Matthew Dillon [Wed, 15 Sep 2010 03:19:17 +0000 (20:19 -0700)]
network - Increase basic mbuf size from 256 to 384 bytes
* Due to the bloat in m_hdr and m_pkthdr the 256-byte mbuf structure
is no longer large enough and there appears to be quite a bit of
legacy code still using m_get() and making assumptions on the
available space without checking actual space.
We have assertions in place to catch these but stabilizing the
system is more important right now.
* Increase the basic mbuf buffer size from 256 to 384 bytes.
Matthew Dillon [Tue, 14 Sep 2010 23:28:53 +0000 (16:28 -0700)]
network - Major netmsg retooling, part 1
* Remove all the netmsg shims and make all pr_usrreqs and some proto->pr_*
requests directly netmsg'd.
* Fix issues with tcp implied connects and tcp6->tcp4 fallbacks with
implied connects.
* Fix an issue with a stack-based udp netmsg (allocate it)
* Consolidate struct ip6protosw and struct protosw into a single
structure and normalize the API functions which differed between
the two (primarily proto->pr_input()).
* Remove protosw->pr_soport()
* Replace varargs protocol *_input() functions (ongoing) with fixed
arguments.
Matthew Dillon [Tue, 14 Sep 2010 22:59:28 +0000 (15:59 -0700)]
vkernel - Improve memory image file startup
* Remove the code that pre-filled a memory image file with zeros. It's
completely worthless, particularly with HAMMER.
* On startup truncate the memory file to 0 and then extend to the
memory size, deleting any backing store from the prior vkernel run.
The new file will start out full of holes.
This greatly improves vkernel startup time.
Sascha Wildner [Tue, 14 Sep 2010 19:48:10 +0000 (21:48 +0200)]
twa(4): Sync with FreeBSD (twa(4) version 3.80.06.003).
Thanks to Xin Li for notifying me of this update.
Tested-by: Damian Lubosch <dl@xiqit.de>
Sascha Wildner [Tue, 14 Sep 2010 18:03:10 +0000 (20:03 +0200)]
ie(4): This driver is ISA only, so remove some unneeded files from SRCS.
Jan Lentfer [Sat, 11 Sep 2010 22:34:08 +0000 (00:34 +0200)]
pf: Make pf work w/ the MPSAFE network stack
add pf_token where appropriate
in pf_socket_lookup() use lwkt_domsg() instead of lwkt_sendmsg()
to make race conditions more unlikely
if_pfsyn.c: re-add lost init code
Jan Lentfer [Sat, 11 Sep 2010 18:50:32 +0000 (20:50 +0200)]
pf: Revert commit 5165ac2
I was too hasty changing byte ordering when trying
to track down a NAT problem
Sepherosa Ziehau [Tue, 14 Sep 2010 14:26:47 +0000 (22:26 +0800)]
ACPI P-State: Force P-State to use the first usable entry in P-State table
It looks like on certain boxes P-State will be set to the last usable
P-State (i.e. lowest frequency)
Sepherosa Ziehau [Tue, 14 Sep 2010 13:43:14 +0000 (21:43 +0800)]
ACPI P-State: When there is no _PSD, create one CPU domain for each CPU
Sepherosa Ziehau [Tue, 14 Sep 2010 13:29:00 +0000 (21:29 +0800)]
test commit
Matthew Dillon [Tue, 14 Sep 2010 01:40:54 +0000 (18:40 -0700)]
network - UDP currently only going to one proto thread
* Adjust udp_addrcpu() to always return cpu 0 for now, the UDP
implementation currently only operates on protocol thread 0.
Matthew Dillon [Tue, 14 Sep 2010 01:40:22 +0000 (18:40 -0700)]
network - Protect soreceive() from backend
* Somehow I missed the token required in soreceive() to protect it from
the backend.
Matthew Dillon [Tue, 14 Sep 2010 00:04:03 +0000 (17:04 -0700)]
network - Fix udp self-referential panic
* udp_ctlinput() can't call domsg, it has to start the chain going with
lwkt_sendmsg().
* Currently udp only runs on protocol thread cpu 0.
Matthew Dillon [Mon, 13 Sep 2010 23:53:56 +0000 (16:53 -0700)]
Merge branch 'master' of ssh://crater.dragonflybsd.org/repository/git/dragonfly
Matthew Dillon [Mon, 13 Sep 2010 23:50:55 +0000 (16:50 -0700)]
kernel - swapoff - regenerate system calls
* Added swapoff, regenerate system calls.
Matthew Dillon [Mon, 13 Sep 2010 23:41:40 +0000 (16:41 -0700)]
Kernel - Implement swapoff
* Generally port of the swapoff implementation from FreeBSD to DragonFly,
with major modifications.
Modifications to handle swapcache issues (VCHR vnodes with VM objects
can have swap associations for swapcache).
* Libkvm changes
So there are two problems with libkvm. The first is not really
swapoff-related - the new sysctl way of reporting numbers bzero'es
swap_max elements in the given swap_ary array. This is in contrast to
the old kvm way, which bzero'es only those elements that will be
actually filled. So if we have 3 swap devices and swap_max is 16, then
the sysctl code will zero out all 16 elements and fill the first 4,
while the old kvm code will zero out exactly 4 elements and fill them.
Since we want to keep API stable (I learned it the hard way :-) ) I
think this fix can be separated out and go to master as a bugfix to the
newly introduced sysctl way of reporting things.
The second problem only shows up if we introduce a swapoff syscall
and enforce using of the old kvm way. It was written with the
assumption that swap devices can only be added, not removed - it
assumes than if I have a swap device with index 3, 4 swap
devices are active. This is not true with swapoff - I can swapon
A, B, C and D, then swapoff B and C and here we are - I have an
active swap device with index 3, but only 2 devices are active.
It turned out to be easier to just rewrite it (based on sysctl way),
because that assumption was rather deep and everything was based on it.
Since along with sysctl way per-device swap accounting was introduced,
the kvm way now uses it instead of scanning blist.
Which brings us to the last change - blist scanning code is now used
only for debugging purposes. getswapinfo_radix() is now called only if
DUMP_TREE flag is set. Pieces that touched swap_ary entries are removed,
swap_ary and swap_max are no longer passed to scanning code.
After all that both ways are now working correctly with the regards to
the swapoff call and the old kvm way (the behaviour is exactly the same,
all boudary cases were tested, API remains the same). The only (minor)
difference is that swapctl numbers are a little bit bigger than kvm way
ones. Thats because kvm way subtracts dmmax (the assumption is that the
first dmmax is never allocated), and sysctl way does not. I tried to fix
this, but it turns out that we need to introduce a dmmax sysctl for that.
So if you want I can add it, but I want to hear from you first (both on
this thing and my changes to libkvm in general).
* Userspace. Add swapoff & adjust manual pages.
Note: Bounty project ($300)
Submitted-by: Ilya Dryomov <idryomov@gmail.com>
Sascha Wildner [Mon, 13 Sep 2010 17:52:26 +0000 (19:52 +0200)]
Fix some mdoc issues in tcp.4's SEE ALSO.
Sascha Wildner [Mon, 13 Sep 2010 17:08:22 +0000 (19:08 +0200)]
Fix two typos in manual pages and messages.
Matthew Dillon [Mon, 13 Sep 2010 15:58:51 +0000 (08:58 -0700)]
build - Fix vkernel installation target for /usr/src/test/vkernel
* Fix the installation target in /usr/src/test/vkernel/Makefile, it
was assuming the old style /boot kernel layout instead of the new.
Matthew Dillon [Mon, 13 Sep 2010 15:20:16 +0000 (08:20 -0700)]
devfs - Fix use-after-free case when making pty's invisible
* Fix a use-after-free case when making a pty devfs node invisible.
The dev_dclose(). Move the test/flag to before the dev_dclose() call.
* Document that the pty code may destroy the device in the dev_dclose()
call, causing the node to become stale.
Reported-by: Francois Tigeot <ftigeot@wolfpond.org>
Reminded-by: sjg
Venkatesh Srinivas [Mon, 13 Sep 2010 11:26:33 +0000 (04:26 -0700)]
Fix !INVARANTS build and reformat token asserts for easy reading.
Matthew Dillon [Mon, 13 Sep 2010 07:08:53 +0000 (00:08 -0700)]
network - Fix multiple MP races (2)
* MEVENT signaling needs the ssb_token as well as the kq_token for now
to prevent blocking inside the predicate. This is a hack for now.
* Add missing porttoken protection in in_pcbremlists().
Reported-by: lentferj
Matthew Dillon [Mon, 13 Sep 2010 05:33:08 +0000 (22:33 -0700)]
network - Fix multiple MP races
* Fix sonewconn() races. sonewconn() was attaching prior to changing
the socket->so_port, relying on the caller to set the socket->so_port.
This resulted in a race where userland wound up with visibility on the
socket and could issue commands, like close(), which would end up going
to the original protocol thread instead of the post-connect protocol thread
which was handling the sonewconn().
Thus the close() could message the backend to detach and compete
against the sonewconn() because the detach message was going to
a different protocol thread.
* When the socket->so_port is changed the inpcb was not being moved
from the old pcbinfo->pcblisthead list to the new one, resulting
in MP races later on during removal.
* Add more debugging kprintf()s.
* Clean up sosetport() use, remove the now-unused *_soport_attach().
Reported-by: Many
Matthew Dillon [Mon, 13 Sep 2010 03:23:22 +0000 (20:23 -0700)]
network - remove the redispatch local
* Remove the resdispatch local variable which is no longer used.
Peter Avalos [Mon, 13 Sep 2010 02:09:43 +0000 (02:09 +0000)]
savecore: Fix typo in comment.
Matthew Dillon [Sun, 12 Sep 2010 20:14:23 +0000 (13:14 -0700)]
network - Add debugging assertions
* Add some assertions to try to catch failure cases earlier.
Matthew Dillon [Sun, 12 Sep 2010 17:30:38 +0000 (10:30 -0700)]
kernel - Fix one-cycle MP race in vshouldmsync()
* vshouldmsync() is the mntvnode fast function, which is called without
any vnode lock. vp->v_object can thus get ripped out from under the
scan function.
Hold vmobj_token through the scan so any pointer accessed via
v_object remains stable (even if no longer related to the vnode
due to the race).
Reported-by: swildner
Sascha Wildner [Sun, 12 Sep 2010 17:12:54 +0000 (19:12 +0200)]
pfctl.8: Adjust date.
Sascha Wildner [Sun, 12 Sep 2010 17:12:36 +0000 (19:12 +0200)]
crashinfo.8: Remove trailing whitespace.
Matthew Dillon [Sun, 12 Sep 2010 04:35:14 +0000 (21:35 -0700)]
network - Assert that the packet's data has not overrrun the buffer in m_free()
* Add an assertion to try to catch subsystems which blow up a mbuf's
data buffer.
Sascha Wildner [Sat, 11 Sep 2010 20:55:05 +0000 (22:55 +0200)]
pfctl(8): Fix some printf issues (and buildworld on x86_64).
Matthew Dillon [Sat, 11 Sep 2010 20:37:21 +0000 (13:37 -0700)]
network - Fix tcp inpcb race
* tcbinfo[cpu].porthashbase was being shared across all the cpus,
creating MP races. Change it so it isn't shared.
Reported-by: "Samuel J. Greear" <sjg@evilcode.net>
Matthew Dillon [Sat, 11 Sep 2010 18:53:57 +0000 (11:53 -0700)]
network - More tokens for ipsec
* Get key_token in more places to fix MP races.
Matthew Dillon [Sat, 11 Sep 2010 18:52:45 +0000 (11:52 -0700)]
kernel - cleanup & assertions in mbuf code
* Assert the mbuf's next/nextpkt fields are NULL on allocation.
Matthew Dillon [Sat, 11 Sep 2010 18:50:56 +0000 (11:50 -0700)]
kernel - Fix mprace in kern_objcache
* mag_purge() is interruptable, do not continue the purge
if the magazine is moved. For example, the magazine could
move to the depot and we would wind up continuing to purge
it without the depot lock.
* Add some temporary debugging
Sascha Wildner [Sat, 11 Sep 2010 16:40:54 +0000 (18:40 +0200)]
Re-add RSS_DEBUG to LINT.
Alex Hornung [Sat, 11 Sep 2010 13:21:23 +0000 (14:21 +0100)]
Fix manual break to debugger
* When manually breaking to debugger, we can't hold any tokens as they
get in the way of kbdmux' lockmgr in an interrupt context.
Alex Hornung [Sat, 11 Sep 2010 12:25:34 +0000 (13:25 +0100)]
savecore,crashinfo - fix several problems
* Fix the savecore rc.d script to only run savecore (and crashinfo) when
there's actually a core dump available.
* Limit the kgdb CPU time to 15 seconds to avoid looping forever if we
have a somewhat broken vmcore.
Venkatesh Srinivas [Sat, 11 Sep 2010 11:16:39 +0000 (04:16 -0700)]
kernel: bzeront() - Switch out loop instruction in i386 bzero for sub/jnz.
Much as I like loop, it has a ~7-9 cycle latency on AMD CPUs. Lets spend
idlezero time actually zeroing...
Sascha Wildner [Sat, 11 Sep 2010 09:40:56 +0000 (11:40 +0200)]
ftp-proxy.8 & pflogd.8: Fix some mdoc issues.
Sascha Wildner [Sat, 11 Sep 2010 08:57:34 +0000 (10:57 +0200)]
Fix LINT build.
I never get why LINT isn't just checked before pushing such huge
changes. Takes just a couple of minutes, really. :)
Matthew Dillon [Sat, 11 Sep 2010 06:02:39 +0000 (23:02 -0700)]
network - Correct double free of mbuf during reboot
* Correct code which was leaving a stale mbuf pointer intact
when flushing the so_rcv sockbuf in a socket. This normally
occured during shutdown/reboot.
Matthew Dillon [Sat, 11 Sep 2010 05:41:30 +0000 (22:41 -0700)]
build - Fix netgraph
* Some source files were missing newly required includes for their
use of the mplock and socketvar2.h inlines.
Reported-by: lentferj
Matthew Dillon [Sat, 11 Sep 2010 11:58:35 +0000 (11:58 +0000)]
network - raw_input needs further protection
* We also need the so_rcv.ssb_token to protect against userland
Samuel J. Greear [Sat, 14 Aug 2010 14:41:35 +0000 (14:41 +0000)]
devfs - Let devfs assume degenerate knotes when a device goes away
* Some devices can disappear while knotes are still active on the klist
embedded in the devices softc or etc., create a mechanism to allow devfs
to assume these degenerate knotes.
* Convert ums to this new way of things
* This does not actually fix the reported panic (where X is holding a mouse
open across detach), panic is moved into the new devfs handler.
* Other devices which could be detached and support the kq interface need to
use this interface as well (umass, ...)
Reported-by: Rumko
Matthew Dillon [Sat, 11 Sep 2010 09:32:24 +0000 (09:32 +0000)]
network - Fixes for wpa, general sockets.
* netisr_characterize() was not properly handling unknown
characterizations (array overflow).
* The raw protocol was not MPSAFE.
* Protect kqinfo->ki_mlist in sowakeup
Matthew Dillon [Fri, 10 Sep 2010 19:09:34 +0000 (12:09 -0700)]
network - Tokenize NFS, fix MP races
* Now that the rest of the network stack is running MPSAFE, poor NFS is
hitting races and other issues because it was depending on the MP lock.
* Recombobulate NFS with tokens, protecting all border crossings:
A global nfs_token is used for the nfs mount list, nfsd list, and
server socket list.
A per-socket token (nfssvc_sock->ns_token) governs each served mount.
A per-mount token (nfsmount->nm_token) governs each client mount.
* Callouts and TCP upcalls are protected. The per-socket TCP upcall
is protected by the nfssvc_sock token.
* The NFS iod thread pairs and nfsd threads now run MPSAFE.
* NFSv3 is now holy-shit fast and can trivially max-out a GigE link
without TSO when the server is not otherwise limited by server-side
disks.
Matthew Dillon [Fri, 10 Sep 2010 16:09:17 +0000 (09:09 -0700)]
Merge branch 'master' of ssh://crater.dragonflybsd.org/repository/git/dragonfly
Jan Lentfer [Fri, 10 Sep 2010 13:39:07 +0000 (15:39 +0200)]
pf: Fix some byte ordering
Venkatesh Srinivas [Fri, 10 Sep 2010 13:03:59 +0000 (06:03 -0700)]
kernel - tmpfs: Set M_NULLOK on node allocations and safe node_init against
null node pointers.
Dragonfly-bug: Still 1726 (tmpfs malloc limit exceeded).
Alex Hornung [Fri, 10 Sep 2010 09:49:59 +0000 (10:49 +0100)]
savecore - Allow automatic text summary generation
* Add two rc variables to automatically generate the text summary for a
core dump every time savecore does its job.
Obtained-From: FreeBSD
Matthew Dillon [Fri, 10 Sep 2010 09:28:37 +0000 (02:28 -0700)]
Merge branch 'master' of ssh://crater.dragonflybsd.org/repository/git/dragonfly
Matthew Dillon [Fri, 10 Sep 2010 09:28:06 +0000 (02:28 -0700)]
Merge branch 'master' of ssh://crater.dragonflybsd.org/repository/git/dragonfly
Matthew Dillon [Fri, 10 Sep 2010 09:27:41 +0000 (02:27 -0700)]
build - Fix kernel build issue
* Missing include for token declaration
Matthew Dillon [Fri, 10 Sep 2010 09:25:58 +0000 (02:25 -0700)]
Merge branch 'master' of ssh://crater.dragonflybsd.org/repository/git/dragonfly
Matthew Dillon [Fri, 10 Sep 2010 04:01:20 +0000 (21:01 -0700)]
network - MP socket free & abort interactions, so_state
* Add so_refs and ref-count the socket structure to deal with
MP races on sofree().
* Ref the socket structure for all soabort() operations (they are usually
asynchronous). The netmsg_pru_abort() handler will sofree() the ref
after calling the protocol stack's abort function.
* Use atomic ops to set and clear bits in so_state, because it is modified
by both the fronttend and the backend.
* Remove numerous critical sections that are no longer effective.
* Protect the accept queues with so_rcv.ssb_token.
* Protect after-the-fact calls to soisdisconnected() with a soreference()
to avoid use-after-free cases.
* Wrap unix domain, mroute, div, raw, and key sockets/protocols with their
own private tokens.
Alex Hornung [Fri, 10 Sep 2010 09:15:57 +0000 (10:15 +0100)]
crashinfo - Add script to gather info from dumps
* Add a script to automatically gather as much info as possible from a
crash dump and save it in text form.
* This is originally from FreeBSD, with a bunch of changes related to
DragonFly.
Obtained-from: FreeBSD
Alex Hornung [Fri, 10 Sep 2010 09:13:05 +0000 (10:13 +0100)]
gdb.kernel - Add commands for tokens
* Add lstok and infotok commands to list some information about the
global mpsafe-ing tokens and show information about any token.
Matthew Dillon [Fri, 10 Sep 2010 09:11:46 +0000 (02:11 -0700)]
build - Fixup world/kernel build
* Fixup the world/kernel build for 64 bit etc. Cleanup recent work.
Matthew Dillon [Fri, 10 Sep 2010 08:51:32 +0000 (01:51 -0700)]
systat - Fix overflow in page calculation on display
* Fix an overflow that caused large amounts of memory to display
incorrectly in the VM page queues column.
* Display the Act/All fields from KB to MB.
Sascha Wildner [Fri, 10 Sep 2010 08:30:37 +0000 (10:30 +0200)]
Hopefully fix mountroot on our LiveCDs.
Before the recent loader changes, the root device to use for CD-ROM
booting (/dev/cd0) was properly passed from the loader to the kernel
so the entry in fstab was never used.
However, with the new loader code this doesn't happen anymore (for
yet unknown reasons), so the fstab entry is tried and fails, because
it is wrong.
So for now, adjust our fstabs to use /dev/cd0.
Alex Hornung [Fri, 10 Sep 2010 08:23:09 +0000 (09:23 +0100)]
gdb.kernel - Fix functions
* Fix a bunch of functions in gdb.kernel, as well as some formatting
issues.
Alex Hornung [Fri, 10 Sep 2010 07:12:15 +0000 (08:12 +0100)]
hardclock - Get proc_token for ksignal
* Try to get the proc_token to be able to call ksignal instead of the
mplock.
* This fixes a panic that occured when using luksFormat.
Reported-By: Matthias Schmidt (matthias@)
Matthias Schmidt [Fri, 10 Sep 2010 06:19:49 +0000 (08:19 +0200)]
Merge branch 'master' of git://git.dragonflybsd.org/dragonfly
Peter Avalos [Fri, 10 Sep 2010 02:33:19 +0000 (16:33 -1000)]
cpdup: Fix off-by-one error.
Submitted-by: dillon@
Venkatesh Srinivas [Fri, 10 Sep 2010 01:26:45 +0000 (18:26 -0700)]
kernel - Convert mpipe global token to per-pipe tokens.
Matthew Dillon [Thu, 9 Sep 2010 22:13:36 +0000 (15:13 -0700)]
network - Completely revamp the netisr / dispatch code
* All netisrs are dispatched MPSAFE (followup in later commits)
* Centralize the protocol threads. There is now just one thread
per cpu managed by the netisr code. No more separate tcp/udp threads
on each cpu.
* Make the mbuf M_HASH/m_pkthdr.hash mechanic the central routing
mechanic for netmsgs.
* Remove the netisr ip_mport and pktinfo_portfn stuff and replace with a
cpufn function which handles M_HASH/m_pkthdr.hash when M_HASH is not
already set.
* Seriously clean up the packet input paths. Adjust ether_input_chain() and
friends to not have to adjust the mbuf forwards and backwards, instead
pass a header offset to the ni_cpufn function. The ip pullup and other
related code will use the offset to determine where the ip header is
within the packet.
Matthias Schmidt [Thu, 9 Sep 2010 20:21:33 +0000 (22:21 +0200)]
mkinitrd.8 - Add history and authors section
Jan Lentfer [Thu, 9 Sep 2010 19:40:31 +0000 (21:40 +0200)]
LINT: readd accidently removed iwi driver
Joe Talbott [Thu, 9 Sep 2010 19:03:34 +0000 (15:03 -0400)]
wlan - Bring iwi back into the fold.
This also fixes LINT builds.
Submitted-By: Johannes Hofmann (jh33)
Alex Hornung [Thu, 9 Sep 2010 18:57:33 +0000 (19:57 +0100)]
cryptsetup - fix hash() frontend function
* hash() in backends.c would always return ENOENT, because the return
variable r wasn't being updated.
* luks wasn't affected by this as it uses a different path, but using
cryptsetup without luks would trigger this.
Reported-by: Matthias Schmidt (matthias@)
Sascha Wildner [Thu, 9 Sep 2010 18:57:03 +0000 (20:57 +0200)]
hier.7: Add some lines about /pfs and /var/hammer.
Suggested-by: matthias
Jan Lentfer [Thu, 9 Sep 2010 18:34:04 +0000 (20:34 +0200)]
pf: add patch from rev 1.553 to patch.c
allow RSTs with th_seq == seqlo +- 1, reduces the amount of 'loose state'
messages, add 'dir=' part to 'loose state' message
Jan Lentfer [Thu, 9 Sep 2010 18:09:03 +0000 (20:09 +0200)]
pf: Add patch from rev 1.554
RSTs sent due to "block return(-rst)" could be routed using
the main routing table instead of an alternate one specified
on the block rule.
Jan Lentfer [Thu, 9 Sep 2010 13:54:56 +0000 (15:54 +0200)]
pf: Update pf.c to fix memory leak
pf.c 1.552.2.1 2007/11/27 fixes a memory leak issue.
This was MFC'd to OpenBSD 4.2
Jan Lentfer [Fri, 27 Aug 2010 20:01:56 +0000 (22:01 +0200)]
pf: Update to OpenBSD 4.2
All sorts of informations is now stored directly in
the mbuf header instead of a seperate mbuf tag. This
brings in a 100% performance increase in comparison
to OpenBSD 4.1. For DragonFly this basically means
this is the same performance as in 2.6, but we are
equal again with OpenBSD's pf data structures.
Necesary additions:
sys/net: add more interface groups related functions
if_creategroup()
if_addgroup()
if_delgroup()
if_getgroup()
if_getgroupmembers()
Imported from OpenBSD
carp: add carp_group_demote_adj()
altq: re-add check of packet tagging
fairq & red support, UDP nat'ing, reassembly fixed by Matthew Dillon
Matthew Dillon [Thu, 9 Sep 2010 16:27:40 +0000 (09:27 -0700)]
network - Make toeplitz the default
Matthew Dillon [Thu, 9 Sep 2010 16:18:27 +0000 (09:18 -0700)]
kernel - Add nfs rpc packet sanity check
* Try to catch potential issues with NFS packet generation.
Temporary.
Venkatesh Srinivas [Thu, 9 Sep 2010 10:51:28 +0000 (03:51 -0700)]
libc: nmalloc - Add a per-thread magazine layer and a malloc slab magazine.
The per-thread magazine layer is based on the Solaris umem/vmem paper; it
creates fixed-sized arrays ('magazines') of buffers and binds them to threads.
Local allocations can often now occur with minimal locking and interference.
The slab magazine is meant to buffer calls calls to mmap when requesting slabs.
In tests with MySQL/sysbench OLTP, it leads to fairly dramatic reductions in
the total number of mmap system calls.
nmalloc now also respects the MALLOC_OPTIONS environment variable:
- The 'U' option generates ktrace entries for all malloc/realloc/free calls
- The 'H' option calls madvise liberally to return pages to the system often
- The 'Z' option zeroes all allocations
This work was developed out-of-tree; the RCS log is on leaf.
Matthew Dillon [Thu, 9 Sep 2010 08:35:59 +0000 (01:35 -0700)]
Merge branch 'master' of ssh://crater.dragonflybsd.org/repository/git/dragonfly
Matthew Dillon [Thu, 9 Sep 2010 08:32:07 +0000 (01:32 -0700)]
kernel - network protocol thread routing
* ip_input() now calls ip_mport() unconditionally and physically compares
the port to &curthread->td_msgport. If they do not match the packet
will be forwarded to the correct protocol thread.
* ip6_input() now unconditionally calls sw6->pr_soport() on the last
received header and forwards the packet to the correct protocol thread.
udp6, tcp6, and rip6 are now marked as terminal protocol (PR_LASTHDR).
Though I'm not sure this is correct they must be marked this way for
the packet to be forwarded to the correct protocol thread.
This fixes an assertion panic when tcp6_input() calls tcp_input().
tcp_input() is expecting to be run on the same protocol thread assigned
to its socket, otherwise callout timers and other entities will get
confused.
Sascha Wildner [Thu, 9 Sep 2010 07:24:02 +0000 (09:24 +0200)]
wi(4): Remove an unused variable.
Matthew Dillon [Thu, 9 Sep 2010 05:53:49 +0000 (22:53 -0700)]
kernel - MPSAFE the protocol drain routines
* The ip fragment drain was not MPSAFE at all. Use a token to protect
the ipq[] queues.
* The tcp reassembly code was only partially MPSAFE due to being
per-cpu. Finish it up. Use atomic ops for the tcp_reass_qsize
global.
* Add port assertions in the TCP input and output paths. If we
are not in the correct thread we panic, period.
* Code cleanup.
Matthew Dillon [Thu, 9 Sep 2010 05:51:03 +0000 (22:51 -0700)]
kernel - Finish MPSAFEing uipc_mbuf.c
* Make the mbuf tracking debug code MPSAFE.
* Get rid of worthless critical sections.
* Code cleanup.
Matthew Dillon [Thu, 9 Sep 2010 05:46:36 +0000 (22:46 -0700)]
kernel - MPSAFE kern_mpipe.c
* Make the mpipe code MPSAFE. The mpipe code is used in numerous places
including the ip reassembly code and some of the tcp code.
* Absolutely required now that the protocol stacks are running MPSAFE.
Matthew Dillon [Thu, 9 Sep 2010 05:43:37 +0000 (22:43 -0700)]
wlan - Update ath, sync rev 1.276@FreeBSD
* Don't delay updating the longcal timer - instead, update the longcal
flag immediately so it's only set once per longcal interval.
Without this, the current AR5416 code will continuously spam NF
calibrations during a periodic calibration if the longcal flag
is set. The longcal flag wouldn't be cleared until the calibration
method indicates that calibrations are "complete".
This could cause numerous problems including stuck beacons.
Taken-From: FreeBSD
Joe Talbott [Thu, 9 Sep 2010 03:30:59 +0000 (23:30 -0400)]
wlan - Add ral and wi back into the build system.
Joe Talbott [Thu, 9 Sep 2010 03:30:00 +0000 (23:30 -0400)]
wlan - Bring ral back into the fold.
*** NB ***
This has not been tested other than that it compiles (due to lack of
hardware) but will be tested within the week.
Joe Talbott [Thu, 9 Sep 2010 03:29:25 +0000 (23:29 -0400)]
wlan - Docuement lksleep changes as well.
Joe Talbott [Thu, 9 Sep 2010 03:28:17 +0000 (23:28 -0400)]
wlan - Bring wi back into the fold.
*** NB ***
This has not yet been tested other than that it compiles (due to lack of
hardware). It will be tested within the week.
Samuel J. Greear [Thu, 9 Sep 2010 01:19:09 +0000 (01:19 +0000)]
userland - dhclient - Fix possible infinite loop from unhandled poll(2) return
* This -probably- doesn't happen, but exercise some defensive programming
here. This will prevent dhclient from running away and eating 100% cpu in the
event of an error on the socket. Plus, we let the user know something went
wrong.
Samuel J. Greear [Thu, 9 Sep 2010 01:12:31 +0000 (01:12 +0000)]
userland - dhclient - Fix possible infinite loop
FreeBSD commit message:
It is possible for bpf to return a length such that:
length != BPF_WORDALIGN(length)
This meeans that it is possible for this to be true:
interface->rbuf_offset > interface->rbuf_len
Handle this case in the test for running out of packets. While
OpenBSD's solution of setting interface->rbuf_len to
BPF_WORDALIGN(length) is safe due to the size of the buffer, I think
this solution results in less hidden assumptions.
This should fix the problem of dhclient running away and consuming 100%
CPU.
PR: bin/102226
Submitted by: Joost Bekkers <joost at jodocus.org>
Reported-by: Many
Obtained-from: FreeBSD
Matthew Dillon [Thu, 9 Sep 2010 00:33:26 +0000 (17:33 -0700)]
kernel - Bring if_alc up to FreeBSD rev 1.16
* Bring in all recent if_alc work from FreeBSD.
Taken-from: FreeBSD
Matthew Dillon [Wed, 8 Sep 2010 22:26:54 +0000 (15:26 -0700)]
wlan - Fix serializer in destroy path, update docs
* Update the README.DRAGONFLY to include tidbits no tsleeps and
firmware loading.
* Fix an issue when a low level driver attempts to destroy a wlan
interface.
Matthew Dillon [Wed, 8 Sep 2010 22:22:14 +0000 (15:22 -0700)]
wlan - Bring iwn back into the fold.
* Note in addition to the usual instructions we also have to release
the wlan_global_serializer across a number of tsleep()s and the
firmware loader.
Joe Talbott [Wed, 8 Sep 2010 17:06:18 +0000 (13:06 -0400)]
wlan - Bring wpi back into the fold.
Joe Talbott [Wed, 8 Sep 2010 17:03:13 +0000 (13:03 -0400)]
wpi - Convert to wlan_global_serializer use.
* Rip out old locking inherited from FreeBSD
* Use wlan_serialize_{enter,exit}() to serialize wlan boundary crossings
per sys/netproto/802_11/README.DRAGONFLY
* Clean up some minor cruft left over from porting.
Matthew Dillon [Wed, 8 Sep 2010 07:20:23 +0000 (00:20 -0700)]
wlan - if_ath driver - Make some adjustments to ath (preliminary)
These adjustments correct some chip races and appear to fix issues
related to running the wlan in AP mode with the atheros driver.
* Fix a bug in ath_txqmove(). This bug was hidden due to a queue
length check in the one place that used the routine but fix it anyway.
* Call stoptxdma() before messing with the beacon linkages rather than
afterwords (ap mode). I'm a bit unclear as to whether the previously
installed beacon should have been allowed to continue to run if no
new beacons are found. For now it isn't.
* Redo the qbusy logic. The old logic had at least 2 chipset/driver
races related to the link field. The new logic makes no assumptions
and only reactivates the txdma if the queue is clearly idle and
we are adding the first (bf) to it. Otherwise leave it to the INT_TX
code to detect where the txdma stopped and restart it at the
appropriate place.
This bit of code needs more work as the INT_TX for tx completion may
be delayed indefinitely (we might need a callout check in there too,
I'm not sure).
* For the moment use MB_WAIT when loading or replentishing the receive ring.
There does not appear to be a proper mechanism to deal with stalls that
might be created if a mbuf fails to allocate.
What we really need here is proper rx ring mbuf replacement logic where
the filled mbuf is NOT removed if no new mbuf can be allocated to take
its place. Using MB_WAIT is a bad hack. It isn't entirely trivial
due to the DMA load and the 32 bit address space restriction.
* Add a few cpu_sfence()s when poking the (bf) link field. This probably
isn't correct.
* Cleanup, add some debugging kprintf()s for a few unexpected conditions
Matthew Dillon [Wed, 8 Sep 2010 07:18:04 +0000 (00:18 -0700)]
wlan - Bug fixes for recent work
* handoff (if in "ap" mode) serializer adjustments.
* interface deletion serializer adjustments.
Matthew Dillon [Wed, 8 Sep 2010 04:40:48 +0000 (21:40 -0700)]
wlan - Disable bgscan by default
* bgscan mode (station scanning while associated) blows up the
Atheros driver (tested w/ AR9280 chipset). Turn it off by default.