|2006-07-04||src||Outstanding||RELENG_6 HEAD||Have adjkerntz(8) exit cleanly and silently when run in a jail(8)'ed environment.|
|2006-06-20||src||Outstanding||RELENG_6 HEAD||Add the appropriate tunables to give jail IPs a netmask that way a jail+if_vlan can be the only IP/alias from a host in a given subnet. Ex:|
cloned_interfaces="vlan598" ifconfig_bge0="media 100baseTX mediaopt full-duplex up" ifconfig_vlan598="vlan 598 vlandev bge0 up" jail_foo_rootdir="/jails/foo.example.com" jail_foo_hostname="foo.example.com" jail_foo_ip="10.42.42.254" jail_foo_netmask="255.255.255.0" jail_foo_interface="vlan598" jail_foo_devfs_enable="YES" jail_foo_mount_enable="YES" ...
|2005-06-02||src||Outstanding||HEAD||Sometimes there are applications that you have to run in Linux compat mode and sometimes those programs are so braindead that they use back asswards syscalls such as setfsuid16() or setfsgid16() that are no-op's on FreeBSD and don't change the execution of the program, but, it's less than desirable to have dmesg(1) filled with these worthless messages. This patch adds a sysctl(8) value compat.linux.syscall_warn_enable that lets one disable all compat warnings.|
|2005-03-22||src||Outstanding||HEAD||Fixes the use of "sufficient" for authentication PAM modules. Required for some pam configurations. Ex: |
|2004-12-09||FreeBSD 5.X||Outstanding||RELENG5||This patch fixes getcwd(3) on RELENG5 on unionfs's.|
|2004-01-13||PostgreSQL||Outstanding||HEAD||On machines that have TCP_NOPUSH defined in netinet/tcp.h, this will change all libpq connections to use setsockopt() TCP_NOPUSH. This hasn't been widely tested in terms of performance, but it _should_ make receiving data faster.|
|2003-09-18||ports||Outstanding||HEAD||The patch supposed by ISC doesn't apply cleanly to the ports/dns/bind9 port. This patch, however, does and is functionally identical.|
|2003-07-20||src||Outstanding||HEAD||Adds a kern.nbuf sysctl, making it possible to see what the nbufs are set to. Also documents kern.nbuf in loader(8) and /boot/loader.help|
|2003-04-16||PostgreSQL||Outstanding||HEAD||Implements the various SHOW commands in psql to appease the whining MySQL masses. PostgreSQL_usability++, MySQL--|
|2003-03-13||doc||Outstanding||HEAD||White space reduction: noticed while fixing a link. Gratuitous tabbing in emacs to indent properly is fine so long as you remove your newly added trailing whitespace.|
|2003-03-12||doc||Outstanding||HEAD||Add the appropriate #include where environ is defined. It's actually stand.h but unistd.h is the correct header to include.|
|2003-03-10||doc||Committed - 2003-03-12||HEAD||This patch updates sendfile(2) to include a TUNING section with some advice along with an update to tuning(7) to reference the updated sendfile(2) man page.|
|2003-03-10||src/doc||Outstanding||HEAD||For a system administrator trying to learn about the kernel and its limits/parameters, it can be quite problematic to have to trudge through the handbook and various man pages scattered about to learn about a specific sysctl MIB that may or may not be documented. This patch addresses this by creating a new man page, sysctl(5), that is used as an index for searching for what man pages contain various references to sysctl MIBs. This man page scales well in that there can be multiple Xr macros and the references will wrap nicely. Maybe tomorrow night I'll create the stub of a man page that can be used for the kernel config and start pushing kernel config bits into various man pages.|
Updates the case of sendfile() when there aren't any sf_buf's available. Instead of calling msleep() and blocking the caller on a socket that has been marked non-blocking, return instantly with EAGAIN. This doesn't provide a mechanism for identifying that there aren't any sf_buf's available. At some point a read only sysctl should be added that lets an administrator know how many sf_buf's are free (max number already exists so it should be trivial for an admin to figure out how many are in use), but that will come at a later date (it's 5:30am, ENOSLEEP). Returning control to the program should dramatically improve the responsiveness of a non-blocking application that uses sendfile() since it gives the application the chance to continue to work. Ideally sf_buf's are freed up for use making it possible to deal with the load described. Currently when this limit is hit, it kills the concurrency of the server that non-blocking IO affords and the throughput of a system drops from __Mbps down to near 0Mbps (I have some cool bandwidth graphs to prove this). With this patch, at the very least the server should be able to continue to send data at __Mbps.
There is a race with this code in that if a non-blocking socket that has had sendfile() called on it where there wasn't an sf_buf available, and it is set back to being a blocking socket, sf_buf_alloc_want will never reach zero and as a result, wakeup_one(&sf_freelist) will be called every time in sf_buf_free(). I'm not sure how best to fix this if it should be, or even what the behavior of wakeup_one(&sf_freelist) will be if it is called when there isn't anything msleep()'ing on sf_freelist.
Another something that I'd like to do is to alert the server admin that he/she's out of sf_buf's the same way that the kernel does when it runs out of nmbclusters.
If someone doesn't believe this is a problem, look at the bandwidth graph from a single box of mine when it hits this limit (remaining traffic is from servers that use writev(), not sendfile()).
UPDATED (2003-03-10): This patch is updated to include Alan Cox's changes to re-use sf_buf's for pages that are already in use by the server. This patch enables re-use of sf_buf's, which is a likely common case for sendfile(2).
|2003-03-04||doc||Committed - 2003-03-04||HEAD||Chump update to note that kern.ipc.nsfbufs can be adjusted at boot time.|
|2003-03-04||src||Committed - 2003-03-04||HEAD||Updates loader.8 and /boot/loader.help (actually help.common) to indicate that kern.ipc.nsfbufs can be updated at boot time.|
|2003-03-03||doc||Committed - 2003-03-04||HEAD +whitespace||Updates the handbook to include a few bits about sf_buf exhaustion and tuning NSFBUFS.|
|2003-02-16||doc||Committed||HEAD||This updates the firewall rules in the Dialup Firewalling article to use IPFW2 and to be substantially more secure (as well as less error prone and simpler).|
|2003-02-14||doc||Committed||HEAD||This patch fundamentally provides a small update to the example .emacs file to add highlighting of whitespace. It also nukes some whitespace and americanizes the spelling of two common words: intepreted -> interpreted and optimise -> optimize.|
|2003-02-11||ports||Outstanding||HEAD||Along with an updated random(6) utility, the following patch is used to randomize the MASTER_SITES for downloading distfiles for ports|
|2003-02-09||src||Committed - 2003-03-05||HEAD||Updates src/sbin/ping to be WARNS=2 compliant|
|2003-02-08||src||Committed - 2003-02-15||-CURRENT||Patch for -CURRENT (should work for 4.X and even probably 3.X). Updates random(6) to be able to completely reorder a file based off of lines or words as specified by isspace(3).|
|2001-02-07||src||Outstanding||HEAD||on all *BSD version, socket buffers contain a list of incoming and/or outgoing mbufs. Unfortunately the list only has a pointer to the head, meaning that all append operations require to scan the full list. The overhead can be very bad in some cases (e.g. small UDP packets), and becomes worse and worse as the socket buffer size increases (which is what one would commonly do when expecting a lot of traffic!). The attached patch implements a tail pointer to the list, so that you can append in constant time. By default, the code works exactly as before -- the tail of the list is reached with the usual linear scan, and the pointer to the tail is only used for comparison purposes to make sure that it yields the same value. If you enable the fast behaviour with sysctl -w kern.ipc.fastscan=1 then the new code takes over and linear scans are replaced by dereferences of the tail pointer. Apart from the obvious benefits of using O(1) instead of O(n) algorithms, your mileage may vary. When the socket buffer is almost always empty (fast receivers) then you have no gain. When the socket buffer is almost always full you also have no gain, because the decision to drop the packet only requires a comparison. However, this code can really avoid trashing in those cases where the queue size oscillates. I'd like to commit this (or similar) code after proper testing, so i'd like people to try it out -- I am reasonably confident about it, and have done a fair amount of testing under heavy udp load, but want to be sure that there are no side effects.|
|1999-06-01||src||Outstanding||A patch to add NEWALIAS and NETBIND to the kernel. Makes adding entire cidr blocks as aliases possible.|
The XSL, XML, and DTD used to generate this page are available for download under the BSDL license.
Copyright © 2003 Sean Chittenden
<seanc@FreeBSD.org>. All rights reserved.|
Last Updated: Sun Sep 14 07:15:00 UTC 2014