODO: aio_fcntl, at least for file-locking - - call on_next_submit callback even fro aio_group - impact unknown, + - avoid some degenerate sorting cases on systems without d_ino, + or when the dentries are mostly sorted by inode already (common + with freshly unpacked tarfiles for example). + - correctly sort dirs by inodes only if we have perfect type + knowledge and DIRS_FIRST is used. + - speed up directory entry sorting further by some 20%. + +3.23 Sat Jun 13 16:57:58 CEST 2009 + - fix off-by-one bug in aio_readdir that was introduced in the + rushed 3.22 release. + +3.22 Sat Jun 13 15:32:40 CEST 2009 + - speed up readdirx sort algorithm slightly. + - bin/treescan was missing from distro tarball. + +3.21 Fri Jun 12 18:45:53 CEST 2009 + - new options --dirs and --files for treescan. + - install bin/treescan by default. + - (libeio) aio_readdir can now be cancelled while executing. + - fix a printf format string for 64 bit systems (could lead + to problems on big endian 64 bit systems). + - do not use qsort() but our own algorithm: glibc initialises + the whole locale and I/O subsystem inside qsort, causing + a stack overflow on 32 bit machines. The new sort uses much less + stack and is more than twice as fast in typical situations. + +3.2 Sun Jun 7 20:30:05 CEST 2009 + - (libeio) pwrite emulation was even more flawed than intended and did + not restore the file offset. + - add aio_readdirx, which can return inode and filetype and sort + the names in various ways. + - unfortunately, utime, chmod, chown on an open file that has just + been written can easily block, which caused aio_copy to block the + process. no more! + - no longer rely on dst path in aio_copy when futime is available. + +3.19 Tue Apr 21 22:05:21 CEST 2009 + - more perl 5.10 workarounds for aio_read and write. + - aio_write no longer modifies the sv (if possible). + - aio_read now works correctly with magic values. + +3.18 Sun Apr 19 21:17:32 CEST 2009 + - better diagnostics when some aio_* functions get passed an illegal fh. + - try to avoid crashes due to incompatible 5.10 API changes. grmbl. + +3.17 Thu Nov 20 08:45:36 CET 2008 + - (libeio) added aio_sync_file_range (untested). + - add aio_busy to @IO::AIO::AIO_REQ. + +3.16 Wed Oct 22 18:28:01 CEST 2008 + - use SvREFCNT_inc instead of SvREFCNT_inc_NN in a non-speed critical + part to improve portability to perl 5.8 (reported by szymon). + +3.15 Mon Oct 13 00:39:55 CEST 2008 + - automatic removal of feeders was broken. + - (libeio) use a more robust method to detect whether a feeder + has added something to the group or not. + +3.1 Thu Oct 2 13:34:40 CEST 2008 + - pre-deref the passed callback object, for increased speed + and decreased memory usage. + - call on_next_submit callback even for aio_group - impact unknown, but seems more correct. + - $req->cb now returns the original callback. - (libeio) pass EIO_GROUP/aio_group requests through the queue without requiring a thread switch. + - (libeio) status code of aio_futime and aio_utime was always passed 0. - do some other ยต-optimisations. 3.07 Sat Aug 2 16:06:13 CEST 2008