* git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-lguest: (31 commits)
lguest: add support for indirect ring entries
lguest: suppress notifications in example Launcher
lguest: try to batch interrupts on network receive
lguest: avoid sending interrupts to Guest when no activity occurs.
lguest: implement deferred interrupts in example Launcher
lguest: remove obsolete LHREQ_BREAK call
lguest: have example Launcher service all devices in separate threads
lguest: use eventfds for device notification
eventfd: export eventfd_signal and eventfd_fget for lguest
lguest: allow any process to send interrupts
lguest: PAE fixes
lguest: PAE support
lguest: Add support for kvm_hypercall4()
lguest: replace hypercall name LHCALL_SET_PMD with LHCALL_SET_PGD
lguest: use native_set_* macros, which properly handle 64-bit entries when PAE is activated
lguest: map switcher with executable page table entries
lguest: fix writev returning short on console output
lguest: clean up length-used value in example launcher
lguest: Segment selectors are 16-bit long. Fix lg_cpu.ss1 definition.
lguest: beyond ARRAY_SIZE of cpu->arch.gdt
...
* git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-virtio:
virtio: enhance id_matching for virtio drivers
virtio: fix id_matching for virtio drivers
virtio: handle short buffers in virtio_rng.
virtio_blk: add missing __dev{init,exit} markings
virtio: indirect ring entries (VIRTIO_RING_F_INDIRECT_DESC)
virtio: teach virtio_has_feature() about transport features
virtio: expose features in sysfs
virtio_pci: optional MSI-X support
virtio_pci: split up vp_interrupt
virtio: find_vqs/del_vqs virtio operations
virtio: add names to virtqueue struct, mapping from devices to queues.
virtio: meet virtio spec by finalizing features before using device
virtio: fix obsolete documentation on probe function
* 'cuse' of git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi/fuse:
CUSE: implement CUSE - Character device in Userspace
fuse: export symbols to be used by CUSE
fuse: update fuse_conn_init() and separate out fuse_conn_kill()
fuse: don't use inode in fuse_file_poll
fuse: don't use inode in fuse_do_ioctl() helper
fuse: don't use inode in fuse_sync_release()
fuse: create fuse_do_open() helper for CUSE
fuse: clean up args in fuse_finish_open() and fuse_release_fill()
fuse: don't use inode in helpers called by fuse_direct_io()
fuse: add members to struct fuse_file
fuse: prepare fuse_direct_io() for CUSE
fuse: clean up fuse_write_fill()
fuse: use struct path in release structure
fuse: misc cleanups
* git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-module-and-param:
module: cleanup FIXME comments about trimming exception table entries.
module: trim exception table on init free.
module: merge module_alloc() finally
uml module: fix uml build process due to this merge
x86 module: merge the rest functions with macros
x86 module: merge the same functions in module_32.c and module_64.c
uvesafb: improve parameter handling.
module_param: allow 'bool' module_params to be bool, not just int.
module_param: add __same_type convenience wrapper for __builtin_types_compatible_p
module_param: split perm field into flags and perm
module_param: invbool should take a 'bool', not an 'int'
cyber2000fb.c: use proper method for stopping unload if CONFIG_ARCH_SHARK
* 'for-2.6.31' of git://git.kernel.org/pub/scm/linux/kernel/git/bart/ide-2.6: (29 commits)
ide: re-implement ide_pci_init_one() on top of ide_pci_init_two()
ide: unexport ide_find_dma_mode()
ide: fix PowerMac bootup oops
ide: skip probe if there are no devices on the port (v2)
sl82c105: add printk() logging facility
ide-tape: fix proc warning
ide: add IDE_DFLAG_NIEN_QUIRK device flag
ide: respect quirk_drives[] list on all controllers
hpt366: enable all quirks for devices on quirk_drives[] list
hpt366: sync quirk_drives[] list with pdc202xx_{new,old}.c
ide: remove superfluous SELECT_MASK() call from do_rw_taskfile()
ide: remove superfluous SELECT_MASK() call from ide_driveid_update()
icside: remove superfluous ->maskproc method
ide-tape: fix IDE_AFLAG_* atomic accesses
ide-tape: change IDE_AFLAG_IGNORE_DSC non-atomically
pdc202xx_old: kill resetproc() method
pdc202xx_old: don't call pdc202xx_reset() on IRQ timeout
pdc202xx_old: use ide_dma_test_irq()
ide: preserve Host Protected Area by default (v2)
ide-gd: implement block device ->set_capacity method (v2)
...
* 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
x86: Provide _sdata in the vmlinux.lds.S file
x86: handle initrd that extends into unusable memory
As explained by Benjamin Herrenschmidt:
Oh and btw, your patch alone doesn't fix powerpc, because it's missing
a whole bunch of GFP_KERNEL's in the arch code... You would have to
grep the entire kernel for things that check slab_is_available() and
even then you'll be missing some.
For example, slab_is_available() didn't always exist, and so in the
early days on powerpc, we used a mem_init_done global that is set form
mem_init() (not perfect but works in practice). And we still have code
using that to do the test.
Therefore, mask out __GFP_WAIT, __GFP_IO, and __GFP_FS in the slab allocators
in early boot code to avoid enabling interrupts.
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Conflicts:
drivers/message/fusion/mptsas.c
fixed up conflict between req->data_len accessors and mptsas driver updates.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Support the VIRTIO_RING_F_INDIRECT_DESC feature.
This is a simple matter of changing the descriptor walking
code to operate on a struct vring_desc* and supplying it
with an indirect table if detected.
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The Guest only really needs to tell us about activity when we're going
to listen to the eventfd: normally, we don't want to know.
So if there are no available buffers, turn on notifications, re-check,
then wait for the Guest to notify us via the eventfd, then turn
notifications off again.
There's enough else going on that the differences are in the noise.
Before: Secs RxKicks TxKicks
1G TCP Guest->Host: 3.94 4686 32815
1M normal pings: 104 142862 1000010
1M 1k pings (-l 120): 57 142026 1000007
After:
1G TCP Guest->Host: 3.76 4691 32811
1M normal pings: 111 142859 997467
1M 1k pings (-l 120): 55 19648 501549
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Rather than triggering an interrupt every time, we only trigger an
interrupt when there are no more incoming packets (or the recv queue
is full).
However, the overhead of doing the select to figure this out is
measurable: 1M pings goes from 98 to 104 seconds, and 1G Guest->Host
TCP goes from 3.69 to 3.94 seconds. It's close to the noise though.
I tested various timeouts, including reducing it as the number of
pending packets increased, timing a 1 gigabyte TCP send from Guest ->
Host and Host -> Guest (GSO disabled, to increase packet rate).
// time tcpblast -o -s 65536 -c 16k 192.168.2.1:9999 > /dev/null
Timeout Guest->Host Pkts/irq Host->Guest Pkts/irq
Before 11.3s 1.0 6.3s 1.0
0 11.7s 1.0 6.6s 23.5
1 17.1s 8.8 8.6s 26.0
1/pending 13.4s 1.9 6.6s 23.8
2/pending 13.6s 2.8 6.6s 24.1
5/pending 14.1s 5.0 6.6s 24.4
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If we track how many buffers we've used, we can tell whether we really
need to interrupt the Guest. This happens as a side effect of
spurious notifications.
Spurious notifications happen because it can take a while before the
Host thread wakes up and sets the VRING_USED_F_NO_NOTIFY flag, and
meanwhile the Guest can more notifications.
A real fix would be to use wake counts, rather than a suppression
flag, but the practical difference is generally in the noise: the
interrupt is usually coalesced into a pending one anyway so we just
save a system call which isn't clearly measurable.
Secs Spurious IRQS
1G TCP Guest->Host: 3.93 58
1M normal pings: 100 72
1M 1k pings (-l 120): 57 492904
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Rather than sending an interrupt on every buffer, we only send an interrupt
when we're about to wait for the Guest to send us a new one. The console
input and network input still send interrupts manually, but the block device,
network and console output queues can simply rely on this logic to send
interrupts to the Guest at the right time.
The patch is cluttered by moving trigger_irq() higher in the code.
In practice, two factors make this optimization less interesting:
(1) we often only get one input at a time, even for networking,
(2) triggering an interrupt rapidly tends to get coalesced anyway.
Before: Secs RxIRQS TxIRQs
1G TCP Guest->Host: 3.72 32784 32771
1M normal pings: 99 1000004 995541
100,000 1k pings (-l 120): 5 49510 49058
After:
1G TCP Guest->Host: 3.69 32809 32769
1M normal pings: 99 1000004 996196
100,000 1k pings (-l 120): 5 52435 52361
(Note the interrupt count on 100k pings goes *up*: see next patch).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We no longer need an efficient mechanism to force the Guest back into
host userspace, as each device is serviced without bothering the main
Guest process (aka. the Launcher).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Currently lguest has three threads: the main Launcher thread, a Waker
thread, and a thread for the block device (because synchronous block
was simply too painful to bear).
The Waker selects() on all the input file descriptors (eg. stdin, net
devices, pipe to the block thread) and when one becomes readable it calls
into the kernel to kick the Launcher thread out into userspace, which
repeats the poll, services the device(s), and then tells the kernel to
release the Waker before re-entering the kernel to run the Guest.
Also, to make a slightly-decent network transmit routine, the Launcher
would suppress further network interrupts while it set a timer: that
signal handler would write to a pipe, which would rouse the Waker
which would prod the Launcher out of the kernel to check the network
device again.
Now we can convert all our virtqueues to separate threads: each one has
a separate eventfd for when the Guest pokes the device, and can trigger
interrupts in the Guest directly.
The linecount shows how much this simplifies, but to really bring it
home, here's an strace analysis of single Guest->Host ping before:
* Guest sends packet, notifies xmit vq, return control to Launcher
* Launcher clears notification flag on xmit ring
* Launcher writes packet to TUN device
writev(4, [{"\0\0\0\0\0\0\0\0\0\0", 10}, {"\366\r\224`\2058\272m\224vf\274\10\0E\0\0T\0\0@\0@\1\265"..., 98}], 2) = 108
* Launcher sets up interrupt for Guest (xmit ring is empty)
write(10, "\2\0\0\0\3\0\0\0", 8) = 0
* Launcher sets up timer for interrupt mitigation
setitimer(ITIMER_REAL, {it_interval={0, 0}, it_value={0, 505}}, NULL) = 0
* Launcher re-runs guest
pread64(10, 0xbfa5f4d4, 4, 0) ...
* Waker notices reply packet in tun device (it was in select)
select(12, [0 3 4 6 11], NULL, NULL, NULL) = 1 (in [4])
* Waker kicks Launcher out of guest:
pwrite64(10, "\3\0\0\0\1\0\0\0", 8, 0) = 0
* Launcher returns from running guest:
... = -1 EAGAIN (Resource temporarily unavailable)
* Launcher looks at input fds:
select(7, [0 3 4 6], NULL, NULL, {0, 0}) = 1 (in [4], left {0, 0})
* Launcher reads pong from tun device:
readv(4, [{"\0\0\0\0\0\0\0\0\0\0", 10}, {"\272m\224vf\274\366\r\224`\2058\10\0E\0\0T\364\26\0\0@"..., 1518}], 2) = 108
* Launcher injects guest notification:
write(10, "\2\0\0\0\2\0\0\0", 8) = 0
* Launcher rechecks fds:
select(7, [0 3 4 6], NULL, NULL, {0, 0}) = 0 (Timeout)
* Launcher clears Waker:
pwrite64(10, "\3\0\0\0\0\0\0\0", 8, 0) = 0
* Launcher reruns Guest:
pread64(10, 0xbfa5f4d4, 4, 0) = ? ERESTARTSYS (To be restarted)
* Signal comes in, uses pipe to wake up Launcher:
--- SIGALRM (Alarm clock) @ 0 (0) ---
write(8, "\0", 1) = 1
sigreturn() = ? (mask now [])
* Waker sees write on pipe:
select(12, [0 3 4 6 11], NULL, NULL, NULL) = 1 (in [6])
* Waker kicks Launcher out of Guest:
pwrite64(10, "\3\0\0\0\1\0\0\0", 8, 0) = 0
* Launcher exits from kernel:
pread64(10, 0xbfa5f4d4, 4, 0) = -1 EAGAIN (Resource temporarily unavailable)
* Launcher looks to see what fd woke it:
select(7, [0 3 4 6], NULL, NULL, {0, 0}) = 1 (in [6], left {0, 0})
* Launcher reads timeout fd, sets notification flag on xmit ring
read(6, "\0", 32) = 1
* Launcher rechecks fds:
select(7, [0 3 4 6], NULL, NULL, {0, 0}) = 0 (Timeout)
* Launcher clears Waker:
pwrite64(10, "\3\0\0\0\0\0\0\0", 8, 0) = 0
* Launcher resumes Guest:
pread64(10, "\0p\0\4", 4, 0) ....
strace analysis of single Guest->Host ping after:
* Guest sends packet, notifies xmit vq, creates event on eventfd.
* Network xmit thread wakes from read on eventfd:
read(7, "\1\0\0\0\0\0\0\0", 8) = 8
* Network xmit thread writes packet to TUN device
writev(4, [{"\0\0\0\0\0\0\0\0\0\0", 10}, {"J\217\232FI\37j\27\375\276\0\304\10\0E\0\0T\0\0@\0@\1\265"..., 98}], 2) = 108
* Network recv thread wakes up from read on tunfd:
readv(4, [{"\0\0\0\0\0\0\0\0\0\0", 10}, {"j\27\375\276\0\304J\217\232FI\37\10\0E\0\0TiO\0\0@\1\214"..., 1518}], 2) = 108
* Network recv thread sets up interrupt for the Guest
write(6, "\2\0\0\0\2\0\0\0", 8) = 0
* Network recv thread goes back to reading tunfd
13:39:42.460285 readv(4, <unfinished ...>
* Network xmit thread sets up interrupt for Guest (xmit ring is empty)
write(6, "\2\0\0\0\3\0\0\0", 8) = 0
* Network xmit thread goes back to reading from eventfd
read(7, <unfinished ...>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Currently, when a Guest wants to perform I/O it calls LHCALL_NOTIFY with
an address: the main Launcher process returns with this address, and figures
out what device to run.
A far nicer model is to let processes bind an eventfd to an address: if we
find one, we simply signal the eventfd.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Davide Libenzi <davidel@xmailserver.org>
lguest wants to attach eventfds to guest notifications, and lguest is
usually a module.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
To: Davide Libenzi <davidel@xmailserver.org>
We currently only allow the Launcher process to send interrupts, but it
as we already send interrupts from the hrtimer, it's a simple matter of
extracting that code into a common set_interrupt routine.
As we switch to a thread per virtqueue, this avoids a bottleneck through the
main Launcher process.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1) j wasn't initialized in setup_pagetables, so they weren't set up for me
causing immediate guest crashes.
2) gpte_addr should not re-read the pmd from the Guest. Especially
not BUG_ON() based on the value. If we ever supported SMP guests,
they could trigger that. And the Launcher could also trigger it
(tho currently root-only).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This version requires that host and guest have the same PAE status.
NX cap is not offered to the guest, yet.
Signed-off-by: Matias Zabaljauregui <zabaljauregui@gmail.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Add support for kvm_hypercall4(); PAE wants it.
Signed-off-by: Matias Zabaljauregui <zabaljauregui at gmail.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
replace LHCALL_SET_PMD with LHCALL_SET_PGD hypercall name
(That's really what it is, and the confusion gets worse with PAE support)
Signed-off-by: Matias Zabaljauregui <zabaljauregui@gmail.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Reported-by: Jeremy Fitzhardinge <jeremy@goop.org>
Some cleanups and replace direct assignment with native_set_* macros which properly handle 64-bit entries when PAE is activated
Signed-off-by: Matias Zabaljauregui <zabaljauregui@gmail.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The "len" field in the used ring for virtio indicates the number of
bytes *written* to the buffer. This means the guest doesn't have to
zero the buffers in advance as it always knows the used length.
Erroneously, the console and network example code puts the length
*read* into that field. The guest ignores it, but it's wrong.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If GDT_ENTRIES were every > 256, this could become a problem.
Signed-off-by: Matias Zabaljauregui <zabaljauregui at gmail.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
18 months ago 5bbf89fc26 changed to loading
bzImages directly, and no longer manually ungzipping them, so we no longer
need libz.
Also, -m32 is useful for those on 64-bit platforms (and harmless on
32-bit).
Reported-by: Ron Minnich <rminnich@gmail.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The downside of the last patch which made restore_flags and irq_enable
check interrupts is that they are now too big to be patched directly
into the callsites, so the C versions are always used.
But the C versions go via PV_CALLEE_SAVE_REGS_THUNK which saves all
the registers. In fact, we don't need any registers in the fast path,
so we can do better than this if we actually code them in assembler.
The results are in the noise, but since it's about the same amount of
code, it's worth applying.
1GB Guest->Host: input(suppressed),output(suppressed)
Before:
Seconds: 0:16.53
Packets: 377268,753673
Interrupts: 22461,24297
Notifications: 1(5245),21303(732370)
Net IRQs triggered: 377023(245),42578(711095)
After:
Seconds: 0:16.48
Packets: 377289,753673
Interrupts: 22281,24465
Notifications: 1(5245),21296(732377)
Net IRQs triggered: 377060(229),42564(711109)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
lguest never checked for pending interrupts when enabling interrupts, and
things still worked. However, it makes a significant difference to TCP
performance, so it's time we fixed it by introducing a pending_irq flag
and checking it on irq_restore and irq_enable.
These two routines are now too big to patch into the 8/10 bytes
patch space, so we drop that code.
Note: The high latency on interrupt delivery had a very curious
effect: once everything else was optimized, networking without GSO was
faster than networking with GSO, since more interrupts were sent and
hence a greater chance of one getting through to the Guest!
Note2: (Almost) Closing the same loophole for iret doesn't have any
measurable effect, so I'm leaving that patch for the moment.
Before:
1GB tcpblast Guest->Host: 30.7 seconds
1GB tcpblast Guest->Host (no GSO): 76.0 seconds
After:
1GB tcpblast Guest->Host: 6.8 seconds
1GB tcpblast Guest->Host (no GSO): 27.8 seconds
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
When the Guest does the LHCALL_HALT hypercall, we go to sleep, expecting
that a timer or the Waker will wake_up_process() us.
But we do it in a stupid way, leaving a classic missing wakeup race.
So split maybe_do_interrupt() into interrupt_pending() and
try_deliver_interrupt(), and check maybe_do_interrupt() and the
"break_out" flag before calling schedule.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2088761152 (lguest: notify on empty) introduced
lguest support for the VIRTIO_F_NOTIFY_ON_EMPTY flag, but in fact it turned on
interrupts all the time.
Because we always process one buffer at a time, the inflight count is always 0
when call trigger_irq and so we always ignore VRING_AVAIL_F_NO_INTERRUPT from
the Guest.
It should be looking to see if there are more buffers in the Guest's queue:
if it's empty, then we force an interrupt.
This makes little difference, since we usually have an empty queue; but
that's the subject of another patch.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The Launcher could be inside the Guest on another CPU; wake_up_process
will do nothing because it is "running". kick_process will knock it
back into our kernel in this case, otherwise we'll miss it until the
next guest exit.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
lguest needs kick_process: wake_up_process() does nothing if a process
is running, which isn't sufficient (we need it in the kernel).
And lguest support is usually modular.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Ingo Molnar <mingo@elte.hu>
Since the Launcher process runs the Guest, it doesn't have to be very
serious about its barriers: the Guest isn't running while we are (Guest
is UP).
Before we change to use threads to service devices, we need to fix this.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Copy from arch/x86/kernel/irqinit_32.c: we don't use the vectors beyond
LGUEST_IRQS (if any), but we might as well set them all.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We hand the /dev/lguest fd everywhere; it's far neater to just make it
a global (it already is, in fact, hidden in the waker_fds struct).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We can't trust the values in the device descriptor table once the
guest has booted, so keep local copies. They could set them to
strange values then cause us to segv (they're 8 bit values, so they
can't make our pointers go too wild).
This becomes more important with the following patches which read them.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This patch allows a virtio driver to use VIRTIO_DEV_ANY_ID for the
device id. This will be used by a test module that can be bound to
any virtio device.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This bug never appeared, since all current virtio drivers use
VIRTIO_DEV_ANY_ID for the vendor field. If a real vendor would be used,
the check in virtio_id_match is wrong - it returns 0 if
id->vendor == dev->id.vendor.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If the device fills less than 4 bytes of our random buffer, we'll
BUG_ON. It's nicer to handle the case where it partially fills the
buffer (the protocol doesn't explicitly bad that).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The remove member of the virtio_driver structure uses __devexit_p(), so
the remove function itself should be marked with __devexit. And where
there be __devexit on the remove, so is there __devinit on the probe.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Add a new feature flag for indirect ring entries. These are ring
entries which point to a table of buffer descriptors.
The idea here is to increase the ring capacity by allowing a larger
effective ring size whereby the ring size dictates the number of
requests that may be outstanding, rather than the size of those
requests.
This should be most effective in the case of block I/O where we can
potentially benefit by concurrently dispatching a large number of
large requests. Even in the simple case of single segment block
requests, this results in a threefold increase in ring capacity.
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Drivers don't add transport features to their table, so we
shouldn't check these with virtio_check_driver_offered_feature().
We could perhaps add an ->offered_feature() virtio_config_op,
but that perhaps that would be overkill for a consitency check
like this.
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>