| 1 |
Revision history for IO::AIO |
| 2 |
|
| 3 |
TODO: better autoconf.pm that can return LIBS etc. |
| 4 |
TODO: aio_cptree/mvtree |
| 5 |
TODO: reduce condvar fairness: schedule hot-cache-threads first? |
| 6 |
TODO: splice/tee/vmsplice? (http://kerneltrap.org/node/6505 http://lwn.net/Articles/178199/) |
| 7 |
TODO: aio_fcntl, at least for file-locking |
| 8 |
TODO: fallocate, if it wouldn't be so useless. |
| 9 |
TODO: IO::AIO::mmap? aio_mincore? |
| 10 |
|
| 11 |
3.4 Sat Jan 2 15:13:04 CET 2010 |
| 12 |
- (libeio) max_poll_time was not properly converted to ticks. |
| 13 |
- clarify cancel_subs description. |
| 14 |
- IO::AIO::sendfile did not actually return the return value |
| 15 |
from the sendfile call. |
| 16 |
- implement aio_msync, aio_mtouch. |
| 17 |
- (libeio) tentatively support darwin in sendfile. |
| 18 |
|
| 19 |
3.31 Thu Nov 12 02:14:29 CET 2009 |
| 20 |
- fix result status documentation of aio_copy, aio_move. |
| 21 |
- speed up object creation considerably by avoiding |
| 22 |
hash lookups. |
| 23 |
|
| 24 |
3.3 Wed Aug 5 13:52:58 CEST 2009 |
| 25 |
- use common::sense. |
| 26 |
- use common schmorp.h header. |
| 27 |
- allow integers (file descriptors) in addition to file handles. |
| 28 |
- take advantage of linux' eventfd (but it seems slower on smp?). |
| 29 |
- use poll not select on posix platforms. |
| 30 |
|
| 31 |
3.261 Wed Jul 1 10:11:51 CEST 2009 |
| 32 |
- more 0S X workarounds (patch by Tokuhiro Matsuno). |
| 33 |
|
| 34 |
3.26 Tue Jun 30 09:33:26 CEST 2009 |
| 35 |
- 0S X of course claims to be posix 2008 but lacks posix_fadvise. |
| 36 |
try to work around this horribly broken OS in a somewhat hackish |
| 37 |
way. might help other os's too. |
| 38 |
|
| 39 |
3.25 Sat Jun 27 05:18:26 CEST 2009 |
| 40 |
- added IO::AIO::fadvise and IO::AIO::sendfile. |
| 41 |
- (libeio) replaced quicksort+insertion sort by a tuned radix |
| 42 |
sort + insertion sort, resulting in comparable runtime (usually |
| 43 |
faster) to the old version, but without any ill side effects on |
| 44 |
degenerated (for quicksort) data. |
| 45 |
- (libeio) correctly sort dirs by inodes if we have perfect type |
| 46 |
knowledge and DIRS_FIRST is used. |
| 47 |
- (libeio) internally the wrong DT_* constants were used, but no |
| 48 |
known system has differing EIO_DT_*/DT_* values, so not an issue. |
| 49 |
- removed a (harmless) assert that was left over on the code |
| 50 |
but should not have been. |
| 51 |
- use more correct types (change IO length from IV to VAL64, |
| 52 |
which makes a difference on systems without 64 bit perls). |
| 53 |
|
| 54 |
3.23 Sat Jun 13 16:57:58 CEST 2009 |
| 55 |
- fix off-by-one bug in aio_readdir that was introduced in the |
| 56 |
rushed 3.22 release. |
| 57 |
|
| 58 |
3.22 Sat Jun 13 15:32:40 CEST 2009 |
| 59 |
- speed up readdirx sort algorithm slightly. |
| 60 |
- bin/treescan was missing from distro tarball. |
| 61 |
|
| 62 |
3.21 Fri Jun 12 18:45:53 CEST 2009 |
| 63 |
- new options --dirs and --files for treescan. |
| 64 |
- install bin/treescan by default. |
| 65 |
- (libeio) aio_readdir can now be cancelled while executing. |
| 66 |
- fix a printf format string for 64 bit systems (could lead |
| 67 |
to problems on big endian 64 bit systems). |
| 68 |
- do not use qsort() but our own algorithm: glibc initialises |
| 69 |
the whole locale and I/O subsystem inside qsort, causing |
| 70 |
a stack overflow on 32 bit machines. The new sort uses much less |
| 71 |
stack and is more than twice as fast in typical situations. |
| 72 |
|
| 73 |
3.2 Sun Jun 7 20:30:05 CEST 2009 |
| 74 |
- (libeio) pwrite emulation was even more flawed than intended and did |
| 75 |
not restore the file offset. |
| 76 |
- add aio_readdirx, which can return inode and filetype and sort |
| 77 |
the names in various ways. |
| 78 |
- unfortunately, utime, chmod, chown on an open file that has just |
| 79 |
been written can easily block, which caused aio_copy to block the |
| 80 |
process. no more! |
| 81 |
- no longer rely on dst path in aio_copy when futime is available. |
| 82 |
|
| 83 |
3.19 Tue Apr 21 22:05:21 CEST 2009 |
| 84 |
- more perl 5.10 workarounds for aio_read and write. |
| 85 |
- aio_write no longer modifies the sv (if possible). |
| 86 |
- aio_read now works correctly with magic values. |
| 87 |
|
| 88 |
3.18 Sun Apr 19 21:17:32 CEST 2009 |
| 89 |
- better diagnostics when some aio_* functions get passed an illegal fh. |
| 90 |
- try to avoid crashes due to incompatible 5.10 API changes. grmbl. |
| 91 |
|
| 92 |
3.17 Thu Nov 20 08:45:36 CET 2008 |
| 93 |
- (libeio) added aio_sync_file_range (untested). |
| 94 |
- add aio_busy to @IO::AIO::AIO_REQ. |
| 95 |
|
| 96 |
3.16 Wed Oct 22 18:28:01 CEST 2008 |
| 97 |
- use SvREFCNT_inc instead of SvREFCNT_inc_NN in a non-speed critical |
| 98 |
part to improve portability to perl 5.8 (reported by szymon). |
| 99 |
|
| 100 |
3.15 Mon Oct 13 00:39:55 CEST 2008 |
| 101 |
- automatic removal of feeders was broken. |
| 102 |
- (libeio) use a more robust method to detect whether a feeder |
| 103 |
has added something to the group or not. |
| 104 |
|
| 105 |
3.1 Thu Oct 2 13:34:40 CEST 2008 |
| 106 |
- pre-deref the passed callback object, for increased speed |
| 107 |
and decreased memory usage. |
| 108 |
- call on_next_submit callback even for aio_group - impact unknown, |
| 109 |
but seems more correct. |
| 110 |
- $req->cb now returns the original callback. |
| 111 |
- (libeio) pass EIO_GROUP/aio_group requests through the queue without |
| 112 |
requiring a thread switch. |
| 113 |
- (libeio) status code of aio_futime and aio_utime was always passed 0. |
| 114 |
- do some other ยต-optimisations. |
| 115 |
|
| 116 |
3.07 Sat Aug 2 16:06:13 CEST 2008 |
| 117 |
- do not include LIBS in autoconf tests. |
| 118 |
|
| 119 |
3.06 Tue Jul 15 12:41:32 CEST 2008 |
| 120 |
- move preadwritelock definition before it's first reference, |
| 121 |
to make it compile again on broken platforms (they still exist) |
| 122 |
or badly cofngiured perls (redhat...), reported by Rob Bloodgood. |
| 123 |
|
| 124 |
3.05 Thu Jun 19 23:23:52 CEST 2008 |
| 125 |
- work around perl overriding readdir etc. with thread-unsafe |
| 126 |
(sic!) versions. |
| 127 |
|
| 128 |
3.04 Wed Jun 18 01:35:38 CEST 2008 |
| 129 |
- (libeio) fix eio_mknod, which confused it's arguments. |
| 130 |
- (libeio) do not use readdir_r, as Ulrich Drepper pointed out that this |
| 131 |
is stupid. |
| 132 |
- (libeio) fix eio__readahead prototype, patch by Jost Krieger. |
| 133 |
- (libeio) fix a bug that could potentially cause IO::AIO |
| 134 |
not to get initialised properly. |
| 135 |
|
| 136 |
3.03 Thu May 29 05:33:30 CEST 2008 |
| 137 |
- (libeio) correctly call pthread_attr_destroy (). |
| 138 |
- (libeio) work around broken bsd headers once more. |
| 139 |
- reduce shared library size again by not including wrappers. |
| 140 |
- max_outstanding could cause poll to enter a busy-waiting loop. |
| 141 |
- document the new IO::AIO::poll_cb result value. |
| 142 |
|
| 143 |
3.02 Mon May 12 02:32:02 CEST 2008 |
| 144 |
- fix a memory leak on aio_readlink. |
| 145 |
- bring back working fchmod. |
| 146 |
- nop and busy now set result to 0. |
| 147 |
- set errno to ENOMEM when allocation fails. |
| 148 |
|
| 149 |
3.01 Sun May 11 03:07:03 CEST 2008 |
| 150 |
- (libeio) make it compile on systems without readahead or readdir_r. |
| 151 |
- (libeio) improve configure check for readahead availability. |
| 152 |
- do not try to link against -lrt. |
| 153 |
- use a separate configure script for IO::AIO (not the libeio one). |
| 154 |
|
| 155 |
3.0 Sun May 11 00:57:14 CEST 2008 |
| 156 |
- added bin/treescan to the distribution. |
| 157 |
- switched to using libeio. |
| 158 |
- LOTS OF INCOMPATIBLE CHANGES: |
| 159 |
- remove signal functionality, it is not worth the effort. |
| 160 |
- max_outstanding no longer returns the previous number of requests. |
| 161 |
- poll_cb no longer returns number of requests. |
| 162 |
|
| 163 |
2.62 Sat Apr 26 13:59:33 CEST 2008 |
| 164 |
- port to solaris perls configured for old posix |
| 165 |
(analysed by Jost Krieger). |
| 166 |
- keep a reference to the perl filehandle object in aio_close, |
| 167 |
so it doesn't get closed prematurely, leading to ugly races. |
| 168 |
|
| 169 |
2.61 Wed Apr 16 18:45:02 CEST 2008 |
| 170 |
- fix treescan output duplication and improve output ordering. |
| 171 |
also display files given on the commandline. |
| 172 |
- use a different algorithm for aio_close that is faster |
| 173 |
and probably has even lower chances of blocking. |
| 174 |
- do our own stack memory management for threads - linux |
| 175 |
allocates outrageous amounts of VM (not so bad unless you |
| 176 |
use mlockall...), which severely limits the number of threads |
| 177 |
on 32-bit arches: stack size is the larger of PTHREAD_STACK_MIN |
| 178 |
and 4096 * sizeof (long) (usually this is 16..64k). |
| 179 |
|
| 180 |
2.6 Sun Mar 30 08:28:11 CEST 2008 |
| 181 |
- added aio_sync. |
| 182 |
- added aio_pathsync. |
| 183 |
- fix prototypes of (void) functions. |
| 184 |
|
| 185 |
2.51 Sat Oct 6 16:04:54 CEST 2007 |
| 186 |
- perlio isn't generally threadsafe, so aio_close could not work reliably. |
| 187 |
aio_close now tries a safe hack that might not be that asynchronous |
| 188 |
(see the manpage for details). |
| 189 |
- discard callback return values as to not grow the stack endlessly |
| 190 |
when poll_cb handles many requests without returning. |
| 191 |
- minor code cleanups. |
| 192 |
|
| 193 |
2.5 Thu Oct 4 14:49:08 CEST 2007 |
| 194 |
- replaced _fd2fh with faster xs code. |
| 195 |
- aio_close will now try to do "the right thing" and thus might |
| 196 |
work sensibly for the very first time. |
| 197 |
|
| 198 |
2.41 Mon Sep 24 21:28:21 CEST 2007 |
| 199 |
- after fork, invest some work to make sure that the poll_fileno |
| 200 |
stays stable (by dup2'ing the new fd over the old one), to make |
| 201 |
it easier for programs/libs that don't control their forking |
| 202 |
behaviour and cannot recreate their watchers. |
| 203 |
|
| 204 |
2.4 Sun Aug 5 18:44:22 CEST 2007 |
| 205 |
- add aio_truncate, aio_chmod, aio_chown, aio_utime (linux |
| 206 |
successfully demonstrated that you can block on futimes...). |
| 207 |
- allow undef as fileoffset for aio_read/write and use read/write(2) |
| 208 |
internally (useful for sockets or O_APPEND handles). |
| 209 |
- allow undef for length in aio_write. |
| 210 |
- negative dataoffsets work as expected now in aio_read/aio_write. |
| 211 |
- use NV instead of UV for 32 bit perls and file offsets, as NVs |
| 212 |
have a larger range then. |
| 213 |
- shared code between BDB and IO::AIO. |
| 214 |
- aio_busy was completely broken. now fixed. |
| 215 |
- readahead emulation now returns something. |
| 216 |
- correctly set errno to ENOSYS on unimplemented functions |
| 217 |
(should never happen, though). |
| 218 |
- large changes to make it partially compile and run on win32, |
| 219 |
but win32 lacks too much functionality, and perl overrides way |
| 220 |
too many functions with crashing versions. |
| 221 |
|
| 222 |
2.33 Tue Jan 23 23:55:41 CET 2007 |
| 223 |
- fix install path (Andreas J. Koenig). |
| 224 |
|
| 225 |
2.32 Mon Jan 22 16:56:23 CET 2007 |
| 226 |
- added aio_rmtree. |
| 227 |
- wow, aio_mkdir was missing. |
| 228 |
- aio_load did return undef on error, not -1. |
| 229 |
- use prefixexp not prefix in autoconf.pm |
| 230 |
(suggested by ... rt.cpan.org unreachable). |
| 231 |
- avoid installing autoconf.pm. |
| 232 |
|
| 233 |
2.31 Sat Jan 6 03:46:02 CET 2007 |
| 234 |
- added aio_load. |
| 235 |
|
| 236 |
2.3 Sat Dec 23 05:48:07 CET 2006 |
| 237 |
- fix off-by-one bug in aio_read, causing memory corruption |
| 238 |
or crashes. |
| 239 |
|
| 240 |
2.21 Fri Dec 22 05:03:38 CET 2006 |
| 241 |
- minor doc updates. |
| 242 |
- minor cleanups. |
| 243 |
- IO::AIO::poll returns request count. |
| 244 |
- undocumented and unfinished async-signal-reporting. |
| 245 |
|
| 246 |
2.2 Tue Oct 31 00:31:00 CET 2006 |
| 247 |
- minor doc updates. |
| 248 |
- added aio_readlink. |
| 249 |
- properly zero-terminate aio_read buffer. |
| 250 |
- fix aio_read/write with nonzero dataoffset. |
| 251 |
- reduced size of aio request structure. |
| 252 |
- better diagnostics when pathnames etc. are not byte-encoded; |
| 253 |
do not leak request in that case, either. |
| 254 |
|
| 255 |
2.1 Sun Oct 29 02:01:57 CET 2006 |
| 256 |
- INCOMPATIBLE CHANGE: replace IO::AIO::poll_some by |
| 257 |
IO::AIO::max_poll_reqs and IO::AIO::max_poll_time. |
| 258 |
- fix a bug in aio_move where it would fail everytime |
| 259 |
it would have to copy files manually. |
| 260 |
- doc updates. |
| 261 |
- start threads only on (very low) demand. |
| 262 |
- end superfluous threads automatically after 10s idling. |
| 263 |
- add IO::AIO::max_idle to control this. |
| 264 |
- some tuning applied to random places. |
| 265 |
- add aio_mknod. |
| 266 |
|
| 267 |
2.0 Thu Oct 26 18:27:58 CEST 2006 |
| 268 |
- minor incompatibility: max_outstanding semantics have changed. |
| 269 |
- major rewrite, consider all functionality EXPERIMENTAL. |
| 270 |
- aio requests can be cancelled. |
| 271 |
- callbacks can be (re-)set later. |
| 272 |
- aio requests can choose between 9 priorities. |
| 273 |
- add aio_group: aio requests can be grouped |
| 274 |
into composite requests. |
| 275 |
- generator interface for groups. |
| 276 |
- add IO::AIO::poll_some. |
| 277 |
- major documentation improvements. |
| 278 |
- very minor bugfixes. |
| 279 |
- add aio_nop for dummy requests. |
| 280 |
- add aio_busy mainly for benchmarking and debugging. |
| 281 |
- use "fast" mutexes on linux. |
| 282 |
- use dynamic (and larger) buffers instead of |
| 283 |
putting them onto the stack. |
| 284 |
- optimise x86 and amd64 a bit. |
| 285 |
- better error checking when ENOMEM. |
| 286 |
- hopefully fix mem and dirp leaks on fork. |
| 287 |
- warn about broken perl malloc. |
| 288 |
- compiles on cygwin (but pread is broken on cygwin |
| 289 |
which is not detected by autoconf, so you have to |
| 290 |
edit autoconf/config.h yourself to enable emulation). |
| 291 |
|
| 292 |
1.8 Sat Jun 24 22:20:44 CEST 2006 |
| 293 |
- add and/or document aio_rename, aio_link, aio_symlink and aio_move. |
| 294 |
|
| 295 |
1.73 Wed Mar 1 22:49:32 CET 2006 |
| 296 |
- codename "jost" |
| 297 |
- AIX/Solaris(?) aren't being helpful again, hardcode a value |
| 298 |
for NAME_MAX and drive with your eyes closed (reported by |
| 299 |
wcooley@nakedape.cc). |
| 300 |
|
| 301 |
1.72 Thu Feb 2 00:40:44 CET 2006 |
| 302 |
- another minor bug in aio_scandir, again, the callback was called |
| 303 |
twice when the directory can't be stat'ed. |
| 304 |
|
| 305 |
1.71 Mon Dec 26 20:20:03 CET 2005 |
| 306 |
- grr, another minor bug in aio_scandir, this time resulting |
| 307 |
in a runtime error. |
| 308 |
|
| 309 |
1.7 Mon Dec 26 19:29:48 CET 2005 |
| 310 |
- aio_scandir errornously called callback twice in some cases. |
| 311 |
- added AnyEvent example. |
| 312 |
|
| 313 |
1.61 Wed Sep 7 19:40:42 CEST 2005 |
| 314 |
- no longer include $Config{libs}, as distros include too much |
| 315 |
garbage in there that is not installed on typical systems |
| 316 |
(for example, debian links perl against -lgdbm and many more |
| 317 |
libraries despite perl not needing it, slowing down startup and |
| 318 |
wasting memory). |
| 319 |
|
| 320 |
1.6 Tue Aug 30 17:44:44 CEST 2005 |
| 321 |
- added aio_readdir, aio_scandir. |
| 322 |
- added aio_sendfile, including sendfile emulation when sendfile |
| 323 |
isn't available. Linux, FreeBSD and HP-UX might work (sendfilev |
| 324 |
support is unavailable on my solaris 8 machine, but preliminary |
| 325 |
code is there. What's lacking is configury stuff and testing...). |
| 326 |
|
| 327 |
1.5 Sat Aug 20 02:32:19 CEST 2005 |
| 328 |
- create the aio threads on demand only (good after forking). |
| 329 |
- somebody nagged about C89 compatibility, so this release should |
| 330 |
be C89 compatible, but no guarantees for future ones. |
| 331 |
- abort when the configure script fails. |
| 332 |
|
| 333 |
1.4 Wed Aug 17 08:07:27 CEST 2005 |
| 334 |
- forgot to recreate the pipe after forking, this could cause |
| 335 |
deadlocks. |
| 336 |
|
| 337 |
1.3 Wed Aug 17 07:25:54 CEST 2005 |
| 338 |
- properly propagate exceptions from callbacks, instead of silently |
| 339 |
eating them. |
| 340 |
- use a different fork algorithm that avoids recursive callback |
| 341 |
invocations and reduces fork latency in busy processes. |
| 342 |
- fix a bug where the buffer scalar in aio_read got assigned the wrong |
| 343 |
length after reading. |
| 344 |
- fix a bug where the child would process some outstanding requests |
| 345 |
from the parent. |
| 346 |
- make a copy of the callback scalar, to avoid some very uncommon |
| 347 |
(but valid) usages where the callback scalar changes. |
| 348 |
- the callback was never freed in aio_read/aio_write. |
| 349 |
- aio_read/aio_write will now set the readonly flag on the data |
| 350 |
scalar for better error reporting. |
| 351 |
|
| 352 |
1.2 Wed Aug 17 01:22:58 CEST 2005 |
| 353 |
- verified on cygwin, linux/amd64+ia32+alpha, solaris. |
| 354 |
- use SvPVbyte for filenames now. You *need* to encode your strings to |
| 355 |
the external filename encoding first. |
| 356 |
- fix the bug where filehandles could leak (again). |
| 357 |
- clarify freebsd situation: make it work and tell me how, or go away |
| 358 |
with your underdocumented, broken and outdated OS. |
| 359 |
- fork'ing should now be officially supported (where linux and |
| 360 |
posix contradict each other, I followed posix). |
| 361 |
- only call read() once per poll, not twice (speedup). |
| 362 |
- new functions: aio_rmdir, aio_symlink. |
| 363 |
- force callback arguments to be undef or a reference, for better |
| 364 |
error catching. |
| 365 |
|
| 366 |
1.1 Sun Aug 7 05:27:23 CEST 2005 |
| 367 |
- added configure script to detect availability of some calls. |
| 368 |
(imperfect, as we cannot detect things like -lrt due to missing |
| 369 |
feedback yet). |
| 370 |
- emulate fdatasync via fsync, pread/pwrite via unsafe read/write |
| 371 |
and readahead via manual reading on platforms lacking those. |
| 372 |
- ported t/03_errors.t+framework by Brad Fitzpatrick |
| 373 |
from Linux::AIO. |
| 374 |
- removed XOPEN_SOURCE because stupid solaris 9 doesn't |
| 375 |
compile with it enabled. Probably breaks others. Oh well. |
| 376 |
- try to be more careful with errno on non-threaded perls. |
| 377 |
- cygwins perl insisted on a matching open mode when morphing |
| 378 |
the fd to an fh. |
| 379 |
- convert sv's to "byte" encoding in aio_read/aio_write. Might |
| 380 |
do that for pathnames, too, in the future, so make sure you |
| 381 |
properly encode your pathnames. |
| 382 |
|
| 383 |
1.0 Fri Jul 22 10:43:30 CEST 2005 |
| 384 |
- aio_open did not reliably deliver (open) filehandles. |
| 385 |
- all aio_*-functions keep a reference to the filehandle so |
| 386 |
it doesn't go away while the aio requests is working on it. |
| 387 |
- disable warnings in AIO.pm. |
| 388 |
|
| 389 |
0.9 Wed Jul 20 23:55:04 CEST 2005 |
| 390 |
- allow recursive calls to poll_cb, use less locking calls, too. |
| 391 |
- fix the Gtk2-example in the SYNOPSIS. |
| 392 |
- use pread/pwrite not pread64/pwrite64. this was simply a leftover |
| 393 |
from Linux::AIO. |
| 394 |
(found by compiling on cygwin. Unfortunately, cygwin declares but |
| 395 |
does not implement pread/pwrite). |
| 396 |
- only include <sys/syscall.h> on linux. |
| 397 |
|
| 398 |
0.5 Tue Jul 12 13:03:09 CEST 2005 |
| 399 |
- removed #include <endian.h>, it's not portable, and we don't |
| 400 |
need it anyway. |
| 401 |
- link against -lrt, as it might contain goodies such as fdatasync, |
| 402 |
and often aio_read, which we do not use yet, but it shouldn't hurt. |
| 403 |
|
| 404 |
0.4 Mon Jul 11 05:28:25 CEST 2005 |
| 405 |
- added IO::AIO::flush, IO::AIO::poll. |
| 406 |
- make callback argument optional for all calls. |
| 407 |
|
| 408 |
0.3 Mon Jul 11 01:42:37 CEST 2005 |
| 409 |
- restructured manpage. |
| 410 |
- added SYNOPSIS with examples for Event, Glib, Tk. |
| 411 |
- remove explicit context support so I don't have to maintain |
| 412 |
it for threaded perls (i.e. it should work now :). |
| 413 |
|
| 414 |
0.2 Sun Jul 10 22:57:12 CEST 2005 |
| 415 |
- return immediately from poll_wait if nreqs==0. |
| 416 |
- get rid of the request pipe and use faster thread-thread |
| 417 |
communications for most requests and responses. |
| 418 |
|
| 419 |
0.1 |
| 420 |
- cloned from Linux::AIO, which is very similar. |
| 421 |
|
