New DragonFlyBSD STABLE iso released

| No TrackBacks |

Shortly after OpenBSD's 3.5 splash web, Matthew Dillon has released a thrid "known stable" iso for DragonFlyBSD. It can be obtained from ftp://ftp.dragonflybsd.org/iso-images/dfly-20040502.iso.gz.

This snapshot included many new features, yet stable:

NEWTOKENs replace old tokens. The new token abstraction provides a serialization mechanism that persists across blocking conditions. This is not a lock, but more like an MP-capable SPL, in that if you block you will lose serialization but then regain it when you wakeup again. This means that newtokens can be used for serialization without having to make lower level subsystems aware of tokens held by higher level subsystems in the call stack. This represents a huge simplification over the FreeBSD-5 mutex model.
Support added for the Silicon Image SATA controller.
DragonFly now supports 16 partitions per slice (up from 8).
Wildcarded sockets have been split from the TCP/UDP connection hash table, and the listen table is now replicated across cpus for lockless access.
UDP sendto() without a connect no longer needs to make a temporary connect inside the kernel, greatly improving UDP performance in this case.
NFS performance has been greatly improved.
Fix some major bugs in the USB/SIM code, greatly improving the reliability of USB disk keys and related devices.
NEWCARD has been brought in from FreeBSD-5.
A bunch of userland scheduler performance issues have been fixed.
Major syscall procedural separation has been completed, separating the user interfacing portion of a syscall from the in-kernel core support, allowing the core support functions to be directly called from emulation code instead of using the stackgap junk.
An optimized MMX and XMM kernel bcopy has been implemented and tested. Most of i386/i386/support.s has been rewritten and a number of FP races in that and npxdna() have been closed.
Brought in the SFBUF facility from FreeBSD-5, including all of Alan Cox's (the FBsd Alan Cox) improvements, plus additional improvements for cpu-localized invlpg calls (and the ability to avoid such calls when they aren't needed).
A major PIPE code revamp has occured, augmenting the SFBUF direct-copy feature with a linear map. Peak standard pipe throughput with 32KB buffers on an AMD64 3200+ now exceeds 4GBytes/sec.
Implement XIO, which is a multi-page buffer wrapper for SFBUFs and will eventually replace linear buffer management in the buffer cache as well as provide linear buffer mappings for other parts of the system.
Added a localized cpu pmap_qenter() called pmap_qenter2() which is capable of maintaining a cpumask, to avoid unnecessary invlpg's on target-side linear maps. It is currently used by the PIPE code and the CAPS code.
Joerg has brought in most of the KOBJ extensions from FreeBSD-5.
Major continuing work by Jeff on the threading and partitioningl of the network stack. Nearly the whole stack is now theoretically MP safe, with only a few niggling issues before we can actually start turning off the MP lock.
The last major element of the LWKT messaging system, lwkt_abortmsg() support, is now in place. Also it is now possible to wait on a port with PCATCH.
Hiten has revamped the TCP protocol statistics, making them per-cpu along the same lines that we have made vmstats per-cpu.
propolice has been turned on by default in GCC3.
CAPS (DragonFly userland IPC messaging) has been revamped to use SFBUFs.
acpica5 now compiles, still needs testing.
SYSTIMERS added, replacing the original hardclock(), statclock(), and other clocks, and is now used as a basis for timing in the system. SYSTIMERS provide a fine-grained, MP-capable, cpu-localizable and distributable timer interrupt abstraction.
Fix RTC write-back issues that were preventing 'ntpdate' changes from being written to the RTC in certain cases.
Finish most of the namecache topology work. We no longer need the v_dd and v_ddid junk. The namecache topology is almost ready for the next major step, which will be namespace locking via namecache rather then via vnode locks.
Add ENOENT caching for NFS, greatly reducing the network overhead (by a factor of 5x!!!) required to support things like buildworld's using an NFS-mounted /usr/src.
Jeff has added predicate messaging requests to the network subsystem, which allows us to convert situations which normally block, like connect(), to use LWKT messages.
Lots of style cleanups, primarily by Chris Pressey.

No TrackBacks

TrackBack URL: https://blog.delphij.net/mt/mt-tb.cgi/265

Monthly Archives

Pages

OpenID accepted here Learn more about OpenID
Powered by Movable Type 5.2.3