3 * **NATA is now the default ata driver.**
4 * **Major changes made to the disk abstraction layer.**
6 * **Significant changes made to disk device minor numbers
8 * **GPT and 64 bit disklabel support added (non booting).**
9 * **The LWP (Userland threading work) is now essentially done.**
10 * **Major work on vinum to remove its roll-your-own disklabel
11 layer and to remove its assumptions about the device layer's
13 * **Introduce libpthread as a softlink allowing libthread_xu or
14 libc_r to be selected. All multithreaded programs are now
15 linked against libpthread. libc_r is the default as
16 there are still a few issues that need to be worked out
18 * **Major improvements to wireless networking drivers.**
20 * **Many new USB and PCI devices added.**
21 * **Numerous new PCI drivers added, particularly w/ networking.**
22 * Userland syslink API is now finished.
23 * Link state changes are now reported through the routing socket.
24 * Signal and proper stop/continue support added to LWPs
25 * Entire kernel is now LWP and thread centric instead of process
27 * Bind 9.3.4 integrated
28 * GCC 4.1.2 integrated, selectable with CCVER. GCC-3 is still
30 * Add an abstraction layer to vn_strategy() to hide DMA limitations
31 of underlying devices.
32 * Fix symbol resolution issues with loaded modules.
33 * Both GCC-3 and GCC-4 are now built by default.
34 * Improve virtual kernel disk and networking support. Multiple
35 disks and network interfaces are now supported.
36 * Implement SMP support for virtual kernels (mainly as a testbed).
37 * Many documentation updates.
38 * Enhance kinfo_proc to return LWP info.
39 * Enhance the virtual kernel's console.
40 * Replace gnu tar with BSD tar.
41 * Fix a single stepping bug that could lead to a panic.
42 * Allow multiple kernels to be supplied on the release ISO and
43 supply a VKERNEL kernel in addition to the normal kernel.
44 * Get rid of the UAREA for good.
45 * Synchronize zoneinfo changes with recent standards changes.
46 * Fix a mbuf leak in sbappendcontrol().
47 * Fix an exec*() syscall memory leak that can lock a machine up
48 when the exec temporary argument space becomes exhausted.
49 * Bring in the latest libarchive, bsdtar, sendmail, awk,
50 ping, file, less, and openssh.
51 * Bring in binutils 2.17.
52 * Support MBs that do not have keyboard controllers by placing
53 an upper limit on the keyboard controller probe timeout.
54 * Remove IPV6 type 0 router header support entirely.
55 * Implement SYSREF for the kernel - this structure and API
56 implements MP-safe reference counting and also associates a
57 unique 64 bit id with the structure allowing the structure to
58 be referenced off-machine. SYSREF is part of the clustering work.
59 * Use SYSREF for struct vmspace and struct vnode, including doing
60 a major revamping of the vnode reference handling code.
61 * Do a major reordering of many SYSINITs and a major reordering
62 of where device configuration occurs in the boot sequence.
63 Device configuration now occurs far later in the boot sequence
64 and has access to most major subsystems including threading
65 and major memory subsystems. Interrupts are now enabled during
66 device configuration. This is somewhat experimental but works
67 great in allowing us to locate ordering issues in various device
68 init functions related to their interrupts.
69 * Implement asynchronous device attachment but default to
71 * Reformulate disk management layer ioctls and remove data
72 pollution. Do not store filesystem-specific data such as
73 the fragment size in the disklabel any more.
74 * Expand the SCSI subsystem to handle 16-byte commands,
75 allowing > 32 bit block addressing for devices that support
77 * Change the 'make upgrade' target to attempt to rebuild the
78 devices found in /dev instead of just rebuilding base devices.
79 * Distributions now include a Makefile in /usr with easy-to-use
80 targets for creating and synchronizing /usr/src and /usr/pkgsrc.
81 * Greatly simplify LWKT messaging within the kernel.
82 * Add virtual CDROM support to the virtual kernel.
83 * Add halt and reboot support to the virtual kernel.
84 * Fix a bug in bus_dmamap_load() that was not properly following
85 boundary and maximum segment size limits.
86 * Save and restore a USB keyboard's translation mode, allowing
87 the keyboard to be un-plugged and re-plugged.
88 * Remove the last source of major SMP TLB invalidations in the
89 fork/exec/exit/wait critical paths, significantly reducing IPI
90 traffic and improving fork/exec performance.
91 * ktrdump now has an option to loop waiting for new input.
92 * Fix mouse jumpiness when emergency interrupts are enabled by
93 not polling PS/2 mouse interrupts.
94 * Fix numerous CAM/USB issues so we do not panic or crash if
95 a USB mass storage device is pulled while still mounted. Make
96 sure umount -f works on pulled devices.
97 * Synchronize the sound subsystem with FreeBSD.
98 * Timeout positive hits in the namecache as well as negative hits
99 for NFS mounts. Otherwise rename-over's can result in a stale
100 namecache that doesn't auto recover.
101 * Add MPLS support to traceroute and add a protocol entry in
102 /etc/protocols for MPLS-in-ip.
103 * Bring in a number of math library fixes from FreeBSD.
104 * **dntpd now properly handles DNS failures, supports DNS pools,
105 and does a quorum consistency check when told to use more then
108 * Fix a bug in the checkpointing code that was preventing a
109 program from being re-checkpointed properly.
110 * Major USB code cleanup.
111 * Adjust dynamic objcache sizing to handle certain degenerate
113 * Synchronize with the Adaptec SCSI code from FreeBSD.
114 * Retire the original ports packaging tools.
117 ## Tue 30 January 2007
119 * **Implement Virtual Kernel support and add a VKERNEL target.**
120 Virtual kernels are run as userland processes and provide a
121 complete (minus hardware drivers) working kernel environment.
122 A simulated disk driver is included and a simulated network
123 driver using the TAP interface is included, allowing network
124 access. See the vkernel(7) manual page for more information.
125 * Introduce getcontext(3), setcontext(3), and friends
129 ## Mon 1 January 2007
131 * Jails are now IPv6 aware and support multiple IP addresses.
132 * Bridging support has been greatly improved.
133 * Many, Many new network drivers and chipsets have been added.
134 * Major infrastructure for 802.x wireless support added (WPA,
135 TX rate control, major API layer for all wireless network
136 devices, better ifconfig integration, and more).
137 * The cpdup program has been greatly enhanced and now supports
138 remote copies via ssh..
139 * Improved support for old-style disklabels.
140 * Kernel sources are now included on release CDs.
141 * Major rewrite of the vnode operations infrastructure, greatly
142 simplifying the layer.
143 * Fix a bug related to POSIX locking.
144 * Tons of documentation updates and code readability cleanups.
145 * Restructure the kernel build to accomodate multiple architectures.
146 Start separting platform components from cpu components.
147 * Synchronize a bunch of AMD64 related items (but we don't support
149 * Major rewrite of the mount glue infrastructure and NULLFS which
150 allows the namecache topology to be shared for multiple NULLFS
151 mounts and also allows arbitrary stacking of read-only and
152 read-write NULLFS mounts with basically zero kernel overhead.
153 * Unconditionally associate a namecache reference with all
154 file descriptors, allowing us to share vnodes across different
156 * NULLFS mounts may now be arbitrarily stacked and distinct paths
157 are no longer required.
158 * Properly dereference mount point roots when unmounting. Raise
159 hell if unmounting a filesystem which still has references.
160 * Do a massive reorganization of the device operations vector and
161 remove many instances where devops functions were dependant on
162 the calling process or thread to obtain cred information.
163 * Do a run through of all system header files to make sure they
164 include any dependancies themselves, instead of requiring the
165 originating source file to include the dependancies first.
166 This greatly reduces the effort required to use kernel header
167 files in new sources.
168 * Synchronize MII support with NetBSD and OpenBSD, including adding
169 standard conforming GMII support. Er. In otherwords, improve
170 driver support for physical network interfaces (phys's).
171 * Remove VOP locking vectors. Make VOP locking functions direct
172 calls (and thus mandatory).
173 * Control access to vnode ref count fields with a spinlock
174 * Fix a dubious construct in usr.bin/tail (no, really! That's
175 what the commit message said!).
176 * Add support for %j to the kernel printf.
177 * Major symbol separation for most kernel functions which
178 inexactly mimic libc functions. e.g. printf -> kprintf,
179 in order to allow user-mode kernels to be linked against libc.
180 * Update to sendmail to 8.13.8.
181 * Update to OpenSSL to 0.9.8d
182 * Update to OpenSSH 4.5p1
183 * Update to Bind 9.3.2-P1
184 * Update to less 3.9.4
185 * Update to awk 20050424
186 * Update to file 4.19
187 * Update to tcpdump 3.9.5
188 * Update to libpcap 0.9.5
189 * Synchronize m4 with FreeBSD
191 * Import bsdtar 1.3.1 and make bsdtar our default tar. GNU tar
192 is still installed as gtar.
193 * Import NetBSD's ftp client under its new name (tnftp).
194 * Bring in GCC-4.1 (setenv CCVER GCC41 to use).
195 * Significant code cleanups to make the world and kernel compile
196 under GCC-4.1. Note that the default compiler is still GCC 3.4
197 (GCC-4.x will probably become the default sometime in 2007).
198 * More work on LWP/THREAD/PROC separation for 1:1 threading support.
199 * Fix a major bug in /bin/make that can cause make (or some child
200 make in a large build) to ignore ^C. make temporarily installs
201 SIG_IGN when polling whether a signal is being ignored or not.
202 Fix the code to not do this.
203 * Fix a long standing signal/fork race that could cause a process
204 group signal to not make it to a newly fork child if it occurs
205 while the fork is in progress.
206 * Synchronize the zoneinfo database with tzdata2006p.
207 * Add an ECC detection device, currently supporting AMD64's memory
209 * Greatly reduce the memory allocated by fsck when fscking
210 filesytems with a huge number of directories (primarily mirors
211 with lots of hardlinked files). Otherwise fsck can run out
212 of memory on such filesystems.
213 * Adjust RCNG to support 'blah=YES/NO' as well as
214 'blah_enable=YES/NO', giving us better compatibility with
215 PkgSrc based RC scripts.
216 * Fix a bug related to ALTQ when setting the algorithm for a queue
217 on which packets are already present.
218 * Do a major clean up of the BUSDMA header file architecture.
219 * We have a new web site layout!
220 * Use spinlocks in the objcache instead of tokens (the critical
221 path is still lockless).
222 * Replace the global VM page hash table with a per-VM-object
224 * A considerable amount of work on IPSEC support has been done.
225 * NATA has been ported from FreeBSD and is currently being tested.
226 * Sync USB support with FreeBSD6 - use task queues to handle
227 operations that cannot be handled from an interrupt thread.
228 * Explore low-speed USB busses during cold boot so the USB keyboard
229 starts working earlier in the boot sequence.
233 * Start introducing glue for the SYSLINK and CCMS infrastructures.
234 (system link protocol and cache coherency management subsystems).
235 * Introduce a system call skeleton and test code for SYSLINK.
236 * Introduce the algorithm and structural topology that will be
237 used for cache coherency into the VFS path.
238 * Introduce a user-managed virtuallized page table infrastructure
239 which can be accessed via mmap(), as part of the infrastructure
240 to support virtual kernels running in userland (MAP_VPAGETABLE).
241 * Introduce vmspace_*() system calls which allows a user process
242 to manage and control multiple VM spaces, as part of the
243 virtualization support effort.
250 * Continued work on LWP/PROC separation.
251 * Continued pkgsrc integration.
252 * Fix refcount bugs in the kernel module loader and unloader.
253 * Lots of netif and serializer cleanups and fixes.
254 * Lots of softupdates, filesystem, and buffer cache related fixes.
255 * Remove more of the old ports-related infrastructure.
256 * Major documentation cleanups.
257 * Major code cleanups, ansification.
258 * Change the system #include topology so that each include file
259 is responsible for #include'ing any dependant include files.
260 * Fix a bug in the PF fragment cache.
261 * Random number generator: Instead of generating entropy from
262 selected interrupts (and none by default), we now generate
263 entropy from all interrupts by default and rate limit it to
264 not interfere with high performance interrupts. Completely
265 * Random number generator: Completely replace the algorithms,
266 remove limitations on returned bytes from /dev/random (which
267 only served to cause programs to not use /dev/random due to its
268 lack of dependability). Add the ability to seed the RNG.
269 Do some automatic initial seeding at boot.
270 * Adjust ssh to find the pkgsrc X11BASE instead of the old ports
272 * Fix some compatibility issues in /bin/sh.
273 * Fix a small number of critical section enter/exit mismatches.
274 * Bring in a bunch of new malloc() features from OpenBSD
275 (guard pages, free page protection, pointer guard, etc).
276 * Clean up the DragonFly build system's automatic .patch handling.
277 * Bring in openssh 4.3p2.
278 * Retire libmsun. It was replaced by NetBSD's libm.
279 * Fix a bug in the NFS timer/retry code.
280 * Fix issues related to wide-char support.
281 * <B>Fix a number of private TSS bugs related to threaded programs.</B>
283 * <B>Completely rewrite the user process scheduler and usched APIs.</B>
284 * <B>Add system calls that allow a blocking/non-blocking flag to be
285 passed independant of the O_NONBLOCK state of the descriptor.</B>
286 * <B>Remove all fcntl(... O_NONBLOCK) calls from libc_r, use the
287 new system calls instead. This solves numerous problems with
288 file descriptors shared between threaded and non-threaded
289 programs getting their non-blocking flag set, and then blowing
290 up the non-threaded program.</B>
291 * Add additional red-black (RB) tree function support for ranged
293 * Get rid of gdb -k, replace with a separately built kgdb, and
294 build a separate libgdb as well.
295 * Implement a VM load heuristic. Remove the full-process SWAP
296 code which never worked well and replace with page-fault
297 rate-limiting code based on the VM load.
298 * Fix a serious bug in adjtime()'s microseconds calculation.
299 * Fix a serious bug in libc's strnstr() function. The function
300 was testing one byte beyond the specified length limit. This
301 can cause a seg-fault when, e.g. using strnstr() on a memory
302 mapped file whos last byte abuts the end of the VM page.
303 * Bring in sendmail 8.13.7.
304 * Bring in SHA256 support from FreeBSD.
305 * Implement a hardlink-mirroring option (-H) in cpdup.
306 * Add missing code needed to detect IPSEC packet replays.
307 * Enable TCP wrappers in sshd.
308 * Restrict recursive DNS queries to localhost by default.
309 * <B>Massive reorganization and rewrite of the 802_11 subsystem,
310 with many pieces taken from FreeBSD.</B>
312 * Fix a number of issues with user-supplied directory offsets
313 blowing up readdir/getdirentries for NFS and UFS.
314 * Normalize internal kernel I/O byte lengths to 'int' and remove
315 a ton of crazy casts to and from unsigned int or size_t.
316 * Remove NQNFS support. The mechanisms are too crude to co-exist
317 with upcoming cache coherency management work and the original
318 implementation hacked up the NFS code pretty severely.
319 * Remove VOP_GETVOBJECT, VOP_DESTROYVOBJECT, and VOP_CREATEVOBJECT.
320 They never lived up to their billing and DragonFly doesn't
321 need them since DragonFly is capable of aliasing vnodes via
322 the namecache. Rearrange the VFS code such that VOP_OPEN is
323 now responsible for assocating a VM object with a vnode.
324 * Formalize open/close counting for the vnode and assert that
326 * Remove the thread_t argument from most VOP and VFS calls.
327 Remove the thread_t argument from many other kernel procedures.
328 * Integrate FreeBSD's new ifconfig(8) utility.
329 * Fix a race condition in the floating point fault code that
330 could sometimes result in a kernel assertion.
331 * Fix a crash in the TCP code when the MTU is too small to
332 support required TCP/IP/IPSEC options and headers.
333 * Separate EXT2 conditionals from the UFS code, copying the files
334 in question to the EXT2 directory instead of trying to
335 conditionalize them. Also remove function hooks and other code
336 mess that had been implemented to allow the UFS code to be
338 * Greatly simplify the lockmgr() API. Remove LK_DRAIN,
339 LK_INTERLOCK, and many other flags. Remove the interlock
341 * Fix a bug in the POSIX locking code (lockf). Actually, completely
342 rewrite the POSIX locking code. The previous code was too
343 complex and mostly unreadable.
344 * Do a major clean up of all *read*() and *write*() system calls,
346 * Replace many instances where LWKT tokens are used with spinlocks.
347 * Make spinlocks panic-friendly. Properly handle the detection
348 of indefinite waits and other deadlock issues.
349 * Improve network performance by embedding the netmsg directly in
350 the mbuf instead of allocating a separate netmsg structure for
352 * <B>Implement both shared and exclusive spinlocks.
353 Implement a much faster shared spinlock. Cache the shared
354 state such that no locked bus cycle operation is required in
357 * Implement msleep(). Use a novel approach to handling the
358 interlock that greatly improves performance over other
360 * Add cpu-binding support to the scheduler and add a system call
361 to access it. A user process can be bound to a subset of cpus.
362 * <B>Prefix all syscall functions in the kernel with 'sys_'
363 to reduce function prototype pollution and incompatibilities,
364 and to eventually support virtualized kernels running in
366 * Port the enhanced SpeedStep driver (EST) for cpu frequency control.
367 * Remove the asynchronous syscall interface. It was an idea before
368 its time. However, keep the formalization of the syscall
369 arguments structures.
370 * Add a facility which statistically records and dumps program
371 counter tracking data for the kernel.
372 * Improve kernel SSE2-based bcopies by calling fnclex instead of
374 * Major BUF/BIO work - <B>make the entire BUF/BIO subsystem BIO
376 * Major BUF/BIO work - <B>get rid of block numbers, use 64 bit
377 byte offsets only.</B>
379 * Major BUF/BIO work - Clean up structures and compartmentalize
380 driver-specific private fields. Rewrite and simplify device
381 and vnode strategy APIs.
382 * Major BUF/BIO work - Remove B_PHYS. There is no longer any
383 differentiation between physical and non-physical I/O at
385 * Major BUF/BIO work - Replace the global buffer cache hash table
386 with a per-vnode RB tree. Add sanity checks. Require that all
387 vnode-based buffers be VMIO backed.
388 * MPSAFE work - <B>Implement the parallel route table algorith.</B>
389 * MPSAFE work - <B>Make the user process scheduler MPSAFE.</B>
390 * MPSAFE work - <B>File descriptor access is now MPSAFE. Many
391 fd related functions, like dup(), close(), etc, are either MPSAFE
392 or mostly MPSAFE.</B>
393 * MPSAFE work - <B>Push the BGL deeper into the kernel call stack
394 for many system calls.</B>
396 * MPSAFE work - <B>Make the process list MPSAFE.</B>
397 * MPSAFE work - <B>Make all cred functions MPSAFE.</B>
398 * NRELEASE - compilable kernel sources are now included on the ISO.
401 ## Sat 7 January 2006
403 * Add the closefrom() system call.
404 * GCC 3.4 is now the default compiler. 2.95.x is no longer
405 supported (it can't handle the new threading storage classes
407 * Import Citrus from NetBSD.
408 * Implement direct TLS support for programs, whether threaded or not.
409 * Major library and user-visible system structure changes
410 (dirent, stat, errno, etc), and other work requires a major
411 library bump for libc and other libraries. libc is now
413 * stat: inode size now 64 bits, nlink now 32 bits. new fields,
415 * dirent: inode size now 64 bits, various fields disentangled from
417 * statfs: new fields, added pad.
418 * Clean up RC scripts that are not used by DragonFly.
419 * Remove the OS keyword requirement for RC scripts.
420 * Add support for unsigned quads to sysctl.
421 * Implement DNTPD, DragonFly's own NTP client time synchronization
423 * Correct a large number of bugs in the third party ntpd code, but
424 for client-side operations we now recommend you use dntpd.
425 * Add a framework for aggregating per-cpu structures for user
427 * Userland TLS (data spaces for threads) support added.
428 * Create a binary library compatibility infrastructure that
429 allows us to install and/or upgrade older revs of shared
430 libraries on newer machines to maintain compatibility with
432 * Fix issues related to the expansion of symbolic links by the
434 * Many, many mdoc cleanups and fixes.
435 * Update cvs, openssl, ssh, sendmail, groff,
436 and other numerous contributed applications.
437 * Bring in a brand new PAM infrastructure.
438 * Introduce pkgsrc support.
439 * Get rid of libmsun.
440 * Implement backwards scanning and partial-transaction handling
443 * FreeBSD-SA-05:06.iir - major disk access vulnerability for IIR
444 * FreeBSD-SA-05:04.ifconf - memory disclosure vulnerability
445 * FreeBSD-SA-05:08.kmem - memory disclosure vulnerability
446 * FreeBSD-SA-05:16.zlib - possible buffer overflow in zlib
447 * FreeBSD-SA-05:18.zlib - possible buffer overflow in zlib
448 * FreeBSD-SA-05:15.tcp - fix TCP RESET window check
449 (DOS attack vulnerability)
450 * ? - a bzip2 vulnerability
452 * Fix a bug in the TCP NewReno algorithm which could result in
453 a large amount of data being unnecessarily retransmitted.
454 * Fix numerous TCP buffering issues.
455 * Implement TCP Appropriate Byte Counting
456 * Bring in ALTQ and reorganize the IF queueing code to remove
457 per-driver dependencies on ALTQ.
458 * Strip away numerous TCP hidden indirections that make code hard
459 to read and understand.
460 * Introduce BPF_MTAP which includes an address family parameter.
461 * Reimplement network polling with a systimer, allowing the
462 frequency to be adjusted on the fly.
463 * Remove the really bad hack that was calling the network polling
464 code from the trap code.
465 * Completely rewrite network polling support.
466 * Make the network IF serializer mandatory for all network device
467 driver interrupts, ioctl's, and if_ callbacks.
468 * Implement a very fast memory object caching infrastructure. This
469 will eventually replace zalloc() (but not yet).
470 * Rewrite the mbuf allocator using the new memory object caching
471 infrastructure. Remove many crazily-large mbuf macros in favor
472 of the new infrastructure.
473 * Convert all remaining uses of the old mbuf m_ext API to the new
474 API. Remove support for the old API.
475 * Reorder the detach sequence in all network drivers. Unhook the
476 interrupt first rather then last.
477 * Fix all instances where an mbuf packet header and mbuf data
478 buffer were being referenced by the wrong name and all instances
479 where the packet header flag was being improperly set or cleared.
480 * Fix a number of mbuf statistics counting bugs.
481 * Fix numerous bugs in ipfw/ipfw2 where m_tag data was not being
482 stored in the right place, resulting in a panic.
483 * Add support for the experimental SCTP protocol.
484 * Fix an issue with cloned interfaces being added twice.
486 * Add a passive IPIQ call for non-time-critical events such as
488 * Add TLS support for threads using the GDT instead of the LDT.
489 * Greatly simplify and demystify the NTP kernel interface. Convert
490 most aspects of the interface over to sysctls.
491 * Implement ranged fsync's in-kernel. This capability will
492 eventually replace the write-behind heuristic.
494 * Introduce MP-safe mountlist scanning code.
495 * Introduce rip-out-safe red-black tree scanning code.
496 * Use the new RB scanning code to get rid of VPLACEMARKER and
497 generally use the new RB scanning code to handle all RB tree
498 scanning in a safe way (allowing the scan code callback to block).
500 * Rename cpu_mb*() functions to cpu_mfence(), cpu_lfence(), and
501 cpu_sfence() to make their function more apparent.
502 * Fix bugs in the LWKT token code related to token references
503 being lost due to a preemption or blocking condition.
504 * Fix bugs in the LWKT rwlock code relating to preemption occuring
505 during token acquisition.
506 * Fix a bug in the LWKT thread queueing and dequeueing code
507 related to a preemption.
508 * Increase the size of the physmap[] array to accomodate newer
509 PCs which have a larger number of memory segments and fix
511 * Use the ACPI timer if present instead of one of the other 8254
512 timers (which are not dependable because BIOS calls might
514 * Change cpu statistics to be accounted for on a per-cpu basis.
515 * Make network routing statistics per-cpu.
516 * Extend the interrupt vector code to pass a frame as a pointer.
517 * Remove the last vestiges of the old mbuf tagging code.
518 * Add a serializer API and code (basically blockable mutexes).
519 * Add interrupt enablement and disablement features to the new
520 serializer module to deal with races against blocked serializer
521 locks when e.g. removing a driver.
522 * Remove bus_{disable,enable}_intr(), it was not generic enough
524 * Remove all spl*() procedures and convert all uses to critical
526 * Do not try to completely halt all cpus when panic()ing as this
527 will likely leave the machine in a state that prevents it from
528 being able to do a dump.
529 * Try to unwind certain conditions when panic()ing from a trap
530 in order to give the machine a better chance to dump its core.
531 * A number of malloc()'s using M_NOWAIT really needed to be
533 * Attempt to avoid a livelocked USB interrupt during boot by
534 delaying the enablement of the EHCI interrupt until after all
535 companion controllers have been attached.
536 * Reimplement the kernel tracepoint facility (KTR) to greatly
537 reduce the complexity of the API as well as remove all hardwired
538 flags and values. In addition, record two levels of call
539 backtrace for each entry, if enabled.
540 * Beef up ktrdump to display symbolic results when possible.
541 * Beef up the slab allocator build with INVARINTS by adding a
542 bitmap to detect duplicate frees and such.
543 * Remove the 16 bit count limit for file descriptors.
544 * Replace the file descriptor allocator with an O(log N)
545 full-on in-place binary search tree.
546 * Allow the initial stack pointer for a use process to be
548 * Fix numerous scheduling issues that could cause the scheduler
549 to lose track of a reschedule request, resulting in poor
550 interactive performance. Rewrite the interactive/batch
552 * Begin to implement a management system to allow multiple
553 userland schedulers to be configured in a system.
554 * Add rm -I and add an alias for interactive shells to use it
555 by default. -I is a less invasive -i.
556 * Fix a bug in the pipe code that was not handling kernel-space
557 writes correctly. Such writes can occur whenever the kernel
558 writes KVM-referenced data to a descriptor, such as that
559 journaling code might do.
560 * Fix many issues with the high level filesystem journaling code.
561 High level journal records are now considered fairly solid.
562 * Implement the transactional features of the high level journaling
563 subsystem by allowing a journaling record to be written prior to
564 the VFS operation being executed, then aborted if the VFS operation
566 * Implement UNDO records for most journaling transaction types.
567 * Implement the journaling code's full-duplex ack protocol feature
568 which allows journals to be broken and restarted without losing
570 * Implement a stat-visible FSMID (filesystem modification id). This
571 identifier changes whenever any modifying operation on the file
572 or directory occurs, and for directories this identifier also
573 changes if anything in the sub-tree under the directory is
574 modified (recursively). The FSMID is synthesized for filesystems
575 which do not implement it directly in order to guarantee its
576 usefulness for at least a subset of operations.
577 * Implement pesistent storage of the FSMID for UFS.
578 * Implement shutdown() support for pipes.
579 * Implement a low level spinlock facility. Basically the
580 implementation gives us an MP-safe critical section type of
581 vehicle. However, being a spinlock the facility may only be
582 used for very short sections of code.
583 * Fix a bug with USB<->CAM communication for USB mass storage
585 * Fix numerous bugs in USB, primarily EHCI.
586 * Fix multiple panics when a fatal trap occurs from an IPI or
587 FAST interrupt. Interlock panics on multiple cpus so only the
588 first is recognized as the 'real' panic.
589 * Add a large number of assertions to the scheduler and interrupt
591 * Fix a critical IPI messaging bug (SMP only).
592 * Do not compile the kernel with the stack protector. The stack
593 protector generates weird incorrect or unexpected code in some
594 cases which interfere with the C<->assembly interactions in the
596 * Various bug fixes to softupdates.
597 * Fix a bitmap scanning bug in UFS which could sometimes result
598 in a sanity check panic, but no data corruption.
599 * Fix a deadlock in UFS's ffs_balloc() related to an incorrect
600 buffer locking order.
601 * Continued work on the buffer cache.
602 * Separate out APIC and ICU interrupt management.
603 * Rewrite the interrupt setup code.
604 * Major rewriting of the VFS directory scanning code. Add a new
605 function vop_write_dirent() to create the dirent for return to
606 userland. The new API is mandatory and filesystem code (not
607 even UFS) may not make assumptions about the size of the
608 userland-returned dirent.
609 * Major cleanup of the device identification method.
610 * Lots of driver updates.
611 * ANSIfy a great deal more of the codebase.
612 * Remove the now obsolete smp_rendezvous() mechanism.
613 * Compile up both the TFTP and the NFS PXE bootp code rather
614 then the (previous) make.conf option to select one or the other.
615 * Convert the lockmgr interlock from a token to a spinlock, also
616 incidently fixing an issue where non-blocking locks would
617 still potentially issue a thread switch.
618 * Fix bugs in the interrupt livelock code.
619 * Rewrite the code handling stopped user processes.
620 * Rewrite tsleep()/wakeup() to be per-cpu and MPSAFE. Reorganize
621 the process states (p_stat), removing a number of states but
622 resynthesizing them in eproc for 'ps'.
623 * Integrate the new if_bridge code from Open/Net/FreeBSD.
624 * Add an emergency interrupt polling feature that can be used
625 to get an otherwise non-working system working.
631 * Fix numerous issues in the new namecache code. The new code is
632 now considered to be solid (Matt).
633 * Fix numerous issues in the NFS code related to cache timeouts
634 not working properly, truncate operations not working properly
635 in certain circumstances, and TCP disconnections due to
636 races in accessing the streamed data. NFS was also not properly
637 setting TCP_NODELAY which resulted in poor performance over
638 TCP connections (Matt).
639 * Split out kernel information access abstraction into libkcore
640 and libkinfo (joerg).
641 * Optimize the interrupt processing code a bit (Matt).
642 * Significant progress in the journaling layer has been made,
643 but it isn't ready for production use yet. The journaling
644 layer is now capable of encoding a forward journaling stream
645 (the undo data is not yet being encoded). Still TODO: undo
646 data, two-way transaction acknowledgement protocol, and
647 emergency swap backing to allow system operation to proceed
648 in the event of a long-term journaling stream stall. (Matt)
649 * Fix numerous issues related to SMP-distributed wildcard
650 listen sockets for TCP. (Jeff, Matt)
651 * Major cleanup of the route table code in preparation for
652 SMP replication work (Jeff)
653 * ALTQ Integration (Joerg)
654 * Fix MTU discovery which appears to have been broken
656 * Miscellanious driver updates. (Various)
657 * Miscellanious ACPI updates.
658 * Fix a security issue with sendfile() (Various)
659 * Lots of work on the make program (Max)
660 * Add TLS support calls to the kernel to create segments suitable
661 for loading into %gs. (Matt, David Xu)
662 * Add TLS support to the dynamic linker (David Xu, Joerg)
663 * Add kernel support for mmap-based userland mutexes
664 * Continuing work on a new threading library (David Xu)
665 * Add jail support for varsyms (Joerg)
666 * Incorporate the Minix MINED editor into the /bin build for use
667 as an emergency editor while in single-user.
668 * Fix an issue with PCMCIA card removal.
669 * Replace the VM map lookup data structures with a red-black tree.
670 * Fix a potentially serious issue with inodes being reused before
671 their associated vnodes have been entirely recycled. This also
672 prevents unnecessary stalls on recycle flushing against new
673 file creation that would otherwise have to block waiting for
674 the inode it chose to finish cycling.
675 * Fix some minor issues with /dev/tty operations.
676 * Fix an issue with the vnode recycler. Certain types of file
677 extractions could cause it to stop oprating due to not being
678 able to clean out the namecache topology. This most often
679 occured when extracting archives with a large number of
681 * Fix two issues that caused the system's time of day tracking to
682 either stop operating or operate incorrectly.
683 * Enhance the checkpointing code to save and restore the
685 * Major work on the XIO and MSFBUF infrastructures as we begin to
686 use them in the kernel more. (Hiten, Matt)
687 * Implement CLOCK_MONOTONIC for the clock_*() system calls.
688 * Fix a bug in the TCP limited transmit code.
689 * Fix a crash that sometimes occurs when using the firewire
695 ## Mon 20 December 2004
697 * The old timeout() API has been completely ripped out and replaced
698 with the newer callout_*() API.
699 * USB support has been synchronized with FreeBSD and NetBSD
700 * USB keyboard detachment and reattachment while X is active no
701 longer messes up the translation mode.
702 * Fix a keyboard lockup issue (the new callout code was not being
703 properly called in an ATKBD hack to deal with lost interrupts).
704 * **TCP SACK support is now considered production stable (Jeff).**
705 * A TCP connection closing bug has been fixed, related to changes
706 we've made to the TCPS state values.
707 * **Expand TCP's header prediction case to handle common
708 window updates (Jeff), greatly improving cpu efficiency.**
709 * **Implement tail-append to sockbufs for the TCP stack, greatly
710 improving cpu efficiencies when dealing with large TCP buffers.**
712 * Lots of miscellanious network layer cleanups (Joerg).
713 * Fix an IPv6 pcb replication issue that was creating problems
714 with e.g. Apache-2.0.
715 * kernel event logging ported from FreeBSD (eirikn).
716 * DCONS (console over firewire) support added (simokawa/from FreeBSD)
717 * Additional GigE drivers added.
718 * Major VFS messaging and interfacing progress. The old namei()
719 and lookup() API has been completely removed. All high level
720 layers now use the new API and run through a compatibility layer
721 to talk to VFSs which still for the most part implement the old
722 VOP calls. Use locked namespaces to protect RENAME, REMOVE,
723 MKDIR, and other calls rather then depending on locked vnodes
724 for those protections.
725 * More VFS work. Keep track of the current directory with a
726 namecache pointer rather then a vnode pointer.
727 * More VFS work. Rewrite the vnode interlock code during
728 deactivation and disposal. Begin consolidating v_lock.
729 * Adjust the boot code to operate more deterministically by always
730 using EDD (linear block number) mode first, only falling back
731 to CHS if EDD fails. Before it would try to use CHS for
732 cylinders < 1024, resulting in non-deterministic operation on
734 * Separate out loader configuration files for BOOTP vs non-BOOTP
736 * IPSEC code moved to netproto/ipsec, update, and cleanups
738 * Minor fixes to PPP.
739 * Performance cleanup of if_em. Fix alignment requirements to
740 reduce instances where bounce buffers are allocated (from FreeBSD).
741 * Kernels built with debug info are now installed with debug info,
742 greatly improving normal enduser's ability to provide useable
743 bug reports. The backup copy of the kernel is stripped when
745 * **A number of bug fixes to the VM system seem to have fixed the few
746 remaining long-term panics in DragonFly. In particular, a very
747 serious bug in contigmalloc() inherited from FreeBSD-4.x has been
748 fixed We now consider DragonFly as stable as FreeBSD-4.x.**
750 * Many minor driver bug fixes here and there, including one to the
751 serial driver which was responsible for machine lockups in
753 * **Major expansion of the checkpoint code API. You can now
754 re-checkpoint programs that have been checkpoint-restored.
755 The system call for checkpointing and checkpoint restore
756 functions is now official. Certain VM area issues have been
757 fixed. And it is now possible (and easy!) to write
758 checkpoint-aware programs.**
759 * Abstract kernel structure access via libkinfo (joerg).
760 * Fix a number of timer issues related to the 8254, sleep/wakeup,
761 and recovery from clock jumps due to high latencies.
762 * Do better ESTALE checking for NFS clients to reduce instances
763 where ESTALE makes it all the way back to the application layer.
764 * Improved polling support for UHCI USB (drhodus).
765 * Fix /boot/loader's handling of extended DOS partitions. There was
766 an off-by-one issue that prevented the boot loader from passing
767 the proper slice number to the kernel in certain cases (walt).
768 * Integration of PF as third firewall (joerg, corecode, Devon O'Dell,
769 from FreeBSD/OpenBSD)
772 ## Sat 18 September 2004
774 * DragonFly has adopted the 3-clause BSD license as its official
776 * NFS - increase the size of the nfsheur hash table as per a
777 Freenix track paper. nfsheur is used for sequential I/O
778 heuristic in NFS. The increase greatly improves clustering
779 under parallel NFS loads.
780 * More VFS work: make vnode locks mandatory, get rid of nolock
781 kludges in NFS, procfs, nullfs, and many other filesystems.
782 Remove the vnode_if.m dynamic dispatch algorithms and replace
783 with a fixed structure (in preparation for upcoming VFS work).
784 Rewrite the VOP table parsing code. Redo vnode/inode hash table
785 interactions to close race conditions (Matt).
786 * Reorganize the boot code to consolidate all fixed ORG directives
787 and other dependancies into a single header file, allowing us
788 to re position the boot code and eventually (not yet) move it
789 out of low BIOS memory. Change the way the BTX code clears BSS
790 to make it more deterministic and more gcc3 compatible.
791 * A major cleanup of ipfilter has been done (Hiten).
792 * Fix a very old but serious bug in the VM system that could result
793 in corrupt user memory when MADV_FREE is used (from Alan Cox).
794 * Rewrite most of the MBUF allocator and support code. Get rid
795 of mb_map and back the mbuf allocator with malloc. Get rid of
796 the old-style m_ext support and replace with new style m_ext
797 support (which is somewhat different then FreeBSD's new code).
798 Cleanup sendfile() to use the new m_ext callback scheme.
799 * Lots minor/major bug fixes, cleanups, new driver support, etc.
800 * More GCC3 work. The system mostly compiles and runs GCC3 builds
801 but the boot/loader code still has some issues (Various people).
802 * Work on the userland schedule. Introduce an 'interactivity'
803 measure in an attempt to do a better job assigning time slices.
804 Fix some scheduler interaction bugs which were sometimes resulting
805 in processes being given a full 1/10 second time slice when they
807 * A major import of the FreeBSD-5 802_11 infrastructure has been
808 accomplished (Joerg).
809 * NDis has been ported from FreeBSD, giving DragonFly access to
810 many more 802.11 devices via windoz device drivers (Matt).
811 * Add thermal control circuit support (Asmodai).
812 * Generally add throttling support to the system.
813 * **Add TCP SACK support (Jeffrey Hsu). This is still considered
816 * Make the syncache per-cpu and dispatch syncache timer events
817 via LWKT messages directly to the appropriate protocol thread
818 (Jeffrey Hsu). This removes all race conditions from the syncache
819 code and makes it 95% MP safe.
820 * **Greatly reduce the number of ACKs generated on a TCP connection
821 going full-out over a GigE (or other fast) interface by delaying
822 the sending of the ACK until all protocol stack packets have been
823 processed. Since GiGE interfaces tend to aggregate 8-12(+)
824 received packets per interrupt, this can cut the ACK rate by 75%
825 (one ack per 8-12 packets instead of one ack per 2 packets), and
826 it does it without violating the TCP spec. The code takes
827 advantage of the protocol thread abstraction used to process
828 TCP packets. (Matt)**
829 * **Greatly reduce the number of pure window updates that occur over
830 a high speed (typ GigE) TCP connection by recognizing that a
831 pure window update is not always necessary when userland has
832 drained the TCP socket buffer. (Jeffrey Hsu)**
833 * Rewrite the callout_*() core infrastructure and rip out the old
834 [un]timeout() API (saving ~800K+ of KVM in the process). The new
835 callout infrastructure uses a DragonFly-friendly per-cpu
836 implementation and is able to guarantee that callouts will occur
837 on the same cpu they were registered on, a feature that the TCP
838 protocol stack threads are going to soon take major advantage of.
839 * Cleanup the link layer broadcast address, consolidating many
840 separate implementations into one ifnet-based implementation
842 * BUF/BIO progress - start working the XIO vm_page mapping code
843 into the system buffer cache (Hiten). Remove b_caller2 and
844 b_driver2 field members from the BUF structure (Hiten) (generally
845 we are trying to remove the non-recursive-friendly driver
846 specific fields from struct buf and friends).
847 * The release went well! There were a few gotchas, such as trying
848 to run dual console output to the serial port causing problems on
849 laptops which did not have serial ports. A bug in the installer
850 was serious enough to have to go to an '1.0A' release a day or two
851 after the 1.0 release. But, generally speaking, the release did its
853 * More USB fixes. Clean up some timer races in USB/CAM interactions
854 related to pulling out USB mass storage cards. Fix a serious bug
855 that could lead to lost transactions and create confusion between
856 the USB code and the device.
857 * Async syscall work: clean up the sendsys2() syscall API into
858 something that's a bit more reasonable (Eirik Nygaard)
859 * Add a generic framework for IOCTL mapping (Simon).
860 * Add VESA mode support, giving us access to bitmapped VESA video
861 modes (Sascha Wildner).
862 * Fix USB keyboard support by giving the USB keyboard preference
863 even if a normal keyboard is detected earlier in the boot process.
864 This is necessary due to hardware/firmware level PS/2 keyboard
865 emulation that many USB chipsets and BIOSes offer.
866 * **Installer Updated: Lots of bug fixes have been made.**
868 * Stability: Spend two weeks stabilizing recent work in preparation
869 for another big push.
874 * Master ISO for the 1.0-RELEASE is now on the FTP site, we
875 release on Monday 12-July-2004.
876 * **Major revamping of the boot code. Support dual-console mode
877 (video and serial). Properly detect missing serial ports
878 (common for laptops). Initialize the serial port to 96008N1.**
879 * Bring ACPICA5 uptodate and integrate it into the build as a
880 KLD, just like FreeBSD-5. Fix additional issues as well.
881 Nearly all of this work was done by YONETANI Tomokazu. Use
882 the latest INTEL code (20040527). This greatly improves
884 * Make the kernel check both CDRom drives for a root filesystem
885 when booted with -C, which allows the boot CD to work in either
886 drive on systems with two drives.
887 * **Integrate the DragonFly Installer into the release build. The
888 DragonFly Installer is a from-scratch design by Chris Pressey,
889 Devon O'Dell, Eirik Nygaard, Hiten Pandya & GeekGod (aka
890 Scott Ullrich). The design incorporates a worker backend and
891 multiply targetable frontends. Currently the console frontend
892 is enabled, but there is also a CGI/WWW frontend which is on the
893 CD but still considered highly experimental.**
895 * Fix a bug in the polling backoff code for the VR device
896 * Implement interrupt livelock detection. When an interrupt
897 livelock is detected the interrupt thread automatically
898 throttles itself to a (sysctl settable) rate. When the interrupt
899 rate drops to half the throttle limit the throttling is removed.
900 This greatly improves debugability, especially on laptops with
901 misrouted interrupts. While it doesn't necessarily fix the
902 broken devices it does tend to allow the system to continue to
903 operate with those devices that *do* still work.
904 * Properly probe for the existance of the serial port in the
905 kernel, which allows us to run a getty on ttyd0 by default for
906 the release CD. Prior to this fix attempting to use ttyd0 would
907 result in a system deadlock.
908 * Bring in some pccard driver improvements from FreeBSD, including
909 increasing the CIS area buffer from 1K to 4K and properly
910 range-checking the CIS parser (which avoids panics and crashes).
911 * Update our DragonFly Copyright and update date specifications
912 on a number of copyrights.
913 * Add support for randomized ephermal source ports.
914 * Do some network driver cleanups.. basically move some common code
915 out of the driver and into ether_ifattach() (work by Joerg).
916 * Try to be more compatible with laptop touchpads whose aux ports
917 return normally illegal values (Eirik Nygaard).
918 * Add common functions for computing the ethernet CRC, work
919 by Joerg, taken from NetBSD.
920 * Add support for additional AGP bridges - taken from FreeBSD-5.
921 * Add support for the 're' network device.
922 * Bring OHCI and EHCI up-to-date with NetBSD.
923 * Miscellaneous driver fixes to ips, usb (ugen), sound support.
924 * Miscellaneous Linux emulation work, by David Rhodus, taken
925 from FreeBSD-SA-04:13.linux.
930 * We are going to release 1.0RC1 tonight.
931 * Implement interrupt livelock detection and rate limiting. It's
932 still a bit raw, but it does the job (easily testable by plugging
933 and unplugging cardbus cards at a high rate). The intent is to
934 try to make the system more survivable from interrupt routing and
935 other boot-time configuration issues, and badly written drivers.
936 * Implement an emergency polling mode for the VR device if its
937 interrupt does not appear to be working.
938 * Implement dual console (screen and serial port) support by
939 default for boot2 and the loader.
944 * Joerg has brought GCC-3.4-20040618 in and it is now hooked up
945 to the build. GCC-3.3 will soon be removed. To use GCC-3.4
946 'setenv CCVER gcc34'.
947 * The world and kernel now builds with gcc-3.4. The kernel builds
948 and runs with gcc-3.4 -O3, but this is not an officially supported
950 * More M_NOWAIT -> M_INTWAIT work. Most of the drivers inherited
951 from FreeBSD make aweful assumptions about M_NOWAIT mallocs
952 which cause them to break under DragonFly.
953 * /usr/bin/ps now reports system thread startup times as the boot
954 time instead of as Jan-1-1970. The p_start field in pstats has
955 been moved to the thread structure so threads can have a start
956 time in the future - Hiten.
957 * Zero the itimers on fork() - Hiten / SUSv3 compliance.
958 * MMX/XMM kernel optimizations are now on by default, greatly
959 improving bcopy/bzero/copyin/copyout performance for large (>4K)
961 * A number of revoke() related panics have been fixed, in particular
962 related to 'script' and other pty-using programs.
963 * The initial MSFBUF scheme (multi-page cached linear buffers)
964 has been committed and is now used for NFS requests. This is
965 part of the continuing work to eventually make I/O devices
966 responsible for any KVM mappings (because most just set up DMA
967 and don't actually have to make any) - Hiten.
968 * Continuing ANSIfication work by several people - Chris Pressey.
969 * Continuing work on the LWKT messaging system. A number of bugs
970 in the lwkt_abortmsg() path have been fixed.
971 * The load average is now calculated properly. Thread sleeps were
972 not being accounted for properly.
973 * Bring in a number of changes from FreeBSD-5: try the elf image
975 * Fix a number of serious ref-counting and ref holding bugs in
976 procfs as part of our use of XIO in procfs - GeekGod and Matt
977 * Use network predicates for accept() and connect() (using the
978 new message abort functionality to handle PCATCH) - Jeff.
979 * Convert netproto/ns to use the pr_usrreqs structure.
980 * Implement markers for traversing PCB lists to fix concurrency
981 problems with sysctl.
982 * Implement a lwkt_setcpu() API function which moves a thread
983 to a particular cpu. This will be used by sysctl to iterate
984 across cpus when collecting per-cpu structural information.
985 * Redo netstat to properly iterate the pcb's across all cpus.
986 * Significant mbuf cleanup - dtom() has now been removed, and
987 a normal malloc() is used for PCB allocations and in other places
988 where mbufs were being abused for structural allocations.
989 * dup_sockaddr() now unconditionally uses M_INTWAIT instead of
990 conditionally using M_NOWAIT, making it more reliable.
991 * Fix a number of USB device ref counting issues and fix issues
992 related to UMASS detaching from CAM while CAM is still active,
994 * Optimize kern_getcwd() some to avoid a string shifting bcopy().
995 * Continued work on asynch syscalls - track pending system calls
996 and make exit1() wait for them (abort support will be forthcoming).
997 * **Add a negative lookup cache for NFS.** This makes a huge
998 difference for things like buildworlds where /usr/src is NFS
999 mounted, reducing (post cached) network bandwidth to 1/10 what
1001 * Add the '-l' option to the 'resident' command, listing all
1002 residented programs and their full paths (if available). -Hiten.
1003 * **Implement the 'rconfig' utility (see the manual page)** - for
1004 automatic search/config-script downloading and execution, which
1005 makes installing a new DFly box from CDBoot a whole lot easier
1006 when you are in a multi-machine environment.
1007 * Revamp the BIO b_dev assignment and revamp the 'disk' layer.
1008 Instead of overloading the raw disk device the disk layer now
1009 creates a new device on top of the raw disk device and takes over
1010 the (user accessible) CDEV (major,minor). The disk layer does
1011 its work and reassigned b_dev to the raw device. biodone() now
1012 unconditionally setes b_dev to NODEV and all I/O ops are required
1013 to initialize b_dev prior to initiating the op. This is
1014 precursor work to our DEV layering and messaging goal.
1015 * Fix the rootfs search to specify the correct unit number rather
1016 then using unit 0, because CDEVSW lookups now require a valid
1017 minor number (CDEVSW's are now registered with a minor number
1018 range and the same major number can be overloaded as long as the
1019 minor ranges do not conflict). Fix by Hiroki Sato.
1020 * Fix a wiring related page table memory leak in the VM system.
1021 * Fix a number of ^T related panics.
1022 * Properly ref-count all devices.
1028 * NEWTOKENs replace old tokens. The new token abstraction provides
1029 a serialization mechanism that persists across blocking conditions.
1030 This is not a lock, but more like an MP-capable SPL, in that
1031 if you block you will lose serialization but then regain it when
1032 you wakeup again. This means that newtokens can be used for
1033 serialization without having to make lower level subsystems
1034 aware of tokens held by higher level subsystems in the call
1035 stack. This represents a huge simplification over the FreeBSD-5
1037 * Support added for the Silicon Image SATA controller.
1038 * DragonFly now supports 16 partitions per slice (up from 8).
1039 * Wildcarded sockets have been split from the TCP/UDP connection
1040 hash table, and the listen table is now replicated across
1041 cpus for lockless access.
1042 * UDP sendto() without a connect no longer needs to make a
1043 temporary connect inside the kernel, greatly improving UDP
1044 performance in this case.
1045 * NFS performance has been greatly improved.
1046 * Fix some major bugs in the USB/SIM code, greatly improving
1047 the reliability of USB disk keys and related devices.
1048 * NEWCARD has been brought in from FreeBSD-5.
1049 * A bunch of userland scheduler performance issues have been fixed.
1050 * Major syscall procedural separation has been completed, separating
1051 the user interfacing portion of a syscall from the in-kernel
1052 core support, allowing the core support functions to be directly
1053 called from emulation code instead of using the stackgap junk.
1054 * An optimized MMX and XMM kernel bcopy has been implemented and
1055 tested. Most of i386/i386/support.s has been rewritten and a
1056 number of FP races in that and npxdna() have been closed.
1057 * Brought in the SFBUF facility from FreeBSD-5, including all of
1058 Alan Cox's (the FBsd Alan Cox) improvements, plus additional
1059 improvements for cpu-localized invlpg calls (and the ability
1060 to avoid such calls when they aren't needed).
1061 * A major PIPE code revamp has occured, augmenting the SFBUF
1062 direct-copy feature with a linear map. Peak standard pipe
1063 throughput with 32KB buffers on an AMD64 3200+ now exceeds
1065 * Implement XIO, which is a multi-page buffer wrapper for SFBUFs
1066 and will eventually replace linear buffer management in the
1067 buffer cache as well as provide linear buffer mappings for other
1068 parts of the system.
1069 * Added a localized cpu pmap_qenter() called pmap_qenter2() which
1070 is capable of maintaining a cpumask, to avoid unnecessary invlpg's
1071 on target-side linear maps. It is currently used by the PIPE
1072 code and the CAPS code.
1073 * Joerg has brought in most of the KOBJ extensions from FreeBSD-5.
1074 * Major continuing work by Jeff on the threading and partitioning
1075 of the network stack. Nearly the whole stack is now
1076 theoretically MP safe, with only a few niggling issues before we
1077 can actually start turning off the MP lock.
1078 * The last major element of the LWKT messaging system,
1079 lwkt_abortmsg() support, is now in place. Also it is now possible
1080 to wait on a port with PCATCH.
1081 * Hiten has revamped the TCP protocol statistics, making them
1082 per-cpu along the same lines that we have made vmstats per-cpu.
1083 * propolice has been turned on by default in GCC3.
1084 * CAPS (DragonFly userland IPC messaging) has been revamped to
1086 * acpica5 now compiles, still needs testing.
1087 * SYSTIMERS added, replacing the original hardclock(), statclock(),
1088 and other clocks, and is now used as a basis for timing in the
1089 system. SYSTIMERS provide a fine-grained, MP-capable,
1090 cpu-localizable and distributable timer interrupt abstraction.
1091 * Fix RTC write-back issues that were preventing 'ntpdate' changes
1092 from being written to the RTC in certain cases.
1093 * Finish most of the namecache topology work. We no longer need
1094 the v_dd and v_ddid junk. The namecache topology is almost
1095 ready for the next major step, which will be namespace locking
1096 via namecache rather then via vnode locks.
1097 * Add ENOENT caching for NFS, greatly reducing the network overhead
1098 (by a factor of 5x!!!) required to support things like
1099 buildworld's using an NFS-mounted /usr/src.
1100 * Jeff has added predicate messaging requests to the network
1101 subsystem, which allows us to convert situations which normally
1102 block, like connect(), to use LWKT messages.
1103 * Lots of style cleanups, primarily by Chris Pressey.
1107 ## Sun 15 February 2004
1109 * Newcard is being integrated.
1110 * A longstanding bug in PCI bus assignments which affects larger
1111 servers has been fixed.
1112 * The IP checksum code has been rewritten and most of it has been
1113 moved to machine-independent sections.
1114 * A general machine-independent CPU synchronization and rendezvous
1115 API has been implemented. Older hardwired IPIs are slowly being
1116 moved to the new API.
1118 * A new 'SysTimer' API has been built which is both MP capable
1119 and fine-grained. 8254 Timer 0 is now being used for fine-grained
1120 timing, while Timer 2 is being used for timebase. Various
1121 hardwired clock distribution code is being slowly moved to the
1122 new API. hardclock and statclock have already been moved.
1123 * A long standing bug in the NTP synchronization code has been fixed.
1124 NTPD should now lock much more quickly.
1125 * Clock interrupt distribution has been rewritten. Along with this,
1126 IPI messaging now has the ability to pass an interrupt frame
1127 pointer to the target function. Most of the old hardwired
1128 clock distribution code has been ripped out.
1129 * nanosleep() is now a fine-grained sleep. After all, what's the
1130 point of having a nanosleep() system call which is only capable
1131 of tick granularity?
1132 * Critical fixes from FreeBSD RELENG_4 integrated by Hiten Pandya.
1133 * Firewire subsystem updated by a patchset from Hidetoshi Shimokawa.
1135 * USB subsystem has been synced with FreeBSD 5 and NetBSD.
1136 * GCC 3.3.3 and Binutils 2.14 have been integrated into base.
1137 * An aggregated Client/Server Directory Services syscall API has
1139 * An amiga-style 'resident' utility program + kernel support has
1140 been implemented, and prelinking support has been removed
1141 (because the resident utility is much better). You can make any
1142 dynamically loaded ELF binary resident with this utility. The
1143 system will load the program and do all shared library mappings
1144 and relocations, then will snapshot the vmspace. All future
1145 executions of the program simply make a copy of the saved
1146 vmspace and skip almost right to main(). Kernel overhead is
1147 fairly low, also. It still isn't as fast as a static binary
1148 but it is considerably faster then non-resident dynamic binaries.
1151 ## Mon 1 December 2003
1153 A great deal of new infrastructure is starting to come to fruition.
1156 * We have a new CD release framework (/usr/src/nrelease) in.
1157 Development on the new framework is proceeding. Basically the
1158 framework is based on a fully functioning live CD system
1159 and will allow us to build a new installation infrastructure
1160 that is capable of leveraging all the features available in
1161 a fully functioning system rather then being forced to use
1163 * A new IPC framework to reliably handle things like password
1164 file lookups is proceeding. This framework is intended to
1165 remove the need for DLL or statically-linked PAM/NSS and at
1166 the same time make maintainance and upgrades easier and more
1168 * The FreeBSD-5 boot infrastructure has been ported and is now
1170 * RCNG from FreeBSD-5 has been ported and is now the default.
1171 * Work is proceeding on bringing in ATAng from FreeBSD-4, with
1172 additional modifications required to guarantee I/O in low
1173 memory situations. That is, it isn't going to be a straight
1174 port. The original ATA code from FreeBSD-4.8, which we call
1175 ATAold, has been given interim fixes to deal with the low memory
1176 and PIO backoff issues so we don't have to rush ATAng.
1179 ## Sat 18 October 2003
1181 Wow, October already! Good progress is being made on several fronts.
1183 * K&R function removal.
1184 * VM function cleanups by Hiten Pandya.
1185 * General kernel API cleanups by David Rhodus.
1186 * Syscall Separation work by David Reese.
1188 * Removal of stackgap code in the Linux Emulation by David Reese.
1189 * Networking work by Jeffrey Hsu.
1190 * Interrupt, Slab Allocator stabilization.
1191 * Introduction of _KERNEL_STRUCTURES ... a better way for
1192 userland programs to access kernel header files rather
1193 then them setting _KERNEL.
1194 * Bring the system uptodate on security issues (David Rhodus, others)
1195 * NFS peformance improvements by David Rhodus and Hiten Pandya.
1197 * GUPROF and kldload work in the kernel by Hiten Pandya.
1198 * Major progress on the checkpointing code in the kernel
1199 primarily by Kip Macy.
1200 * All work through this moment has been stabilized with major
1201 input from David Rhodus.
1204 Matt's current focus continues to be on rewriting the namecache code.
1205 Several intermediate commits have already been made but the big changes
1208 Galen has started experimenting with userland threads, by porting the
1209 LWKT subsystem (which is mostly self contained) to userland.
1211 ## Wed 27 August 2003 - Slab Allocator, __P removal
1214 * DragonFly now has slab allocator for the kernel! The allocator is
1215 about 1/3 the size of FreeBSD-5's slab allocator and features per-cpu
1216 isolation, mutexless operation, cache sensitivity (locality of reference),
1217 and optimized zeroing code.
1219 * The core of the slab allocator is MP safe but at the moment we still
1220 use the malloc_type structure for statistics reporting which is not
1221 yet MP safe, and the backing store (KVM routines) are not MP safe.
1222 Even, so making the whole thing MP safe is not expected to be difficult.
1225 * Robert Garrett has made great process removing __P(), Jeffrey has been
1226 working on the nework stack, David has been backporting bug fixes from
1227 FreeBSD and doing other cleanups, and Hiten and Jeroen have been
1228 investigating documentation and source code reorganization.
1231 ## Sun 10 August 2003 - Source Reorganization
1233 * A major source tree reorganization has been accomplished, including
1234 separation of device drivers by functionality, moving filesystems into
1235 a vfs/ subdirectory, and the removal of the modules/ subdirectory with
1236 an intent to integrate the module makefiles into the normal sys/
1239 * Work on syscall messaging is ongoing and we will soon hopefully have
1240 a fully working demonstration of asynch messaging.
1243 ## 09 July 2003 to 22 July 2003 - Misc work, DEV messaging
1245 * A large number of commits related to the messaging infrastructure have
1246 been made, and DEV has been message encapsulated (though it still runs
1249 * Announced the official start of work on the userland threading API:
1250 <a href="misc/uthread.txt">misc/uthread.txt</a>
1254 ## 27 June 2003 to 09 July 2003 - MP operation
1256 This section contains work completed so far. Some items have not yet
1257 been integrated into the next section up.
1259 * **(done)** Get it so user processes can run simultaniously
1260 on multiple cpus. User processes are MP, the kernel is still
1261 mostly UP via the MP lock.
1262 * **(done)** Run normal interrupts and software interrupts
1263 in their own threads.
1265 * **(done)** Implement interrupt preemption with a
1266 block-return-to-original-thread feature (which is more
1267 like BSDI and less like 5.x).
1268 * **(done)** Finish separating the userland scheduler
1269 from LWKT. The userland scheduler now schedules one LWKT
1270 thread per cpu. Additionally, the realtime, normal, and
1271 idle priority queues work as expected and do not create
1272 priority inversions in the kernel. Implement a strict
1273 priority+rr model for LWKTs and assign priorities to
1274 interrupts, software interrupts, and user processes running
1275 in kernel mode. Deal with ps, systat, and top. Fix
1276 bugs in the sysctl used to retrieve processes. Add threads
1277 to the information returned by said sysctl.
1278 * Replace the UIO structure with a linked list of VM objects,
1279 offsets, and ranges, which will serve for both system and
1281 * Move kernel memory management into its own thread(s),
1282 then pull in the SLAB allocator from DUX3, implement a
1283 per-cpu cache, and get rid of zalloc*().
1284 * Implement virtual cpus, primarily for testing purposes.
1286 * (done) Separate scheduling of user processes from the
1288 i.e. only schedule one user process in the LWKT scheduler at
1290 * Change over to a flat 64 bit I/O path.
1291 * (done) Get real SMP working again... implement the token
1292 transfer matrix between cpus, keep the BGL for the moment but
1293 physically move the count into the thread structure so it
1294 doesn't add (much) to our switching overhead.
1295 * Fix BUF/BIO and turn I/O initiation into a message-passing
1296 subsystem. Move all DEVices to their own threads and
1297 implement as message-passing. VFS needs a major overhaul
1298 before VFS devices can be moved into their own threads to
1299 the reentrant nature of VFS.
1302 ## 17 June 2003 to 26 June 2003 - Add light weight kernel threads to the tree.
1304 This work has been completed. It involved creating a clearly defined
1305 thread/process abstraction.
1308 * **(done)** embed a thread structure in the proc structure.
1309 * **(done)** replace the curproc global with curthread, create
1310 macros to mimic the old curproc in C code.
1311 * **(done)** Add idlethread. curthread is never NULL now.
1312 * **(done)** replace the npxproc global with npxthread.
1314 * **(done)** Separate the thread structure from the proc structure.
1315 * **(done)** remove the curpcb global. Access it via curthread.
1316 ('curthread' will be the only global that needs to be
1317 changed when switching threads. Move the PCB to the end
1318 of the thread stack.
1319 * **(done)** npxproc becomes npxthread.
1320 * **(done)** cleanup globaldata access.
1322 * **(done)** Separate the heavy weight scheduler from the thread
1323 scheduler and make the low level switching function operate
1324 directly on threads and only threads. Heavy weight process
1325 switching (involving things like user_ret, timestamps,
1326 and so forth) will occur as an abstraction on top of the
1327 LWKT scheduler. swtch.s is almost there already.
1328 The LWKT switcher only messes with basic registers and
1329 ignores the things that are only required by full blown
1330 processes ( debug regs, FP, common_tss, and CR3 ). The heavy
1331 weight scheduler that deals with full blown process contexts
1332 handles all save/restore elements. LWKT switch times are
1334 * **(done)** change all system cals from (proc,uap) to (uap).
1335 * **(done)** change the device interface to take threads instead of
1336 procs (d_thread_t is now a thread pointer). Change the
1337 select interface to take thread arguments instead of procs
1338 (interface itself still needs to be fixed). Consolidate
1339 p_cred into p_ucred. Consolidate p_prison into p_ucred.
1340 Change suser*() to take ucreds instead of processes.
1341 * **(done)** Move kernel stack management to the thread structure.
1342 Note: the kernel stack may not be swapped. Move the pcb to
1343 the top of the kernel stack and point to it from
1344 the thread via td_pcb (similar to FreeBSD-5).
1346 * **(done)** Get rid of the idiotic microuptime and related
1347 crap from the critical path in the switching code.
1348 Implement statistical time statistics from the stat clock
1349 interrupt that works for both threads and processes without
1350 ruining switching performance (very necessary since there
1351 is going to be far more kernel->kernel done with this
1352 design, but it also gets rid of a serious eyesore that
1353 has been in the FreeBSD tree for a long time).
1354 * **(done)** Translate most proc pointers into thread pointers
1355 in the VFS and DEV subsystems (which took all day to do).
1356 * **(done)** Cleanup VFS and BUF. Remove creds from calls that should
1357 not pass them, such as close, getattr, vtruncbuf, fsync, and
1358 createvobject. Remove creds from struct buf. All of the
1359 above operations can be run in any context, passing the
1360 originator's cred makes no sense.
1361 * **(95% done and tested)** Remove all use of **curproc** and
1362 **curthread** from the VFS
1363 and DEV subsystems in preparation for moving them to
1364 serializable threads. Add thread arguments as necessary.
1366 * **(95% done and tested)** Finish up using thread references
1367 for all subsystems that
1368 have no business talking to a process. e.g. VM, BIO.
1370 * **(done)**Make tsleep/wakeup work with pure threads.
1371 * **(done, needs more testing. buildworld succeeds on test box)**
1372 Move kernel threads (which are implemented as heavy-weight
1373 processes to LWKT threads).
1376 ## 16 June 2003 - Completed repository fork, $Tag change, and
1379 * Creating a new repository required a bit of work. First the RELENG_4
1380 tree had to be checked out. Then a new cvs repository was created
1381 with a new cvs tag. Then a combination of scripts and manual work
1382 (taking all day) was used to move the $FreeBSD tags into comments
1383 and add a new tag to all the source files to represent the new
1384 repository. I decided to cleanup a large number of FBSDID and
1385 rcsid[] declarations at the same time, basically moving all tag
1386 descriptions into comments with the intent that a linking support
1387 program would be written later to collect tags for placement in
1388 binaries. Third party tags in static declarations which contained
1389 copyright information were retained as static declarations.
1391 * Some minor adjustments to the syscall generator was required, and
1392 I also decided to physically remove UUCP from the tree.
1394 * Finally, buildworld and buildkernel were made to work again and
1395 the result was checked in as rev 1.2 (rev 1.1 being the original