Update rcu_dereference() primitives to use new lockdep-based
checking. The rcu_dereference() in __in6_dev_get() may be
protected either by rcu_read_lock() or RTNL, per Eric Dumazet.
The rcu_dereference() in __sk_free() is protected by the fact
that it is never reached if an update could change it. Check
for this by using rcu_dereference_check() to verify that the
struct sock's ->sk_wmem_alloc counter is zero.
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: laijs@cn.fujitsu.com
Cc: dipankar@in.ibm.com
Cc: mathieu.desnoyers@polymtl.ca
Cc: josh@joshtriplett.org
Cc: dvhltc@us.ibm.com
Cc: niv@us.ibm.com
Cc: peterz@infradead.org
Cc: rostedt@goodmis.org
Cc: Valdis.Kletnieks@vt.edu
Cc: dhowells@redhat.com
LKML-Reference: <1266887105-1528-5-git-send-email-paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Traffic (tcp) doesnot start on a vlan interface when gro is enabled.
Even the tcp handshake was not taking place.
This is because, the eth_type_trans call before the netif_receive_skb
in napi_gro_finish() resets the skb->dev to napi->dev from the previously
set vlan netdev interface. This causes the ip_route_input to drop the
incoming packet considering it as a packet coming from a martian source.
I could repro this on 2.6.32.7 (stable) and 2.6.33-rc7.
With this fix, the traffic starts and the test runs fine on both vlan
and non-vlan interfaces.
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Patrick McHardy <kaber@trash.net>
Signed-off-by: Ajit Khaparde <ajitk@serverengines.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The wireless sysfs methods like the rest of the networking sysfs
methods are removed with the rtnl_lock held and block until
the existing methods stop executing. So use rtnl_trylock
and restart_syscall so that the code continues to work.
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Yuck. It turns out that when we restart sysctls we were restarting
with the values already changed. Which unfortunately meant that
the second time through we thought there was no change and skipped
all kinds of work, despite the fact that there was indeed a change.
I have fixed this the simplest way possible by restoring the changed
values when we restart the sysctl write.
One of my coworkers spotted this bug when after disabling forwarding
on an interface pings were still forwarded.
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
xfrm_state_clone calls kfree instead of xfrm_state_put to free
a failed state. Depending on the state of the failed state, it
can cause leaks to things like module references.
All states should be freed by xfrm_state_put past the point of
xfrm_init_state.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
When ipcomp_tunnel_attach fails we will call ipcomp_destroy twice.
This may lead to double-frees on certain structures.
As there is no reason to explicitly call ipcomp_destroy, this patch
removes it from ipcomp*.c and lets the standard xfrm_state destruction
take place.
This is based on the discovery and patch by Alexey Dobriyan.
Tested-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
For hardware with IEEE80211_HW_HAS_RATE_CONTROL the rate controller is not
initialized. However, calling functions such as ieee80211_beacon_get result
in the rate_control_get_rate function getting called, which is accessing
(in this case uninitialized) rate control structures unconditionally.
Fix by exiting the function before setting the rates for HW with
IEEE80211_HW_HAS_RATE_CONTROL set. The initialization of the ieee80211_tx_info
struct is intentionally still executed.
Signed-off-by: Juuso Oikarinen <juuso.oikarinen@nokia.com>
Reviewed-by: Kalle Valo <kalle.valo@nokia.com>
Cc: stable@kernel.org
Signed-off-by: John W. Linville <linville@tuxdriver.com>
Currently we treat IGMPv3 reports as if it were an IGMPv2/v1 report.
This is broken as IGMPv3 reports are formatted differently. So we
end up suppressing a bogus multicast group (which should be harmless
as long as the leading reserved field is zero).
In fact, IGMPv3 does not allow membership report suppression so
we should simply ignore IGMPv3 membership reports as a host.
This patch does exactly that. I kept the case statement for it
so people won't accidentally add it back thinking that we overlooked
this case.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make sure, that TCP has a nonzero RTT estimation after three-way
handshake. Currently, a listening TCP has a value of 0 for srtt,
rttvar and rto right after the three-way handshake is completed
with TCP timestamps disabled.
This will lead to corrupt RTO recalculation and retransmission
flood when RTO is recalculated on backoff reversion as introduced
in "Revert RTO on ICMP destination unreachable"
(f1ecd5d9e7).
This behaviour can be provoked by connecting to a server which
"responds first" (like SMTP) and rejecting every packet after
the handshake with dest-unreachable, which will lead to softirq
load on the server (up to 30% per socket in some tests).
Thanks to Ilpo Jarvinen for providing debug patches and to
Denys Fedoryshchenko for reporting and testing.
Changes since v3: Removed bad characters in patchfile.
Reported-by: Denys Fedoryshchenko <denys@visp.net.lb>
Signed-off-by: Damian Lukowski <damian@tvk.rwth-aachen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The action modules have been prefixed with 'act_', but the Kconfig
description was not changed.
Signed-off-by: Jan Luebbe <jluebbe@debian.org>
Acked-by: Jamal Hadi Salim <hadi@cyberus.ca>
Signed-off-by: David S. Miller <davem@davemloft.net>
restructure client create code to handle error cases better and
only cleanup initialized portions of the stack.
Signed-off-by: Venkateswararao Jujjuri <jvrao@us.ibm.com>
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>
Kernel bugzilla #15239
On some workloads, it is quite possible to get a huge dst list to
process in dst_gc_task(), and trigger soft lockup detection.
Fix is to call cond_resched(), as we run in process context.
Reported-by: Pawel Staszewski <pstaszewski@itcare.pl>
Tested-by: Pawel Staszewski <pstaszewski@itcare.pl>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Options pointer is being moved before calling kfree() which seems
to cause problems. This uses a separate pointer to track and free
original allocation.
Signed-off-by: Venkateswararao Jujjuri <jvrao@us.ibm.com>
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>w
Reinette found the reason for the warnings that
happened occasionally when a hw-offloaded scan
finished; her description of the problem:
mac80211 will defer the handling of scan requests if it is
busy with management work at the time. The scan requests
are deferred and run after the work has completed. When
this occurs there are currently two problems.
* The scan request for hardware scan is not fully populated
with the band and channels to scan not initialized.
* When the scan is queued the state is not correctly updated
to reflect that a scan is in progress. The problem here is
that when the driver completes the scan and calls
ieee80211_scan_completed() a warning will be triggered
since mac80211 was not aware that a scan was in progress.
The reason is that the queued scan work will start
the hw scan right away when the hw_scan_req struct
has already been allocated. However, in the first
pass it will not have been filled, which happens
at the same time as setting the bits. To fix this,
simply move the allocation after the pending work
test as well, so that the first iteration of the
scan work will call __ieee80211_start_scan() even
in the hardware scan case.
Bug-identified-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
We only reply to probe request if either the requested SSID is the
broadcast SSID or if the requested SSID matches our own SSID. This
latter case was not properly handled since we were replying to different
SSID with the same length as our own SSID.
Signed-off-by: Benoit Papillault <benoit.papillault@free.fr>
Cc: stable@kernel.org
Signed-off-by: John W. Linville <linville@tuxdriver.com>
stat structures contain a size prefix. In our twstat messages
we were including the size of the size prefix in the prefix, which is not
what the protocol wants, and Inferno servers would complain.
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>
If the user specifies a transport and we can't find it, we failed back
to the default trainsport silently. This patch will make the code
complain more loudly and return an error code.
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>
The 9p virtio transport was not updating its connection status correctly
preventing it from being able to mount the server.
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>
As noticed by Jon Masters <jonathan@jonmasters.org>, the conntrack hash
size is global and not per namespace, but modifiable at runtime through
/sys/module/nf_conntrack/hashsize. Changing the hash size will only
resize the hash in the current namespace however, so other namespaces
will use an invalid hash size. This can cause crashes when enlarging
the hashsize, or false negative lookups when shrinking it.
Move the hash size into the per-namespace data and only use the global
hash size to initialize the per-namespace value when instanciating a
new namespace. Additionally restrict hash resizing to init_net for
now as other namespaces are not handled currently.
Cc: stable@kernel.org
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
As per C99 6.2.4(2) when temporary table data goes out of scope,
the behaviour is undefined:
if (compat) {
struct foo tmp;
...
private = &tmp;
}
[dereference private]
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Cc: stable@kernel.org
Signed-off-by: Patrick McHardy <kaber@trash.net>
Expectation hashtable size was simply glued to a variable with no code
to rehash expectations, so it was a bug to allow writing to it.
Make "expect_hashsize" readonly.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Cc: stable@kernel.org
Signed-off-by: Patrick McHardy <kaber@trash.net>
nf_conntrack_cachep is currently shared by all netns instances, but
because of SLAB_DESTROY_BY_RCU special semantics, this is wrong.
If we use a shared slab cache, one object can instantly flight between
one hash table (netns ONE) to another one (netns TWO), and concurrent
reader (doing a lookup in netns ONE, 'finding' an object of netns TWO)
can be fooled without notice, because no RCU grace period has to be
observed between object freeing and its reuse.
We dont have this problem with UDP/TCP slab caches because TCP/UDP
hashtables are global to the machine (and each object has a pointer to
its netns).
If we use per netns conntrack hash tables, we also *must* use per netns
conntrack slab caches, to guarantee an object can not escape from one
namespace to another one.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
[Patrick: added unique slab name allocation]
Cc: stable@kernel.org
Signed-off-by: Patrick McHardy <kaber@trash.net>
As discovered by Jon Masters <jonathan@jonmasters.org>, the "untracked"
conntrack, which is located in the data section, might be accidentally
freed when a new namespace is instantiated while the untracked conntrack
is attached to a skb because the reference count it re-initialized.
The best fix would be to use a seperate untracked conntrack per
namespace since it includes a namespace pointer. Unfortunately this is
not possible without larger changes since the namespace is not easily
available everywhere we need it. For now move the untracked conntrack
initialization to the init_net setup function to make sure the reference
count is not re-initialized and handle cleanup in the init_net cleanup
function to make sure namespaces can exit properly while the untracked
conntrack is in use in other namespaces.
Cc: stable@kernel.org
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
The report descriptor is read by user space (via the Service
Discovery Protocol), so it is only available during the ioctl
to connect. However, the HID probe function that needs the
descriptor might not be called until a specific module is
loaded. Keep a copy of the descriptor so it is available for
later use.
Signed-off-by: Michael Poole <mdpoole@troilus.org>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
Add missing try_to_freeze() to one of the pktgen_thread_worker() code
paths so that it doesn't block suspend/hibernation.
Fixes http://bugzilla.kernel.org/show_bug.cgi?id=15006
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Reported-and-tested-by: Ciprian Dorin Craciun <ciprian.craciun@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
One ioctl has been forgotten when the BKL was push down into irnet_ppp
ioctl function.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@holoscopio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add the missing unlock_kernel in one ioctl operation.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@holoscopio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When in sniff mode with a long interval time (1.28s) it can take 4+ seconds
to establish a SCO link. Fix by requesting active mode before requesting
SCO connection. This improves SCO setup time to ~500ms.
Bluetooth headsets that use a long interval time, and exhibit the long
SCO connection time include Motorola H790, HX1 and H17. They have a
CSR 2.1 chipset.
Verified this behavior and fix with host Bluetooth chipsets: BCM4329 and
TI1271.
2009-10-13 14:17:46.183722 > HCI Event: Mode Change (0x14) plen 6
status 0x00 handle 1 mode 0x02 interval 2048
Mode: Sniff
2009-10-13 14:17:53.436285 < HCI Command: Setup Synchronous Connection (0x01|0x0028) plen 17
handle 1 voice setting 0x0060
2009-10-13 14:17:53.445593 > HCI Event: Command Status (0x0f) plen 4
Setup Synchronous Connection (0x01|0x0028) status 0x00 ncmd 1
2009-10-13 14:17:57.788855 > HCI Event: Synchronous Connect Complete 0x2c) plen 17
status 0x00 handle 257 bdaddr 00:1A:0E:F1:A4:7F type eSCO
Air mode: CVSD
Signed-off-by: Nick Pelly <npelly@google.com>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
This fixes commit (38ff3e6bb9) ("dccp_probe:
Fix module load dependencies between dccp and dccp_probe", from 15 Jan).
It fixes the construction of the first argument of try_then_request_module(),
where only valid return codes from the first argument should be returned.
What we do now is assign the result of register_jprobe() to ret, without
the side effect of the comparison.
Acked-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes a bug introduced in commit de4ef86cfc
("dccp: fix dccp rmmod when kernel configured to use slub", 17 Jan): the
vsnprintf used sizeof(slab_name_fmt), which became truncated to 4 bytes, since
slab_name_fmt is now a 4-byte pointer and no longer a 32-character array.
This lead to error messages such as
FATAL: Error inserting dccp: No buffer space available
>> kernel: [ 1456.341501] kmem_cache_create: duplicate cache cci
generated due to the truncation after the 3rd character.
Fixed for the moment by introducing a symbolic constant. Tested to fix the bug.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
1. After sock_register() returns, it's possible to create sockets,
even if module still not initialized fully (blame generic module code
for that!)
2. Consequently, pfkey_create() can be called with pfkey_net_id still not
initialized which will BUG_ON in net_generic():
kernel BUG at include/net/netns/generic.h:43!
3. During netns shutdown, netns ops should be unregistered after
key manager unregistered because key manager calls can be triggered
from xfrm_user module:
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
pfkey_broadcast+0x111/0x210 [af_key]
pfkey_send_notify+0x16a/0x300 [af_key]
km_state_notify+0x41/0x70
xfrm_flush_sa+0x75/0x90 [xfrm_user]
4. Unregister netns ops after socket ops just in case and for symmetry.
Reported by Luca Tettamanti.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Tested-by: Luca Tettamanti <kronos.it@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 2da31939a4, support
for Bluetooth hid_output_raw_report was added, but it pushes
the data to the interrupt channel instead of the contol one.
This patch makes hid_output_raw_report use the control channel
instead. Using the interrupt channel was a mistake.
Signed-off-by: Bastien Nocera <hadess@hadess.net>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
Move skb_clone after error confition check so it is not going
potentially out of the scope.
Signed-off-by: Andrei Emeltchenko <andrei.emeltchenko@nokia.com>
Acked-by: Gustavo F. Padovan <padovan@profusion.mobi>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
Trivial fix for double free of SKB pointer with kfree_skb to
make code simplier and cleaner. Remove unused variable err.
Signed-off-by: Andrei Emeltchenko <andrei.emeltchenko@nokia.com>
Acked-by: Gustavo F. Padovan <padovan@profusion.mobi>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
I got below kernel oops when I try to bring down the network interface if
ftrace is enabled. The root cause is drv_ampdu_action() is passed with a
NULL ssn pointer in the BA session tear down case. We need to check and
avoid dereferencing it in trace entry assignment.
BUG: unable to handle kernel NULL pointer dereference
Modules linked in: at (null)
IP: [<f98fe02a>] ftrace_raw_event_drv_ampdu_action+0x10a/0x160 [mac80211]
*pde = 00000000
Oops: 0000 [#1] SMP DEBUG_PAGEALLOC
[...]
Call Trace:
[<f98fdf20>] ? ftrace_raw_event_drv_ampdu_action+0x0/0x160 [mac80211]
[<f98dac4c>] ? __ieee80211_stop_rx_ba_session+0xfc/0x220 [mac80211]
[<f98d97fb>] ? ieee80211_sta_tear_down_BA_sessions+0x3b/0x50 [mac80211]
[<f98dc6f6>] ? ieee80211_set_disassoc+0xe6/0x230 [mac80211]
[<f98dc6ac>] ? ieee80211_set_disassoc+0x9c/0x230 [mac80211]
[<f98dcbb8>] ? ieee80211_mgd_deauth+0x158/0x170 [mac80211]
[<f98e4bdb>] ? ieee80211_deauth+0x1b/0x20 [mac80211]
[<f8987f49>] ? __cfg80211_mlme_deauth+0xe9/0x120 [cfg80211]
[<f898b870>] ? __cfg80211_disconnect+0x170/0x1d0 [cfg80211]
Cc: Johannes Berg <johannes@sipsolutions.net>
Cc: stable@kernel.org
Signed-off-by: Zhu Yi <yi.zhu@intel.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
The commit 0b5ccb2(title:ipv6: reassembly: use seperate reassembly queues for
conntrack and local delivery) has broken the saddr&&daddr member of
nf_ct_frag6_queue when creating new queue. And then hash value
generated by nf_hashfn() was not equal with that generated by fq_find().
So, a new received fragment can't be inserted to right queue.
The patch fixes the bug with adding member of user to nf_ct_frag6_queue structure.
Signed-off-by: Shan Wei <shanwei@cn.fujitsu.com>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
By rounding up the buffer size to power of 2, several expensive
modulus operations can be avoided. This patch also solves a bug where
the gap need when ring gets full was not being accounted for.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
GC is non-existent in netns, so after you hit GC threshold, no new
dst entries will be created until someone triggers cleanup in init_net.
Make xfrm4_dst_ops and xfrm6_dst_ops per-netns.
This is not done in a generic way, because it woule waste
(AF_MAX - 2) * sizeof(struct dst_ops) bytes per-netns.
Reorder GC threshold initialization so it'd be done before registering
XFRM policies.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Bruno Prémont found commit 9793241fe9
(vlan: Precise RX stats accounting) added a regression for non
hw accelerated vlans.
[ 26.390576] BUG: unable to handle kernel NULL pointer dereference at (null)
[ 26.396369] IP: [<df856b89>] vlan_skb_recv+0x89/0x280 [8021q]
vlan_dev_info() was used with original device, instead of
skb->dev. Also spotted by Américo Wang.
Reported-By: Bruno Prémont <bonbons@linux-vserver.org>
Tested-By: Bruno Prémont <bonbons@linux-vserver.org>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
"ip xfrm state|policy count" report SA/SP count from init_net,
not from netns of caller process.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
In a string like "v:SIP/2.0..." it was checking for !isalpha('S') when it
meant to be inspecting the ':'.
Patch by Greg Alexander <greqcs@galexander.org>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Hey all-
I was tinkering with dccp recently and noticed that I BUG halted the
kernel when I rmmod-ed the dccp module. The bug halt occured because the page
that I passed to kfree failed the PageCompound and PageSlab test in the slub
implementation of kfree. I tracked the problem down to the following set of
events:
1) dccp, unlike all other uses of kmem_cache_create, allocates a string
dynamically when registering a slab cache. This allocated string is freed when
the cache is destroyed.
2) Normally, (1) is not an issue, but when Slub is in use, it is possible that
caches are 'merged'. This process causes multiple caches of simmilar
configuration to use the same cache data structure. When this happens, the new
name of the cache is effectively dropped.
3) (2) results in kmem_cache_name returning an ambigous value (i.e.
ccid_kmem_cache_destroy, which uses this fuction to retrieve the name pointer
for freeing), is no longer guaranteed that the string it assigned is what is
returned.
4) If such merge event occurs, ccid_kmem_cache_destroy frees the wrong pointer,
which trips over the BUG in the slub implementation of kfree (since its likely
not a slab allocation, but rather a pointer into the static string table
section.
So, what to do about this. At first blush this is pretty clearly a leak in the
information that slub owns, and as such a slub bug. Unfortunately, theres no
really good way to fix it, without exposing slub specific implementation details
to the generic slab interface. Also, even if we could fix this in slub cleanly,
I think the RCU free option would force us to do lots of string duplication, not
only in slub, but in every slab allocator. As such, I'd like to propose this
solution. Basically, I just move the storage for the kmem cache name to the
ccid_operations structure. In so doing, we don't have to do the kstrdup or
kfree when we allocate/free the various caches for dccp, and so we avoid the
problem, by storing names with static memory, rather than heap, the way all
other calls to kmem_cache_create do.
I've tested this out myself here, and it solves the problem quite well.
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Paul E. McKenney