Get rid of the Rx ring and replace it with a pair of queues instead. One
queue gets the packets that are in-sequence and are ready for processing by
recvmsg(); the other queue gets the out-of-sequence packets for addition to
the first queue as the holes get filled.
The annotation ring is removed and replaced with a SACK table. The SACK
table has the bits set that correspond exactly to the sequence number of
the packet being acked. The SACK ring is copied when an ACK packet is
being assembled and rotated so that the first ACK is in byte 0.
Flow control handling is altered so that packets that are moved to the
in-sequence queue are hard-ACK'd even before they're consumed - and then
the Rx window size in the ACK packet (rsize) is shrunk down to compensate
(even going to 0 if the window is full).
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Clean up the rxrpc_propose_ACK() function. If deferred PING ACK proposal
is split out, it's only really needed for deferred DELAY ACKs. All other
ACKs, bar terminal IDLE ACK are sent immediately. The deferred IDLE ACK
submission can be handled by conversion of a DELAY ACK into an IDLE ACK if
there's nothing to be SACK'd.
Also, because there's a delay between an ACK being generated and being
transmitted, it's possible that other ACKs of the same type will be
generated during that interval. Apart from the ACK time and the serial
number responded to, most of the ACK body, including window and SACK
parameters, are not filled out till the point of transmission - so we can
avoid generating a new ACK if there's one pending that will cover the SACK
data we need to convey.
Therefore, don't propose a new DELAY or IDLE ACK for a call if there's one
already pending.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
AF_RXRPC's listen() handler lets you set the backlog up to 32 (if you bump
up the sysctl), but whilst the preallocation circular buffers have 32 slots
in them, one of them has to be a dead slot because we're using CIRC_CNT().
This means that listen(rxrpc_sock, 32) will cause an oops when the socket
is closed because rxrpc_service_prealloc_one() allocated one too many calls
and rxrpc_discard_prealloc() won't then be able to get rid of them because
it'll think the ring is empty. rxrpc_release_calls_on_socket() then tries
to abort them, but oopses because call->peer isn't yet set.
Fix this by setting the maximum backlog to RXRPC_BACKLOG_MAX - 1 to match
the ring capacity.
BUG: kernel NULL pointer dereference, address: 0000000000000086
...
RIP: 0010:rxrpc_send_abort_packet+0x73/0x240 [rxrpc]
Call Trace:
<TASK>
? __wake_up_common_lock+0x7a/0x90
? rxrpc_notify_socket+0x8e/0x140 [rxrpc]
? rxrpc_abort_call+0x4c/0x60 [rxrpc]
rxrpc_release_calls_on_socket+0x107/0x1a0 [rxrpc]
rxrpc_release+0xc9/0x1c0 [rxrpc]
__sock_release+0x37/0xa0
sock_close+0x11/0x20
__fput+0x89/0x240
task_work_run+0x59/0x90
do_exit+0x319/0xaa0
Fixes: 00e907127e ("rxrpc: Preallocate peers, conns and calls for incoming service requests")
Reported-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-afs@lists.infradead.org
Link: https://lists.infradead.org/pipermail/linux-afs/2022-March/005079.html
Signed-off-by: David S. Miller <davem@davemloft.net>
Rewrite the rxrpc client connection manager so that it can support multiple
connections for a given security key to a peer. The following changes are
made:
(1) For each open socket, the code currently maintains an rbtree with the
connections placed into it, keyed by communications parameters. This
is tricky to maintain as connections can be culled from the tree or
replaced within it. Connections can require replacement for a number
of reasons, e.g. their IDs span too great a range for the IDR data
type to represent efficiently, the call ID numbers on that conn would
overflow or the conn got aborted.
This is changed so that there's now a connection bundle object placed
in the tree, keyed on the same parameters. The bundle, however, does
not need to be replaced.
(2) An rxrpc_bundle object can now manage the available channels for a set
of parallel connections. The lock that manages this is moved there
from the rxrpc_connection struct (channel_lock).
(3) There'a a dummy bundle for all incoming connections to share so that
they have a channel_lock too. It might be better to give each
incoming connection its own bundle. This bundle is not needed to
manage which channels incoming calls are made on because that's the
solely at whim of the client.
(4) The restrictions on how many client connections are around are
removed. Instead, a previous patch limits the number of client calls
that can be allocated. Ordinarily, client connections are reaped
after 2 minutes on the idle queue, but when more than a certain number
of connections are in existence, the reaper starts reaping them after
2s of idleness instead to get the numbers back down.
It could also be made such that new call allocations are forced to
wait until the number of outstanding connections subsides.
Signed-off-by: David Howells <dhowells@redhat.com>
The MSCC bug fix in 'net' had to be slightly adjusted because the
register accesses are done slightly differently in net-next.
Signed-off-by: David S. Miller <davem@davemloft.net>
rxrpc currently uses a fixed 4s retransmission timeout until the RTT is
sufficiently sampled. This can cause problems with some fileservers with
calls to the cache manager in the afs filesystem being dropped from the
fileserver because a packet goes missing and the retransmission timeout is
greater than the call expiry timeout.
Fix this by:
(1) Copying the RTT/RTO calculation code from Linux's TCP implementation
and altering it to fit rxrpc.
(2) Altering the various users of the RTT to make use of the new SRTT
value.
(3) Replacing the use of rxrpc_resend_timeout to use the calculated RTO
value instead (which is needed in jiffies), along with a backoff.
Notes:
(1) rxrpc provides RTT samples by matching the serial numbers on outgoing
DATA packets that have the RXRPC_REQUEST_ACK set and PING ACK packets
against the reference serial number in incoming REQUESTED ACK and
PING-RESPONSE ACK packets.
(2) Each packet that is transmitted on an rxrpc connection gets a new
per-connection serial number, even for retransmissions, so an ACK can
be cross-referenced to a specific trigger packet. This allows RTT
information to be drawn from retransmitted DATA packets also.
(3) rxrpc maintains the RTT/RTO state on the rxrpc_peer record rather than
on an rxrpc_call because many RPC calls won't live long enough to
generate more than one sample.
(4) The calculated SRTT value is in units of 8ths of a microsecond rather
than nanoseconds.
The (S)RTT and RTO values are displayed in /proc/net/rxrpc/peers.
Fixes: 17926a7932 ([AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both"")
Signed-off-by: David Howells <dhowells@redhat.com>
- add SPDX header;
- adjust title markup;
- use autonumbered list markups;
- mark code blocks and literals as such;
- mark tables as such;
- adjust identation, whitespaces and blank lines where needed;
- add to networking/index.rst.
Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the sysctl code the proc_dointvec_minmax() function is often used to
validate the user supplied value between an allowed range. This
function uses the extra1 and extra2 members from struct ctl_table as
minimum and maximum allowed value.
On sysctl handler declaration, in every source file there are some
readonly variables containing just an integer which address is assigned
to the extra1 and extra2 members, so the sysctl range is enforced.
The special values 0, 1 and INT_MAX are very often used as range
boundary, leading duplication of variables like zero=0, one=1,
int_max=INT_MAX in different source files:
$ git grep -E '\.extra[12].*&(zero|one|int_max)' |wc -l
248
Add a const int array containing the most commonly used values, some
macros to refer more easily to the correct array member, and use them
instead of creating a local one for every object file.
This is the bloat-o-meter output comparing the old and new binary
compiled with the default Fedora config:
# scripts/bloat-o-meter -d vmlinux.o.old vmlinux.o
add/remove: 2/2 grow/shrink: 0/2 up/down: 24/-188 (-164)
Data old new delta
sysctl_vals - 12 +12
__kstrtab_sysctl_vals - 12 +12
max 14 10 -4
int_max 16 - -16
one 68 - -68
zero 128 28 -100
Total: Before=20583249, After=20583085, chg -0.00%
[mcroce@redhat.com: tipc: remove two unused variables]
Link: http://lkml.kernel.org/r/20190530091952.4108-1-mcroce@redhat.com
[akpm@linux-foundation.org: fix net/ipv6/sysctl_net_ipv6.c]
[arnd@arndb.de: proc/sysctl: make firmware loader table conditional]
Link: http://lkml.kernel.org/r/20190617130014.1713870-1-arnd@arndb.de
[akpm@linux-foundation.org: fix fs/eventpoll.c]
Link: http://lkml.kernel.org/r/20190430180111.10688-1-mcroce@redhat.com
Signed-off-by: Matteo Croce <mcroce@redhat.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public licence as published by
the free software foundation either version 2 of the licence or at
your option any later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 114 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190520170857.552531963@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The static int 'zero' is defined but is never used hence it is
redundant and can be removed. The use of this variable was removed
with commit a158bdd324 ("rxrpc: Fix call timeouts").
Cleans up clang warning:
warning: 'zero' defined but not used [-Wunused-const-variable=]
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix the rxrpc call expiration timeouts and make them settable from
userspace. By analogy with other rx implementations, there should be three
timeouts:
(1) "Normal timeout"
This is set for all calls and is triggered if we haven't received any
packets from the peer in a while. It is measured from the last time
we received any packet on that call. This is not reset by any
connection packets (such as CHALLENGE/RESPONSE packets).
If a service operation takes a long time, the server should generate
PING ACKs at a duration that's substantially less than the normal
timeout so is to keep both sides alive. This is set at 1/6 of normal
timeout.
(2) "Idle timeout"
This is set only for a service call and is triggered if we stop
receiving the DATA packets that comprise the request data. It is
measured from the last time we received a DATA packet.
(3) "Hard timeout"
This can be set for a call and specified the maximum lifetime of that
call. It should not be specified by default. Some operations (such
as volume transfer) take a long time.
Allow userspace to set/change the timeouts on a call with sendmsg, using a
control message:
RXRPC_SET_CALL_TIMEOUTS
The data to the message is a number of 32-bit words, not all of which need
be given:
u32 hard_timeout; /* sec from first packet */
u32 idle_timeout; /* msec from packet Rx */
u32 normal_timeout; /* msec from data Rx */
This can be set in combination with any other sendmsg() that affects a
call.
Signed-off-by: David Howells <dhowells@redhat.com>
Keep that call timeouts as ktimes rather than jiffies so that they can be
expressed as functions of RTT.
Signed-off-by: David Howells <dhowells@redhat.com>
In addition to sending a PING ACK to gain RTT data, we can set the
RXRPC_REQUEST_ACK flag on a DATA packet and get a REQUESTED-ACK ACK. The
ACK packet contains the serial number of the packet it is in response to,
so we can look through the Tx buffer for a matching DATA packet.
This requires that the data packets be stamped with the time of
transmission as a ktime rather than having the resend_at time in jiffies.
This further requires the resend code to do the resend determination in
ktimes and convert to jiffies to set the timer.
Signed-off-by: David Howells <dhowells@redhat.com>
call->rx_winsize should be initialised to the sysctl setting and the sysctl
setting should be limited to the maximum we want to permit. Further, we
need to place this in the ACK info instead of the sysctl setting.
Furthermore, discard the idea of accepting the subpackets of a jumbo packet
that lie beyond the receive window when the first packet of the jumbo is
within the window. Just discard the excess subpackets instead. This
allows the receive window to be opened up right to the buffer size less one
for the dead slot.
Signed-off-by: David Howells <dhowells@redhat.com>
rxrpc calls shouldn't hold refs on the sock struct. This was done so that
the socket wouldn't go away whilst the call was in progress, such that the
call could reach the socket's queues.
However, we can mark the socket as requiring an RCU release and rely on the
RCU read lock.
To make this work, we do:
(1) rxrpc_release_call() removes the call's call user ID. This is now
only called from socket operations and not from the call processor:
rxrpc_accept_call() / rxrpc_kernel_accept_call()
rxrpc_reject_call() / rxrpc_kernel_reject_call()
rxrpc_kernel_end_call()
rxrpc_release_calls_on_socket()
rxrpc_recvmsg()
Though it is also called in the cleanup path of
rxrpc_accept_incoming_call() before we assign a user ID.
(2) Pass the socket pointer into rxrpc_release_call() rather than getting
it from the call so that we can get rid of uninitialised calls.
(3) Fix call processor queueing to pass a ref to the work queue and to
release that ref at the end of the processor function (or to pass it
back to the work queue if we have to requeue).
(4) Skip out of the call processor function asap if the call is complete
and don't requeue it if the call is complete.
(5) Clean up the call immediately that the refcount reaches 0 rather than
trying to defer it. Actual deallocation is deferred to RCU, however.
(6) Don't hold socket refs for allocated calls.
(7) Use the RCU read lock when queueing a message on a socket and treat
the call's socket pointer according to RCU rules and check it for
NULL.
We also need to use the RCU read lock when viewing a call through
procfs.
(8) Transmit the final ACK/ABORT to a client call in rxrpc_release_call()
if this hasn't been done yet so that we can then disconnect the call.
Once the call is disconnected, it won't have any access to the
connection struct and the UDP socket for the call work processor to be
able to send the ACK. Terminal retransmission will be handled by the
connection processor.
(9) Release all calls immediately on the closing of a socket rather than
trying to defer this. Incomplete calls will be aborted.
The call refcount model is much simplified. Refs are held on the call by:
(1) A socket's user ID tree.
(2) A socket's incoming call secureq and acceptq.
(3) A kernel service that has a call in progress.
(4) A queued call work processor. We have to take care to put any call
that we failed to queue.
(5) sk_buffs on a socket's receive queue. A future patch will get rid of
this.
Whilst we're at it, we can do:
(1) Get rid of the RXRPC_CALL_EV_RELEASE event. Release is now done
entirely from the socket routines and never from the call's processor.
(2) Get rid of the RXRPC_CALL_DEAD state. Calls now end in the
RXRPC_CALL_COMPLETE state.
(3) Get rid of the rxrpc_call::destroyer work item. Calls are now torn
down when their refcount reaches 0 and then handed over to RCU for
final cleanup.
(4) Get rid of the rxrpc_call::deadspan timer. Calls are cleaned up
immediately they're finished with and don't hang around.
Post-completion retransmission is handled by the connection processor
once the call is disconnected.
(5) Get rid of the dead call expiry setting as there's no longer a timer
to set.
(6) rxrpc_destroy_all_calls() can just check that the call list is empty.
Signed-off-by: David Howells <dhowells@redhat.com>
Improve the management and caching of client rxrpc connection objects.
From this point, client connections will be managed separately from service
connections because AF_RXRPC controls the creation and re-use of client
connections but doesn't have that luxury with service connections.
Further, there will be limits on the numbers of client connections that may
be live on a machine. No direct restriction will be placed on the number
of client calls, excepting that each client connection can support a
maximum of four concurrent calls.
Note that, for a number of reasons, we don't want to simply discard a
client connection as soon as the last call is apparently finished:
(1) Security is negotiated per-connection and the context is then shared
between all calls on that connection. The context can be negotiated
again if the connection lapses, but that involves holding up calls
whilst at least two packets are exchanged and various crypto bits are
performed - so we'd ideally like to cache it for a little while at
least.
(2) If a packet goes astray, we will need to retransmit a final ACK or
ABORT packet. To make this work, we need to keep around the
connection details for a little while.
(3) The locally held structures represent some amount of setup time, to be
weighed against their occupation of memory when idle.
To this end, the client connection cache is managed by a state machine on
each connection. There are five states:
(1) INACTIVE - The connection is not held in any list and may not have
been exposed to the world. If it has been previously exposed, it was
discarded from the idle list after expiring.
(2) WAITING - The connection is waiting for the number of client conns to
drop below the maximum capacity. Calls may be in progress upon it
from when it was active and got culled.
The connection is on the rxrpc_waiting_client_conns list which is kept
in to-be-granted order. Culled conns with waiters go to the back of
the queue just like new conns.
(3) ACTIVE - The connection has at least one call in progress upon it, it
may freely grant available channels to new calls and calls may be
waiting on it for channels to become available.
The connection is on the rxrpc_active_client_conns list which is kept
in activation order for culling purposes.
(4) CULLED - The connection got summarily culled to try and free up
capacity. Calls currently in progress on the connection are allowed
to continue, but new calls will have to wait. There can be no waiters
in this state - the conn would have to go to the WAITING state
instead.
(5) IDLE - The connection has no calls in progress upon it and must have
been exposed to the world (ie. the EXPOSED flag must be set). When it
expires, the EXPOSED flag is cleared and the connection transitions to
the INACTIVE state.
The connection is on the rxrpc_idle_client_conns list which is kept in
order of how soon they'll expire.
A connection in the ACTIVE or CULLED state must have at least one active
call upon it; if in the WAITING state it may have active calls upon it;
other states may not have active calls.
As long as a connection remains active and doesn't get culled, it may
continue to process calls - even if there are connections on the wait
queue. This simplifies things a bit and reduces the amount of checking we
need do.
There are a couple flags of relevance to the cache:
(1) EXPOSED - The connection ID got exposed to the world. If this flag is
set, an extra ref is added to the connection preventing it from being
reaped when it has no calls outstanding. This flag is cleared and the
ref dropped when a conn is discarded from the idle list.
(2) DONT_REUSE - The connection should be discarded as soon as possible and
should not be reused.
This commit also provides a number of new settings:
(*) /proc/net/rxrpc/max_client_conns
The maximum number of live client connections. Above this number, new
connections get added to the wait list and must wait for an active
conn to be culled. Culled connections can be reused, but they will go
to the back of the wait list and have to wait.
(*) /proc/net/rxrpc/reap_client_conns
If the number of desired connections exceeds the maximum above, the
active connection list will be culled until there are only this many
left in it.
(*) /proc/net/rxrpc/idle_conn_expiry
The normal expiry time for a client connection, provided there are
fewer than reap_client_conns of them around.
(*) /proc/net/rxrpc/idle_conn_fast_expiry
The expedited expiry time, used when there are more than
reap_client_conns of them around.
Note that I combined the Tx wait queue with the channel grant wait queue to
save space as only one of these should be in use at once.
Note also that, for the moment, the service connection cache still uses the
old connection management code.
Signed-off-by: David Howells <dhowells@redhat.com>
The rxrpc_transport struct is now redundant, given that the rxrpc_peer
struct is now per peer port rather than per peer host, so get rid of it.
Service connection lists are transferred to the rxrpc_peer struct, as is
the conn_lock. Previous patches moved the client connection handling out
of the rxrpc_transport struct and discarded the connection bundling code.
Signed-off-by: David Howells <dhowells@redhat.com>
Limit the socket incoming call backlog queue size so that a remote client
can't pump in sufficient new calls that the server runs out of memory. Note
that this is partially theoretical at the moment since whilst the number of
calls is limited, the number of packets trying to set up new calls is not.
This will be addressed in a later patch.
If the caller of listen() specifies a backlog INT_MAX, then they get the
current maximum; anything else greater than max_backlog or anything
negative incurs EINVAL.
The limit on the maximum queue size can be set by:
echo N >/proc/sys/net/rxrpc/max_backlog
where 4<=N<=32.
Further, set the default backlog to 0, requiring listen() to be called
before we start actually queueing new calls. Whilst this kind of is a
change in the UAPI, the caller can't actually *accept* new calls anyway
unless they've first called listen() to put the socket into the LISTENING
state - thus the aforementioned new calls would otherwise just sit there,
eating up kernel memory. (Note that sockets that don't have a non-zero
service ID bound don't get incoming calls anyway.)
Given that the default backlog is now 0, make the AFS filesystem call
kernel_listen() to set the maximum backlog for itself.
Possible improvements include:
(1) Trimming a too-large backlog to max_backlog when listen is called.
(2) Trimming the backlog value whenever the value is used so that changes
to max_backlog are applied to an open socket automatically. Note that
the AFS filesystem opens one socket and keeps it open for extended
periods, so would miss out on changes to max_backlog.
(3) Having a separate setting for the AFS filesystem.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace all "unsigned" types with "unsigned int" types.
Reported-by: David Miller <davem@davemloft.net>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The upper bound of the defined range for rx_mtu is being set in the same
member as the lower bound (extra1) rather than the correct place (extra2).
I'm not entirely sure why this compiles.
Signed-off-by: David Howells <dhowells@redhat.com>
Expose RxRPC parameters via sysctls to control the Rx window size, the Rx MTU
maximum size and the number of packets that can be glued into a jumbo packet.
More info added to Documentation/networking/rxrpc.txt.
Signed-off-by: David Howells <dhowells@redhat.com>
Add sysctls for configuring RxRPC protocol handling, specifically controls on
delays before ack generation, the delay before resending a packet, the maximum
lifetime of a call and the expiration times of calls, connections and
transports that haven't been recently used.
More info added in Documentation/networking/rxrpc.txt.
Signed-off-by: David Howells <dhowells@redhat.com>
Delete the old RxRPC code as it's now no longer used.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The semantic effect of insert_at_head is that it would allow new registered
sysctl entries to override existing sysctl entries of the same name. Which is
pain for caching and the proc interface never implemented.
I have done an audit and discovered that none of the current users of
register_sysctl care as (excpet for directories) they do not register
duplicate sysctl entries.
So this patch simply removes the support for overriding existing entries in
the sys_sysctl interface since no one uses it or cares and it makes future
enhancments harder.
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: David Howells <dhowells@redhat.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Andi Kleen <ak@muc.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Corey Minyard <minyard@acm.org>
Cc: Neil Brown <neilb@suse.de>
Cc: "John W. Linville" <linville@tuxdriver.com>
Cc: James Bottomley <James.Bottomley@steeleye.com>
Cc: Jan Kara <jack@ucw.cz>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: Mark Fasheh <mark.fasheh@oracle.com>
Cc: David Chinner <dgc@sgi.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Patrick McHardy <kaber@trash.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
