6 years agomacvtap: read vnet_hdr_size once
Willem de Bruijn [Fri, 3 Feb 2017 23:20:49 +0000 (18:20 -0500)]
macvtap: read vnet_hdr_size once

[ Upstream commit 837585a5375c38d40361cfe64e6fd11e1addb936 ]

When IFF_VNET_HDR is enabled, a virtio_net header must precede data.
Data length is verified to be greater than or equal to expected header
length tun->vnet_hdr_sz before copying.

Macvtap functions read the value once, but unless READ_ONCE is used,
the compiler may ignore this and read multiple times. Enforce a single
read and locally cached value to avoid updates between test and use.

Signed-off-by: Willem de Bruijn <>
Suggested-by: Eric Dumazet <>
Acked-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: BAckported to 3.2:
 - Use ACCESS_ONCE() instead of READ_ONCE()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agotun: read vnet_hdr_sz once
Willem de Bruijn [Fri, 3 Feb 2017 23:20:48 +0000 (18:20 -0500)]
tun: read vnet_hdr_sz once

[ Upstream commit e1edab87faf6ca30cd137e0795bc73aa9a9a22ec ]

When IFF_VNET_HDR is enabled, a virtio_net header must precede data.
Data length is verified to be greater than or equal to expected header
length tun->vnet_hdr_sz before copying.

Read this value once and cache locally, as it can be updated between
the test and use (TOCTOU).

Signed-off-by: Willem de Bruijn <>
Reported-by: Dmitry Vyukov <>
CC: Eric Dumazet <>
Acked-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Use ACCESS_ONCE() instead of READ_ONCE()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agotun: Fix TUN_PKT_STRIP setting
Herbert Xu [Sun, 2 Nov 2014 20:30:14 +0000 (04:30 +0800)]
tun: Fix TUN_PKT_STRIP setting

commit 2eb783c43e7cf807a45899c10ed556b6dc116625 upstream.

We set the flag TUN_PKT_STRIP if the user buffer provided is too
small to contain the entire packet plus meta-data.  However, this
has been broken ever since we added GSO meta-data.  VLAN acceleration
also has the same problem.

This patch fixes this by taking both into account when setting the

The fact that this has been broken for six years without anyone
realising means that nobody actually uses this flag.

Fixes: f43798c27684 ("tun: Allow GSO using virtio_net_hdr")
Signed-off-by: Herbert Xu <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - No VLAN acceleration support
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agotcp: fix 0 divide in __tcp_select_window()
Eric Dumazet [Wed, 1 Feb 2017 16:33:53 +0000 (08:33 -0800)]
tcp: fix 0 divide in __tcp_select_window()

[ Upstream commit 06425c308b92eaf60767bc71d359f4cbc7a561f8 ]

syszkaller fuzzer was able to trigger a divide by zero, when
TCP window scaling is not enabled.

SO_RCVBUF can be used not only to increase sk_rcvbuf, also
to decrease it below current receive buffers utilization.

If mss is negative or 0, just return a zero TCP window.

Signed-off-by: Eric Dumazet <>
Reported-by: Dmitry Vyukov <>
Acked-by: Neal Cardwell <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv6: pointer math error in ip6_tnl_parse_tlv_enc_lim()
Dan Carpenter [Wed, 1 Feb 2017 08:46:32 +0000 (11:46 +0300)]
ipv6: pointer math error in ip6_tnl_parse_tlv_enc_lim()

[ Upstream commit 63117f09c768be05a0bf465911297dc76394f686 ]

Casting is a high precedence operation but "off" and "i" are in terms of
bytes so we need to have some parenthesis here.

Fixes: fbfa743a9d2a ("ipv6: fix ip6_tnl_parse_tlv_enc_lim()")
Signed-off-by: Dan Carpenter <>
Acked-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv6: fix ip6_tnl_parse_tlv_enc_lim()
Eric Dumazet [Tue, 24 Jan 2017 00:43:06 +0000 (16:43 -0800)]
ipv6: fix ip6_tnl_parse_tlv_enc_lim()

[ Upstream commit fbfa743a9d2a0ffa24251764f10afc13eb21e739 ]

This function suffers from multiple issues.

First one is that pskb_may_pull() may reallocate skb->head,
so the 'raw' pointer needs either to be reloaded or not used at all.

Second issue is that NEXTHDR_DEST handling does not validate
that the options are present in skb->data, so we might read
garbage or access non existent memory.

With help from Willem de Bruijn.

Signed-off-by: Eric Dumazet <>
Reported-by: Dmitry Vyukov <>
Cc: Willem de Bruijn <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agocan: Fix kernel panic at security_sock_rcv_skb
Eric Dumazet [Fri, 27 Jan 2017 16:11:44 +0000 (08:11 -0800)]
can: Fix kernel panic at security_sock_rcv_skb

[ Upstream commit f1712c73714088a7252d276a57126d56c7d37e64 ]

Zhang Yanmin reported crashes [1] and provided a patch adding a
synchronize_rcu() call in can_rx_unregister()

The main problem seems that the sockets themselves are not RCU

If CAN uses RCU for delivery, then sockets should be freed only after
one RCU grace period.

Recent kernels could use sock_set_flag(sk, SOCK_RCU_FREE), but let's
ease stable backports with the following fix instead.

BUG: unable to handle kernel NULL pointer dereference at (null)
IP: [<ffffffff81495e25>] selinux_socket_sock_rcv_skb+0x65/0x2a0

Call Trace:
 [<ffffffff81485d8c>] security_sock_rcv_skb+0x4c/0x60
 [<ffffffff81d55771>] sk_filter+0x41/0x210
 [<ffffffff81d12913>] sock_queue_rcv_skb+0x53/0x3a0
 [<ffffffff81f0a2b3>] raw_rcv+0x2a3/0x3c0
 [<ffffffff81f06eab>] can_rcv_filter+0x12b/0x370
 [<ffffffff81f07af9>] can_receive+0xd9/0x120
 [<ffffffff81f07beb>] can_rcv+0xab/0x100
 [<ffffffff81d362ac>] __netif_receive_skb_core+0xd8c/0x11f0
 [<ffffffff81d36734>] __netif_receive_skb+0x24/0xb0
 [<ffffffff81d37f67>] process_backlog+0x127/0x280
 [<ffffffff81d36f7b>] net_rx_action+0x33b/0x4f0
 [<ffffffff810c88d4>] __do_softirq+0x184/0x440
 [<ffffffff81f9e86c>] do_softirq_own_stack+0x1c/0x30
 [<ffffffff810c76fb>] do_softirq.part.18+0x3b/0x40
 [<ffffffff810c8bed>] do_softirq+0x1d/0x20
 [<ffffffff81d30085>] netif_rx_ni+0xe5/0x110
 [<ffffffff8199cc87>] slcan_receive_buf+0x507/0x520
 [<ffffffff8167ef7c>] flush_to_ldisc+0x21c/0x230
 [<ffffffff810e3baf>] process_one_work+0x24f/0x670
 [<ffffffff810e44ed>] worker_thread+0x9d/0x6f0
 [<ffffffff810e4450>] ? rescuer_thread+0x480/0x480
 [<ffffffff810ebafc>] kthread+0x12c/0x150
 [<ffffffff81f9ccef>] ret_from_fork+0x3f/0x70

Reported-by: Zhang Yanmin <>
Signed-off-by: Eric Dumazet <>
Acked-by: Oliver Hartkopp <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: socket: fix recvmmsg not returning error from sock_error
Maxime Jayat [Tue, 21 Feb 2017 17:35:51 +0000 (18:35 +0100)]
net: socket: fix recvmmsg not returning error from sock_error

[ Upstream commit e623a9e9dec29ae811d11f83d0074ba254aba374 ]

Commit 34b88a68f26a ("net: Fix use after free in the recvmmsg exit path"),
changed the exit path of recvmmsg to always return the datagrams
variable and modified the error paths to set the variable to the error
code returned by recvmsg if necessary.

However in the case sock_error returned an error, the error code was
then ignored, and recvmmsg returned 0.

Change the error path of recvmmsg to correctly return the error code
of sock_error.

The bug was triggered by using recvmmsg on a CAN interface which was
not up. Linux 4.6 and later return 0 in this case while earlier
releases returned -ENETDOWN.

Fixes: 34b88a68f26a ("net: Fix use after free in the recvmmsg exit path")
Signed-off-by: Maxime Jayat <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv6: addrconf: Avoid addrconf_disable_change() using RCU read-side lock
Kefeng Wang [Thu, 19 Jan 2017 08:26:21 +0000 (16:26 +0800)]
ipv6: addrconf: Avoid addrconf_disable_change() using RCU read-side lock

[ Upstream commit 03e4deff4987f79c34112c5ba4eb195d4f9382b0 ]

Just like commit 4acd4945cd1e ("ipv6: addrconf: Avoid calling
netdevice notifiers with RCU read-side lock"), it is unnecessary
to make addrconf_disable_change() use RCU iteration over the
netdev list, since it already holds the RTNL lock, or we may meet
Illegal context switch in RCU read-side critical section.

Signed-off-by: Kefeng Wang <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoigmp: Make igmp group member RFC 3376 compliant
Michal Tesar [Mon, 2 Jan 2017 13:38:36 +0000 (14:38 +0100)]
igmp: Make igmp group member RFC 3376 compliant

[ Upstream commit 7ababb782690e03b78657e27bd051e20163af2d6 ]

5.2. Action on Reception of a Query

 When a system receives a Query, it does not respond immediately.
 Instead, it delays its response by a random amount of time, bounded
 by the Max Resp Time value derived from the Max Resp Code in the
 received Query message.  A system may receive a variety of Queries on
 different interfaces and of different kinds (e.g., General Queries,
 Group-Specific Queries, and Group-and-Source-Specific Queries), each
 of which may require its own delayed response.

 Before scheduling a response to a Query, the system must first
 consider previously scheduled pending responses and in many cases
 schedule a combined response.  Therefore, the system must be able to
 maintain the following state:

 o A timer per interface for scheduling responses to General Queries.

 o A per-group and interface timer for scheduling responses to Group-
   Specific and Group-and-Source-Specific Queries.

 o A per-group and interface list of sources to be reported in the
   response to a Group-and-Source-Specific Query.

 When a new Query with the Router-Alert option arrives on an
 interface, provided the system has state to report, a delay for a
 response is randomly selected in the range (0, [Max Resp Time]) where
 Max Resp Time is derived from Max Resp Code in the received Query
 message.  The following rules are then used to determine if a Report
 needs to be scheduled and the type of Report to schedule.  The rules
 are considered in order and only the first matching rule is applied.

 1. If there is a pending response to a previous General Query
    scheduled sooner than the selected delay, no additional response
    needs to be scheduled.

 2. If the received Query is a General Query, the interface timer is
    used to schedule a response to the General Query after the
    selected delay.  Any previously pending response to a General
    Query is canceled.

Currently the timer is rearmed with new random expiration time for
every incoming query regardless of possibly already pending report.
Which is not aligned with the above RFE.
It also might happen that higher rate of incoming queries can
postpone the report after the expiration time of the first query
causing group membership loss.

Now the per interface general query timer is rearmed only
when there is no pending report already scheduled on that interface or
the newly selected expiration time is before the already pending
scheduled report.

Signed-off-by: Michal Tesar <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agodrop_monitor: consider inserted data in genlmsg_end
Reiter Wolfgang [Tue, 3 Jan 2017 00:39:10 +0000 (01:39 +0100)]
drop_monitor: consider inserted data in genlmsg_end

[ Upstream commit 3b48ab2248e61408910e792fe84d6ec466084c1a ]

Final nlmsg_len field update must reflect inserted net_dm_drop_point

This patch depends on previous patch:
"drop_monitor: add missing call to genlmsg_end"

Signed-off-by: Reiter Wolfgang <>
Acked-by: Neil Horman <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agodrop_monitor: add missing call to genlmsg_end
Reiter Wolfgang [Sat, 31 Dec 2016 20:11:57 +0000 (21:11 +0100)]
drop_monitor: add missing call to genlmsg_end

[ Upstream commit 4200462d88f47f3759bdf4705f87e207b0f5b2e4 ]

Update nlmsg_len field with genlmsg_end to enable userspace processing
using nlmsg_next helper. Also adds error handling.

Signed-off-by: Reiter Wolfgang <>
Acked-by: Neil Horman <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv6: handle -EFAULT from skb_copy_bits
Dave Jones [Thu, 22 Dec 2016 16:16:22 +0000 (11:16 -0500)]
ipv6: handle -EFAULT from skb_copy_bits

[ Upstream commit a98f91758995cb59611e61318dddd8a6956b52c3 ]

By setting certain socket options on ipv6 raw sockets, we can confuse the
length calculation in rawv6_push_pending_frames triggering a BUG_ON.

RIP: 0010:[<ffffffff817c6390>] [<ffffffff817c6390>] rawv6_sendmsg+0xc30/0xc40
RSP: 0018:ffff881f6c4a7c18  EFLAGS: 00010282
RAX: 00000000fffffff2 RBX: ffff881f6c681680 RCX: 0000000000000002
RDX: ffff881f6c4a7cf8 RSI: 0000000000000030 RDI: ffff881fed0f6a00
RBP: ffff881f6c4a7da8 R08: 0000000000000000 R09: 0000000000000009
R10: ffff881fed0f6a00 R11: 0000000000000009 R12: 0000000000000030
R13: ffff881fed0f6a00 R14: ffff881fee39ba00 R15: ffff881fefa93a80

Call Trace:
 [<ffffffff8118ba23>] ? unmap_page_range+0x693/0x830
 [<ffffffff81772697>] inet_sendmsg+0x67/0xa0
 [<ffffffff816d93f8>] sock_sendmsg+0x38/0x50
 [<ffffffff816d982f>] SYSC_sendto+0xef/0x170
 [<ffffffff816da27e>] SyS_sendto+0xe/0x10
 [<ffffffff81002910>] do_syscall_64+0x50/0xa0
 [<ffffffff817f7cbc>] entry_SYSCALL64_slow_path+0x25/0x25

Handle by jumping to the failure path if skb_copy_bits gets an EFAULT.


#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <netinet/in.h>

#define LEN 504

int main(int argc, char* argv[])
int fd;
int zero = 0;
char buf[LEN];

memset(buf, 0, LEN);

fd = socket(AF_INET6, SOCK_RAW, 7);

setsockopt(fd, SOL_IPV6, IPV6_CHECKSUM, &zero, 4);
setsockopt(fd, SOL_IPV6, IPV6_DSTOPTS, &buf, LEN);

sendto(fd, buf, 1, 0, (struct sockaddr *) buf, 110);

Signed-off-by: Dave Jones <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonetvsc: reduce maximum GSO size
stephen hemminger [Tue, 6 Dec 2016 21:43:54 +0000 (13:43 -0800)]
netvsc: reduce maximum GSO size

[ Upstream commit a50af86dd49ee1851d1ccf06dd0019c05b95e297 ]

Hyper-V (and Azure) support using NVGRE which requires some extra space
for encapsulation headers. Because of this the largest allowed TSO
packet is reduced.

For older releases, hard code a fixed reduced value.  For next release,
there is a better solution which uses result of host offload

Signed-off-by: Stephen Hemminger <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust filename, context]
Signed-off-by: Ben Hutchings <>
6 years agonet/dccp: fix use-after-free in dccp_invalid_packet
Eric Dumazet [Mon, 28 Nov 2016 14:26:49 +0000 (06:26 -0800)]
net/dccp: fix use-after-free in dccp_invalid_packet

[ Upstream commit 648f0c28df282636c0c8a7a19ca3ce5fc80a39c3 ]

pskb_may_pull() can reallocate skb->head, we need to reload dh pointer
in dccp_invalid_packet() or risk use after free.

Bug found by Andrey Konovalov using syzkaller.

Signed-off-by: Eric Dumazet <>
Reported-by: Andrey Konovalov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet/sched: pedit: make sure that offset is valid
Amir Vadai [Mon, 28 Nov 2016 10:56:40 +0000 (12:56 +0200)]
net/sched: pedit: make sure that offset is valid

[ Upstream commit 95c2027bfeda21a28eb245121e6a249f38d0788e ]

Add a validation function to make sure offset is valid:
1. Not below skb head (could happen when offset is negative).
2. Validate both 'offset' and 'at'.

Signed-off-by: Amir Vadai <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: sky2: Fix shutdown crash
Jeremy Linton [Thu, 17 Nov 2016 15:14:25 +0000 (09:14 -0600)]
net: sky2: Fix shutdown crash

[ Upstream commit 06ba3b2133dc203e1e9bc36cee7f0839b79a9e8b ]

The sky2 frequently crashes during machine shutdown with:

sky2_get_stats+0x60/0x3d8 [sky2]

This is caused by the sky2 being called after it has been shutdown.
A previous thread about this can be found here:

An alternative fix is to assure that IFF_UP gets cleared by
calling dev_close() during shutdown. This is similar to what the
bnx2/tg3/xgene and maybe others are doing to assure that the driver
isn't being called following _shutdown().

Signed-off-by: Jeremy Linton <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoip6_tunnel: disable caching when the traffic class is inherited
Paolo Abeni [Wed, 16 Nov 2016 15:26:46 +0000 (16:26 +0100)]
ip6_tunnel: disable caching when the traffic class is inherited

[ Upstream commit b5c2d49544e5930c96e2632a7eece3f4325a1888 ]

If an ip6 tunnel is configured to inherit the traffic class from
the inner header, the dst_cache must be disabled or it will foul
the policy routing.

The issue is apprently there since at leat Linux-2.6.12-rc2.

Reported-by: Liam McBirnie <>
Cc: Liam McBirnie <>
Acked-by: Hannes Frederic Sowa <>
Signed-off-by: Paolo Abeni <>
Signed-off-by: David S. Miller <>
Signed-off-by: Jiri Slaby <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agosock: fix sendmmsg for partial sendmsg
Soheil Hassas Yeganeh [Fri, 4 Nov 2016 19:36:49 +0000 (15:36 -0400)]
sock: fix sendmmsg for partial sendmsg

[ Upstream commit 3023898b7d4aac65987bd2f485cc22390aae6f78 ]

Do not send the next message in sendmmsg for partial sendmsg

sendmmsg assumes that it can continue sending the next message
when the return value of the individual sendmsg invocations
is positive. It results in corrupting the data for TCP,
SCTP, and UNIX streams.

For example, sendmmsg([["abcd"], ["efgh"]]) can result in a stream
of "aefgh" if the first sendmsg invocation sends only the first
byte while the second sendmsg goes through.

Datagram sockets either send the entire datagram or fail, so
this patch affects only sockets of type SOCK_STREAM and

Fixes: 228e548e6020 ("net: Add sendmmsg socket system call")
Signed-off-by: Soheil Hassas Yeganeh <>
Signed-off-by: Eric Dumazet <>
Signed-off-by: Willem de Bruijn <>
Signed-off-by: Neal Cardwell <>
Acked-by: Maciej Żenczykowski <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: we don't have the iov_iter API, so make
 ___sys_sendmsg() calculate and write back the remaining length]
Signed-off-by: Ben Hutchings <>
6 years agosctp: assign assoc_id earlier in __sctp_connect
Marcelo Ricardo Leitner [Thu, 3 Nov 2016 19:03:41 +0000 (17:03 -0200)]
sctp: assign assoc_id earlier in __sctp_connect

[ Upstream commit 7233bc84a3aeda835d334499dc00448373caf5c0 ]

sctp_wait_for_connect() currently already holds the asoc to keep it
alive during the sleep, in case another thread release it. But Andrey
Konovalov and Dmitry Vyukov reported an use-after-free in such

Problem is that __sctp_connect() doesn't get a ref on the asoc and will
do a read on the asoc after calling sctp_wait_for_connect(), but by then
another thread may have closed it and the _put on sctp_wait_for_connect
will actually release it, causing the use-after-free.

Fix is, instead of doing the read after waiting for the connect, do it
before so, and avoid this issue as the socket is still locked by then.
There should be no issue on returning the asoc id in case of failure as
the application shouldn't trust on that number in such situations

This issue doesn't exist in sctp_sendmsg() path.

Reported-by: Dmitry Vyukov <>
Reported-by: Andrey Konovalov <>
Tested-by: Andrey Konovalov <>
Signed-off-by: Marcelo Ricardo Leitner <>
Reviewed-by: Xin Long <>
Acked-by: Neil Horman <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv6: dccp: fix out of bound access in dccp_v6_err()
Eric Dumazet [Thu, 3 Nov 2016 03:30:48 +0000 (20:30 -0700)]
ipv6: dccp: fix out of bound access in dccp_v6_err()

[ Upstream commit 1aa9d1a0e7eefcc61696e147d123453fc0016005 ]

dccp_v6_err() does not use pskb_may_pull() and might access garbage.

We only need 4 bytes at the beginning of the DCCP header, like TCP,
so the 8 bytes pulled in icmpv6_notify() are more than enough.

Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: use offsetof() + sizeof() instead of
Signed-off-by: Ben Hutchings <>
6 years agodccp: fix out of bound access in dccp_v4_err()
Eric Dumazet [Thu, 3 Nov 2016 02:00:40 +0000 (19:00 -0700)]
dccp: fix out of bound access in dccp_v4_err()

[ Upstream commit 6706a97fec963d6cb3f7fc2978ec1427b4651214 ]

dccp_v4_err() does not use pskb_may_pull() and might access garbage.

We only need 4 bytes at the beginning of the DCCP header, like TCP,
so the 8 bytes pulled in icmp_socket_deliver() are more than enough.

This patch might allow to process more ICMP messages, as some routers
are still limiting the size of reflected bytes to 28 (RFC 792), instead
of extended lengths (RFC 1812

Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: use offsetof() + sizeof() instead of
Signed-off-by: Ben Hutchings <>
6 years agodccp: do not send reset to already closed sockets
Eric Dumazet [Thu, 3 Nov 2016 01:04:24 +0000 (18:04 -0700)]
dccp: do not send reset to already closed sockets

[ Upstream commit 346da62cc186c4b4b1ac59f87f4482b47a047388 ]

Andrey reported following warning while fuzzing with syzkaller

WARNING: CPU: 1 PID: 21072 at net/dccp/proto.c:83 dccp_set_state+0x229/0x290
Kernel panic - not syncing: panic_on_warn set ...

CPU: 1 PID: 21072 Comm: syz-executor Not tainted 4.9.0-rc1+ #293
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
 ffff88003d4c7738 ffffffff81b474f4 0000000000000003 dffffc0000000000
 ffffffff844f8b00 ffff88003d4c7804 ffff88003d4c7800 ffffffff8140c06a
 0000000041b58ab3 ffffffff8479ab7d ffffffff8140beae ffffffff8140cd00
Call Trace:
 [<     inline     >] __dump_stack lib/dump_stack.c:15
 [<ffffffff81b474f4>] dump_stack+0xb3/0x10f lib/dump_stack.c:51
 [<ffffffff8140c06a>] panic+0x1bc/0x39d kernel/panic.c:179
 [<ffffffff8111125c>] __warn+0x1cc/0x1f0 kernel/panic.c:542
 [<ffffffff8111144c>] warn_slowpath_null+0x2c/0x40 kernel/panic.c:585
 [<ffffffff8389e5d9>] dccp_set_state+0x229/0x290 net/dccp/proto.c:83
 [<ffffffff838a0aa2>] dccp_close+0x612/0xc10 net/dccp/proto.c:1016
 [<ffffffff8316bf1f>] inet_release+0xef/0x1c0 net/ipv4/af_inet.c:415
 [<ffffffff82b6e89e>] sock_release+0x8e/0x1d0 net/socket.c:570
 [<ffffffff82b6e9f6>] sock_close+0x16/0x20 net/socket.c:1017
 [<ffffffff815256ad>] __fput+0x29d/0x720 fs/file_table.c:208
 [<ffffffff81525bb5>] ____fput+0x15/0x20 fs/file_table.c:244
 [<ffffffff811727d8>] task_work_run+0xf8/0x170 kernel/task_work.c:116
 [<     inline     >] exit_task_work include/linux/task_work.h:21
 [<ffffffff8111bc53>] do_exit+0x883/0x2ac0 kernel/exit.c:828
 [<ffffffff811221fe>] do_group_exit+0x10e/0x340 kernel/exit.c:931
 [<ffffffff81143c94>] get_signal+0x634/0x15a0 kernel/signal.c:2307
 [<ffffffff81054aad>] do_signal+0x8d/0x1a30 arch/x86/kernel/signal.c:807
 [<ffffffff81003a05>] exit_to_usermode_loop+0xe5/0x130
 [<     inline     >] prepare_exit_to_usermode arch/x86/entry/common.c:190
 [<ffffffff81006298>] syscall_return_slowpath+0x1a8/0x1e0
 [<ffffffff83fc1a62>] entry_SYSCALL_64_fastpath+0xc0/0xc2
Dumping ftrace buffer:
   (ftrace buffer empty)
Kernel Offset: disabled

Fix this the same way we did for TCP in commit 565b7b2d2e63
("tcp: do not send reset to already closed sockets")

Signed-off-by: Eric Dumazet <>
Reported-by: Andrey Konovalov <>
Tested-by: Andrey Konovalov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: mangle zero checksum in skb_checksum_help()
Eric Dumazet [Sat, 29 Oct 2016 18:02:36 +0000 (11:02 -0700)]
net: mangle zero checksum in skb_checksum_help()

[ Upstream commit 4f2e4ad56a65f3b7d64c258e373cb71e8d2499f4 ]

Sending zero checksum is ok for TCP, but not for UDP.

UDPv6 receiver should by default drop a frame with a 0 checksum,
and UDPv4 would not verify the checksum and might accept a corrupted

Simply replace such checksum by 0xffff, regardless of transport.

This error was caught on SIT tunnels, but seems generic.

Signed-off-by: Eric Dumazet <>
Cc: Maciej Żenczykowski <>
Cc: Willem de Bruijn <>
Acked-by: Maciej Żenczykowski <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: clear sk_err_soft in sk_clone_lock()
Eric Dumazet [Fri, 28 Oct 2016 20:40:24 +0000 (13:40 -0700)]
net: clear sk_err_soft in sk_clone_lock()

[ Upstream commit e551c32d57c88923f99f8f010e89ca7ed0735e83 ]

At accept() time, it is possible the parent has a non zero
sk_err_soft, leftover from a prior error.

Make sure we do not leave this value in the child, as it
makes future getsockopt(SO_ERROR) calls quite unreliable.

Signed-off-by: Eric Dumazet <>
Acked-by: Soheil Hassas Yeganeh <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: sctp, forbid negative length
Jiri Slaby [Fri, 21 Oct 2016 12:13:24 +0000 (14:13 +0200)]
net: sctp, forbid negative length

[ Upstream commit a4b8e71b05c27bae6bad3bdecddbc6b68a3ad8cf ]

Most of getsockopt handlers in net/sctp/socket.c check len against
sizeof some structure like:
        if (len < sizeof(int))
                return -EINVAL;

On the first look, the check seems to be correct. But since len is int
and sizeof returns size_t, int gets promoted to unsigned size_t too. So
the test returns false for negative lengths. Yes, (-1 < sizeof(long)) is

Fix this in sctp by explicitly checking len < 0 before any getsockopt
handler is called.

Note that sctp_getsockopt_events already handled the negative case.
Since we added the < 0 check elsewhere, this one can be removed.

If not checked, this is the result:
UBSAN: Undefined behaviour in ../mm/page_alloc.c:2722:19
shift exponent 52 is too large for 32-bit type 'int'
CPU: 1 PID: 24535 Comm: syz-executor Not tainted 4.8.1-0-syzkaller #1
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 04/01/2014
 0000000000000000 ffff88006d99f2a8 ffffffffb2f7bdea 0000000041b58ab3
 ffffffffb4363c14 ffffffffb2f7bcde ffff88006d99f2d0 ffff88006d99f270
 0000000000000000 0000000000000000 0000000000000034 ffffffffb5096422
Call Trace:
 [<ffffffffb3051498>] ? __ubsan_handle_shift_out_of_bounds+0x29c/0x300
 [<ffffffffb273f0e4>] ? kmalloc_order+0x24/0x90
 [<ffffffffb27416a4>] ? kmalloc_order_trace+0x24/0x220
 [<ffffffffb2819a30>] ? __kmalloc+0x330/0x540
 [<ffffffffc18c25f4>] ? sctp_getsockopt_local_addrs+0x174/0xca0 [sctp]
 [<ffffffffc18d2bcd>] ? sctp_getsockopt+0x10d/0x1b0 [sctp]
 [<ffffffffb37c1219>] ? sock_common_getsockopt+0xb9/0x150
 [<ffffffffb37be2f5>] ? SyS_getsockopt+0x1a5/0x270

Signed-off-by: Jiri Slaby <>
Cc: Vlad Yasevich <>
Cc: Neil Horman <>
Cc: "David S. Miller" <>
Acked-by: Neil Horman <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agotcp: fix wrong checksum calculation on MTU probing
Douglas Caetano dos Santos [Thu, 22 Sep 2016 18:52:04 +0000 (15:52 -0300)]
tcp: fix wrong checksum calculation on MTU probing

[ Upstream commit 2fe664f1fcf7c4da6891f95708a7a56d3c024354 ]

With TCP MTU probing enabled and offload TX checksumming disabled,
tcp_mtu_probe() calculated the wrong checksum when a fragment being copied
into the probe's SKB had an odd length. This was caused by the direct use
of skb_copy_and_csum_bits() to calculate the checksum, as it pads the
fragment being copied, if needed. When this fragment was not the last, a
subsequent call used the previous checksum without considering this

The effect was a stale connection in one way, as even retransmissions
wouldn't solve the problem, because the checksum was never recalculated for
the full SKB length.

Signed-off-by: Douglas Caetano dos Santos <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: avoid sk_forward_alloc overflows
Eric Dumazet [Thu, 15 Sep 2016 15:48:46 +0000 (08:48 -0700)]
net: avoid sk_forward_alloc overflows

[ Upstream commit 20c64d5cd5a2bdcdc8982a06cb05e5e1bd851a3d ]

A malicious TCP receiver, sending SACK, can force the sender to split
skbs in write queue and increase its memory usage.

Then, when socket is closed and its write queue purged, we might
overflow sk_forward_alloc (It becomes negative)

sk_mem_reclaim() does nothing in this case, and more than 2GB
are leaked from TCP perspective (tcp_memory_allocated is not changed)

Then warnings trigger from inet_sock_destruct() and
sk_stream_kill_queues() seeing a not zero sk_forward_alloc

All TCP stack can be stuck because TCP is under memory pressure.

A simple fix is to preemptively reclaim from sk_mem_uncharge().

This makes sure a socket wont have more than 2 MB forward allocated,
after burst and idle period.

Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agotcp: fix overflow in __tcp_retransmit_skb()
Eric Dumazet [Thu, 15 Sep 2016 15:12:33 +0000 (08:12 -0700)]
tcp: fix overflow in __tcp_retransmit_skb()

[ Upstream commit ffb4d6c8508657824bcef68a36b2a0f9d8c09d10 ]

If a TCP socket gets a large write queue, an overflow can happen
in a test in __tcp_retransmit_skb() preventing all retransmits.

The flow then stalls and resets after timeouts.


sysctl -w net.core.wmem_max=1000000000
netperf -H dest -- -s 1000000000

Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: fix sk_mem_reclaim_partial()
Eric Dumazet [Fri, 15 May 2015 19:39:25 +0000 (12:39 -0700)]
net: fix sk_mem_reclaim_partial()

commit 1a24e04e4b50939daa3041682b38b82c896ca438 upstream.

sk_mem_reclaim_partial() goal is to ensure each socket has
one SK_MEM_QUANTUM forward allocation. This is needed both for
performance and better handling of memory pressure situations in
follow up patches.

SK_MEM_QUANTUM is currently a page, but might be reduced to 4096 bytes
as some arches have 64KB pages.

Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Keep using atomic_long_sub() directly, not sk_memory_allocated_sub()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoipmr/ip6mr: Initialize the last assert time of mfc entries.
Tom Goff [Thu, 23 Jun 2016 20:11:57 +0000 (16:11 -0400)]
ipmr/ip6mr: Initialize the last assert time of mfc entries.

[ Upstream commit 70a0dec45174c976c64b4c8c1d0898581f759948 ]

This fixes wrong-interface signaling on 32-bit platforms for entries
created when jiffies > 2^31 + MFC_ASSERT_THRESH.

Signed-off-by: Tom Goff <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: bridge: fix old ioctl unlocked net device walk
Nikolay Aleksandrov [Wed, 4 May 2016 14:18:45 +0000 (16:18 +0200)]
net: bridge: fix old ioctl unlocked net device walk

[ Upstream commit 31ca0458a61a502adb7ed192bf9716c6d05791a5 ]

get_bridge_ifindices() is used from the old "deviceless" bridge ioctl
calls which aren't called with rtnl held. The comment above says that it is
called with rtnl but that is not really the case.
Here's a sample output from a test ASSERT_RTNL() which I put in
get_bridge_ifindices and executed "brctl show":
[  957.422726] RTNL: assertion failed at net/bridge//br_ioctl.c (30)
[  957.422925] CPU: 0 PID: 1862 Comm: brctl Tainted: G        W  O
4.6.0-rc4+ #157
[  957.423009] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996),
BIOS 1.8.1-20150318_183358- 04/01/2014
[  957.423009]  0000000000000000 ffff880058adfdf0 ffffffff8138dec5
[  957.423009]  ffffffff81ce8380 ffff880058adfe58 ffffffffa05ead32
[  957.423009]  00007ffec1a444b0 0000000000000400 ffff880053c19130
[  957.423009] Call Trace:
[  957.423009]  [<ffffffff8138dec5>] dump_stack+0x85/0xc0
[  957.423009]  [<ffffffffa05ead32>]
br_ioctl_deviceless_stub+0x212/0x2e0 [bridge]
[  957.423009]  [<ffffffff81515beb>] sock_ioctl+0x22b/0x290
[  957.423009]  [<ffffffff8126ba75>] do_vfs_ioctl+0x95/0x700
[  957.423009]  [<ffffffff8126c159>] SyS_ioctl+0x79/0x90
[  957.423009]  [<ffffffff8163a4c0>] entry_SYSCALL_64_fastpath+0x23/0xc1

Since it only reads bridge ifindices, we can use rcu to safely walk the net
device list. Also remove the wrong rtnl comment above.

Signed-off-by: Nikolay Aleksandrov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agosch_dsmark: update backlog as well
WANG Cong [Thu, 25 Feb 2016 22:55:03 +0000 (14:55 -0800)]
sch_dsmark: update backlog as well

[ Upstream commit bdf17661f63a79c3cb4209b970b1cc39e34f7543 ]

Similarly, we need to update backlog too when we update qlen.

Cc: Jamal Hadi Salim <>
Signed-off-by: Cong Wang <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: open-code qdisc_qstats_backlog_{inc,dec}()]
Signed-off-by: Ben Hutchings <>
6 years agosch_htb: update backlog as well
WANG Cong [Thu, 25 Feb 2016 22:55:02 +0000 (14:55 -0800)]
sch_htb: update backlog as well

[ Upstream commit 431e3a8e36a05a37126f34b41aa3a5a6456af04e ]

We saw qlen!=0 but backlog==0 on our production machine:

qdisc htb 1: dev eth0 root refcnt 2 r2q 10 default 1 direct_packets_stat 0 ver 3.17
 Sent 172680457356 bytes 222469449 pkt (dropped 0, overlimits 123575834 requeues 0)
 backlog 0b 72p requeues 0

The problem is we only count qlen for HTB qdisc but not backlog.
We need to update backlog too when we update qlen, so that we
can at least know the average packet length.

Cc: Jamal Hadi Salim <>
Acked-by: Jamal Hadi Salim <>
Signed-off-by: Cong Wang <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Open-code qdisc_qstats_backlog_{inc,dec}()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoroute: do not cache fib route info on local routes with oif
Chris Friesen [Fri, 8 Apr 2016 21:21:30 +0000 (15:21 -0600)]
route: do not cache fib route info on local routes with oif

[ Upstream commit d6d5e999e5df67f8ec20b6be45e2229455ee3699 ]

For local routes that require a particular output interface we do not want
to cache the result.  Caching the result causes incorrect behaviour when
there are multiple source addresses on the interface.  The end result
being that if the intended recipient is waiting on that interface for the
packet he won't receive it because it will be delivered on the loopback
interface and the IP_PKTINFO ipi_ifindex will be set to the loopback
interface as well.

This can be tested by running a program such as "dhcp_release" which
attempts to inject a packet on a particular interface so that it is
received by another program on the same board.  The receiving process
should see an IP_PKTINFO ipi_ifndex value of the source interface
(e.g., eth1) instead of the loopback interface (e.g., lo).  The packet
will still appear on the loopback interface in tcpdump but the important
aspect is that the CMSG info is correct.

Sample dhcp_release command line:

   dhcp_release eth1 02:11:33:22:44:66

Signed-off-by: Allain Legacy <>
Signed off-by: Chris Friesen <>
Reviewed-by: Julian Anastasov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agodecnet: Do not build routes to devices without decnet private data.
David S. Miller [Mon, 11 Apr 2016 03:01:30 +0000 (23:01 -0400)]
decnet: Do not build routes to devices without decnet private data.

[ Upstream commit a36a0d4008488fa545c74445d69eaf56377d5d4e ]

In particular, make sure we check for decnet private presence
for loopback devices.

Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agolib/vsprintf.c: improve sanity check in vsnprintf()
Rasmus Villemoes [Thu, 12 Feb 2015 23:01:39 +0000 (15:01 -0800)]
lib/vsprintf.c: improve sanity check in vsnprintf()

commit 2aa2f9e21e4eb25c720b2e7d80f8929638f6ad73 upstream.

On 64 bit, size may very well be huge even if bit 31 happens to be 0.
Somehow it doesn't feel right that one can pass a 5 GiB buffer but not a
3 GiB one.  So cap at INT_MAX as was probably the intention all along.
This is also the made-up value passed by sprintf and vsprintf.

Signed-off-by: Rasmus Villemoes <>
Cc: Jiri Kosina <>
Cc: Randy Dunlap <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
Cc: Willy Tarreau <>
6 years agoirda: Fix lockdep annotations in hashbin_delete().
David S. Miller [Fri, 17 Feb 2017 21:19:39 +0000 (16:19 -0500)]
irda: Fix lockdep annotations in hashbin_delete().

commit 4c03b862b12f980456f9de92db6d508a4999b788 upstream.

A nested lock depth was added to the hasbin_delete() code but it
doesn't actually work some well and results in tons of lockdep splats.

Fix the code instead to properly drop the lock around the operation
and just keep peeking the head of the hashbin queue.

Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoFix missing sanity check in /dev/sg
Al Viro [Sun, 19 Feb 2017 07:15:27 +0000 (07:15 +0000)]
Fix missing sanity check in /dev/sg

commit 137d01df511b3afe1f05499aea05f3bafc0fb221 upstream.

What happens is that a write to /dev/sg is given a request with non-zero
->iovec_count combined with zero ->dxfer_len.  Or with ->dxferp pointing
to an array full of empty iovecs.

Having write permission to /dev/sg shouldn't be equivalent to the
ability to trigger BUG_ON() while holding spinlocks...

Found by Dmitry Vyukov and syzkaller.

[ The BUG_ON() got changed to a WARN_ON_ONCE(), but this fixes the
  underlying issue.  - Linus ]

Signed-off-by: Al Viro <>
Reported-by: Dmitry Vyukov <>
Reviewed-by: Christoph Hellwig <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: we're not using iov_iter, but can check the
 byte length after truncation]
Signed-off-by: Ben Hutchings <>
6 years agopacket: Do not call fanout_release from atomic contexts
Anoob Soman [Wed, 15 Feb 2017 20:25:39 +0000 (20:25 +0000)]
packet: Do not call fanout_release from atomic contexts

commit 2bd624b4611ffee36422782d16e1c944d1351e98 upstream.

Commit 6664498280cf ("packet: call fanout_release, while UNREGISTERING a
netdev"), unfortunately, introduced the following issues.

1. calling mutex_lock(&fanout_mutex) (fanout_release()) from inside
rcu_read-side critical section. rcu_read_lock disables preemption, most often,
which prohibits calling sleeping functions.

[  ] include/linux/rcupdate.h:560 Illegal context switch in RCU read-side critical section!
[  ]
[  ] rcu_scheduler_active = 1, debug_locks = 0
[  ] 4 locks held by ovs-vswitchd/1969:
[  ]  #0:  (cb_lock){++++++}, at: [<ffffffff8158a6c9>] genl_rcv+0x19/0x40
[  ]  #1:  (ovs_mutex){+.+.+.}, at: [<ffffffffa04878ca>] ovs_vport_cmd_del+0x4a/0x100 [openvswitch]
[  ]  #2:  (rtnl_mutex){+.+.+.}, at: [<ffffffff81564157>] rtnl_lock+0x17/0x20
[  ]  #3:  (rcu_read_lock){......}, at: [<ffffffff81614165>] packet_notifier+0x5/0x3f0
[  ]
[  ] Call Trace:
[  ]  [<ffffffff813770c1>] dump_stack+0x85/0xc4
[  ]  [<ffffffff810c9077>] lockdep_rcu_suspicious+0x107/0x110
[  ]  [<ffffffff810a2da7>] ___might_sleep+0x57/0x210
[  ]  [<ffffffff810a2fd0>] __might_sleep+0x70/0x90
[  ]  [<ffffffff8162e80c>] mutex_lock_nested+0x3c/0x3a0
[  ]  [<ffffffff810de93f>] ? vprintk_default+0x1f/0x30
[  ]  [<ffffffff81186e88>] ? printk+0x4d/0x4f
[  ]  [<ffffffff816106dd>] fanout_release+0x1d/0xe0
[  ]  [<ffffffff81614459>] packet_notifier+0x2f9/0x3f0

2. calling mutex_lock(&fanout_mutex) inside spin_lock(&po->bind_lock).
"sleeping function called from invalid context"

[  ] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:620
[  ] in_atomic(): 1, irqs_disabled(): 0, pid: 1969, name: ovs-vswitchd
[  ] INFO: lockdep is turned off.
[  ] Call Trace:
[  ]  [<ffffffff813770c1>] dump_stack+0x85/0xc4
[  ]  [<ffffffff810a2f52>] ___might_sleep+0x202/0x210
[  ]  [<ffffffff810a2fd0>] __might_sleep+0x70/0x90
[  ]  [<ffffffff8162e80c>] mutex_lock_nested+0x3c/0x3a0
[  ]  [<ffffffff816106dd>] fanout_release+0x1d/0xe0
[  ]  [<ffffffff81614459>] packet_notifier+0x2f9/0x3f0

3. calling dev_remove_pack(&fanout->prot_hook), from inside
spin_lock(&po->bind_lock) or rcu_read-side critical-section. dev_remove_pack()
-> synchronize_net(), which might sleep.

[  ] BUG: scheduling while atomic: ovs-vswitchd/1969/0x00000002
[  ] INFO: lockdep is turned off.
[  ] Call Trace:
[  ]  [<ffffffff813770c1>] dump_stack+0x85/0xc4
[  ]  [<ffffffff81186274>] __schedule_bug+0x64/0x73
[  ]  [<ffffffff8162b8cb>] __schedule+0x6b/0xd10
[  ]  [<ffffffff8162c5db>] schedule+0x6b/0x80
[  ]  [<ffffffff81630b1d>] schedule_timeout+0x38d/0x410
[  ]  [<ffffffff810ea3fd>] synchronize_sched_expedited+0x53d/0x810
[  ]  [<ffffffff810ea6de>] synchronize_rcu_expedited+0xe/0x10
[  ]  [<ffffffff8154eab5>] synchronize_net+0x35/0x50
[  ]  [<ffffffff8154eae3>] dev_remove_pack+0x13/0x20
[  ]  [<ffffffff8161077e>] fanout_release+0xbe/0xe0
[  ]  [<ffffffff81614459>] packet_notifier+0x2f9/0x3f0

4. fanout_release() races with calls from different CPU.

To fix the above problems, remove the call to fanout_release() under
rcu_read_lock(). Instead, call __dev_remove_pack(&fanout->prot_hook) and
netdev_run_todo will be happy that &dev->ptype_specific list is empty. In order
to achieve this, I moved dev_{add,remove}_pack() out of fanout_{add,release} to
__fanout_{link,unlink}. So, call to {,__}unregister_prot_hook() will make sure
fanout->prot_hook is removed as well.

Fixes: 6664498280cf ("packet: call fanout_release, while UNREGISTERING a netdev")
Reported-by: Eric Dumazet <>
Signed-off-by: Anoob Soman <>
Acked-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Don't call fanout_release_data()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agopacket: call fanout_release, while UNREGISTERING a netdev
Anoob Soman [Wed, 5 Oct 2016 14:12:54 +0000 (15:12 +0100)]
packet: call fanout_release, while UNREGISTERING a netdev

commit 6664498280cf17a59c3e7cf1a931444c02633ed1 upstream.

If a socket has FANOUT sockopt set, a new proto_hook is registered
as part of fanout_add(). When processing a NETDEV_UNREGISTER event in
af_packet, __fanout_unlink is called for all sockets, but prot_hook which was
registered as part of fanout_add is not removed. Call fanout_release, on a
NETDEV_UNREGISTER, which removes prot_hook and removes fanout from the

This fixes BUG_ON(!list_empty(&dev->ptype_specific)) in netdev_run_todo()

Signed-off-by: Anoob Soman <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agovfs: fix uninitialized flags in splice_to_pipe()
Miklos Szeredi [Thu, 16 Feb 2017 16:49:02 +0000 (17:49 +0100)]
vfs: fix uninitialized flags in splice_to_pipe()

commit 5a81e6a171cdbd1fa8bc1fdd80c23d3d71816fac upstream.

Flags (PIPE_BUF_FLAG_PACKET, PIPE_BUF_FLAG_GIFT) could remain on the
unused part of the pipe ring buffer.  Previously splice_to_pipe() left
the flags value alone, which could result in incorrect behavior.

Uninitialized flags appears to have been there from the introduction of
the splice syscall.

Signed-off-by: Miklos Szeredi <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: adjust context, indentation]
Signed-off-by: Ben Hutchings <>
6 years agonet: xilinx_emaclite: fix receive buffer overflow
Anssi Hannula [Tue, 14 Feb 2017 17:11:44 +0000 (19:11 +0200)]
net: xilinx_emaclite: fix receive buffer overflow

commit cd224553641848dd17800fe559e4ff5d208553e8 upstream.

xilinx_emaclite looks at the received data to try to determine the
Ethernet packet length but does not properly clamp it if
proto_type == ETH_P_IP or 1500 < proto_type <= 1518, causing a buffer
overflow and a panic via skb_panic() as the length exceeds the allocated
skb size.

Fix those cases.

Also add an additional unconditional check with WARN_ON() at the end.

Signed-off-by: Anssi Hannula <>
Fixes: bb81b2ddfa19 ("net: add Xilinx emac lite device driver")
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agosiano: make it work again with CONFIG_VMAP_STACK
Mauro Carvalho Chehab [Tue, 14 Feb 2017 19:47:57 +0000 (17:47 -0200)]
siano: make it work again with CONFIG_VMAP_STACK

commit f9c85ee67164b37f9296eab3b754e543e4e96a1c upstream.

Reported as a Kaffeine bug:

The USB control messages require DMA to work. We cannot pass
a stack-allocated buffer, as it is not warranted that the
stack would be into a DMA enabled area.

On Kernel 4.9, the default is to not accept DMA on stack anymore
on x86 architecture. On other architectures, this has been a
requirement since Kernel 2.2. So, after this patch, this driver
should likely work fine on all archs.

Tested with USB ID 2040:5510: Hauppauge Windham

Signed-off-by: Mauro Carvalho Chehab <>
[bwh: Backported to 3.2:
 - s/sms_msg_hdr/SmsMsgHdr_ST/
 - Adjust filename, context]
Signed-off-by: Ben Hutchings <>
6 years agopacket: fix races in fanout_add()
Eric Dumazet [Tue, 14 Feb 2017 17:03:51 +0000 (09:03 -0800)]
packet: fix races in fanout_add()

commit d199fab63c11998a602205f7ee7ff7c05c97164b upstream.

Multiple threads can call fanout_add() at the same time.

We need to grab fanout_mutex earlier to avoid races that could
lead to one thread freeing po->rollover that was set by another thread.

Do the same in fanout_release(), for peace of mind, and to help us
finding lockdep issues earlier.

Fixes: dc99f600698d ("packet: Add fanout support.")
Fixes: 0648ab70afe6 ("packet: rollover prepare: per-socket state")
Signed-off-by: Eric Dumazet <>
Cc: Willem de Bruijn <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - No rollover queue stats
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agofutex: Move futex_init() to core_initcall
Yang Yang [Fri, 30 Dec 2016 08:17:55 +0000 (16:17 +0800)]
futex: Move futex_init() to core_initcall

commit 25f71d1c3e98ef0e52371746220d66458eac75bc upstream.

The UEVENT user mode helper is enabled before the initcalls are executed
and is available when the root filesystem has been mounted.

The user mode helper is triggered by device init calls and the executable
might use the futex syscall.

futex_init() is marked __initcall which maps to device_initcall, but there
is no guarantee that futex_init() is invoked _before_ the first device init
call which triggers the UEVENT user mode helper.

If the user mode helper uses the futex syscall before futex_init() then the
syscall crashes with a NULL pointer dereference because the futex subsystem
has not been initialized yet.

Move futex_init() to core_initcall so futexes are initialized before the
root filesystem is mounted and the usermode helper becomes available.

[ tglx: Rewrote changelog ]

Signed-off-by: Yang Yang <>
Cc: Peter Zijlstra <>
Signed-off-by: Thomas Gleixner <>
Signed-off-by: Ben Hutchings <>
6 years agonet/llc: avoid BUG_ON() in skb_orphan()
Eric Dumazet [Sun, 12 Feb 2017 22:03:52 +0000 (14:03 -0800)]
net/llc: avoid BUG_ON() in skb_orphan()

commit 8b74d439e1697110c5e5c600643e823eb1dd0762 upstream.

It seems nobody used LLC since linux-3.12.

Fortunately fuzzers like syzkaller still know how to run this code,
otherwise it would be no fun.

Setting skb->sk without skb->destructor leads to all kinds of
bugs, we now prefer to be very strict about it.

Ideally here we would use skb_set_owner() but this helper does not exist yet,
only CAN seems to have a private helper for that.

Fixes: 376c7311bdb6 ("net: add a temporary sanity check in skb_orphan()")
Signed-off-by: Eric Dumazet <>
Reported-by: Andrey Konovalov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet/sock: Add sock_efree() function
Ben Hutchings [Fri, 3 Mar 2017 02:32:07 +0000 (02:32 +0000)]
net/sock: Add sock_efree() function

Extracted from commit 62bccb8cdb69 ("net-timestamp: Make the clone operation
stand-alone from phy timestamping").

Signed-off-by: Ben Hutchings <>
6 years agol2tp: do not use udp_ioctl()
Eric Dumazet [Fri, 10 Feb 2017 00:15:52 +0000 (16:15 -0800)]
l2tp: do not use udp_ioctl()

commit 72fb96e7bdbbdd4421b0726992496531060f3636 upstream.

udp_ioctl(), as its name suggests, is used by UDP protocols,
but is also used by L2TP :(

L2TP should use its own handler, because it really does not
look the same.

SIOCINQ for instance should not assume UDP checksum or headers.

Thanks to Andrey and syzkaller team for providing the report
and a nice reproducer.

While crashes only happen on recent kernels (after commit
7c13f97ffde6 ("udp: do fwd memory scheduling on dequeue")), this
probably needs to be backported to older kernels.

Fixes: 7c13f97ffde6 ("udp: do fwd memory scheduling on dequeue")
Fixes: 85584672012e ("udp: Fix udp_poll() and ioctl()")
Signed-off-by: Eric Dumazet <>
Reported-by: Andrey Konovalov <>
Acked-by: Paolo Abeni <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Drop the IPv6 change
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoxen-netfront: Delete rx_refill_timer in xennet_disconnect_backend()
Boris Ostrovsky [Mon, 30 Jan 2017 17:45:46 +0000 (12:45 -0500)]
xen-netfront: Delete rx_refill_timer in xennet_disconnect_backend()

commit 74470954857c264168d2b5a113904cf0cfd27d18 upstream.

rx_refill_timer should be deleted as soon as we disconnect from the
backend since otherwise it is possible for the timer to go off before
we get to xennet_destroy_queues(). If this happens we may dereference
queue->rx.sring which is set to NULL in xennet_disconnect_backend().

Signed-off-by: Boris Ostrovsky <>
Reviewed-by: Juergen Gross <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: there's only one RX queue here, and del_timer_sync()
 was called from xennet_remove() but that's also too late]
Signed-off-by: Ben Hutchings <>
6 years agoscsi: zfcp: fix use-after-free by not tracing WKA port open/close on failed send
Steffen Maier [Wed, 8 Feb 2017 14:34:22 +0000 (15:34 +0100)]
scsi: zfcp: fix use-after-free by not tracing WKA port open/close on failed send

commit 2dfa6688aafdc3f74efeb1cf05fb871465d67f79 upstream.

Dan Carpenter kindly reported:
The patch d27a7cb91960: "zfcp: trace on request for open and close of
WKA port" from Aug 10, 2016, leads to the following static checker

drivers/s390/scsi/zfcp_fsf.c:1615 zfcp_fsf_open_wka_port()
warn: 'req' was already freed.

  1609          zfcp_fsf_start_timer(req, ZFCP_FSF_REQUEST_TIMEOUT);
  1610          retval = zfcp_fsf_req_send(req);
  1611          if (retval)
  1612                  zfcp_fsf_req_free(req);

  1613  out:
  1614          spin_unlock_irq(&qdio->req_q_lock);
  1615          if (req && !IS_ERR(req))
  1616                  zfcp_dbf_rec_run_wka("fsowp_1", wka_port, req->req_id);
Use after free.

  1617          return retval;
  1618  }

Same thing for zfcp_fsf_close_wka_port() as well.

Rather than relying on req being NULL (or ERR_PTR) for all cases where
we don't want to trace or should not trace,
simply check retval which is unconditionally initialized with -EIO != 0
and it can only become 0 on successful retval = zfcp_fsf_req_send(req).
With that we can also remove the then again unnecessary unconditional
initialization of req which was introduced with that earlier commit.

Reported-by: Dan Carpenter <>
Suggested-by: Benjamin Block <>
Signed-off-by: Steffen Maier <>
Fixes: d27a7cb91960 ("zfcp: trace on request for open and close of WKA port")
Reviewed-by: Benjamin Block <>
Reviewed-by: Jens Remus <>
Signed-off-by: Martin K. Petersen <>
Signed-off-by: Ben Hutchings <>
6 years agoping: fix a null pointer dereference
WANG Cong [Tue, 7 Feb 2017 20:59:46 +0000 (12:59 -0800)]
ping: fix a null pointer dereference

commit 73d2c6678e6c3af7e7a42b1e78cd0211782ade32 upstream.

Andrey reported a kernel crash:

  general protection fault: 0000 [#1] SMP KASAN
  Dumping ftrace buffer:
     (ftrace buffer empty)
  Modules linked in:
  CPU: 2 PID: 3880 Comm: syz-executor1 Not tainted 4.10.0-rc6+ #124
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
  task: ffff880060048040 task.stack: ffff880069be8000
  RIP: 0010:ping_v4_push_pending_frames net/ipv4/ping.c:647 [inline]
  RIP: 0010:ping_v4_sendmsg+0x1acd/0x23f0 net/ipv4/ping.c:837
  RSP: 0018:ffff880069bef8b8 EFLAGS: 00010206
  RAX: dffffc0000000000 RBX: ffff880069befb90 RCX: 0000000000000000
  RDX: 0000000000000018 RSI: ffff880069befa30 RDI: 00000000000000c2
  RBP: ffff880069befbb8 R08: 0000000000000008 R09: 0000000000000000
  R10: 0000000000000002 R11: 0000000000000000 R12: ffff880069befab0
  R13: ffff88006c624a80 R14: ffff880069befa70 R15: 0000000000000000
  FS:  00007f6f7c716700(0000) GS:ffff88006de00000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00000000004a6f28 CR3: 000000003a134000 CR4: 00000000000006e0
  Call Trace:
   inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744
   sock_sendmsg_nosec net/socket.c:635 [inline]
   sock_sendmsg+0xca/0x110 net/socket.c:645
   SYSC_sendto+0x660/0x810 net/socket.c:1687
   SyS_sendto+0x40/0x50 net/socket.c:1655

This is because we miss a check for NULL pointer for skb_peek() when
the queue is empty. Other places already have the same check.

Fixes: c319b4d76b9e ("net: ipv4: add IPPROTO_ICMP socket kind")
Reported-by: Andrey Konovalov <>
Tested-by: Andrey Konovalov <>
Signed-off-by: Cong Wang <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agobtrfs: fix btrfs_compat_ioctl failures on non-compat ioctls
Jeff Mahoney [Tue, 7 Feb 2017 00:39:09 +0000 (19:39 -0500)]
btrfs: fix btrfs_compat_ioctl failures on non-compat ioctls

commit 2a362249187a8d0f6d942d6e1d763d150a296f47 upstream.

Commit 4c63c2454ef incorrectly assumed that returning -ENOIOCTLCMD would
cause the native ioctl to be called.  The ->compat_ioctl callback is
expected to handle all ioctls, not just compat variants.  As a result,
when using 32-bit userspace on 64-bit kernels, everything except those
three ioctls would return -ENOTTY.

Fixes: 4c63c2454ef ("btrfs: bugfix: handle FS_IOC32_{GETFLAGS,SETFLAGS,GETVERSION} in btrfs_ioctl")
Signed-off-by: Jeff Mahoney <>
Reviewed-by: David Sterba <>
Signed-off-by: David Sterba <>
Signed-off-by: Ben Hutchings <>
6 years agoALSA: seq: Fix race at creating a queue
Takashi Iwai [Wed, 8 Feb 2017 11:35:39 +0000 (12:35 +0100)]
ALSA: seq: Fix race at creating a queue

commit 4842e98f26dd80be3623c4714a244ba52ea096a8 upstream.

When a sequencer queue is created in snd_seq_queue_alloc(),it adds the
new queue element to the public list before referencing it.  Thus the
queue might be deleted before the call of snd_seq_queue_use(), and it
results in the use-after-free error, as spotted by syzkaller.

The fix is to reference the queue object at the right time.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
6 years agocatc: Use heap buffer for memory size test
Ben Hutchings [Sat, 4 Feb 2017 16:57:04 +0000 (16:57 +0000)]
catc: Use heap buffer for memory size test

commit 2d6a0e9de03ee658a9adc3bfb2f0ca55dff1e478 upstream.

Allocating USB buffers on the stack is not portable, and no longer
works on x86_64 (with VMAP_STACK enabled as per default).

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Ben Hutchings <>
Signed-off-by: David S. Miller <>
6 years agocatc: Combine failure cleanup code in catc_probe()
Ben Hutchings [Sat, 4 Feb 2017 16:56:56 +0000 (16:56 +0000)]
catc: Combine failure cleanup code in catc_probe()

commit d41149145f98fe26dcd0bfd1d6cc095e6e041418 upstream.

Signed-off-by: Ben Hutchings <>
Signed-off-by: David S. Miller <>
6 years agortl8150: Use heap buffers for all register access
Ben Hutchings [Sat, 4 Feb 2017 16:56:32 +0000 (16:56 +0000)]
rtl8150: Use heap buffers for all register access

commit 7926aff5c57b577ab0f43364ff0c59d968f6a414 upstream.

Allocating USB buffers on the stack is not portable, and no longer
works on x86_64 (with VMAP_STACK enabled as per default).

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Ben Hutchings <>
Signed-off-by: David S. Miller <>
6 years agoALSA: seq: Don't handle loop timeout at snd_seq_pool_done()
Takashi Iwai [Mon, 6 Feb 2017 14:09:48 +0000 (15:09 +0100)]
ALSA: seq: Don't handle loop timeout at snd_seq_pool_done()

commit 37a7ea4a9b81f6a864c10a7cb0b96458df5310a3 upstream.

snd_seq_pool_done() syncs with closing of all opened threads, but it
aborts the wait loop with a timeout, and proceeds to the release
resource even if not all threads have been closed.  The timeout was 5
seconds, and if you run a crazy stuff, it can exceed easily, and may
result in the access of the invalid memory address -- this is what
syzkaller detected in a bug report.

As a fix, let the code graduate from naiveness, simply remove the loop

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: deleted log statement is slightly different]
Signed-off-by: Ben Hutchings <>
6 years agomac80211: Fix adding of mesh vendor IEs
Thorsten Horstmann [Fri, 3 Feb 2017 13:38:29 +0000 (14:38 +0100)]
mac80211: Fix adding of mesh vendor IEs

commit da7061c82e4a1bc6a5e134ef362c86261906c860 upstream.

The function ieee80211_ie_split_vendor doesn't return 0 on errors. Instead
it returns any offset < ielen when WLAN_EID_VENDOR_SPECIFIC is found. The
return value in mesh_add_vendor_ies must therefore be checked against
ifmsh->ie_len and not 0. Otherwise all ifmsh->ie starting with
WLAN_EID_VENDOR_SPECIFIC will be rejected.

Fixes: 082ebb0c258d ("mac80211: fix mesh beacon format")
Signed-off-by: Thorsten Horstmann <>
Signed-off-by: Mathias Kretschmer <>
Signed-off-by: Simon Wunderlich <>
[ Add commit message]
Signed-off-by: Sven Eckelmann <>
Signed-off-by: Johannes Berg <>
Signed-off-by: Ben Hutchings <>
6 years agonetlabel: out of bound access in cipso_v4_validate()
Eric Dumazet [Fri, 3 Feb 2017 08:03:26 +0000 (00:03 -0800)]
netlabel: out of bound access in cipso_v4_validate()

commit d71b7896886345c53ef1d84bda2bc758554f5d61 upstream.

syzkaller found another out of bound access in ip_options_compile(),
or more exactly in cipso_v4_validate()

Fixes: 20e2a8648596 ("cipso: handle CIPSO options correctly when NetLabel is disabled")
Fixes: 446fda4f2682 ("[NetLabel]: CIPSOv4 engine")
Signed-off-by: Eric Dumazet <>
Reported-by: Dmitry Vyukov <>
Cc: Paul Moore <>
Acked-by: Paul Moore <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agomm, fs: check for fatal signals in do_generic_file_read()
Michal Hocko [Fri, 3 Feb 2017 21:13:29 +0000 (13:13 -0800)]
mm, fs: check for fatal signals in do_generic_file_read()

commit 5abf186a30a89d5b9c18a6bf93a2c192c9fd52f6 upstream.

do_generic_file_read() can be told to perform a large request from
userspace.  If the system is under OOM and the reading task is the OOM
victim then it has an access to memory reserves and finishing the full
request can lead to the full memory depletion which is dangerous.  Make
sure we rather go with a short read and allow the killed task to

Signed-off-by: Michal Hocko <>
Reviewed-by: Christoph Hellwig <>
Cc: Tetsuo Handa <>
Cc: Al Viro <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: pl2303: add ATEN device ID
Marcel J.E. Mol [Mon, 30 Jan 2017 18:26:40 +0000 (19:26 +0100)]
USB: serial: pl2303: add ATEN device ID

commit d07830db1bdb254e4b50d366010b219286b8c937 upstream.

Seems that ATEN serial-to-usb devices using pl2303 exist with
different device ids. This patch adds a missing device ID so it
is recognised by the driver.

Signed-off-by: Marcel J.E. Mol <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agoARM: 8643/3: arm/ptrace: Preserve previous registers for short regset write
Dave Martin [Wed, 18 Jan 2017 16:11:56 +0000 (17:11 +0100)]
ARM: 8643/3: arm/ptrace: Preserve previous registers for short regset write

commit 228dbbfb5d77f8e047b2a1d78da14b7158433027 upstream.

Ensure that if userspace supplies insufficient data to
PTRACE_SETREGSET to fill all the registers, the thread's old
registers are preserved.

Fixes: 5be6f62b0059 ("ARM: 6883/1: ptrace: Migrate to regsets framework")
Signed-off-by: Dave Martin <>
Acked-by: Russell King <>
Signed-off-by: Russell King <>
Signed-off-by: Ben Hutchings <>
6 years agocan: bcm: fix hrtimer/tasklet termination in bcm op removal
Oliver Hartkopp [Wed, 18 Jan 2017 20:30:51 +0000 (21:30 +0100)]
can: bcm: fix hrtimer/tasklet termination in bcm op removal

commit a06393ed03167771246c4c43192d9c264bc48412 upstream.

When removing a bcm tx operation either a hrtimer or a tasklet might run.
As the hrtimer triggers its associated tasklet and vice versa we need to
take care to mutually terminate both handlers.

Reported-by: Michael Josenhans <>
Signed-off-by: Oliver Hartkopp <>
Tested-by: Michael Josenhans <>
Signed-off-by: Marc Kleine-Budde <>
Signed-off-by: Ben Hutchings <>
6 years agoparisc: Don't use BITS_PER_LONG in userspace-exported swab.h header
Helge Deller [Sat, 28 Jan 2017 10:52:02 +0000 (11:52 +0100)]
parisc: Don't use BITS_PER_LONG in userspace-exported swab.h header

commit 2ad5d52d42810bed95100a3d912679d8864421ec upstream.

In swab.h the "#if BITS_PER_LONG > 32" breaks compiling userspace programs if
BITS_PER_LONG is #defined by userspace with the sizeof() compiler builtin.

Solve this problem by using __BITS_PER_LONG instead.  Since we now
#include asm/bitsperlong.h avoid further potential userspace pollution
by moving the #define of SHIFT_PER_LONG to bitops.h which is not
exported to userspace.

This patch unbreaks compiling qemu on hppa/parisc.

Signed-off-by: Helge Deller <>
[bwh: Backported to 3.2: adjust filenames]
Signed-off-by: Ben Hutchings <>
6 years agosysctl: fix proc_doulongvec_ms_jiffies_minmax()
Eric Dumazet [Thu, 26 Jan 2017 02:20:55 +0000 (18:20 -0800)]
sysctl: fix proc_doulongvec_ms_jiffies_minmax()

commit ff9f8a7cf935468a94d9927c68b00daae701667e upstream.

We perform the conversion between kernel jiffies and ms only when
exporting kernel value to user space.

We need to do the opposite operation when value is written by user.

Only matters when HZ != 1000

Signed-off-by: Eric Dumazet <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: Add quirk for WORLDE easykey.25 MIDI keyboard
Lukáš Lalinský [Fri, 20 Jan 2017 18:46:34 +0000 (19:46 +0100)]
USB: Add quirk for WORLDE easykey.25 MIDI keyboard

commit d9b2997e4a0a874e452df7cdd7de5a54502bd0aa upstream.

Add a quirk for WORLDE easykey.25 MIDI keyboard (idVendor=0218,
idProduct=0401). The device reports that it has config string
descriptor at index 3, but when the system selects the configuration
and tries to get the description, it returns a -EPROTO error,
the communication restarts and this keeps repeating over and over again.
Not requesting the string descriptor makes the device work correctly.

Relevant info from Wireshark:


    bLength: 9
    bDescriptorType: 0x02 (CONFIGURATION)
    wTotalLength: 101
    bNumInterfaces: 2
    bConfigurationValue: 1
    iConfiguration: 3
    Configuration bmAttributes: 0xc0  SELF-POWERED  NO REMOTE-WAKEUP
        1... .... = Must be 1: Must be 1 for USB 1.1 and higher
        .1.. .... = Self-Powered: This device is SELF-POWERED
        ..0. .... = Remote Wakeup: This device does NOT support remote wakeup
    bMaxPower: 50  (100mA)


     45 0.369104       host                  2.38.0                USB      64     GET DESCRIPTOR Request STRING


URB setup
    bmRequestType: 0x80
        1... .... = Direction: Device-to-host
        .00. .... = Type: Standard (0x00)
        ...0 0000 = Recipient: Device (0x00)
    bRequest: GET DESCRIPTOR (6)
    Descriptor Index: 0x03
    bDescriptorType: 0x03
    Language Id: English (United States) (0x0409)
    wLength: 255

     46 0.369255       2.38.0                host                  USB      64     GET DESCRIPTOR Response STRING[Malformed Packet]


Frame 46: 64 bytes on wire (512 bits), 64 bytes captured (512 bits) on interface 0
    [Source: 2.38.0]
    [Destination: host]
    URB id: 0xffff88021f62d480
    URB type: URB_COMPLETE ('C')
    URB transfer type: URB_CONTROL (0x02)
    Endpoint: 0x80, Direction: IN
    Device: 38
    URB bus id: 2
    Device setup request: not relevant ('-')
    Data: present (0)
    URB sec: 1484896277
    URB usec: 455031
    URB status: Protocol error (-EPROTO) (-71)
    URB length [bytes]: 0
    Data length [bytes]: 0
    [Request in: 45]
    [Time from request: 0.000151000 seconds]
    Unused Setup Header
    Interval: 0
    Start frame: 0
    Copy of Transfer Flags: 0x00000200
    Number of ISO descriptors: 0
[Malformed Packet: USB]
    [Expert Info (Error/Malformed): Malformed Packet (Exception occurred)]
        [Malformed Packet (Exception occurred)]
        [Severity level: Error]
        [Group: Malformed]

Signed-off-by: Lukáš Lalinský <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
6 years agomm/huge_memory.c: respect FOLL_FORCE/FOLL_COW for thp
Keno Fischer [Tue, 24 Jan 2017 23:17:48 +0000 (15:17 -0800)]
mm/huge_memory.c: respect FOLL_FORCE/FOLL_COW for thp

commit 8310d48b125d19fcd9521d83b8293e63eb1646aa upstream.

In commit 19be0eaffa3a ("mm: remove gup_flags FOLL_WRITE games from
__get_user_pages()"), the mm code was changed from unsetting FOLL_WRITE
after a COW was resolved to setting the (newly introduced) FOLL_COW
instead.  Simultaneously, the check in gup.c was updated to still allow
writes with FOLL_FORCE set if FOLL_COW had also been set.

However, a similar check in huge_memory.c was forgotten.  As a result,
remote memory writes to ro regions of memory backed by transparent huge
pages cause an infinite loop in the kernel (handle_mm_fault sets
FOLL_COW and returns 0 causing a retry, but follow_trans_huge_pmd bails
out immidiately because `(flags & FOLL_WRITE) && !pmd_write(*pmd)` is

While in this state the process is stil SIGKILLable, but little else
works (e.g.  no ptrace attach, no other signals).  This is easily
reproduced with the following code (assuming thp are set to always):

    #include <assert.h>
    #include <fcntl.h>
    #include <stdint.h>
    #include <stdio.h>
    #include <string.h>
    #include <sys/mman.h>
    #include <sys/stat.h>
    #include <sys/types.h>
    #include <sys/wait.h>
    #include <unistd.h>

    #define TEST_SIZE 5 * 1024 * 1024

    int main(void) {
      int status;
      pid_t child;
      int fd = open("/proc/self/mem", O_RDWR);
      void *addr = mmap(NULL, TEST_SIZE, PROT_READ,
                        MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
      assert(addr != MAP_FAILED);
      pid_t parent_pid = getpid();
      if ((child = fork()) == 0) {
        void *addr2 = mmap(NULL, TEST_SIZE, PROT_READ | PROT_WRITE,
                           MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
        assert(addr2 != MAP_FAILED);
        memset(addr2, 'a', TEST_SIZE);
        pwrite(fd, addr2, TEST_SIZE, (uintptr_t)addr);
        return 0;
      assert(child == waitpid(child, &status, 0));
      assert(WIFEXITED(status) && WEXITSTATUS(status) == 0);
      return 0;

Fix this by updating follow_trans_huge_pmd in huge_memory.c analogously
to the update in gup.c in the original commit.  The same pattern exists
in follow_devmap_pmd.  However, we should not be able to reach that
check with FOLL_COW set, so add WARN_ONCE to make sure we notice if we
ever do.

[ coding-style fixes]
Signed-off-by: Keno Fischer <>
Acked-by: Kirill A. Shutemov <>
Cc: Greg Thelen <>
Cc: Nicholas Piggin <>
Cc: Willy Tarreau <>
Cc: Oleg Nesterov <>
Cc: Kees Cook <>
Cc: Andy Lutomirski <>
Cc: Michal Hocko <>
Cc: Hugh Dickins <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2:
 - Drop change to follow_devmap_pmd()
 - pmd_dirty() is not available; check the page flags as in
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: option: add device ID for HP lt2523 (Novatel E371)
Bjørn Mork [Tue, 24 Jan 2017 09:31:18 +0000 (10:31 +0100)]
USB: serial: option: add device ID for HP lt2523 (Novatel E371)

commit 5d03a2fd2292e71936c4235885c35ccc3c94695b upstream.

Yet another laptop vendor rebranded Novatel E371.

Signed-off-by: Bjørn Mork <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agocrypto: api - Clear CRYPTO_ALG_DEAD bit before registering an alg
Salvatore Benedetto [Fri, 13 Jan 2017 11:54:08 +0000 (11:54 +0000)]
crypto: api - Clear CRYPTO_ALG_DEAD bit before registering an alg

commit d6040764adcb5cb6de1489422411d701c158bb69 upstream.

Make sure CRYPTO_ALG_DEAD bit is cleared before proceeding with
the algorithm registration. This fixes qat-dh registration when
driver is restarted

Signed-off-by: Salvatore Benedetto <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Ben Hutchings <>
6 years agoplatform/x86: intel_mid_powerbtn: Set IRQ_ONESHOT
Andy Shevchenko [Thu, 19 Jan 2017 16:39:40 +0000 (18:39 +0200)]
platform/x86: intel_mid_powerbtn: Set IRQ_ONESHOT

commit 5a00b6c2438460b870a451f14593fc40d3c7edf6 upstream.

The commit 1c6c69525b40 ("genirq: Reject bogus threaded irq requests")
starts refusing misconfigured interrupt handlers. This makes
intel_mid_powerbtn not working anymore.

Add a mandatory flag to a threaded IRQ request in the driver.

Fixes: 1c6c69525b40 ("genirq: Reject bogus threaded irq requests")
Signed-off-by: Andy Shevchenko <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agopowerpc: Ignore reserved field in DCSR and PVR reads and writes
Anton Blanchard [Thu, 19 Jan 2017 03:19:10 +0000 (14:19 +1100)]
powerpc: Ignore reserved field in DCSR and PVR reads and writes

commit 178f358208ceb8b38e5cff3f815e0db4a6a70a07 upstream.

IBM bit 31 (for the rest of us - bit 0) is a reserved field in the
instruction definition of mtspr and mfspr. Hardware is encouraged to
(and does) ignore it.

As a result, if userspace executes an mtspr DSCR with the reserved bit
set, we get a DSCR facility unavailable exception. The kernel fails to
match against the expected value/mask, and we silently return to
userspace to try and re-execute the same mtspr DSCR instruction. We
loop forever until the process is killed.

We should do something here, and it seems mirroring what hardware does
is the better option vs killing the process. While here, relax the
matching of mfspr PVR too.

Signed-off-by: Anton Blanchard <>
Signed-off-by: Michael Ellerman <>
[bwh: Backported to 3.2: drop changes to PPC_INST_M{F,T}SPR_DSCR_USER_MASK]
Signed-off-by: Ben Hutchings <>
6 years agopowerpc/ptrace: Preserve previous fprs/vsrs on short regset write
Dave Martin [Thu, 5 Jan 2017 16:50:57 +0000 (16:50 +0000)]
powerpc/ptrace: Preserve previous fprs/vsrs on short regset write

commit 99dfe80a2a246c600440a815741fd2e74a8b4977 upstream.

Ensure that if userspace supplies insufficient data to PTRACE_SETREGSET
to fill all the registers, the thread's old registers are preserved.

Fixes: c6e6771b87d4 ("powerpc: Introduce VSX thread_struct and CONFIG_VSX")
Signed-off-by: Dave Martin <>
Signed-off-by: Michael Ellerman <>
[bwh: Backported to 3.2:
 - fpscr and fpr are direct members of struct thread_struct
 - Use memcpy() for fpscr, like the reverse copy below, to avoid type error
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agonbd: only set MSG_MORE when we have more to send
Josef Bacik [Thu, 19 Jan 2017 21:08:49 +0000 (16:08 -0500)]
nbd: only set MSG_MORE when we have more to send

commit d61b7f972dab2a7d187c38254845546dfc8eed85 upstream.

A user noticed that write performance was horrible over loopback and we
traced it to an inversion of when we need to set MSG_MORE.  It should be
set when we have more bvec's to send, not when we are on the last bvec.
This patch made the test go from 20 iops to 78k iops.

Signed-off-by: Josef Bacik <>
Fixes: 429a787be679 ("nbd: fix use-after-free of rq/bio in the xmit path")
Signed-off-by: Jens Axboe <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agonbd: fix use-after-free of rq/bio in the xmit path
Jens Axboe [Thu, 17 Nov 2016 19:30:37 +0000 (12:30 -0700)]
nbd: fix use-after-free of rq/bio in the xmit path

commit 429a787be6793554ee02aacc7e1f11ebcecc4453 upstream.

For writes, we can get a completion in while we're still iterating
the request and bio chain. If that happens, we're reading freed
memory and we can crash.

Break out after the last segment and avoid having the iterator
read freed memory.

Reviewed-by: Josef Bacik <>
Signed-off-by: Jens Axboe <>
[bwh: Backported to 3.2:
 - bio_for_each_segment() uses iterator of type int
 - Open-code bio_iter_last()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoceph: fix bad endianness handling in parse_reply_info_extra
Jeff Layton [Thu, 12 Jan 2017 19:42:41 +0000 (14:42 -0500)]
ceph: fix bad endianness handling in parse_reply_info_extra

commit 6df8c9d80a27cb587f61b4f06b57e248d8bc3f86 upstream.

sparse says:

    fs/ceph/mds_client.c:291:23: warning: restricted __le32 degrades to integer
    fs/ceph/mds_client.c:293:28: warning: restricted __le32 degrades to integer
    fs/ceph/mds_client.c:294:28: warning: restricted __le32 degrades to integer
    fs/ceph/mds_client.c:296:28: warning: restricted __le32 degrades to integer

The op value is __le32, so we need to convert it before comparing it.

Signed-off-by: Jeff Layton <>
Reviewed-by: Sage Weil <>
Signed-off-by: Ilya Dryomov <>
[bwh: Backported to 3.2: only filelock and directory replies are handled]
Signed-off-by: Ben Hutchings <>
6 years agocan: ti_hecc: add missing prepare and unprepare of the clock
Yegor Yefremov [Wed, 18 Jan 2017 10:35:57 +0000 (11:35 +0100)]
can: ti_hecc: add missing prepare and unprepare of the clock

commit befa60113ce7ea270cb51eada28443ca2756f480 upstream.

In order to make the driver work with the common clock framework, this
patch converts the clk_enable()/clk_disable() to

Also add error checking for clk_prepare_enable().

Signed-off-by: Yegor Yefremov <>
Signed-off-by: Marc Kleine-Budde <>
Signed-off-by: Ben Hutchings <>
6 years agoubifs: Fix journal replay wrt. xattr nodes
Richard Weinberger [Tue, 10 Jan 2017 10:49:40 +0000 (11:49 +0100)]
ubifs: Fix journal replay wrt. xattr nodes

commit 1cb51a15b576ee325d527726afff40947218fd5e upstream.

When replaying the journal it can happen that a journal entry points to
a garbage collected node.
This is the case when a power-cut occurred between a garbage collect run
and a commit. In such a case nodes have to be read using the failable
read functions to detect whether the found node matches what we expect.

One corner case was forgotten, when the journal contains an entry to
remove an inode all xattrs have to be removed too. UBIFS models xattr
like directory entries, so the TNC code iterates over
all xattrs of the inode and removes them too. This code re-uses the
functions for walking directories and calls ubifs_tnc_next_ent().
ubifs_tnc_next_ent() expects to be used only after the journal and
aborts when a node does not match the expected result. This behavior can
render an UBIFS volume unmountable after a power-cut when xattrs are

Fix this issue by using failable read functions in ubifs_tnc_next_ent()
too when replaying the journal.
Fixes: 1e51764a3c2ac05a ("UBIFS: add new flash file system")
Reported-by: Rock Lee <>
Reviewed-by: David Gstir <>
Signed-off-by: Richard Weinberger <>
Signed-off-by: Ben Hutchings <>
6 years agonet/mlx4_core: Fix racy CQ (Completion Queue) free
Jack Morgenstein [Mon, 16 Jan 2017 16:31:37 +0000 (18:31 +0200)]
net/mlx4_core: Fix racy CQ (Completion Queue) free

commit 291c566a28910614ce42d0ffe82196eddd6346f4 upstream.

In function mlx4_cq_completion() and mlx4_cq_event(), the
radix_tree_lookup requires a rcu_read_lock.
This is mandatory: if another core frees the CQ, it could
run the radix_tree_node_rcu_free() call_rcu() callback while
its being used by the radix tree lookup function.

Additionally, in function mlx4_cq_event(), since we are adding
the rcu lock around the radix-tree lookup, we no longer need to take
the spinlock. Also, the synchronize_irq() call for the async event
eliminates the need for incrementing the cq reference count in

Other changes:
1. In function mlx4_cq_free(), replace spin_lock_irq with spin_lock:
   we no longer take this spinlock in the interrupt context.
   The spinlock here, therefore, simply protects against different
   threads simultaneously invoking mlx4_cq_free() for different cq's.

2. In function mlx4_cq_free(), we move the radix tree delete to before
   the synchronize_irq() calls. This guarantees that we will not
   access this cq during any subsequent interrupts, and therefore can
   safely free the CQ after the synchronize_irq calls. The rcu_read_lock
   in the interrupt handlers only needs to protect against corrupting the
   radix tree; the interrupt handlers may access the cq outside the
   rcu_read_lock due to the synchronize_irq calls which protect against
   premature freeing of the cq.

3. In function mlx4_cq_event(), we change the mlx_warn message to mlx4_dbg.

4. We leave the cq reference count mechanism in place, because it is
   still needed for the cq completion tasklet mechanism.

Fixes: 6d90aa5cf17b ("net/mlx4_core: Make sure there are no pending async events when freeing CQ")
Fixes: 225c7b1feef1 ("IB/mlx4: Add a driver Mellanox ConnectX InfiniBand adapters")
Signed-off-by: Jack Morgenstein <>
Signed-off-by: Matan Barak <>
Signed-off-by: Tariq Toukan <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agosvcrpc: don't leak contexts on PROC_DESTROY
J. Bruce Fields [Mon, 9 Jan 2017 22:15:18 +0000 (17:15 -0500)]
svcrpc: don't leak contexts on PROC_DESTROY

commit 78794d1890708cf94e3961261e52dcec2cc34722 upstream.

Context expiry times are in units of seconds since boot, not unix time.

The use of get_seconds() here therefore sets the expiry time decades in
the future.  This prevents timely freeing of contexts destroyed by
client RPC_GSS_PROC_DESTROY requests.  We'd still free them eventually
(when the module is unloaded or the container shut down), but a lot of
contexts could pile up before then.

Fixes: c5b29f885afe "sunrpc: use seconds since boot in expiry cache"
Reported-by: Andy Adamson <>
Signed-off-by: J. Bruce Fields <>
Signed-off-by: Ben Hutchings <>
6 years agoi2c: fix kernel memory disclosure in dev interface
Vlad Tsyrklevich [Mon, 9 Jan 2017 15:53:36 +0000 (22:53 +0700)]
i2c: fix kernel memory disclosure in dev interface

commit 30f939feaeee23e21391cfc7b484f012eb189c3c upstream.

i2c_smbus_xfer() does not always fill an entire block, allowing
kernel stack memory disclosure through the temp variable. Clear
it before it's read to.

Signed-off-by: Vlad Tsyrklevich <>
Signed-off-by: Wolfram Sang <>
Signed-off-by: Ben Hutchings <>
6 years agommc: mxs-mmc: Fix additional cycles after transmission stop
Stefan Wahren [Thu, 5 Jan 2017 19:24:04 +0000 (19:24 +0000)]
mmc: mxs-mmc: Fix additional cycles after transmission stop

commit 01167c7b9cbf099c69fe411a228e4e9c7104e123 upstream.

According to the code the intention is to append 8 SCK cycles
instead of 4 at end of a MMC_STOP_TRANSMISSION command. But this
will never happened because it's an AC command not an ADTC command.
So fix this by moving the statement into the right function.

Signed-off-by: Stefan Wahren <>
Fixes: e4243f13d10e (mmc: mxs-mmc: add mmc host driver for i.MX23/28)
Signed-off-by: Ulf Hansson <>
Signed-off-by: Ben Hutchings <>
6 years agoxhci: fix deadlock at host remove by running watchdog correctly
Mathias Nyman [Wed, 11 Jan 2017 15:10:34 +0000 (17:10 +0200)]
xhci: fix deadlock at host remove by running watchdog correctly

commit d6169d04097fd9ddf811e63eae4e5cd71e6666e2 upstream.

If a URB is killed while the host is removed we can end up in a situation
where the hub thread takes the roothub device lock, and waits for
the URB to be given back by xhci-hcd, blocking the host remove code.

xhci-hcd tries to stop the endpoint and give back the urb, but can't
as the host is removed from PCI bus at the same time, preventing the normal
way of giving back urb.

Instead we need to rely on the stop command timeout function to give back
the urb. This xhci_stop_endpoint_command_watchdog() timeout function
used a XHCI_STATE_DYING flag to indicate if the timeout function is already
running, but later this flag has been taking into use in other places to
mark that xhci is dying.

Remove checks for XHCI_STATE_DYING in xhci_urb_dequeue. We are still
checking that reading from pci state does not return 0xffffffff or that
host is not halted before trying to stop the endpoint.

This whole area of stopping endpoints, giving back URBs, and the wathdog
timeout need rework, this fix focuses on solving a specific deadlock
issue that we can then send to stable before any major rework.

Signed-off-by: Mathias Nyman <>
Signed-off-by: Greg Kroah-Hartman <>
[bwh: Backported to 3.2: the checks look slightly different]
Signed-off-by: Ben Hutchings <>
6 years agogro: use min_t() in skb_gro_reset_offset()
Eric Dumazet [Wed, 11 Jan 2017 03:52:43 +0000 (19:52 -0800)]
gro: use min_t() in skb_gro_reset_offset()

commit 7cfd5fd5a9813f1430290d20c0fead9b4582a307 upstream.

On 32bit arches, (skb->end - skb->data) is not 'unsigned int',
so we shall use min_t() instead of min() to avoid a compiler error.

Fixes: 1272ce87fa01 ("gro: Enter slow-path if there is no tailroom")
Reported-by: kernel test robot <>
Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix control-message error handling
Johan Hovold [Fri, 6 Jan 2017 18:15:18 +0000 (19:15 +0100)]
USB: serial: ch341: fix control-message error handling

commit 2d5a9c72d0c4ac73cf97f4b7814ed6c44b1e49ae upstream.

A short control transfer would currently fail to be detected, something
which could lead to stale buffer data being used as valid input.

Check for short transfers, and make sure to log any transfer errors.

Note that this also avoids leaking heap data to user space (TIOCMGET)
and the remote device (break control).

Fixes: 6ce76104781a ("USB: Driver for CH341 USB-serial adaptor")
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agovme: Fix wrong pointer utilization in ca91cx42_slave_get
Augusto Mecking Caringi [Tue, 10 Jan 2017 10:45:00 +0000 (10:45 +0000)]
vme: Fix wrong pointer utilization in ca91cx42_slave_get

commit c8a6a09c1c617402cc9254b2bc8da359a0347d75 upstream.

In ca91cx42_slave_get function, the value pointed by vme_base pointer is
set through:

*vme_base = ioread32(bridge->base + CA91CX42_VSI_BS[i]);

So it must be dereferenced to be used in calculation of pci_base:

*pci_base = (dma_addr_t)*vme_base + pci_offset;

This bug was caught thanks to the following gcc warning:

drivers/vme/bridges/vme_ca91cx42.c: In function ‘ca91cx42_slave_get’:
drivers/vme/bridges/vme_ca91cx42.c:467:14: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
*pci_base = (dma_addr_t)vme_base + pci_offset;

Signed-off-by: Augusto Mecking Caringi <>
Acked-By: Martyn Welch <>
Signed-off-by: Greg Kroah-Hartman <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
6 years agosysrq: attach sysrq handler correctly for 32-bit kernel
Akinobu Mita [Thu, 5 Jan 2017 17:14:16 +0000 (02:14 +0900)]
sysrq: attach sysrq handler correctly for 32-bit kernel

commit 802c03881f29844af0252b6e22be5d2f65f93fd0 upstream.

The sysrq input handler should be attached to the input device which has
a left alt key.

On 32-bit kernels, some input devices which has a left alt key cannot
attach sysrq handler.  Because the keybit bitmap in struct input_device_id
for sysrq is not correctly initialized.  KEY_LEFTALT is 56 which is
greater than BITS_PER_LONG on 32-bit kernels.

I found this problem when using a matrix keypad device which defines
a KEY_LEFTALT (56) but doesn't have a KEY_O (24 == 56%32).

Cc: Jiri Slaby <>
Signed-off-by: Akinobu Mita <>
Acked-by: Dmitry Torokhov <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
6 years agoocfs2: fix crash caused by stale lvb with fsdlm plugin
Eric Ren [Wed, 11 Jan 2017 00:57:33 +0000 (16:57 -0800)]
ocfs2: fix crash caused by stale lvb with fsdlm plugin

commit e7ee2c089e94067d68475990bdeed211c8852917 upstream.

The crash happens rather often when we reset some cluster nodes while
nodes contend fiercely to do truncate and append.

The crash backtrace is below:

   dlm: C21CBDA5E0774F4BA5A9D4F317717495: dlm_recover_grant 1 locks on 971 resources
   dlm: C21CBDA5E0774F4BA5A9D4F317717495: dlm_recover 9 generation 5 done: 4 ms
   ocfs2: Begin replay journal (node 318952601, slot 2) on device (253,18)
   ocfs2: End replay journal (node 318952601, slot 2) on device (253,18)
   ocfs2: Beginning quota recovery on device (253,18) for slot 2
   ocfs2: Finishing quota recovery on device (253,18) for slot 2
   (truncate,30154,1):ocfs2_truncate_file:470 ERROR: bug expression: le64_to_cpu(fe->i_size) != i_size_read(inode)
   (truncate,30154,1):ocfs2_truncate_file:470 ERROR: Inode 290321, inode i_size = 732 != di i_size = 937, i_flags = 0x1
   ------------[ cut here ]------------
   kernel BUG at /usr/src/linux/fs/ocfs2/file.c:470!
   invalid opcode: 0000 [#1] SMP
   Modules linked in: ocfs2_stack_user(OEN) ocfs2(OEN) ocfs2_nodemanager ocfs2_stackglue(OEN) quota_tree dlm(OEN) configfs fuse sd_mod    iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi af_packet iscsi_ibft iscsi_boot_sysfs softdog xfs libcrc32c ppdev parport_pc pcspkr parport      joydev virtio_balloon virtio_net i2c_piix4 acpi_cpufreq button processor ext4 crc16 jbd2 mbcache ata_generic cirrus virtio_blk ata_piix               drm_kms_helper ahci syscopyarea libahci sysfillrect sysimgblt fb_sys_fops ttm floppy libata drm virtio_pci virtio_ring uhci_hcd virtio ehci_hcd       usbcore serio_raw usb_common sg dm_multipath dm_mod scsi_dh_rdac scsi_dh_emc scsi_dh_alua scsi_mod autofs4
   Supported: No, Unsupported modules are loaded
   CPU: 1 PID: 30154 Comm: truncate Tainted: G           OE   N  4.4.21-69-default #1
   Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.8.1-0-g4adadbd-20151112_172657-sheep25 04/01/2014
   task: ffff88004ff6d240 ti: ffff880074e68000 task.ti: ffff880074e68000
   RIP: 0010:[<ffffffffa05c8c30>]  [<ffffffffa05c8c30>] ocfs2_truncate_file+0x640/0x6c0 [ocfs2]
   RSP: 0018:ffff880074e6bd50  EFLAGS: 00010282
   RAX: 0000000000000074 RBX: 000000000000029e RCX: 0000000000000000
   RDX: 0000000000000001 RSI: 0000000000000246 RDI: 0000000000000246
   RBP: ffff880074e6bda8 R08: 000000003675dc7a R09: ffffffff82013414
   R10: 0000000000034c50 R11: 0000000000000000 R12: ffff88003aab3448
   R13: 00000000000002dc R14: 0000000000046e11 R15: 0000000000000020
   FS:  00007f839f965700(0000) GS:ffff88007fc80000(0000) knlGS:0000000000000000
   CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
   CR2: 00007f839f97e000 CR3: 0000000036723000 CR4: 00000000000006e0
   Call Trace:
     ocfs2_setattr+0x698/0xa90 [ocfs2]
   Code: 24 28 ba d6 01 00 00 48 c7 c6 30 43 62 a0 8b 41 2c 89 44 24 08 48 8b 41 20 48 c7 c1 78 a3 62 a0 48 89 04 24 31 c0 e8 a0 97 f9 ff <0f> 0b 3d 00 fe ff ff 0f 84 ab fd ff ff 83 f8 fc 0f 84 a2 fd ff
   RIP  [<ffffffffa05c8c30>] ocfs2_truncate_file+0x640/0x6c0 [ocfs2]

It's because ocfs2_inode_lock() get us stale LVB in which the i_size is
not equal to the disk i_size.  We mistakenly trust the LVB because the
underlaying fsdlm dlm_lock() doesn't set lkb_sbflags with
DLM_SBF_VALNOTVALID properly for us.  But, why?

The current code tries to downconvert lock without DLM_LKF_VALBLK flag
to tell o2cb don't update RSB's LVB if it's a PR->NULL conversion, even
if the lock resource type needs LVB.  This is not the right way for

The fsdlm plugin behaves different on DLM_LKF_VALBLK, it depends on
DLM_LKF_VALBLK to decide if we care about the LVB in the LKB.  If
DLM_LKF_VALBLK is not set, fsdlm will skip recovering RSB's LVB from
this lkb and set the right DLM_SBF_VALNOTVALID appropriately when node
failure happens.

The following diagram briefly illustrates how this crash happens:

RSB1 is inode metadata lock resource with LOCK_TYPE_USES_LVB;

The 1st round:

             Node1                                    Node2
                                                  RSB1(master): NULL->EX
ocfs2_downconvert_lock(PR->NULL, set_lvb==0)
  ocfs2_dlm_lock(no DLM_LKF_VALBLK)

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

dlm_lock(no DLM_LKF_VALBLK)
  convert_lock(overwrite lkb->lkb_exflags
               with no DLM_LKF_VALBLK)

RSB1: NULL                                        RSB1: EX
                                                  reset Node2

/* The LVB is not trustable if the node with EX fails and
 * no lock >= PR is left. We should set RSB_VALNOTVALID for RSB1.

 if(!(kb_exflags & DLM_LKF_VALBLK)) /* This means we miss the chance to
           return;                   * to invalid the LVB here.

The 2nd round:

         Node 1                                Node2
RSB1(become master from recovery)

    /* dlm_lock() return the stale lvb without setting DLM_SBF_VALNOTVALID */
    ocfs2_meta_lvb_is_trustable() return 1 /* so we don't refresh inode from disk */
      mlog_bug_on_msg(disk isize != i_size_read(inode))  /* crash! */

The fix is quite straightforward.  We keep to set DLM_LKF_VALBLK flag
for dlm_lock() if the lock resource type needs LVB and the fsdlm plugin
is uesed.

Signed-off-by: Eric Ren <>
Reviewed-by: Joseph Qi <>
Cc: Mark Fasheh <>
Cc: Joel Becker <>
Cc: Junxiao Bi <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agogro: Disable frag0 optimization on IPv6 ext headers
Herbert Xu [Tue, 10 Jan 2017 20:24:15 +0000 (12:24 -0800)]
gro: Disable frag0 optimization on IPv6 ext headers

commit 57ea52a865144aedbcd619ee0081155e658b6f7d upstream.

The GRO fast path caches the frag0 address.  This address becomes
invalid if frag0 is modified by pskb_may_pull or its variants.
So whenever that happens we must disable the frag0 optimization.

This is usually done through the combination of gro_header_hard
and gro_header_slow, however, the IPv6 extension header path did
the pulling directly and would continue to use the GRO fast path

This patch fixes it by disabling the fast path when we enter the
IPv6 extension header path.

Fixes: 78a478d0efd9 ("gro: Inline skb_gro_header and cache frag0 virtual address")
Reported-by: Slava Shwartsman <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
6 years agogro: Enter slow-path if there is no tailroom
Herbert Xu [Tue, 10 Jan 2017 20:24:01 +0000 (12:24 -0800)]
gro: Enter slow-path if there is no tailroom

commit 1272ce87fa017ca4cf32920764d879656b7a005a upstream.

The GRO path has a fast-path where we avoid calling pskb_may_pull
and pskb_expand by directly accessing frag0.  However, this should
only be done if we have enough tailroom in the skb as otherwise
we'll have to expand it later anyway.

This patch adds the check by capping frag0_len with the skb tailroom.

Fixes: cb18978cbf45 ("gro: Open-code final pskb_may_pull")
Reported-by: Slava Shwartsman <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoARM: 8634/1: hw_breakpoint: blacklist Scorpion CPUs
Mark Rutland [Fri, 6 Jan 2017 12:12:47 +0000 (13:12 +0100)]
ARM: 8634/1: hw_breakpoint: blacklist Scorpion CPUs

commit ddc37832a1349f474c4532de381498020ed71d31 upstream.

On APQ8060, the kernel crashes in arch_hw_breakpoint_init, taking an
undefined instruction trap within write_wb_reg. This is because Scorpion
CPUs erroneously appear to set DBGPRSR.SPD when WFI is issued, even if
the core is not powered down. When DBGPRSR.SPD is set, breakpoint and
watchpoint registers are treated as undefined.

It's possible to trigger similar crashes later on from userspace, by
requesting the kernel to install a breakpoint or watchpoint, as we can
go idle at any point between the reset of the debug registers and their
later use. This has always been the case.

Given that this has always been broken, no-one has complained until now,
and there is no clear workaround, disable hardware breakpoints and
watchpoints on Scorpion to avoid these issues.

Signed-off-by: Mark Rutland <>
Reported-by: Linus Walleij <>
Reviewed-by: Stephen Boyd <>
Acked-by: Will Deacon <>
Cc: Russell King <>
Signed-off-by: Russell King <>
[bwh: Backported to 3.2:
 - Open-code read_cpuid_part()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix baud rate and line-control handling
Johan Hovold [Fri, 6 Jan 2017 18:15:16 +0000 (19:15 +0100)]
USB: serial: ch341: fix baud rate and line-control handling

commit 55fa15b5987db22b4f35d3f0798928c126be5f1c upstream.

Revert to using direct register writes to set the divisor and
line-control registers.

A recent change switched to using the init vendor command to update
these registers, something which also enabled support for CH341A
devices. It turns out that simply setting bit 7 in the divisor register
is sufficient to support CH341A and specifically prevent data from being
buffered until a full endpoint-size packet (32 bytes) has been received.

Using the init command also had the side-effect of temporarily
deasserting the DTR/RTS signals on every termios change (including
initialisation on open) something which for example could cause problems
in setups where DTR is used to trigger a reset.

Fixes: 4e46c410e050 ("USB: serial: ch341: reinitialize chip on
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix resume after reset
Johan Hovold [Fri, 6 Jan 2017 18:15:14 +0000 (19:15 +0100)]
USB: serial: ch341: fix resume after reset

commit ce5e292828117d1b71cbd3edf9e9137cf31acd30 upstream.

Fix reset-resume handling which failed to resubmit the read and
interrupt URBs, thereby leaving a port that was open before suspend in a
broken state until closed and reopened.

Fixes: 1ded7ea47b88 ("USB: ch341 serial: fix port number changed after
Fixes: 2bfd1c96a9fb ("USB: serial: ch341: remove reset_resume callback")
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2:
 - Move initialisation of 'serial' up to make this work
 - Delete the call to usb_serial_resume() that was still present and
   would be redundant with usb_serial_generic_resume()
 - Open-code tty_port_initialized()]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix open error handling
Johan Hovold [Fri, 6 Jan 2017 18:15:13 +0000 (19:15 +0100)]
USB: serial: ch341: fix open error handling

commit f2950b78547ffb8475297ada6b92bc2d774d5461 upstream.

Make sure to stop the interrupt URB before returning on errors during

Fixes: 664d5df92e88 ("USB: usb-serial ch341: support for DTR/RTS/CTS")
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix modem-control and B0 handling
Johan Hovold [Fri, 6 Jan 2017 18:15:12 +0000 (19:15 +0100)]
USB: serial: ch341: fix modem-control and B0 handling

commit 030ee7ae52a46a2be52ccc8242c4a330aba8d38e upstream.

The modem-control signals are managed by the tty-layer during open and
should not be asserted prematurely when set_termios is called from
driver open.

Also make sure that the signals are asserted only when changing speed
from B0.

Fixes: 664d5df92e88 ("USB: usb-serial ch341: support for DTR/RTS/CTS")
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix open and resume after B0
Johan Hovold [Fri, 6 Jan 2017 18:15:11 +0000 (19:15 +0100)]
USB: serial: ch341: fix open and resume after B0

commit a20047f36e2f6a1eea4f1fd261aaa55882369868 upstream.

The private baud_rate variable is used to configure the port at open and
reset-resume and must never be set to (and left at) zero or reset-resume
and all further open attempts will fail.

Fixes: aa91def41a7b ("USB: ch341: set tty baud speed according to tty
Fixes: 664d5df92e88 ("USB: usb-serial ch341: support for DTR/RTS/CTS")
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: fix initial modem-control state
Johan Hovold [Fri, 6 Jan 2017 18:15:10 +0000 (19:15 +0100)]
USB: serial: ch341: fix initial modem-control state

commit 4e2da44691cffbfffb1535f478d19bc2dca3e62b upstream.

DTR and RTS will be asserted by the tty-layer when the port is opened
and deasserted on close (if HUPCL is set). Make sure the initial state
is not-asserted before the port is first opened as well.

Fixes: 664d5df92e88 ("USB: usb-serial ch341: support for DTR/RTS/CTS")
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: reinitialize chip on reconfiguration
Aidan Thornton [Sat, 22 Oct 2016 21:02:24 +0000 (22:02 +0100)]
USB: serial: ch341: reinitialize chip on reconfiguration

commit 4e46c410e050bcac36deadbd8e20449d078204e8 upstream.

Changing the LCR register after initialization does not seem to be reliable
on all chips (particularly not on CH341A). Restructure initialization and
configuration to always reinit the chip on configuration changes instead and
pass the LCR register value directly to the initialization command.

(Note that baud rates above 500kbaud are incorrect, but they're incorrect in
the same way both before and after this patch at least on the CH340G. Fixing
this isn't a priority as higher baud rates don't seem that reliable anyway.)

Cleaned-up version of a patch by Grigori Goronzy

Signed-off-by: Aidan Thornton <>
Reviewed-by: Grigori Goronzy <>
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: tty_struct::termios is a pointer, not a struct]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: ch341: add register and USB request definitions
Aidan Thornton [Sat, 22 Oct 2016 21:02:23 +0000 (22:02 +0100)]
USB: serial: ch341: add register and USB request definitions

commit 6fde8d29b0424f292a4ec5dbce01458ad759a41f upstream.

No functional changes, this just gives names to some registers and USB
requests based on Grigori Goronzy's work and WinChipTech's Linux driver
(which reassuringly agree), then uses them in place of magic numbers.
This also renames the misnamed BREAK2 register (actually UART config)

Signed-off-by: Aidan Thornton <>
Reviewed-by: Grigori Goronzy <>
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: ch341: set tty baud speed according to tty struct
Nicolas PLANEL [Sun, 1 Mar 2015 18:47:22 +0000 (13:47 -0500)]
USB: ch341: set tty baud speed according to tty struct

commit aa91def41a7bb1fd65492934ce6bea19202b6080 upstream.

The ch341_set_baudrate() function initialize the device baud speed
according to the value on priv->baud_rate. By default the ch341_open() set
it to a hardcoded value (DEFAULT_BAUD_RATE 9600). Unfortunately, the
tty_struct is not initialized with the same default value. (usually 56700)

This means that the tty_struct and the device baud rate generator are not
synchronized after opening the port.

Fixup is done by calling ch341_set_termios() if tty exist.
Remove unnecessary variable priv->baud_rate setup as it's already done by
Remove unnecessary call to ch341_set_{handshake,baudrate}() in
ch341_open() as there already called in ch341_configure() and

Signed-off-by: Nicolas PLANEL <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>