It seems commit fda9ef5d67 introduced a RCU
protection for sk_filter(), without a rcu_dereference()
Either we need a rcu_dereference(), either a comment should explain why we
dont need it. I vote for the former.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
alg_key_len is the length in bits of the key, not in bytes.
Best way to fix this is to move alg_len() function from net/xfrm/xfrm_user.c
to include/net/xfrm.h, and to use it in xfrm_algo_clone()
alg_len() is renamed to xfrm_alg_len() because of its global exposition.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
For cards that initially have the MAC address stored in reverse order,
the forcedeth driver uses a flag to signal whether the address was
already corrected, so that it is not reversed again on a subsequent
probe.
Unfortunately this flag, which is stored in a register of the card,
seems to get lost during suspend, resulting in the MAC address being
reversed again. To fix that, the MAC address needs to be written back
in reversed order before we suspend and the flag needs to be reset.
The flag is still required because at least kexec will never write
back the reversed address and thus needs to know what state the card
is in.
Signed-off-by: Bjrn Steinbrink <B.Steinbrink@gmx.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
usb_control_msg was changed long ago (2.6.12-pre) to take milliseconds
instead of jiffies. Oddly, mcs7830 wasn't added until 2.6.19-rc3.
Signed-off-by: Russ Dill <Russ.Dill@asu.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
lro_mgr->features contains a bitmask of LRO_F_* values which are
defined as power of two, not as bit indexes.
They must be checked with x&LRO_F_FOO, not with test_bit(LRO_F_FOO,&x).
Signed-off-by: Brice Goglin <Brice.Goglin@inria.fr>
Acked-by: Andrew Gallatin <gallatin@myri.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Both NetLabel and SELinux (other LSMs may grow to use it as well) rely
on the 'iif' field to determine the receiving network interface of
inbound packets. Unfortunately, at present this field is not
preserved across a skb clone operation which can lead to garbage
values if the cloned skb is sent back through the network stack. This
patch corrects this problem by properly copying the 'iif' field in
__skb_clone() and removing the 'iif' field assignment from
skb_act_clone() since it is no longer needed.
Also, while we are here, put the assignments in the same order as the
offsets to reduce cacheline bounces.
Signed-off-by: Paul Moore <paul.moore@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Back in 2.6.12-pre, usb_start_wait_urb was switched over to take
milliseconds instead of jiffies. kaweth.c was never updated to match.
Signed-off-by: Russ Dill <Russ.Dill@asu.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
Unfortunately Jeb decided to move away from our group. We wish Jeb
good luck with his new group!
Reordered people a bit so most active team members are on top.
Signed-off-by: Auke Kok <auke-jan.h.kok@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This makes the ->poll() routines of the E100, E1000, E1000E, IXGB, and
IXGBE drivers complete ->poll() consistently.
Now they will all break out when the amount of RX work done is less
than 'budget'.
At a later time, we may want put back code to include the TX work as
well (as at least one other NAPI driver does, but by in large NAPI
drivers do not do this). But if so, it should be done consistently
across the board to all of these drivers.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Auke Kok <auke-jan.h.kok@intel.com>
This finally adds the code in net_rx_action() to break out of the
->poll()'ing loop when a napi_disable() is found to be pending.
Now, even if a device is being flooded with packets it can be cleanly
brought down.
Signed-off-by: David S. Miller <davem@davemloft.net>
When we add the generic napi_disable_pending() breakout
logic to net_rx_action() it means that napi_disable()
can cause NAPI poll interrupt events to be disabled.
And this is exactly what we want. If a napi_disable()
is pending, and we are looping in the ->poll(), we want
->poll() event interrupts to stay disabled and we want
to complete the NAPI poll ASAP.
When ->poll() break out during device down was being handled on a
per-driver basis, often these drivers would turn interrupts back on
when '!netif_running()' was detected.
And this would just cause a reschedule of the NAPI ->poll() in the
interrupt handler before the napi_disable() could get in there and
grab the NAPI_STATE_SCHED bit.
The vast majority of drivers don't care if napi_disable() might have
the side effect of disabling NAPI ->poll() event interrupts. In all
such cases, when a napi_disable() is performed, the driver just
disabled interrupts or is about to.
However there were three exceptions to this in PCNET32, R8169, and
SKY2. To fix those cases, at the subsequent napi_enable() points, I
added code to ensure that the ->poll() interrupt events are enabled in
the hardware.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Don Fry <pcnet32@verizon.net>
Currently mac80211 fails silently when trying to set a nonexistent
rate. Return an error instead.
Signed-Off-By: Andy Lutomirski <luto@myrealbox.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
This will make sure that always the correct core is selected, even if
there are both a PCI and PCI-E core on a PCI or PCI-E card.
Signed-off-by: Michael Buesch <mb@bu3sch.de>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
Drivers do this to try to break out of the ->poll()'ing loop
when the device is being brought administratively down.
Now that we have a napi_disable() "pending" state we are going
to solve that problem generically.
Signed-off-by: David S. Miller <davem@davemloft.net>
Create a bit to signal that a napi_disable() is in progress.
This sets up infrastructure such that net_rx_action() can generically
break out of the ->poll() loop on a NAPI context that has a pending
napi_disable() yet is being bombed with packets (and thus would
otherwise poll endlessly and not allow the napi_disable() to finish).
Now, what napi_disable() does is first set the NAPI_STATE_DISABLE bit
(to indicate that a disable is pending), then it polls for the
NAPI_STATE_SCHED bit, and once the NAPI_STATE_SCHED bit is acquired
the NAPI_STATE_DISABLE bit is cleared. Here, the test_and_set_bit()
provides the necessary memory barrier between the various bitops.
napi_schedule_prep() now tests for a pending disable as it's first
action and won't try to obtain the NAPI_STATE_SCHED bit if a disable
is pending.
As a result, we can remove the netif_running() check in
netif_rx_schedule_prep() because the NAPI disable pending state serves
this purpose. And, it does so in a NAPI centric manner which is what
we really want.
Signed-off-by: David S. Miller <davem@davemloft.net>
It is pointless, because everything that can make a device go away
will do a napi_disable() first.
The main impetus behind this is that now we can legally do a NAPI
completion in generic code like net_rx_action() which a following
changeset needs to do. net_rx_action() can only perform actions
in NAPI centric ways, because there may be a one to many mapping
between NAPI contexts and network devices (SKY2 is one example).
We also want to get rid of this because it's an extra atomic in the
NAPI paths, and also because it is one of the last instances where the
NAPI interfaces care about net devices.
The one remaining netdev detail the NAPI stuff cares about is the
netif_running() check which will be killed off in a subsequent
changeset.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes the parsing of the RX data header channel field.
The current code parses the header incorrectly and passes a wrong
channel number and frequency for each frame to mac80211.
The FIXMEs added by this patch don't matter for now as the code
where they live won't get executed anyway. They will be fixed later.
Signed-off-by: Michael Buesch <mb@bu3sch.de>
Signed-off-by: John W. Linville <linville@tuxdriver.com>
easy to trigger as user with sfuzz.
irda_create() is quiet on unknown sock->type,
match this behaviour for SOCK_DGRAM unknown protocol
Signed-off-by: maximilian attems <max@stro.at>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some recent changes completely removed accounting for the FORWARD_TSN
parameter length in the INIT and INIT-ACK chunk. This is wrong and
should be restored.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When processing an unexpected INIT chunk, we do not need to
do any preservation of the old AUTH parameters. In fact,
doing such preservations will nullify AUTH and allow connection
stealing.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The even should be called SCTP_AUTHENTICATION_INDICATION.
Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This should fix the kernel warn/oops reported while routing.
The tulip driver has a fencepost bug with new NAPI in 2.6.24
It has an off by one bug if a full quantum is reached.
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
meth didn't set a valid mac address during probing, but later during
open. Newer kernel refuse to open device with 00:00:00:00:00:00 as mac
address -> dead ethernet. This patch sets the mac address in the probe
function and uses only the mac address from the netdevice struct when
setting up the hardware.
Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
This was missed when commit e2ac455a18
fixed the compile errors in drivers/net/netx-eth.c caused by
commit 09f75cd7bf.
Signed-off-by: Adrian Bunk <adrian.bunk@movial.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
The recent changes for ip command line processing fixed some problems
but unfortunately broke some common usage scenarios. In current
2.6.24-rc6 the following command line results in no IP address
assignment, which is surely a regression:
ip=10.0.2.15::10.0.2.2:255.255.255.0::eth0:off
Please find below a patch that works for all cases I can find.
Signed-off-by: Amos Waterland <apw@us.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We currently check that iph->ihl is bounded by the real length and that
the real length is greater than the minimum IP header length. However,
we did not check the caes where iph->ihl is less than the minimum IP
header length.
This breaks because some ip_fast_csum implementations assume that which
is quite reasonable.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is possible for the TX ring to have packets sit in it for unbounded
amounts of time.
The only way to defer TX interrupts in the chip is to periodically set
"mark" bits, when processing of a TX descriptor with the mark bit set
is complete it triggers the interrupt for the TX queue's LDG.
A consequence of this kind of scheme is that if packet flow suddenly
stops, the remaining TX packets will just sit there.
If this happens, since those packets could be charged to TCP socket
send queues, such sockets could get stuck.
The simplest solution is to divorce the socket ownership of the packet
once the device takes the SKB, by using skb_orphan() in
niu_start_xmit().
In hindsight, it would have been much nicer if the chip provided two
interrupt sources for TX (like basically every other ethernet chip
does). Namely, keep the "mark" bit, but also signal the LDG when the
TX queue becomes completely empty. That way there is no need to have
a deadlock breaker like this.
Signed-off-by: David S. Miller <davem@davemloft.net>
niu_slowpath_interrupt() expects values to be setup in lp->{v0,v1,v2}
but they aren't. That's only done by niu_schedule_napi() which is
done later in the interrupt path.
If niu_rx_error() returns zero, and v0 is clear, hit the
RX_DMA_CTL_STATE register with a RX_DMA_CTL_STAT_MEX.
Only emit verbose RX error logs if a fatal channel or port error is
signalled. Other cases will be recorded into statistics by
niu_log_rxchan_errors().
Signed-off-by: David S. Miller <davem@davemloft.net>
Roland Westrelin did a great analysis of a long standing thinko in the
return path of futex_lock_pi.
While we fixed the lock steal case long ago, which was easy to trigger,
we never had a test case which exposed this problem and stupidly never
thought about the reverse lock stealing scenario and the return to user
space with a stale state.
When a blocked tasks returns from rt_mutex_timed_locked without holding
the rt_mutex (due to a signal or timeout) and at the same time the task
holding the futex is releasing the futex and assigning the ownership of
the futex to the returning task, then it might happen that a third task
acquires the rt_mutex before the final rt_mutex_trylock() of the
returning task happens under the futex hash bucket lock. The returning
task returns to user space with ETIMEOUT or EINTR, but the user space
futex value is assigned to this task. The task which acquired the
rt_mutex fixes the user space futex value right after the hash bucket
lock has been released by the returning task, but for a short period of
time the user space value is wrong.
Detailed description is available at:
https://bugzilla.redhat.com/show_bug.cgi?id=400541
The fix for this is the same as we do when the rt_mutex was acquired by
a higher priority task via lock stealing from the designated new owner.
In that case we already fix the user space value and the internal
pi_state up before we return. This mechanism can be used to fixup the
above corner case as well. When the returning task, which failed to
acquire the rt_mutex, notices that it is the designated owner of the
futex, then it fixes up the stale user space value and the pi_state,
before returning to user space. This happens with the futex hash bucket
lock held, so the task which acquired the rt_mutex is guaranteed to be
blocked on the hash bucket lock. We can access the rt_mutex owner, which
gives us the pid of the new owner, safely here as the owner is not able
to modify (release) it while waiting on the hash bucket lock.
Rename the "curr" argument of fixup_pi_state_owner() to "newowner" to
avoid confusion with current and add the check for the stale state into
the failure path of rt_mutex_trylock() in the return path of
unlock_futex_pi(). If the situation is detected use
fixup_pi_state_owner() to assign everything to the owner of the
rt_mutex.
Pointed-out-and-tested-by: Roland Westrelin <roland.westrelin@sun.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Cleaning out all the incorrect 'no change made' checks for termios
settings showed up a problem with the PL2303. The hardware here seems to
lose sync and bits if you tell it to make no changes. This shows up with
a real world application.
To fix this the driver check for meaningful hardware changes is restored
but doing the tests correctly and as a tty layer function so it doesn't
get duplicated wrongly everywhere if other drivers turn out to need it.
Signed-off-by: Alan Cox <alan@redhat.com>
Tested-by: Mirko Parthey <mirko.parthey@informatik.tu-chemnitz.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
hfs seems prone to bad things when it encounters on disk corruption. Many
values are read from disk, and used as lengths to memcpy, as an example.
This patch fixes up several of these problematic cases.
o sanity check the on-disk maximum key lengths on mount
(these are set to a defined value at mkfs time and shouldn't differ)
o check on-disk node keylens against the maximum key length for each tree
o fix hfs_btree_open so that going out via free_tree: doesn't wind
up in hfs_releasepage, which wants to follow the very pointer
we were trying to set up:
HFS_SB(sb)->cat_tree = hfs_btree_open()
...
failure gets to hfs_releasepage and tries
to follow HFS_SB(sb)->cat_tree
Tested with the fsfuzzer; it survives more than it used to.
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Cc: Roman Zippel <zippel@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When using FLAT_MEMORY and ARCH_PFN_OFFSET is not 0, the kernel crashes in
memmap_init_zone(). This bug got introduced by commit
c713216dee
Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Bob Picco <bob.picco@hp.com>
Cc: Dave Hansen <haveblue@us.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: Andi Kleen <ak@muc.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: "Keith Mannthey" <kmannth@gmail.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Rework this functions so that gcc-3.2 can successfully perform
constant-folding.
Signed-off-by: Jean Delvare <khali@linux-fr.org>
Cc: Jaroslav Kysela <perex@suse.cz>
Cc: Takashi Iwai <tiwai@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
gcc 3.2 has a hard time coping with the code in dmi_id_init():
drivers/built-in.o(.init.text+0x789e): In function `dmi_id_init':
: undefined reference to `__you_cannot_kmalloc_that_much'
make: *** [.tmp_vmlinux1] Error 1
Moving half of the code to a separate function seems to help. This is a
no-op for gcc 4.1 which will successfully inline the code anyway.
Signed-off-by: Jean Delvare <khali@linux-fr.org>
Cc: Dave Airlie <airlied@linux.ie>
Tested-by: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch adds the array length of "free_area.free_list" to the vmcoreinfo
data so that makedumpfile (dump filtering command) can exclude all free pages
in linux-2.6.24.
makedumpfile creates a small dumpfile by excluding unnecessary pages for the
analysis. To distinguish unnecessary pages, makedumpfile gets the vmcoreinfo
data which has the minimum debugging information only for dump filtering.
In 2.6.24-rc1 or later, the free_area.free_list is an array which has one list
for each migrate types instead of a single list. makedumpfile needs the array
length of "free_area.free_list" and the vmcoreinfo data should contain it.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Tested-by: Ken'ichi Ohmichi <oomichi@mxs.nes.nec.co.jp>
Acked-by: Simon Horman <horms@verge.net.au>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch corrects some erroneous dentry handling in eCryptfs.
If there is a problem creating the lower file, then there is nothing that
the persistent lower file can do to really help us. This patch makes a
vfs_create() failure in the lower filesystem always lead to an
unconditional do_create failure in eCryptfs.
Under certain sequences of operations, the eCryptfs dentry can remain in
the dcache after an unlink. This patch calls d_drop() on the eCryptfs
dentry to correct this.
eCryptfs has no business calling d_delete() directly on a lower
filesystem's dentry. This patch removes the call to d_delete() on the
lower persistent file's dentry in ecryptfs_destroy_inode().
(Thanks to David Kleikamp, Eric Sandeen, and Jeff Moyer for helping
identify and resolve this issue)
Signed-off-by: Michael Halcrow <mhalcrow@us.ibm.com>
Cc: Dave Kleikamp <shaggy@austin.ibm.com>
Cc: Eric Sandeen <sandeen@redhat.com>
Cc: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The use of get_zeroed_page() with __GFP_HIGHMEM is invalid. Use
alloc_page() with __GFP_ZERO instead of invalid get_zeroed_page().
(This patch is only compile tested)
Cc: Carsten Otte <cotte@de.ibm.com>
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Acked-by: Hugh Dickins <hugh@veritas.com>
Acked-by: Carsten Otte <cotte@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We currently do not wait for the block from the missing device to be
computed from parity before copying data to the new stripe layout.
The change in the raid6 code is not techincally needed as we don't delay
data block recovery in the same way for raid6 yet. But making the change
now is safer long-term.
This bug exists in 2.6.23 and 2.6.24-rc
Cc: <stable@kernel.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 664cceb009 changed the parameters of
the function make_key_ref(). The macros that are used in case CONFIG_KEY
is not defined did not change.
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Sebastian Siewior <sebastian@breakpoint.cc>
Acked-by: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
On large partition, scanning the free clusters is very slow if users
doesn't use "usefree" option.
For optimizing it, this patch uses sb_breadahead() to read of FAT
sectors. On some user's 15GB partition, this patch improved it very
much (1min => 600ms).
The following is the result of 2GB partition on my machine.
without patch:
root@devron (/)# time df -h > /dev/null
real 0m1.202s
user 0m0.000s
sys 0m0.440s
with patch:
root@devron (/)# time df -h > /dev/null
real 0m0.378s
user 0m0.012s
sys 0m0.168s
Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix a glitch reported by lockdep in the spi_bitbang code: it needs to
consistently block IRQs when holding that spinlock.
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
With CPU_HOTPLUG=n:
WARNING: vmlinux.o(.text+0x104f8): Section mismatch: reference to .init.text:fork_idle (between
'do_fork_idle' and 'lapic_timer_broadcast')
do_fork_idle() needs to be __cpuinit. It can be static as well.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband:
IB/srp: Release transport before removing host
IB/mlx4: Fix value of pkey_index in QP1 completions
MAINTAINERS: Update Sean Hefty's email address
The documented call sequence for removing a host is to call the
transport xxx_remove_host() prior to scsi_remove_host(). The SRP
transport used to crash when that order was followed, but as it is now
fixed, use the documented order.
Signed-off-by: David Dillow <dillowda@ornl.gov>
Acked-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix the value of pkey_index in completions to get a valid value for
GSI QPs. Without this fix, incoming GSI packets on port 2 get an
invalid P_Key index in the completion, which prevents the MAD layer
from sending back a response, which can make the second port of
ConnectX HCAs completely useless.
Signed-off-by: Dotan Barak <dotanb@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Eric Dumazet