6 years agox86/PCI: Mark Broadwell-EP Home Agent 1 as having non-compliant BARs
Prarit Bhargava [Wed, 11 May 2016 16:27:15 +0000 (12:27 -0400)]
x86/PCI: Mark Broadwell-EP Home Agent 1 as having non-compliant BARs

commit da77b67195de1c65bef4908fa29967c4d0af2da2 upstream.

Commit b894157145e4 ("x86/PCI: Mark Broadwell-EP Home Agent & PCU as having
non-compliant BARs") marked Home Agent 0 & PCU has having non-compliant
BARs.  Home Agent 1 also has non-compliant BARs.

Mark Home Agent 1 as having non-compliant BARs so the PCI core doesn't
touch them.

The problem with these devices is documented in the Xeon v4 specification

  BDF2          PCI BARs in the Home Agent Will Return Non-Zero Values
                During Enumeration

  Problem:      During system initialization the Operating System may access
                the standard PCI BARs (Base Address Registers).  Due to
                this erratum, accesses to the Home Agent BAR registers (Bus
                1; Device 18; Function 0,4; Offsets (0x14-0x24) will return
                non-zero values.

  Implication:  The operating system may issue a warning.  Intel has not
                observed any functional failures due to this erratum.

Fixes: b894157145e4 ("x86/PCI: Mark Broadwell-EP Home Agent & PCU as having non-compliant BARs")
Signed-off-by: Prarit Bhargava <>
Signed-off-by: Bjorn Helgaas <>
CC: Thomas Gleixner <>
CC: Ingo Molnar <>
CC: "H. Peter Anvin" <>
CC: Andi Kleen <>
Signed-off-by: Ben Hutchings <>
6 years agonet/mlx4_core: Fix access to uninitialized index
Tariq Toukan [Sun, 15 May 2016 07:21:26 +0000 (10:21 +0300)]
net/mlx4_core: Fix access to uninitialized index

commit 2bb07e155bb3e0c722c806723f737cf8020961ef upstream.

Prevent using uninitialized or negative index when handling
steering entries.

Fixes: b12d93d63c32 ('mlx4: Add support for promiscuous mode in the new steering model.')
Signed-off-by: Tariq Toukan <>
Reported-by: Dan Carpenter <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agommc: mmc: Fix partition switch timeout for some eMMCs
Adrian Hunter [Thu, 5 May 2016 05:12:28 +0000 (08:12 +0300)]
mmc: mmc: Fix partition switch timeout for some eMMCs

commit 1c447116d017a98c90f8f71c8c5a611e0aa42178 upstream.

Some eMMCs set the partition switch timeout too low.

Now typically eMMCs are considered a critical component (e.g. because
they store the root file system) and consequently are expected to be
reliable.  Thus we can neglect the use case where eMMCs can't switch
reliably and we might want a lower timeout to facilitate speedy

Although we could employ a quirk for the cards that are affected (if
we could identify them all), as described above, there is little
benefit to having a low timeout, so instead simply set a minimum

The minimum is set to 300ms somewhat arbitrarily - the examples that
have been seen had a timeout of 10ms but were sometimes taking 60-70ms.

Signed-off-by: Adrian Hunter <>
Signed-off-by: Ulf Hansson <>
Signed-off-by: Ben Hutchings <>
6 years agosched/loadavg: Fix loadavg artifacts on fully idle and on fully loaded systems
Vik Heyndrickx [Thu, 28 Apr 2016 18:46:28 +0000 (20:46 +0200)]
sched/loadavg: Fix loadavg artifacts on fully idle and on fully loaded systems

commit 20878232c52329f92423d27a60e48b6a6389e0dd upstream.

Systems show a minimal load average of 0.00, 0.01, 0.05 even when they
have no load at all.

Uptime and /proc/loadavg on all systems with kernels released during the
last five years up until kernel version 4.6-rc5, show a 5- and 15-minute
minimum loadavg of 0.01 and 0.05 respectively. This should be 0.00 on
idle systems, but the way the kernel calculates this value prevents it
from getting lower than the mentioned values.

Likewise but not as obviously noticeable, a fully loaded system with no
processes waiting, shows a maximum 1/5/15 loadavg of 1.00, 0.99, 0.95
(multiplied by number of cores).

Once the (old) load becomes 93 or higher, it mathematically can never
get lower than 93, even when the active (load) remains 0 forever.
This results in the strange 0.00, 0.01, 0.05 uptime values on idle
systems.  Note: 93/2048 = 0.0454..., which rounds up to 0.05.

It is not correct to add a 0.5 rounding (=1024/2048) here, since the
result from this function is fed back into the next iteration again,
so the result of that +0.5 rounding value then gets multiplied by
(2048-2037), and then rounded again, so there is a virtual "ghost"
load created, next to the old and active load terms.

By changing the way the internally kept value is rounded, that internal
value equivalent now can reach 0.00 on idle, and 1.00 on full load. Upon
increasing load, the internally kept load value is rounded up, when the
load is decreasing, the load value is rounded down.

The modified code was tested on nohz=off and nohz kernels. It was tested
on vanilla kernel 4.6-rc5 and on centos 7.1 kernel 3.10.0-327. It was
tested on single, dual, and octal cores system. It was tested on virtual
hosts and bare hardware. No unwanted effects have been observed, and the
problems that the patch intended to fix were indeed gone.

Tested-by: Damien Wyart <>
Signed-off-by: Vik Heyndrickx <>
Signed-off-by: Peter Zijlstra (Intel) <>
Cc: Doug Smythies <>
Cc: Linus Torvalds <>
Cc: Mike Galbraith <>
Cc: Peter Zijlstra <>
Cc: Thomas Gleixner <>
Fixes: 0f004f5a696a ("sched: Cure more NO_HZ load average woes")
Signed-off-by: Ingo Molnar <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
6 years agopowerpc/mm/hash64: Fix subpage protection with 4K HPTE config
Michael Ellerman [Fri, 6 May 2016 06:47:12 +0000 (16:47 +1000)]
powerpc/mm/hash64: Fix subpage protection with 4K HPTE config

commit aac55d7573c5d46ed9a62818d5d3e69dd2060105 upstream.

With Linux page size of 64K and hardware only supporting 4K HPTE, if we
use subpage protection, we always fail for the subpage 0 as shown
below (using the selftest subpage_prot test):

  520175565:  (4520111850): Failed at 0x3fffad4b0000 (p=13,sp=0,w=0), want=fault, got=pass !
  4520890210: (4520826495): Failed at 0x3fffad5b0000 (p=29,sp=0,w=0), want=fault, got=pass !
  4521574251: (4521510536): Failed at 0x3fffad6b0000 (p=45,sp=0,w=0), want=fault, got=pass !
  4522258324: (4522194609): Failed at 0x3fffad7b0000 (p=61,sp=0,w=0), want=fault, got=pass !

This is because hash preload wrongly inserts the HPTE entry for subpage
0 without looking at the subpage protection information.

Fix it by teaching should_hash_preload() not to preload if we have
subpage protection configured for that range.

It appears this has been broken since it was introduced in 2008.

Fixes: fa28237cfcc5 ("[POWERPC] Provide a way to protect 4k subpages when using 64k pages")
Signed-off-by: Aneesh Kumar K.V <>
[mpe: Rework into should_hash_preload() to avoid build fails w/SLICES=n]
Signed-off-by: Michael Ellerman <>
Signed-off-by: Ben Hutchings <>
6 years agopowerpc/mm/hash64: Factor out hash preload psize check
Michael Ellerman [Fri, 6 May 2016 06:46:00 +0000 (16:46 +1000)]
powerpc/mm/hash64: Factor out hash preload psize check

commit 8bbc9b7b001eaab8abf7e9e24edf1bb285c8d825 upstream.

Currently we have a check in hash_preload() against the psize, which is
only included when CONFIG_PPC_MM_SLICES is enabled. We want to expand
this check in a subsequent patch, so factor it out to allow that. As a
bonus it removes the #ifdef in the C code.

Unfortunately we can't put this in the existing CONFIG_PPC_MM_SLICES
block because it would require a forward declaration.

Signed-off-by: Michael Ellerman <>
Signed-off-by: Ben Hutchings <>
6 years agoata: sata_dwc_460ex: remove incorrect locking
Mans Rullgard [Tue, 26 Apr 2016 09:03:02 +0000 (12:03 +0300)]
ata: sata_dwc_460ex: remove incorrect locking

commit 55e610cdd28c0ad3dce0652030c0296d549673f3 upstream.

This lock is already taken in ata_scsi_queuecmd() a few levels up the
call stack so attempting to take it here is an error.  Moreover, it is
pointless in the first place since it only protects a single, atomic

Enabling lock debugging gives the following output:

[ INFO: possible recursive locking detected ]
4.4.0-rc5+ #189 Not tainted
kworker/u2:3/37 is trying to acquire lock:
 (&(&host->lock)->rlock){-.-...}, at: [<90283294>] sata_dwc_exec_command_by_tag.constprop.14+0x44/0x8c

but task is already holding lock:
 (&(&host->lock)->rlock){-.-...}, at: [<902761ac>] ata_scsi_queuecmd+0x2c/0x330

other info that might help us debug this:
 Possible unsafe locking scenario:


 *** DEADLOCK ***
 May be due to missing lock nesting notation

4 locks held by kworker/u2:3/37:
 #0:  ("events_unbound"){.+.+.+}, at: [<9003a0a4>] process_one_work+0x12c/0x430
 #1:  ((&entry->work)){+.+.+.}, at: [<9003a0a4>] process_one_work+0x12c/0x430
 #2:  (&bdev->bd_mutex){+.+.+.}, at: [<9011fd54>] __blkdev_get+0x50/0x380
 #3:  (&(&host->lock)->rlock){-.-...}, at: [<902761ac>] ata_scsi_queuecmd+0x2c/0x330

stack backtrace:
CPU: 0 PID: 37 Comm: kworker/u2:3 Not tainted 4.4.0-rc5+ #189
Workqueue: events_unbound async_run_entry_fn
Stack : 90b38e30 00000021 00000003 9b2a6040 00000000 9005f3f0 904fc8dc 00000025
        906b96e4 00000000 90528648 9b3336c4 904fc8dc 9009bf18 00000002 00000004
        00000000 00000000 9b3336c4 9b3336e4 904fc8dc 9003d074 00000000 90500000
        9005e738 00000000 00000000 00000000 00000000 00000000 00000000 00000000
        6e657665 755f7374 756f626e 0000646e 00000000 00000000 9b00ca00 9b025000
Call Trace:
[<90009d6c>] show_stack+0x88/0xa4
[<90057744>] __lock_acquire+0x1ce8/0x2154
[<900583e4>] lock_acquire+0x64/0x8c
[<9045ff10>] _raw_spin_lock_irqsave+0x54/0x78
[<90283294>] sata_dwc_exec_command_by_tag.constprop.14+0x44/0x8c
[<90283484>] sata_dwc_qc_issue+0x1a8/0x24c
[<9026b39c>] ata_qc_issue+0x1f0/0x410
[<90273c6c>] ata_scsi_translate+0xb4/0x200
[<90276234>] ata_scsi_queuecmd+0xb4/0x330
[<9025800c>] scsi_dispatch_cmd+0xd0/0x128
[<90259934>] scsi_request_fn+0x58c/0x638
[<901a3e50>] __blk_run_queue+0x40/0x5c
[<901a83d4>] blk_queue_bio+0x27c/0x28c
[<901a5914>] generic_make_request+0xf0/0x188
[<901a5a54>] submit_bio+0xa8/0x194
[<9011adcc>] submit_bh_wbc.isra.23+0x15c/0x17c
[<9011c908>] block_read_full_page+0x3e4/0x428
[<9009e2e0>] do_read_cache_page+0xac/0x210
[<9009fd90>] read_cache_page+0x18/0x24
[<901bbd18>] read_dev_sector+0x38/0xb0
[<901bd174>] msdos_partition+0xb4/0x5c0
[<901bcb8c>] check_partition+0x140/0x274
[<901bba60>] rescan_partitions+0xa0/0x2b0
[<9011ff68>] __blkdev_get+0x264/0x380
[<901201ac>] blkdev_get+0x128/0x36c
[<901b9378>] add_disk+0x3c0/0x4bc
[<90268268>] sd_probe_async+0x100/0x224
[<90043a44>] async_run_entry_fn+0x50/0x124
[<9003a11c>] process_one_work+0x1a4/0x430
[<9003a4f4>] worker_thread+0x14c/0x4fc
[<900408f4>] kthread+0xd0/0xe8
[<90004338>] ret_from_kernel_thread+0x14/0x1c

Fixes: 62936009f35a ("[libata] Add 460EX on-chip SATA driver, sata_dwc_460ex")
Tested-by: Christian Lamparter <>
Signed-off-by: Mans Rullgard <>
Signed-off-by: Tejun Heo <>
Signed-off-by: Ben Hutchings <>
6 years agogcov: disable tree-loop-im to reduce stack usage
Arnd Bergmann [Mon, 25 Apr 2016 15:35:30 +0000 (17:35 +0200)]
gcov: disable tree-loop-im to reduce stack usage

commit c87bf431448b404a6ef5fbabd74c0e3e42157a7f upstream.

Enabling CONFIG_GCOV_PROFILE_ALL produces us a lot of warnings like

lib/lz4/lz4hc_compress.c: In function 'lz4_compresshcctx':
lib/lz4/lz4hc_compress.c:514:1: warning: the frame size of 1504 bytes is larger than 1024 bytes [-Wframe-larger-than=]

After some investigation, I found that this behavior started with gcc-4.9,
and opened
A suggested workaround for it is to use the -fno-tree-loop-im
flag that turns off one of the optimization stages in gcc, so the
code runs a little slower but does not use excessive amounts
of stack.

We could make this conditional on the gcc version, but I could not
find an easy way to do this in Kbuild and the benefit would be
fairly small, given that most of the gcc version in production are
affected now.

I'm marking this for 'stable' backports because it addresses a bug
with code generation in gcc that exists in all kernel versions
with the affected gcc releases.

Signed-off-by: Arnd Bergmann <>
Acked-by: Peter Oberparleiter <>
Signed-off-by: Michal Marek <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoMIPS: Adjust set_pte() SMP fix to handle R10000_LLSC_WAR
Joshua Kinard [Mon, 7 Sep 2015 10:42:30 +0000 (06:42 -0400)]
MIPS: Adjust set_pte() SMP fix to handle R10000_LLSC_WAR

commit 128639395b2ceacc6a56a0141d0261012bfe04d3 upstream.

Update the recent changes to set_pte() that were added in 46011e6ea392
to handle R10000_LLSC_WAR, and format the assembly to match other areas
of the MIPS tree using the same WAR.

This also incorporates a patch recently sent in my Markos Chandras,
"Remove local LL/SC preprocessor variants", so that patch doesn't need
to be applied if this one is accepted.

Signed-off-by: Joshua Kinard <>
Fixes: 46011e6ea392 ("MIPS: Make set_pte() SMP safe.)
Cc: David Daney <>
Cc: Linux/MIPS <>
Signed-off-by: Ralf Baechle <>
[bwh: Backported to 3.2:
 - Use {LL,SC}_INSN not __{LL,SC}
 - Use literal arch=r4000 instead of MIPS_ISA_ARCH_LEVEL since R6 is not
Signed-off-by: Ben Hutchings <>
6 years agoACPI / sysfs: fix error code in get_status()
Dan Carpenter [Thu, 5 May 2016 13:23:04 +0000 (16:23 +0300)]
ACPI / sysfs: fix error code in get_status()

commit f18ebc211e259d4f591e39e74b2aa2de226c9a1d upstream.

The problem with ornamental, do-nothing gotos is that they lead to
"forgot to set the error code" bugs.  We should be returning -EINVAL
here but we don't.  It leads to an uninitalized variable in

    drivers/acpi/sysfs.c:603 counter_show()
    error: uninitialized symbol 'status'.

Fixes: 1c8fce27e275 (ACPI: introduce drivers/acpi/sysfs.c)
Signed-off-by: Dan Carpenter <>
Signed-off-by: Rafael J. Wysocki <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: option: add even more ZTE device ids
Lei Liu [Wed, 4 May 2016 08:34:22 +0000 (16:34 +0800)]
USB: serial: option: add even more ZTE device ids

commit 74d2a91aec97ab832790c9398d320413ad185321 upstream.

Add even more ZTE device ids.

Signed-off-by: lei liu <>
[johan: rebase and replace commit message ]
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: option: add more ZTE device ids
lei liu [Tue, 3 May 2016 21:44:19 +0000 (14:44 -0700)]
USB: serial: option: add more ZTE device ids

commit f0d09463c59c2d764a6c6d492cbe6d2c77f27153 upstream.

More ZTE device ids.

Signed-off-by: lei liu <>
[properly sort them - gregkh]
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Johan Hovold <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agotty: vt, return error when con_startup fails
Jiri Slaby [Tue, 3 May 2016 15:05:54 +0000 (17:05 +0200)]
tty: vt, return error when con_startup fails

commit 6798df4c5fe0a7e6d2065cf79649a794e5ba7114 upstream.

When csw->con_startup() fails in do_register_con_driver, we return no
error (i.e. 0). This was changed back in 2006 by commit 3e795de763.
Before that we used to return -ENODEV.

So fix the return value to be -ENODEV in that case again.

Fixes: 3e795de763 ("VT binding: Add binding/unbinding support for the VT console")
Signed-off-by: Jiri Slaby <>
Reported-by: "Dan Carpenter" <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
6 years agochar: Drop bogus dependency of DEVPORT on !M68K
Geert Uytterhoeven [Mon, 11 Apr 2016 08:40:55 +0000 (10:40 +0200)]
char: Drop bogus dependency of DEVPORT on !M68K

commit 309124e2648d668a0c23539c5078815660a4a850 upstream.

According to full-history-linux commit d3794f4fa7c3edc3 ("[PATCH] M68k
update (part 25)"), port operations are allowed on m68k if CONFIG_ISA is

However, commit 153dcc54df826d2f ("[PATCH] mem driver: fix conditional
on isa i/o support") accidentally changed an "||" into an "&&",
disabling it completely on m68k. This logic was retained when
introducing the DEVPORT symbol in commit 4f911d64e04a44c4 ("Make
/dev/port conditional on config symbol").

Drop the bogus dependency on !M68K to fix this.

Fixes: 153dcc54df826d2f ("[PATCH] mem driver: fix conditional on isa i/o support")
Signed-off-by: Geert Uytterhoeven <>
Tested-by: Al Stone <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
6 years agoext4: fix hang when processing corrupted orphaned inode list
Theodore Ts'o [Sat, 30 Apr 2016 04:48:54 +0000 (00:48 -0400)]
ext4: fix hang when processing corrupted orphaned inode list

commit c9eb13a9105e2e418f72e46a2b6da3f49e696902 upstream.

If the orphaned inode list contains inode #5, ext4_iget() returns a
bad inode (since the bootloader inode should never be referenced
directly).  Because of the bad inode, we end up processing the inode
repeatedly and this hangs the machine.

This can be reproduced via:

   mke2fs -t ext4 /tmp/foo.img 100
   debugfs -w -R "ssv last_orphan 5" /tmp/foo.img
   mount -o loop /tmp/foo.img /mnt

(But don't do this if you are using an unpatched kernel if you care
about the system staying functional.  :-)

This bug was found by the port of American Fuzzy Lop into the kernel
to find file system problems[1].  (Since it *only* happens if inode #5
shows up on the orphan list --- 3, 7, 8, etc. won't do it, it's not
surprising that AFL needed two hours before it found it.)


Reported by: Vegard Nossum <>
Signed-off-by: Theodore Ts'o <>
Signed-off-by: Ben Hutchings <>
6 years agoaacraid: Fix for aac_command_thread hang
Raghava Aditya Renukunta [Tue, 26 Apr 2016 06:31:57 +0000 (23:31 -0700)]
aacraid: Fix for aac_command_thread hang

commit fc4bf75ea300a5e62a2419f89dd0e22189dd7ab7 upstream.

Typically under error conditions, it is possible for aac_command_thread()
to miss the wakeup from kthread_stop() and go back to sleep, causing it
to hang aac_shutdown.

In the observed scenario, the adapter is not functioning correctly and so
aac_fib_send() never completes (or time-outs depending on how it was
called). Shortly after aac_command_thread() starts it performs
aac_fib_send(SendHostTime) which hangs. When aac_probe_one
/aac_get_adapter_info send time outs, kthread_stop is called which breaks
the command thread out of it's hang.

The code will still go back to sleep in schedule_timeout() without
checking kthread_should_stop() so it causes aac_probe_one to hang until
the schedule_timeout() which is 30 minutes.

Fixed by: Adding another kthread_should_stop() before schedule_timeout()
Signed-off-by: Raghava Aditya Renukunta <>
Reviewed-by: Johannes Thumshirn <>
Signed-off-by: Martin K. Petersen <>
Signed-off-by: Ben Hutchings <>
6 years agobtrfs: bugfix: handle FS_IOC32_{GETFLAGS,SETFLAGS,GETVERSION} in btrfs_ioctl
Luke Dashjr [Thu, 29 Oct 2015 08:22:21 +0000 (08:22 +0000)]
btrfs: bugfix: handle FS_IOC32_{GETFLAGS,SETFLAGS,GETVERSION} in btrfs_ioctl

commit 4c63c2454eff996c5e27991221106eb511f7db38 upstream.

32-bit ioctl uses these rather than the regular FS_IOC_* versions. They can
be handled in btrfs using the same code. Without this, 32-bit {ch,ls}attr

Signed-off-by: Luke Dashjr <>
Reviewed-by: Josef Bacik <>
Reviewed-by: David Sterba <>
Signed-off-by: David Sterba <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agocrypto: s5p-sss - fix incorrect usage of scatterlists api
Marek Szyprowski [Tue, 26 Apr 2016 07:29:26 +0000 (09:29 +0200)]
crypto: s5p-sss - fix incorrect usage of scatterlists api

commit d1497977fecb9acce05988d6322ad415ef93bb39 upstream.

sg_dma_len() macro can be used only on scattelists which are mapped, so
all calls to it before dma_map_sg() are invalid. Replace them by proper
check for direct sg segment length read.

Fixes: a49e490c7a8a ("crypto: s5p-sss - add S5PV210 advanced crypto engine support")
Fixes: 9e4a1100a445 ("crypto: s5p-sss - Handle unaligned buffers")
Signed-off-by: Marek Szyprowski <>
Reviewed-by: Krzysztof Kozlowski <>
Acked-by: Vladimir Zapolskiy <>
Signed-off-by: Herbert Xu <>
[bwh: Backported to 3.2: unaligned DMA is unsupported so there is a different
 set of calls to replace]
Signed-off-by: Ben Hutchings <>
6 years agoalpha/PCI: Call iomem_is_exclusive() for IORESOURCE_MEM, but not IORESOURCE_IO
Bjorn Helgaas [Mon, 25 Apr 2016 20:59:50 +0000 (15:59 -0500)]
alpha/PCI: Call iomem_is_exclusive() for IORESOURCE_MEM, but not IORESOURCE_IO

commit c20e128030caf0537d5e906753eac1c28fefdb75 upstream.

The alpha pci_mmap_resource() is used for both IORESOURCE_MEM and
IORESOURCE_IO resources, but iomem_is_exclusive() is only applicable for

Call iomem_is_exclusive() only for IORESOURCE_MEM resources, and do it
earlier to match the generic version of pci_mmap_resource().

Fixes: 10a0ef39fbd1 ("PCI/alpha: pci sysfs resources")
Signed-off-by: Bjorn Helgaas <>
CC: Ivan Kokshaysky <>
Signed-off-by: Ben Hutchings <>
6 years agoPCI: Supply CPU physical address (not bus address) to iomem_is_exclusive()
Bjorn Helgaas [Fri, 8 Apr 2016 00:15:14 +0000 (17:15 -0700)]
PCI: Supply CPU physical address (not bus address) to iomem_is_exclusive()

commit ca620723d4ff9ea7ed484eab46264c3af871b9ae upstream.

iomem_is_exclusive() requires a CPU physical address, but on some arches we
supplied a PCI bus address instead.

On most arches, pci_resource_to_user(res) returns "res->start", which is a
CPU physical address.  But on microblaze, mips, powerpc, and sparc, it
returns the PCI bus address corresponding to "res->start".

The result is that pci_mmap_resource() may fail when it shouldn't (if the
bus address happens to match an existing resource), or it may succeed when
it should fail (if the resource is exclusive but the bus address doesn't
match it).

Call iomem_is_exclusive() with "res->start", which is always a CPU physical
address, not the result of pci_resource_to_user().

Fixes: e8de1481fd71 ("resource: allow MMIO exclusivity for device drivers")
Suggested-by: Yinghai Lu <>
Signed-off-by: Bjorn Helgaas <>
CC: Arjan van de Ven <>
Signed-off-by: Ben Hutchings <>
6 years agocrypto: s5p-sss - Fix missed interrupts when working with 8 kB blocks
Krzysztof Kozlowski [Fri, 22 Apr 2016 12:15:23 +0000 (14:15 +0200)]
crypto: s5p-sss - Fix missed interrupts when working with 8 kB blocks

commit 79152e8d085fd64484afd473ef6830b45518acba upstream.

The tcrypt testing module on Exynos5422-based Odroid XU3/4 board failed on
testing 8 kB size blocks:

$ sudo modprobe tcrypt sec=1 mode=500
testing speed of async ecb(aes) (ecb-aes-s5p) encryption
test 0 (128 bit key, 16 byte blocks): 21971 operations in 1 seconds (351536 bytes)
test 1 (128 bit key, 64 byte blocks): 21731 operations in 1 seconds (1390784 bytes)
test 2 (128 bit key, 256 byte blocks): 21932 operations in 1 seconds (5614592 bytes)
test 3 (128 bit key, 1024 byte blocks): 21685 operations in 1 seconds (22205440 bytes)
test 4 (128 bit key, 8192 byte blocks):

This was caused by a race issue of missed BRDMA_DONE ("Block cipher
Receiving DMA") interrupt. Device starts processing the data in DMA mode
immediately after setting length of DMA block: receiving (FCBRDMAL) or
transmitting (FCBTDMAL). The driver sets these lengths from interrupt
handler through s5p_set_dma_indata() function (or xxx_setdata()).

However the interrupt handler was first dealing with receive buffer
(dma-unmap old, dma-map new, set receive block length which starts the
operation), then with transmit buffer and finally was clearing pending
interrupts (FCINTPEND). Because of the time window between setting
receive buffer length and clearing pending interrupts, the operation on
receive buffer could end already and driver would miss new interrupt.

User manual for Exynos5422 confirms in example code that setting DMA
block lengths should be the last operation.

The tcrypt hang could be also observed in following blocked-task dmesg:

INFO: task modprobe:258 blocked for more than 120 seconds.
      Not tainted 4.6.0-rc4-next-20160419-00005-g9eac8b7b7753-dirty #42
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
modprobe        D c06b09d8     0   258    256 0x00000000
[<c06b09d8>] (__schedule) from [<c06b0f24>] (schedule+0x40/0xac)
[<c06b0f24>] (schedule) from [<c06b49f8>] (schedule_timeout+0x124/0x178)
[<c06b49f8>] (schedule_timeout) from [<c06b17fc>] (wait_for_common+0xb8/0x144)
[<c06b17fc>] (wait_for_common) from [<bf0013b8>] (test_acipher_speed+0x49c/0x740 [tcrypt])
[<bf0013b8>] (test_acipher_speed [tcrypt]) from [<bf003e8c>] (do_test+0x2240/0x30ec [tcrypt])
[<bf003e8c>] (do_test [tcrypt]) from [<bf008048>] (tcrypt_mod_init+0x48/0xa4 [tcrypt])
[<bf008048>] (tcrypt_mod_init [tcrypt]) from [<c010177c>] (do_one_initcall+0x3c/0x16c)
[<c010177c>] (do_one_initcall) from [<c0191ff0>] (do_init_module+0x5c/0x1ac)
[<c0191ff0>] (do_init_module) from [<c0185610>] (load_module+0x1a30/0x1d08)
[<c0185610>] (load_module) from [<c0185ab0>] (SyS_finit_module+0x8c/0x98)
[<c0185ab0>] (SyS_finit_module) from [<c01078c0>] (ret_fast_syscall+0x0/0x3c)

Fixes: a49e490c7a8a ("crypto: s5p-sss - add S5PV210 advanced crypto engine support")
Signed-off-by: Krzysztof Kozlowski <>
Tested-by: Marek Szyprowski <>
Signed-off-by: Herbert Xu <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoath5k: Change led pin configuration for compaq c700 laptop
Joseph Salisbury [Mon, 14 Mar 2016 18:51:48 +0000 (14:51 -0400)]
ath5k: Change led pin configuration for compaq c700 laptop

commit 7b9bc799a445aea95f64f15e0083cb19b5789abe upstream.

Commit 09c9bae26b0d3c9472cb6ae45010460a2cee8b8d ("ath5k: add led pin
configuration for compaq c700 laptop") added a pin configuration for the Compaq
c700 laptop.  However, the polarity of the led pin is reversed.  It should be
red for wifi off and blue for wifi on, but it is the opposite.  This bug was
reported in the following bug report:

Fixes: 09c9bae26b0d3c9472cb6ae45010460a2cee8b8d ("ath5k: add led pin configuration for compaq c700 laptop")
Signed-off-by: Joseph Salisbury <>
Signed-off-by: Kalle Valo <>
Signed-off-by: Ben Hutchings <>
6 years agoregmap: cache: Fix typo in cache_bypass parameter description
Andrew F. Davis [Wed, 23 Mar 2016 14:26:33 +0000 (09:26 -0500)]
regmap: cache: Fix typo in cache_bypass parameter description

commit 267c85860308d36bc163c5573308cd024f659d7c upstream.

Setting the flag 'cache_bypass' will bypass the cache not the hardware.
Fix this comment here.

Fixes: 0eef6b0415f5 ("regmap: Fix doc comment")
Signed-off-by: Andrew F. Davis <>
Signed-off-by: Mark Brown <>
Signed-off-by: Ben Hutchings <>
6 years agoLinux 3.2.81 v3.2.81
Ben Hutchings [Wed, 15 Jun 2016 20:28:15 +0000 (21:28 +0100)]
Linux 3.2.81

6 years agonet: fix a kernel infoleak in x25 module
Kangjie Lu [Sun, 8 May 2016 16:10:14 +0000 (12:10 -0400)]
net: fix a kernel infoleak in x25 module

commit 79e48650320e6fba48369fccf13fd045315b19b8 upstream.

Stack object "dte_facilities" is allocated in x25_rx_call_request(),
which is supposed to be initialized in x25_negotiate_facilities.
However, 5 fields (8 bytes in total) are not initialized. This
object is then copied to userland via copy_to_user, thus infoleak

Signed-off-by: Kangjie Lu <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: fix infoleak in rtnetlink
Kangjie Lu [Tue, 3 May 2016 20:46:24 +0000 (16:46 -0400)]
net: fix infoleak in rtnetlink

commit 5f8e44741f9f216e33736ea4ec65ca9ac03036e6 upstream.

The stack object “map” has a total size of 32 bytes. Its last 4
bytes are padding generated by compiler. These padding bytes are
not initialized and sent out via “nla_put”.

Signed-off-by: Kangjie Lu <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context, indentation]
Signed-off-by: Ben Hutchings <>
6 years agonet: fix infoleak in llc
Kangjie Lu [Tue, 3 May 2016 20:35:05 +0000 (16:35 -0400)]
net: fix infoleak in llc

commit b8670c09f37bdf2847cc44f36511a53afc6161fd upstream.

The stack object “info” has a total size of 12 bytes. Its last byte
is padding which is not initialized and leaked via “put_cmsg”.

Signed-off-by: Kangjie Lu <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonf_conntrack: avoid kernel pointer value leak in slab name
Linus Torvalds [Sat, 14 May 2016 18:11:44 +0000 (11:11 -0700)]
nf_conntrack: avoid kernel pointer value leak in slab name

commit 31b0b385f69d8d5491a4bca288e25e63f1d945d0 upstream.

The slab name ends up being visible in the directory structure under
/sys, and even if you don't have access rights to the file you can see
the filenames.

Just use a 64-bit counter instead of the pointer to the 'net' structure
to generate a unique name.

This code will go away in 4.7 when the conntrack code moves to a single
kmemcache, but this is the backportable simple solution to avoiding
leaking kernel pointers to user space.

Fixes: 5b3501faa874 ("netfilter: nf_conntrack: per netns nf_conntrack_cachep")
Signed-off-by: Linus Torvalds <>
Acked-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoget_rock_ridge_filename(): handle malformed NM entries
Al Viro [Thu, 5 May 2016 20:25:35 +0000 (16:25 -0400)]
get_rock_ridge_filename(): handle malformed NM entries

commit 99d825822eade8d827a1817357cbf3f889a552d6 upstream.

Payloads of NM entries are not supposed to contain NUL.  When we run
into such, only the part prior to the first NUL goes into the
concatenation (i.e. the directory entry name being encoded by a bunch
of NM entries).  We do stop when the amount collected so far + the
claimed amount in the current NM entry exceed 254.  So far, so good,
but what we return as the total length is the sum of *claimed*
sizes, not the actual amount collected.  And that can grow pretty
large - not unlimited, since you'd need to put CE entries in
between to be able to get more than the maximum that could be
contained in one isofs directory entry / continuation chunk and
we are stop once we'd encountered 32 CEs, but you can get about 8Kb
easily.  And that's what will be passed to readdir callback as the
name length.  8Kb __copy_to_user() from a buffer allocated by

Signed-off-by: Al Viro <>
Signed-off-by: Ben Hutchings <>
6 years agoparisc: fix a bug when syscall number of tracee is __NR_Linux_syscalls
Dmitry V. Levin [Wed, 27 Apr 2016 01:56:11 +0000 (04:56 +0300)]
parisc: fix a bug when syscall number of tracee is __NR_Linux_syscalls

commit f0b22d1bb2a37a665a969e95785c75a4f49d1499 upstream.

Do not load one entry beyond the end of the syscall table when the
syscall number of a traced process equals to __NR_Linux_syscalls.
Similar bug with regular processes was fixed by commit 3bb457af4fa8
("[PARISC] Fix bug when syscall nr is __NR_Linux_syscalls").

This bug was found by strace test suite.

Signed-off-by: Dmitry V. Levin <>
Acked-by: Helge Deller <>
Signed-off-by: Helge Deller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoproc: prevent accessing /proc/<PID>/environ until it's ready
Mathias Krause [Thu, 5 May 2016 23:22:26 +0000 (16:22 -0700)]
proc: prevent accessing /proc/<PID>/environ until it's ready

commit 8148a73c9901a8794a50f950083c00ccf97d43b3 upstream.

If /proc/<PID>/environ gets read before the envp[] array is fully set up
in create_{aout,elf,elf_fdpic,flat}_tables(), we might end up trying to
read more bytes than are actually written, as env_start will already be
set but env_end will still be zero, making the range calculation
underflow, allowing to read beyond the end of what has been written.

Fix this as it is done for /proc/<PID>/cmdline by testing env_end for
zero.  It is, apparently, intentionally set last in create_*_tables().

This bug was found by the PaX size_overflow plugin that detected the
arithmetic underflow of 'this_len = env_end - (env_start + src)' when
env_end is still zero.

The expected consequence is that userland trying to access
/proc/<PID>/environ of a not yet fully set up process may get
inconsistent data as we're in the middle of copying in the environment

Signed-off-by: Mathias Krause <>
Cc: Emese Revfy <>
Cc: Pax Team <>
Cc: Al Viro <>
Cc: Mateusz Guzik <>
Cc: Alexey Dobriyan <>
Cc: Cyrill Gorcunov <>
Cc: Jarod Wilson <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agocrypto: hash - Fix page length clamping in hash walk
Herbert Xu [Wed, 4 May 2016 09:52:56 +0000 (17:52 +0800)]
crypto: hash - Fix page length clamping in hash walk

commit 13f4bb78cf6a312bbdec367ba3da044b09bf0e29 upstream.

The crypto hash walk code is broken when supplied with an offset
greater than or equal to PAGE_SIZE.  This patch fixes it by adjusting
walk->pg and walk->offset when this happens.

Reported-by: Steffen Klassert <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Ben Hutchings <>
6 years agoACPICA: Dispatcher: Update thread ID for recursive method calls
Prarit Bhargava [Wed, 4 May 2016 05:48:56 +0000 (13:48 +0800)]
ACPICA: Dispatcher: Update thread ID for recursive method calls

commit 93d68841a23a5779cef6fb9aa0ef32e7c5bd00da upstream.

ACPICA commit 7a3bd2d962f221809f25ddb826c9e551b916eb25

Set the mutex owner thread ID.
Original patch from: Prarit Bhargava <>

Signed-off-by: Prarit Bhargava <>
Tested-by: Andy Lutomirski <> # On a Dell XPS 13 9350
Signed-off-by: Bob Moore <>
Signed-off-by: Lv Zheng <>
Signed-off-by: Rafael J. Wysocki <>
Signed-off-by: Ben Hutchings <>
6 years agodrm/radeon: make sure vertical front porch is at least 1
Alex Deucher [Mon, 2 May 2016 22:53:27 +0000 (18:53 -0400)]
drm/radeon: make sure vertical front porch is at least 1

commit 3104b8128d4d646a574ed9d5b17c7d10752cd70b upstream.

hw doesn't like a 0 value.

Signed-off-by: Alex Deucher <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoMinimal fix-up of bad hashing behavior of hash_64()
Linus Torvalds [Mon, 2 May 2016 19:46:42 +0000 (12:46 -0700)]
Minimal fix-up of bad hashing behavior of hash_64()

commit 689de1d6ca95b3b5bd8ee446863bf81a4883ea25 upstream.

This is a fairly minimal fixup to the horribly bad behavior of hash_64()
with certain input patterns.

In particular, because the multiplicative value used for the 64-bit hash
was intentionally bit-sparse (so that the multiply could be done with
shifts and adds on architectures without hardware multipliers), some
bits did not get spread out very much.  In particular, certain fairly
common bit ranges in the input (roughly bits 12-20: commonly with the
most information in them when you hash things like byte offsets in files
or memory that have block factors that mean that the low bits are often
zero) would not necessarily show up much in the result.

There's a bigger patch-series brewing to fix up things more completely,
but this is the fairly minimal fix for the 64-bit hashing problem.  It
simply picks a much better constant multiplier, spreading the bits out a
lot better.

NOTE! For 32-bit architectures, the bad old hash_64() remains the same
for now, since 64-bit multiplies are expensive.  The bigger hashing
cleanup will replace the 32-bit case with something better.

The new constants were picked by George Spelvin who wrote that bigger
cleanup series.  I just picked out the constants and part of the comment
from that series.

Cc: George Spelvin <>
Cc: Thomas Gleixner <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoMake hash_64() use a 64-bit multiply when appropriate
Linus Torvalds [Sat, 13 Sep 2014 18:24:03 +0000 (11:24 -0700)]
Make hash_64() use a 64-bit multiply when appropriate

commit 23d0db76ffa13ffb95229946e4648568c3c29db5 upstream.

The hash_64() function historically does the multiply by the
GOLDEN_RATIO_PRIME_64 number with explicit shifts and adds, because
unlike the 32-bit case, gcc seems unable to turn the constant multiply
into the more appropriate shift and adds when required.

However, that means that we generate those shifts and adds even when the
architecture has a fast multiplier, and could just do it better in

Use the now-cleaned-up CONFIG_ARCH_HAS_FAST_MULTIPLIER (together with
"is it a 64-bit architecture") to decide whether to use an integer
multiply or the explicit sequence of shift/add instructions.

Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
[bwh: This has no immediate effect in 3.2 because nothing defines
 CONFIG_ARCH_HAS_FAST_MULTIPLIER. However the following fix removes
 that condition.]

6 years agoEDAC: i7core, sb_edac: Don't return NOTIFY_BAD from mce_decoder callback
Tony Luck [Fri, 29 Apr 2016 13:42:25 +0000 (15:42 +0200)]
EDAC: i7core, sb_edac: Don't return NOTIFY_BAD from mce_decoder callback

commit c4fc1956fa31003bfbe4f597e359d751568e2954 upstream.

Both of these drivers can return NOTIFY_BAD, but this terminates
processing other callbacks that were registered later on the chain.
Since the driver did nothing to log the error it seems wrong to prevent
other interested parties from seeing it. E.g. neither of them had even
bothered to check the type of the error to see if it was a memory error
before the return NOTIFY_BAD.

Signed-off-by: Tony Luck <>
Acked-by: Aristeu Rozanski <>
Acked-by: Mauro Carvalho Chehab <>
Cc: linux-edac <>
Signed-off-by: Borislav Petkov <>
Signed-off-by: Ben Hutchings <>
6 years agomm/huge_memory: replace VM_NO_THP VM_BUG_ON with actual VMA check
Konstantin Khlebnikov [Thu, 28 Apr 2016 23:18:32 +0000 (16:18 -0700)]
mm/huge_memory: replace VM_NO_THP VM_BUG_ON with actual VMA check

commit 3486b85a29c1741db99d0c522211c82d2b7a56d0 upstream.

Khugepaged detects own VMAs by checking vm_file and vm_ops but this way
it cannot distinguish private /dev/zero mappings from other special
mappings like /dev/hpet which has no vm_ops and popultes PTEs in mmap.

This fixes false-positive VM_BUG_ON and prevents installing THP where
they are not expected.

Fixes: 78f11a255749 ("mm: thp: fix /dev/zero MAP_PRIVATE and vm_flags cleanups")
Signed-off-by: Konstantin Khlebnikov <>
Reported-by: Dmitry Vyukov <>
Acked-by: Vlastimil Babka <>
Acked-by: Kirill A. Shutemov <>
Cc: Dmitry Vyukov <>
Cc: Andrea Arcangeli <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2:
 - The assertions use VM_BUG_ON() and also check is_linear_pfn_mapping();
   keep that check
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agothp: introduce hugepage_vma_check()
Bob Liu [Wed, 12 Dec 2012 00:00:39 +0000 (16:00 -0800)]
thp: introduce hugepage_vma_check()

commit fa475e517adb422cb3492e636195f9b2c0d009c8 upstream.

Multiple places do the same check.

Signed-off-by: Bob Liu <>
Cc: Andrea Arcangeli <>
Cc: Michal Hocko <>
Cc: Minchan Kim <>
Cc: Ni zhan Chen <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2:
 - Also move the is_linear_pfn_mapping() test
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoIB/security: Restrict use of the write() interface
Jason Gunthorpe [Mon, 11 Apr 2016 01:13:13 +0000 (19:13 -0600)]
IB/security: Restrict use of the write() interface

commit e6bd18f57aad1a2d1ef40e646d03ed0f2515c9e3 upstream.

The drivers/infiniband stack uses write() as a replacement for
bi-directional ioctl().  This is not safe. There are ways to
trigger write calls that result in the return structure that
is normally written to user space being shunted off to user
specified kernel memory instead.

For the immediate repair, detect and deny suspicious accesses to
the write API.

For long term, update the user space libraries and the kernel API
to something that doesn't present the same security vulnerabilities
(likely a structured ioctl() interface).

The impacted uAPI interfaces are generally only available if
hardware from drivers/infiniband is installed in the system.

Reported-by: Jann Horn <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Jason Gunthorpe <>
[ Expanded check to all known write() entry points ]
Signed-off-by: Doug Ledford <>
[bwh: Backported to 3.2:
 - Drop changes to hfi1
 - include/rdma/ib.h didn't exist, so create it with the usual header guard
   and include it in drivers/infiniband/core/ucma.c
 - ipath_write() has the same problem, so add the same restriction there]
Signed-off-by: Ben Hutchings <>
6 years agodrm/i915: Fix system resume if PCI device remained enabled
Imre Deak [Mon, 18 Apr 2016 11:45:54 +0000 (14:45 +0300)]
drm/i915: Fix system resume if PCI device remained enabled

commit dab9a2663f4e688106c041f7cd2797a721382f0a upstream.

During system resume we depended on pci_enable_device() also putting the
device into PCI D0 state. This won't work if the PCI device was already
enabled but still in D3 state. This is because pci_enable_device() is
refcounted and will not change the HW state if called with a non-zero
refcount. Leaving the device in D3 will make all subsequent device
accesses fail.

This didn't cause a problem most of the time, since we resumed with an
enable refcount of 0. But it fails at least after module reload because
after that we also happen to leak a PCI device enable reference: During
probing we call drm_get_pci_dev() which will enable the PCI device, but
during device removal drm_put_dev() won't disable it. This is a bug of
its own in DRM core, but without much harm as it only leaves the PCI
device enabled. Fixing it is also a bit more involved, due to DRM
mid-layering and because it affects non-i915 drivers too. The fix in
this patch is valid regardless of the problem in DRM core.

- Add a code comment about the relation of this fix to the freeze/thaw
  vs. the suspend/resume phases. (Ville)
- Add a code comment about the inconsistent ordering of set power state
  and device enable calls. (Chris)

CC: Ville Syrjälä <>
CC: Chris Wilson <>
Signed-off-by: Imre Deak <>
Reviewed-by: Ville Syrjälä <>
(cherry picked from commit 44410cd0bfb26bde9288da34c190cc9267d42a20)
Signed-off-by: Jani Nikula <>
[bwh: Backported to 3.2:
 - Return error code directly
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: cp210x: add Straizona Focusers device ids
Jasem Mutlaq [Tue, 19 Apr 2016 07:38:27 +0000 (10:38 +0300)]
USB: serial: cp210x: add Straizona Focusers device ids

commit 613ac23a46e10d4d4339febdd534fafadd68e059 upstream.

Adding VID:PID for Straizona Focusers to cp210x driver.

Signed-off-by: Jasem Mutlaq <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agoUSB: serial: cp210x: add ID for Link ECU
Mike Manning [Mon, 18 Apr 2016 12:13:23 +0000 (12:13 +0000)]
USB: serial: cp210x: add ID for Link ECU

commit 1d377f4d690637a0121eac8701f84a0aa1e69a69 upstream.

The Link ECU is an aftermarket ECU computer for vehicles that provides
full tuning abilities as well as datalogging and displaying capabilities
via the USB to Serial adapter built into the device.

Signed-off-by: Mike Manning <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
6 years agobatman-adv: Fix broadcast/ogm queue limit on a removed interface
Linus Lüssing [Fri, 11 Mar 2016 13:04:49 +0000 (14:04 +0100)]
batman-adv: Fix broadcast/ogm queue limit on a removed interface

commit c4fdb6cff2aa0ae740c5f19b6f745cbbe786d42f upstream.

When removing a single interface while a broadcast or ogm packet is
still pending then we will free the forward packet without releasing the
queue slots again.

This patch is supposed to fix this issue.

Fixes: 6d5808d4ae1b ("batman-adv: Add missing hardif_free_ref in forw_packet_free")
Signed-off-by: Linus Lüssing <>
[ fix conflicts with current version]
Signed-off-by: Sven Eckelmann <>
Signed-off-by: Marek Lindner <>
Signed-off-by: Antonio Quartulli <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agobatman-adv: Reduce refcnt of removed router when updating route
Sven Eckelmann [Sun, 20 Mar 2016 11:27:53 +0000 (12:27 +0100)]
batman-adv: Reduce refcnt of removed router when updating route

commit d1a65f1741bfd9c69f9e4e2ad447a89b6810427d upstream.

_batadv_update_route rcu_derefences orig_ifinfo->router outside of a
spinlock protected region to print some information messages to the debug
log. But this pointer is not checked again when the new pointer is assigned
in the spinlock protected region. Thus is can happen that the value of
orig_ifinfo->router changed in the meantime and thus the reference counter
of the wrong router gets reduced after the spinlock protected region.

Just rcu_dereferencing the value of orig_ifinfo->router inside the spinlock
protected region (which also set the new pointer) is enough to get the
correct old router object.

Fixes: e1a5382f978b ("batman-adv: Make orig_node->router an rcu protected pointer")
Signed-off-by: Sven Eckelmann <>
Signed-off-by: Marek Lindner <>
Signed-off-by: Antonio Quartulli <>
[bwh: Backported to 3.2: s/orig_ifinfo/orig_node/]
Signed-off-by: Ben Hutchings <>
6 years agobatman-adv: Check skb size before using encapsulated ETH+VLAN header
Sven Eckelmann [Fri, 26 Feb 2016 16:56:13 +0000 (17:56 +0100)]
batman-adv: Check skb size before using encapsulated ETH+VLAN header

commit c78296665c3d81f040117432ab9e1cb125521b0c upstream.

The encapsulated ethernet and VLAN header may be outside the received
ethernet frame. Thus the skb buffer size has to be checked before it can be
parsed to find out if it encapsulates another batman-adv packet.

Fixes: 420193573f11 ("batman-adv: softif bridge loop avoidance")
Signed-off-by: Sven Eckelmann <>
Signed-off-by: Marek Lindner <>
Signed-off-by: Antonio Quartulli <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agox86/mm/xen: Suppress hugetlbfs in PV guests
Jan Beulich [Thu, 21 Apr 2016 06:27:04 +0000 (00:27 -0600)]
x86/mm/xen: Suppress hugetlbfs in PV guests

commit 103f6112f253017d7062cd74d17f4a514ed4485c upstream.

Huge pages are not normally available to PV guests. Not suppressing
hugetlbfs use results in an endless loop of page faults when user mode
code tries to access a hugetlbfs mapped area (since the hypervisor
denies such PTEs to be created, but error indications can't be
propagated out of xen_set_pte_at(), just like for various of its
siblings), and - once killed in an oops like this:

  kernel BUG at .../fs/hugetlbfs/inode.c:428!
  invalid opcode: 0000 [#1] SMP
  RIP: e030:[<ffffffff811c333b>]  [<ffffffff811c333b>] remove_inode_hugepages+0x25b/0x320
  Call Trace:
   [<ffffffff811c3415>] hugetlbfs_evict_inode+0x15/0x40
   [<ffffffff81167b3d>] evict+0xbd/0x1b0
   [<ffffffff8116514a>] __dentry_kill+0x19a/0x1f0
   [<ffffffff81165b0e>] dput+0x1fe/0x220
   [<ffffffff81150535>] __fput+0x155/0x200
   [<ffffffff81079fc0>] task_work_run+0x60/0xa0
   [<ffffffff81063510>] do_exit+0x160/0x400
   [<ffffffff810637eb>] do_group_exit+0x3b/0xa0
   [<ffffffff8106e8bd>] get_signal+0x1ed/0x470
   [<ffffffff8100f854>] do_signal+0x14/0x110
   [<ffffffff810030e9>] prepare_exit_to_usermode+0xe9/0xf0
   [<ffffffff814178a5>] retint_user+0x8/0x13

This is CVE-2016-3961 / XSA-174.

Reported-by: Vitaly Kuznetsov <>
Signed-off-by: Jan Beulich <>
Cc: Andrew Morton <>
Cc: Andy Lutomirski <>
Cc: Boris Ostrovsky <>
Cc: Borislav Petkov <>
Cc: Brian Gerst <>
Cc: David Vrabel <>
Cc: Denys Vlasenko <>
Cc: H. Peter Anvin <>
Cc: Juergen Gross <>
Cc: Linus Torvalds <>
Cc: Luis R. Rodriguez <>
Cc: Peter Zijlstra <>
Cc: Thomas Gleixner <>
Cc: Toshi Kani <>
Cc: xen-devel <>
Signed-off-by: Ingo Molnar <>
Signed-off-by: Ben Hutchings <>
6 years agos390/hugetlb: add hugepages_supported define
Dominik Dingel [Fri, 17 Jul 2015 23:23:39 +0000 (16:23 -0700)]
s390/hugetlb: add hugepages_supported define

commit 7f9be77555bb2e52de84e9dddf7b4eb20cc6e171 upstream.

On s390 we only can enable hugepages if the underlying hardware/hypervisor
also does support this.  Common code now would assume this to be
signaled by setting HPAGE_SHIFT to 0.  But on s390, where we only
support one hugepage size, there is a link between HPAGE_SHIFT and

So instead of setting HPAGE_SHIFT to 0, we will implement the check for
the hardware capability.

Signed-off-by: Dominik Dingel <>
Acked-by: Martin Schwidefsky <>
Cc: Heiko Carstens <>
Cc: Christian Borntraeger <>
Cc: Michael Holzheu <>
Cc: Gerald Schaefer <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
6 years agomm: hugetlb: allow hugepages_supported to be architecture specific
Dominik Dingel [Fri, 17 Jul 2015 23:23:37 +0000 (16:23 -0700)]
mm: hugetlb: allow hugepages_supported to be architecture specific

commit 2531c8cf56a640cd7d17057df8484e570716a450 upstream.

s390 has a constant hugepage size, by setting HPAGE_SHIFT we also change
e.g. the pageblock_order, which should be independent in respect to
hugepage support.

With this patch every architecture is free to define how to check
for hugepage support.

Signed-off-by: Dominik Dingel <>
Acked-by: Martin Schwidefsky <>
Cc: Heiko Carstens <>
Cc: Christian Borntraeger <>
Cc: Michael Holzheu <>
Cc: Gerald Schaefer <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
6 years agohugetlb: ensure hugepage access is denied if hugepages are not supported
Nishanth Aravamudan [Tue, 6 May 2014 19:50:00 +0000 (12:50 -0700)]
hugetlb: ensure hugepage access is denied if hugepages are not supported

commit 457c1b27ed56ec472d202731b12417bff023594a upstream.

Currently, I am seeing the following when I `mount -t hugetlbfs /none
/dev/hugetlbfs`, and then simply do a `ls /dev/hugetlbfs`.  I think it's
related to the fact that hugetlbfs is properly not correctly setting
itself up in this state?:

  Unable to handle kernel paging request for data at address 0x00000031
  Faulting instruction address: 0xc000000000245710
  Oops: Kernel access of bad area, sig: 11 [#1]
  SMP NR_CPUS=2048 NUMA pSeries

In KVM guests on Power, in a guest not backed by hugepages, we see the

  AnonHugePages:         0 kB
  HugePages_Total:       0
  HugePages_Free:        0
  HugePages_Rsvd:        0
  HugePages_Surp:        0
  Hugepagesize:         64 kB

HPAGE_SHIFT == 0 in this configuration, which indicates that hugepages
are not supported at boot-time, but this is only checked in
hugetlb_init().  Extract the check to a helper function, and use it in a
few relevant places.

This does make hugetlbfs not supported (not registered at all) in this
environment.  I believe this is fine, as there are no valid hugepages
and that won't change at runtime.

[ use pr_info(), per Mel]
[ fix build when HPAGE_SHIFT is undefined]
Signed-off-by: Nishanth Aravamudan <>
Reviewed-by: Aneesh Kumar K.V <>
Acked-by: Mel Gorman <>
Cc: Randy Dunlap <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2:
 - Drop changes to hugetlb_show_meminfo()
 - Adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoatl2: Disable unimplemented scatter/gather feature
Ben Hutchings [Wed, 20 Apr 2016 22:23:08 +0000 (23:23 +0100)]
atl2: Disable unimplemented scatter/gather feature

commit f43bfaeddc79effbf3d0fcb53ca477cca66f3db8 upstream.

atl2 includes NETIF_F_SG in hw_features even though it has no support
for non-linear skbs.  This bug was originally harmless since the
driver does not claim to implement checksum offload and that used to
be a requirement for SG.

Now that SG and checksum offload are independent features, if you
explicitly enable SG *and* use one of the rare protocols that can use
SG without checkusm offload, this potentially leaks sensitive
information (before you notice that it just isn't working).  Therefore
this obscure bug has been designated CVE-2016-2117.

Reported-by: Justin Yackoski <>
Signed-off-by: Ben Hutchings <>
Fixes: ec5f06156423 ("net: Kill link between CSUM and SG features.")
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]

6 years agopowerpc: scan_features() updates incorrect bits for REAL_LE
Anton Blanchard [Fri, 15 Apr 2016 02:06:13 +0000 (12:06 +1000)]
powerpc: scan_features() updates incorrect bits for REAL_LE

commit 6997e57d693b07289694239e52a10d2f02c3a46f upstream.

The REAL_LE feature entry in the ibm_pa_feature struct is missing an MMU
feature value, meaning all the remaining elements initialise the wrong

This means instead of checking for byte 5, bit 0, we check for byte 0,
bit 0, and then we incorrectly set the CPU feature bit as well as MMU
feature bit 1 and CPU user feature bits 0 and 2 (5).

Checking byte 0 bit 0 (IBM numbering), means we're looking at the
"Memory Management Unit (MMU)" feature - ie. does the CPU have an MMU.
In practice that bit is set on all platforms which have the property.

This means we set CPU_FTR_REAL_LE always. In practice that seems not to
matter because all the modern cpus which have this property also
implement REAL_LE, and we've never needed to disable it.

We're also incorrectly setting MMU feature bit 1, which is:

  #define MMU_FTR_TYPE_8xx 0x00000002

Luckily the only place that looks for MMU_FTR_TYPE_8xx is in Book3E
code, which can't run on the same cpus as scan_features(). So this also
doesn't matter in practice.

Finally in the CPU user feature mask, we're setting bits 0 and 2. Bit 2
is not currently used, and bit 0 is:

  #define PPC_FEATURE_PPC_LE 0x00000001

Which says the CPU supports the old style "PPC Little Endian" mode.
Again this should be harmless in practice as no 64-bit CPUs implement
that mode.

Fix the code by adding the missing initialisation of the MMU feature.

Also add a comment marking CPU user feature bit 2 (0x4) as reserved. It
would be unsafe to start using it as old kernels incorrectly set it.

Fixes: 44ae3ab3358e ("powerpc: Free up some CPU feature bits by moving out MMU-related features")
Signed-off-by: Anton Blanchard <>
[mpe: Flesh out changelog, add comment reserving 0x4]
Signed-off-by: Michael Ellerman <>
[bwh: Backported to 3.2: adjust filename, context]
Signed-off-by: Ben Hutchings <>
6 years agoInput: pmic8xxx-pwrkey - fix algorithm for converting trigger delay
Stephen Boyd [Sun, 17 Apr 2016 12:21:42 +0000 (05:21 -0700)]
Input: pmic8xxx-pwrkey - fix algorithm for converting trigger delay

commit eda5ecc0a6b865561997e177c393f0b0136fe3b7 upstream.

The trigger delay algorithm that converts from microseconds to
the register value looks incorrect. According to most of the PMIC
documentation, the equation is

delay (Seconds) = (1 / 1024) * 2 ^ (x + 4)

except for one case where the documentation looks to have a
formatting issue and the equation looks like

delay (Seconds) = (1 / 1024) * 2 x + 4

Most likely this driver was written with the improper
documentation to begin with. According to the downstream sources
the valid delays are from 2 seconds to 1/64 second, and the
latter equation just doesn't make sense for that. Let's fix the
algorithm and the range check to match the documentation and the
downstream sources.

Reported-by: Bjorn Andersson <>
Fixes: 92d57a73e410 ("input: Add support for Qualcomm PMIC8XXX power key")
Signed-off-by: Stephen Boyd <>
Tested-by: John Stultz <>
Acked-by: Bjorn Andersson <>
Signed-off-by: Dmitry Torokhov <>
[bwh: Backported to 3.2: use pdata->kpd_trigger_delay_us not kpd_delay]
Signed-off-by: Ben Hutchings <>
6 years agousb: hcd: out of bounds access in for_each_companion
Robert Dobrowolski [Thu, 24 Mar 2016 10:30:07 +0000 (03:30 -0700)]
usb: hcd: out of bounds access in for_each_companion

commit e86103a75705c7c530768f4ffaba74cf382910f2 upstream.

On BXT platform Host Controller and Device Controller figure as
same PCI device but with different device function. HCD should
not pass data to Device Controller but only to Host Controllers.
Checking if companion device is Host Controller, otherwise skip.

Signed-off-by: Robert Dobrowolski <>
Acked-by: Alan Stern <>
Signed-off-by: Greg Kroah-Hartman <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoUSB: uas: Add a new NO_REPORT_LUNS quirk
Hans de Goede [Tue, 12 Apr 2016 10:27:09 +0000 (12:27 +0200)]
USB: uas: Add a new NO_REPORT_LUNS quirk

commit 1363074667a6b7d0507527742ccd7bbed5e3ceaa upstream.

Add a new NO_REPORT_LUNS quirk and set it for Seagate drives with
an usb-id of: 0bc2:331a, as these will fail to respond to a
REPORT_LUNS command.

Reported-and-tested-by: David Webb <>
Signed-off-by: Hans de Goede <>
Acked-by: Alan Stern <>
Signed-off-by: Greg Kroah-Hartman <>
[bwh: Backported to 3.2:
 - Adjust context
 - Drop the UAS changes]
Signed-off-by: Ben Hutchings <>
6 years agousb: xhci: fix wild pointers in xhci_mem_cleanup
Lu Baolu [Fri, 8 Apr 2016 13:25:09 +0000 (16:25 +0300)]
usb: xhci: fix wild pointers in xhci_mem_cleanup

commit 71504062a7c34838c3fccd92c447f399d3cb5797 upstream.

This patch fixes some wild pointers produced by xhci_mem_cleanup.
These wild pointers will cause system crash if xhci_mem_cleanup()
is called twice.

Reported-and-tested-by: Pengcheng Li <>
Signed-off-by: Lu Baolu <>
Signed-off-by: Mathias Nyman <>
Signed-off-by: Greg Kroah-Hartman <>
[bwh: Backported to 3.2: there's no xhci_hcd::ext_caps field to clear]
Signed-off-by: Ben Hutchings <>
6 years agousb: xhci: applying XHCI_PME_STUCK_QUIRK to Intel BXT B0 host
Rafal Redzimski [Fri, 8 Apr 2016 13:25:05 +0000 (16:25 +0300)]
usb: xhci: applying XHCI_PME_STUCK_QUIRK to Intel BXT B0 host

commit 0d46faca6f887a849efb07c1655b5a9f7c288b45 upstream.

Broxton B0 also requires XHCI_PME_STUCK_QUIRK.
Adding PCI device ID for Broxton B and adding to quirk.

Signed-off-by: Rafal Redzimski <>
Signed-off-by: Robert Dobrowolski <>
Signed-off-by: Mathias Nyman <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
6 years agonl80211: check netlink protocol in socket release notification
Dmitry Ivanov [Wed, 6 Apr 2016 14:23:18 +0000 (17:23 +0300)]
nl80211: check netlink protocol in socket release notification

commit 8f815cdde3e550e10c2736990d791f60c2ce43eb upstream.

A non-privileged user can create a netlink socket with the same port_id as
used by an existing open nl80211 netlink socket (e.g. as used by a hostapd
process) with a different protocol number.

Closing this socket will then lead to the notification going to nl80211's
socket release notification handler, and possibly cause an action such as
removing a virtual interface.

Fix this issue by checking that the netlink protocol is NETLINK_GENERIC.
Since generic netlink has no notifier chain of its own, we can't fix the
problem more generically.

Fixes: 026331c4d9b5 ("cfg80211/mac80211: allow registering for and sending action frames")
Signed-off-by: Dmitry Ivanov <>
[rewrite commit message]
Signed-off-by: Johannes Berg <>
Signed-off-by: Ben Hutchings <>
6 years agokvm: x86: do not leak guest xcr0 into host interrupt handlers
David Matlack [Wed, 30 Mar 2016 19:24:47 +0000 (12:24 -0700)]
kvm: x86: do not leak guest xcr0 into host interrupt handlers

commit fc5b7f3bf1e1414bd4e91db6918c85ace0c873a5 upstream.

An interrupt handler that uses the fpu can kill a KVM VM, if it runs
under the following conditions:
 - the guest's xcr0 register is loaded on the cpu
 - the guest's fpu context is not loaded
 - the host is using eagerfpu

Note that the guest's xcr0 register and fpu context are not loaded as
part of the atomic world switch into "guest mode". They are loaded by
KVM while the cpu is still in "host mode".

Usage of the fpu in interrupt context is gated by irq_fpu_usable(). The
interrupt handler will look something like this:

if (irq_fpu_usable()) {

        [... code that uses the fpu ...]


As long as the guest's fpu is not loaded and the host is using eager
fpu, irq_fpu_usable() returns true (interrupted_kernel_fpu_idle()
returns true). The interrupt handler proceeds to use the fpu with
the guest's xcr0 live.

kernel_fpu_begin() saves the current fpu context. If this uses
XSAVE[OPT], it may leave the xsave area in an undesirable state.
According to the SDM, during XSAVE bit i of XSTATE_BV is not modified
if bit i is 0 in xcr0. So it's possible that XSTATE_BV[i] == 1 and
xcr0[i] == 0 following an XSAVE.

kernel_fpu_end() restores the fpu context. Now if any bit i in
XSTATE_BV == 1 while xcr0[i] == 0, XRSTOR generates a #GP. The
fault is trapped and SIGSEGV is delivered to the current process.

Only pre-4.2 kernels appear to be vulnerable to this sequence of
events. Commit 653f52c ("kvm,x86: load guest FPU context more eagerly")
from 4.2 forces the guest's fpu to always be loaded on eagerfpu hosts.

This patch fixes the bug by keeping the host's xcr0 loaded outside
of the interrupts-disabled region where KVM switches into guest mode.

Suggested-by: Andy Lutomirski <>
Signed-off-by: David Matlack <>
[Move load after goto cancel_injection. - Paolo]
Signed-off-by: Paolo Bonzini <>
[bwh: Backported to 3.2:
 - Adjust context
 - Drop change in__kvm_set_xcr()]
Signed-off-by: Ben Hutchings <>
6 years agolibahci: save port map for forced port map
Srinivas Kandagatla [Fri, 1 Apr 2016 07:52:56 +0000 (08:52 +0100)]
libahci: save port map for forced port map

commit 2fd0f46cb1b82587c7ae4a616d69057fb9bd0af7 upstream.

In usecases where force_port_map is used saved_port_map is never set,
resulting in not programming the PORTS_IMPL register as part of initial
config. This patch fixes this by setting it to port_map even in case
where force_port_map is used, making it more inline with other parts of
the code.

Fixes: 566d1827df2e ("libata: disable forced PORTS_IMPL for >= AHCI 1.3")
Signed-off-by: Srinivas Kandagatla <>
Acked-by: Tejun Heo <>
Reviewed-by: Andy Gross <>
Signed-off-by: Tejun Heo <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoInput: gtco - fix crash on detecting device without endpoints
Vladis Dronov [Thu, 31 Mar 2016 17:53:42 +0000 (10:53 -0700)]
Input: gtco - fix crash on detecting device without endpoints

commit 162f98dea487206d9ab79fc12ed64700667a894d upstream.

The gtco driver expects at least one valid endpoint. If given malicious
descriptors that specify 0 for the number of endpoints, it will crash in
the probe function. Ensure there is at least one endpoint on the interface
before using it.

Also let's fix a minor coding style issue.

The full correct report of this issue can be found in the public
Red Hat Bugzilla:

Reported-by: Ralf Spenneberg <>
Signed-off-by: Vladis Dronov <>
Signed-off-by: Dmitry Torokhov <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoipmi: fix timeout calculation when bmc is disconnected
Xie XiuQi [Fri, 24 Jan 2014 20:00:52 +0000 (14:00 -0600)]
ipmi: fix timeout calculation when bmc is disconnected

commit e21404dc0ac7ac971c1e36274b48bb460463f4e5 upstream.

Loading ipmi_si module while bmc is disconnected, we found the timeout
is longer than 5 secs.  Actually it takes about 3 mins and 20

error message as below:
  Dec 12 19:08:59 linux kernel: IPMI BT: timeout in RD_WAIT [ ] 1 retries left
  Dec 12 19:08:59 linux kernel: BT: write 4 bytes seq=0x01 03 18 00 01
  Dec 12 19:12:19 linux kernel: IPMI BT: timeout in RD_WAIT [ ]
  Dec 12 19:12:19 linux kernel: failed 2 retries, sending error response
  Dec 12 19:12:19 linux kernel: IPMI: BT reset (takes 5 secs)
  Dec 12 19:12:19 linux kernel: IPMI BT: flag reset [ ]

Function wait_for_msg_done() use schedule_timeout_uninterruptible(1) to
sleep 1 tick, so we should subtract jiffies_to_usecs(1) instead of 100
usecs from timeout.

Reported-by: Hu Shiyuan <>
Signed-off-by: Xie XiuQi <>
Signed-off-by: Corey Minyard <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
6 years agox86, sparse: Do not force removal of __user when calling copy_to/from_user_nocheck()
Steven Rostedt [Fri, 3 Jan 2014 21:45:00 +0000 (16:45 -0500)]
x86, sparse: Do not force removal of __user when calling copy_to/from_user_nocheck()

commit df90ca969035d3f6c95044e272f75bf417b14245 upstream.

Commit ff47ab4ff3cdd "x86: Add 1/2/4/8 byte optimization to 64bit
__copy_{from,to}_user_inatomic" added a "_nocheck" call in between
the copy_to/from_user() and copy_user_generic(). As both the
normal and nocheck versions of theses calls use the proper __user
annotation, a typecast to remove it should not be added.
This causes sparse to spin out the following warnings:

arch/x86/include/asm/uaccess_64.h:207:47: warning: incorrect type in argument 2 (different address spaces)
arch/x86/include/asm/uaccess_64.h:207:47:    expected void const [noderef] <asn:1>*src
arch/x86/include/asm/uaccess_64.h:207:47:    got void const *<noident>
arch/x86/include/asm/uaccess_64.h:207:47: warning: incorrect type in argument 2 (different address spaces)
arch/x86/include/asm/uaccess_64.h:207:47:    expected void const [noderef] <asn:1>*src
arch/x86/include/asm/uaccess_64.h:207:47:    got void const *<noident>
arch/x86/include/asm/uaccess_64.h:207:47: warning: incorrect type in argument 2 (different address spaces)
arch/x86/include/asm/uaccess_64.h:207:47:    expected void const [noderef] <asn:1>*src
arch/x86/include/asm/uaccess_64.h:207:47:    got void const *<noident>
arch/x86/include/asm/uaccess_64.h:207:47: warning: incorrect type in argument 2 (different address spaces)
arch/x86/include/asm/uaccess_64.h:207:47:    expected void const [noderef] <asn:1>*src
arch/x86/include/asm/uaccess_64.h:207:47:    got void const *<noident>

Cc: Andi Kleen <>
Signed-off-by: Steven Rostedt <>
Signed-off-by: H. Peter Anvin <>
Signed-off-by: Ben Hutchings <>
Cc: Jaccon Bastiaansen <>
Cc: Thomas Gleixner <>
Cc: Peter Zijlstra <>
6 years agox86: Add 1/2/4/8 byte optimization to 64bit __copy_{from,to}_user_inatomic
Andi Kleen [Fri, 16 Aug 2013 21:17:19 +0000 (14:17 -0700)]
x86: Add 1/2/4/8 byte optimization to 64bit __copy_{from,to}_user_inatomic

commit ff47ab4ff3cddfa7bc1b25b990e24abe2ae474ff upstream.

The 64bit __copy_{from,to}_user_inatomic always called
copy_from_user_generic, but skipped the special optimizations for 1/2/4/8
byte accesses.

This especially hurts the futex call, which accesses the 4 byte futex
user value with a complicated fast string operation in a function call,
instead of a single movl.

Use __copy_{from,to}_user for _inatomic instead to get the same
optimizations. The only problem was the might_fault() in those functions.
So move that to new wrapper and call __copy_{f,t}_user_nocheck()
from *_inatomic directly.

32bit already did this correctly by duplicating the code.

Signed-off-by: Andi Kleen <>
Signed-off-by: H. Peter Anvin <>
Signed-off-by: Ben Hutchings <>
Cc: Jaccon Bastiaansen <>
Cc: Thomas Gleixner <>
Cc: Peter Zijlstra <>
6 years agocrypto: gcm - Fix rfc4543 decryption crash
Herbert Xu [Fri, 18 Mar 2016 14:42:40 +0000 (22:42 +0800)]
crypto: gcm - Fix rfc4543 decryption crash

This bug has already bee fixed upstream since 4.2.  However, it
was fixed during the AEAD conversion so no fix was backported to
the older kernels.

When we do an RFC 4543 decryption, we will end up writing the
ICV beyond the end of the dst buffer.  This should lead to a
crash but for some reason it was never noticed.

This patch fixes it by only writing back the ICV for encryption.

Fixes: d733ac90f9fe ("crypto: gcm - fix rfc4543 to handle async...")
Reported-by: Patrick Meyer <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
6 years agocrypto: gcm - fix rfc4543 to handle async crypto correctly
Jussi Kivilinna [Sun, 7 Apr 2013 13:43:46 +0000 (16:43 +0300)]
crypto: gcm - fix rfc4543 to handle async crypto correctly

commit d733ac90f9fe8ac284e523f9920b507555b12f6d upstream.

If the gcm cipher used by rfc4543 does not complete request immediately,
the authentication tag is not copied to destination buffer. Patch adds
correct async logic for this case.

Signed-off-by: Jussi Kivilinna <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Ben Hutchings <>
6 years agox86/microcode/amd: Do not overwrite final patch levels
Borislav Petkov [Mon, 12 Oct 2015 09:22:42 +0000 (11:22 +0200)]
x86/microcode/amd: Do not overwrite final patch levels

commit 0399f73299f1b7e04de329050f7111b362b7eeb5 upstream.

A certain number of patch levels of applied microcode should not
be overwritten by the microcode loader, otherwise bad things
will happen.

Check those and abort update if the current core has one of
those final patch levels applied by the BIOS. 32-bit needs
special handling, of course.

See for more

Tested-by: Peter Kirchgeßner <>
Signed-off-by: Borislav Petkov <>
Cc: Borislav Petkov <>
Cc: H. Peter Anvin <>
Cc: Linus Torvalds <>
Cc: Peter Zijlstra <>
Cc: Thomas Gleixner <>
Cc: Tony Luck <>
Signed-off-by: Ingo Molnar <>
Signed-off-by: Ben Hutchings <>
Cc: Henrique de Moraes Holschuh <>
6 years agox86/microcode/amd: Extract current patch level read to a function
Borislav Petkov [Mon, 12 Oct 2015 09:22:41 +0000 (11:22 +0200)]
x86/microcode/amd: Extract current patch level read to a function

commit 2eff73c0a11f19ff082a566e3429fbaaca7b8e7b upstream.

Pave the way for checking the current patch level of the
microcode in a core. We want to be able to do stuff depending on
the patch level - in this case decide whether to update or not.
But that will be added in a later patch.

Drop unused local var uci assignment, while at it.

Integrate a fix for 32-bit and CONFIG_PARAVIRT from Takashi Iwai:

 Use native_rdmsr() in check_current_patch_level() because with
 CONFIG_PARAVIRT enabled and on 32-bit, where we run before
 paging has been enabled, we cannot deref pv_info yet. Or we
 could, but we'd need to access its physical address. This way of
 fixing it is simpler. See: for the background.

Signed-off-by: Borislav Petkov <>
Cc: Borislav Petkov <>
Cc: H. Peter Anvin <>
Cc: Linus Torvalds <>
Cc: Peter Zijlstra <>
Cc: Takashi Iwai <>:
Cc: Thomas Gleixner <>
Cc: Tony Luck <>
Signed-off-by: Ingo Molnar <>
Signed-off-by: Ben Hutchings <>
Cc: Henrique de Moraes Holschuh <>
6 years agoRevert "net: validate variable length ll headers"
Ben Hutchings [Sun, 1 May 2016 16:00:43 +0000 (18:00 +0200)]
Revert "net: validate variable length ll headers"

This reverts commit b5518429e70cd783b8ca52335456172c1a0589f6, which was
commit 2793a23aacbd754dbbb5cb75093deb7e4103bace upstream.  It is
pointless unless af_packet calls the new function.

Signed-off-by: Ben Hutchings <>
6 years agoRevert "ax25: add link layer header validation function"
Ben Hutchings [Sun, 1 May 2016 15:59:44 +0000 (17:59 +0200)]
Revert "ax25: add link layer header validation function"

This reverts commit 0954b59d9f4b2dcc59f28d1f64c3a21062a64372, which was
commit ea47781c26510e5d97f80f9aceafe9065bd5e3aa upstream.  It is
pointless unless af_packet calls the new function.

Signed-off-by: Ben Hutchings <>
6 years agoLinux 3.2.80 v3.2.80
Ben Hutchings [Sat, 30 Apr 2016 22:05:28 +0000 (00:05 +0200)]
Linux 3.2.80

6 years agonetfilter: x_tables: fix unconditional helper
Florian Westphal [Tue, 22 Mar 2016 17:02:52 +0000 (18:02 +0100)]
netfilter: x_tables: fix unconditional helper

commit 54d83fc74aa9ec72794373cb47432c5f7fb1a309 upstream.

Ben Hawkes says:

 In the mark_source_chains function (net/ipv4/netfilter/ip_tables.c) it
 is possible for a user-supplied ipt_entry structure to have a large
 next_offset field. This field is not bounds checked prior to writing a
 counter value at the supplied offset.

Problem is that mark_source_chains should not have been called --
the rule doesn't have a next entry, so its supposed to return
an absolute verdict of either ACCEPT or DROP.

However, the function conditional() doesn't work as the name implies.
It only checks that the rule is using wildcard address matching.

However, an unconditional rule must also not be using any matches
(no -m args).

The underflow validator only checked the addresses, therefore
passing the 'unconditional absolute verdict' test, while
mark_source_chains also tested for presence of matches, and thus
proceeeded to the next (not-existent) rule.

Unify this so that all the callers have same idea of 'unconditional rule'.

Reported-by: Ben Hawkes <>
Signed-off-by: Florian Westphal <>
Signed-off-by: Pablo Neira Ayuso <>
Signed-off-by: Ben Hutchings <>
6 years agojme: Fix device PM wakeup API usage
Guo-Fu Tseng [Sat, 5 Mar 2016 00:11:56 +0000 (08:11 +0800)]
jme: Fix device PM wakeup API usage

commit 81422e672f8181d7ad1ee6c60c723aac649f538f upstream.

According to Documentation/power/devices.txt

The driver should not use device_set_wakeup_enable() which is the policy
for user to decide.

Using device_init_wakeup() to initialize dev->power.should_wakeup and
dev->power.can_wakeup on driver initialization.

And use device_may_wakeup() on suspend to decide if WoL function should
be enabled on NIC.

Reported-by: Diego Viola <>
Signed-off-by: Guo-Fu Tseng <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agojme: Do not enable NIC WoL functions on S0
Guo-Fu Tseng [Sat, 5 Mar 2016 00:11:55 +0000 (08:11 +0800)]
jme: Do not enable NIC WoL functions on S0

commit 0772a99b818079e628a1da122ac7ee023faed83e upstream.

Otherwise it might be back on resume right after going to suspend in
some hardware.

Reported-by: Diego Viola <>
Signed-off-by: Guo-Fu Tseng <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoipv6: Count in extension headers in skb->network_header
Jakub Sitnicki [Tue, 5 Apr 2016 16:41:08 +0000 (18:41 +0200)]
ipv6: Count in extension headers in skb->network_header

[ Upstream commit 3ba3458fb9c050718b95275a3310b74415e767e2 ]

When sending a UDPv6 message longer than MTU, account for the length
of fragmentable IPv6 extension headers in skb->network_header offset.
Same as we do in alloc_new_skb path in __ip6_append_data().

This ensures that later on __ip6_make_skb() will make space in
headroom for fragmentable extension headers:

/* move skb->data to ip header from ext header */
if (skb->data < skb_network_header(skb))
__skb_pull(skb, skb_network_offset(skb));

Prevents a splat due to skb_under_panic:

skbuff: skb_under_panic: text:ffffffff8143397b len:2126 put:14 \
head:ffff880005bacf50 data:ffff880005bacf4a tail:0x48 end:0xc0 dev:lo
------------[ cut here ]------------
kernel BUG at net/core/skbuff.c:104!
invalid opcode: 0000 [#1] KASAN
CPU: 0 PID: 160 Comm: reproducer Not tainted 4.6.0-rc2 #65
Call Trace:
 [<ffffffff813eb7b9>] skb_push+0x79/0x80
 [<ffffffff8143397b>] eth_header+0x2b/0x100
 [<ffffffff8141e0d0>] neigh_resolve_output+0x210/0x310
 [<ffffffff814eab77>] ip6_finish_output2+0x4a7/0x7c0
 [<ffffffff814efe3a>] ip6_output+0x16a/0x280
 [<ffffffff815440c1>] ip6_local_out+0xb1/0xf0
 [<ffffffff814f1115>] ip6_send_skb+0x45/0xd0
 [<ffffffff81518836>] udp_v6_send_skb+0x246/0x5d0
 [<ffffffff8151985e>] udpv6_sendmsg+0xa6e/0x1090

Reported-by: Ji Jianwen <>
Signed-off-by: Jakub Sitnicki <>
Acked-by: Hannes Frederic Sowa <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoipv4: l2tp: fix a potential issue in l2tp_ip_recv
Haishuang Yan [Sun, 3 Apr 2016 14:09:23 +0000 (22:09 +0800)]
ipv4: l2tp: fix a potential issue in l2tp_ip_recv

[ Upstream commit 5745b8232e942abd5e16e85fa9b27cc21324acf0 ]

pskb_may_pull() can change skb->data, so we have to load ptr/optr at the
right place.

Signed-off-by: Haishuang Yan <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoxfrm: Fix crash observed during device unregistration and decryption [Thu, 24 Mar 2016 04:39:50 +0000 (22:39 -0600)]
xfrm: Fix crash observed during device unregistration and decryption

[ Upstream commit 071d36bf21bcc837be00cea55bcef8d129e7f609 ]

A crash is observed when a decrypted packet is processed in receive
path. get_rps_cpus() tries to dereference the skb->dev fields but it
appears that the device is freed from the poison pattern.

[<ffffffc000af58ec>] get_rps_cpu+0x94/0x2f0
[<ffffffc000af5f94>] netif_rx_internal+0x140/0x1cc
[<ffffffc000af6094>] netif_rx+0x74/0x94
[<ffffffc000bc0b6c>] xfrm_input+0x754/0x7d0
[<ffffffc000bc0bf8>] xfrm_input_resume+0x10/0x1c
[<ffffffc000ba6eb8>] esp_input_done+0x20/0x30
[<ffffffc0000b64c8>] process_one_work+0x244/0x3fc
[<ffffffc0000b7324>] worker_thread+0x2f8/0x418
[<ffffffc0000bb40c>] kthread+0xe0/0xec

     |    dev = 0xFFFFFFC08B688000,
     |    skb = 0xFFFFFFC0C76AAC00 -> (
     |      dev = 0xFFFFFFC08B688000 -> (
     |        name =
     |        name_hlist = (next = 0xAAAAAAAAAAAAAAAA, pprev =

Following are the sequence of events observed -

- Encrypted packet in receive path from netdevice is queued
- Encrypted packet queued for decryption (asynchronous)
- Netdevice brought down and freed
- Packet is decrypted and returned through callback in esp_input_done
- Packet is queued again for process in network stack using netif_rx

Since the device appears to have been freed, the dereference of
skb->dev in get_rps_cpus() leads to an unhandled page fault

Fix this by holding on to device reference when queueing packets
asynchronously and releasing the reference on call back return.

v2: Make the change generic to xfrm as mentioned by Steffen and
update the title to xfrm

Suggested-by: Herbert Xu <>
Signed-off-by: Jerome Stanislaus <>
Signed-off-by: Subash Abhinov Kasiviswanathan <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoqlge: Fix receive packets drop.
Manish Chopra [Tue, 15 Mar 2016 11:13:45 +0000 (07:13 -0400)]
qlge: Fix receive packets drop.

[ Upstream commit 2c9a266afefe137bff06bbe0fc48b4d3b3cb348c ]

When running small packets [length < 256 bytes] traffic, packets were
being dropped due to invalid data in those packets which were
delivered by the driver upto the stack. Using pci_dma_sync_single_for_cpu
ensures copying latest and updated data into skb from the receive buffer.

Signed-off-by: Sony Chacko <>
Signed-off-by: Manish Chopra <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agofarsync: fix off-by-one bug in fst_add_one
Arnd Bergmann [Mon, 14 Mar 2016 14:18:35 +0000 (15:18 +0100)]
farsync: fix off-by-one bug in fst_add_one

[ Upstream commit e725a66c0202b5f36c2f9d59d26a65c53bbf21f7 ]

gcc-6 finds an out of bounds access in the fst_add_one function
when calculating the end of the mmio area:

drivers/net/wan/farsync.c: In function 'fst_add_one':
drivers/net/wan/farsync.c:418:53: error: index 2 denotes an offset greater than size of 'u8[2][8192] {aka unsigned char[2][8192]}' [-Werror=array-bounds]
 #define BUF_OFFSET(X)   (BFM_BASE + offsetof(struct buf_window, X))
include/linux/compiler-gcc.h:158:21: note: in definition of macro '__compiler_offsetof'
  __builtin_offsetof(a, b)
drivers/net/wan/farsync.c:418:37: note: in expansion of macro 'offsetof'
 #define BUF_OFFSET(X)   (BFM_BASE + offsetof(struct buf_window, X))
drivers/net/wan/farsync.c:2519:36: note: in expansion of macro 'BUF_OFFSET'
                                  + BUF_OFFSET ( txBuffer[i][NUM_TX_BUFFER][0]);

The warning is correct, but not critical because this appears
to be a write-only variable that is set by each WAN driver but
never accessed afterwards.

I'm taking the minimal fix here, using the correct pointer by
pointing 'mem_end' to the last byte inside of the register area
as all other WAN drivers do, rather than the first byte outside of
it. An alternative would be to just remove the mem_end member

Signed-off-by: Arnd Bergmann <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agomacvtap: always pass ethernet header in linear
Willem de Bruijn [Tue, 8 Mar 2016 20:18:54 +0000 (15:18 -0500)]
macvtap: always pass ethernet header in linear

[ Upstream commit 8e2ad4113ce4671686740f808ff2795395c39eef ]

The stack expects link layer headers in the skb linear section.
Macvtap can create skbs with llheader in frags in edge cases:
when (IFF_VNET_HDR is off or vnet_hdr.hdr_len < ETH_HLEN) and
prepad + len > PAGE_SIZE and vnet_hdr.flags has no or bad csum.

Add checks to ensure linear is always at least ETH_HLEN.
At this point, len is already ensured to be >= ETH_HLEN.

For backwards compatiblity, rounds up short vnet_hdr.hdr_len.
This differs from tap and packet, which return an error.

Fixes b9fb9ee07e67 ("macvtap: add GSO/csum offload support")
Signed-off-by: Willem de Bruijn <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: don't use macvtap16_to_cpu()]
Signed-off-by: Ben Hutchings <>
6 years agosh_eth: fix NULL pointer dereference in sh_eth_ring_format()
Sergei Shtylyov [Mon, 7 Mar 2016 22:36:28 +0000 (01:36 +0300)]
sh_eth: fix NULL pointer dereference in sh_eth_ring_format()

[ Upstream commit c1b7fca65070bfadca94dd53a4e6b71cd4f69715 ]

In a low memory situation, if netdev_alloc_skb() fails on a first RX ring
loop iteration  in sh_eth_ring_format(), 'rxdesc' is still NULL.  Avoid
kernel oops by adding the 'rxdesc' check after the loop.

Reported-by: Wolfram Sang <>
Signed-off-by: Sergei Shtylyov <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoax25: add link layer header validation function
Willem de Bruijn [Thu, 10 Mar 2016 02:58:33 +0000 (21:58 -0500)]
ax25: add link layer header validation function

[ Upstream commit ea47781c26510e5d97f80f9aceafe9065bd5e3aa ]

As variable length protocol, AX25 fails link layer header validation
tests based on a minimum length. header_ops.validate allows protocols
to validate headers that are shorter than hard_header_len. Implement
this callback for AX25.

See also

Signed-off-by: Willem de Bruijn <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agonet: validate variable length ll headers
Willem de Bruijn [Thu, 10 Mar 2016 02:58:32 +0000 (21:58 -0500)]
net: validate variable length ll headers

[ Upstream commit 2793a23aacbd754dbbb5cb75093deb7e4103bace ]

Netdevice parameter hard_header_len is variously interpreted both as
an upper and lower bound on link layer header length. The field is
used as upper bound when reserving room at allocation, as lower bound
when validating user input in PF_PACKET.

Clarify the definition to be maximum header length. For validation
of untrusted headers, add an optional validate member to header_ops.

Allow bypassing of validation by passing CAP_SYS_RAWIO, for instance
for deliberate testing of corrupt input. In this case, pad trailing
bytes, as some device drivers expect completely initialized headers.

See also

Signed-off-by: Willem de Bruijn <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: net_device has inline comments instead of kernel-doc]
Signed-off-by: Ben Hutchings <>
6 years agocdc_ncm: toggle altsetting to force reset before setup
Bjørn Mork [Thu, 3 Mar 2016 21:20:53 +0000 (22:20 +0100)]
cdc_ncm: toggle altsetting to force reset before setup

[ Upstream commit 48906f62c96cc2cd35753e59310cb70eb08cc6a5 ]

Some devices will silently fail setup unless they are reset first.
This is necessary even if the data interface is already in
altsetting 0, which it will be when the device is probed for the
first time.  Briefly toggling the altsetting forces a function
reset regardless of the initial state.

This fixes a setup problem observed on a number of Huawei devices,
appearing to operate in NTB-32 mode even if we explicitly set them
to NTB-16 mode.

Signed-off-by: Bjørn Mork <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: hard-code 1 for data_altsetting]
Signed-off-by: Ben Hutchings <>
6 years agosctp: lack the check for ports in sctp_v6_cmp_addr
Xin Long [Sun, 28 Feb 2016 02:03:51 +0000 (10:03 +0800)]
sctp: lack the check for ports in sctp_v6_cmp_addr

[ Upstream commit 40b4f0fd74e46c017814618d67ec9127ff20f157 ]

As the member .cmp_addr of sctp_af_inet6, sctp_v6_cmp_addr should also check
the port of addresses, just like sctp_v4_cmp_addr, cause it's invoked by

Now sctp_v6_cmp_addr just check the port when two addresses have different
family, and lack the port check for two ipv6 addresses. that will make
sctp_hash_cmp() cannot work well.

so fix it by adding ports comparison in sctp_v6_cmp_addr().

Signed-off-by: Xin Long <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: jme: fix suspend/resume on JMC260
Diego Viola [Tue, 23 Feb 2016 15:04:04 +0000 (12:04 -0300)]
net: jme: fix suspend/resume on JMC260

[ Upstream commit ee50c130c82175eaa0820c96b6d3763928af2241 ]

The JMC260 network card fails to suspend/resume because the call to
jme_start_irq() was too early, moving the call to jme_start_irq() after
the call to jme_reset_link() makes it work.

Prior this change suspend/resume would fail unless /sys/power/pm_async=0
was explicitly specified.

Relevant bug report:

Signed-off-by: Diego Viola <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoserial: sh-sci: Remove cpufreq notifier to fix crash/deadlock
Geert Uytterhoeven [Mon, 4 Apr 2016 14:36:25 +0000 (16:36 +0200)]
serial: sh-sci: Remove cpufreq notifier to fix crash/deadlock

commit ff1cab374ad98f4b9f408525ca9c08992b4ed784 upstream.

The BSP team noticed that there is spin/mutex lock issue on sh-sci when
CPUFREQ is used.  The issue is that the notifier function may call
mutex_lock() while the spinlock is held, which can lead to a BUG().
This may happen if CPUFREQ is changed while another CPU calls

Taking the spinlock was added to the notifier function in commit
e552de2413edad1a ("sh-sci: add platform device private data"), to
protect the list of serial ports against modification during traversal.
At that time the Common Clock Framework didn't exist yet, and
clk_get_rate() just returned clk->rate without taking a mutex.
Note that since commit d535a2305facf9b4 ("serial: sh-sci: Require a
device per port mapping."), there's no longer a list of serial ports to
traverse, and taking the spinlock became superfluous.

To fix the issue, just remove the cpufreq notifier:
  1. The notifier doesn't work correctly: all it does is update the
     stored clock rate; it does not update the divider in the hardware.
     The divider will only be updated when calling sci_set_termios().
     I believe this was broken back in 2004, when the old
     drivers/char/sh-sci.c driver (where the notifier did update the
     divider) was replaced by drivers/serial/sh-sci.c (where the
     notifier just updated port->uartclk).
     Cfr. full-history-linux commits 6f8deaef2e9675d9 ("[PATCH] sh: port
     sh-sci driver to the new API") and 3f73fe878dc9210a ("[PATCH]
     Remove old sh-sci driver").
  2. On modern SoCs, the sh-sci parent clock rate is no longer related
     to the CPU clock rate anyway, so using a cpufreq notifier is

Signed-off-by: Geert Uytterhoeven <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agobio: return EINTR if copying to user space got interrupted
Hannes Reinecke [Fri, 12 Feb 2016 08:39:15 +0000 (09:39 +0100)]
bio: return EINTR if copying to user space got interrupted

commit 2d99b55d378c996b9692a0c93dd25f4ed5d58934 upstream.

Commit 35dc248383bbab0a7203fca4d722875bc81ef091 introduced a check for
current->mm to see if we have a user space context and only copies data
if we do. Now if an IO gets interrupted by a signal data isn't copied
into user space any more (as we don't have a user space context) but
user space isn't notified about it.

This patch modifies the behaviour to return -EINTR from bio_uncopy_user()
to notify userland that a signal has interrupted the syscall, otherwise
it could lead to a situation where the caller may get a buffer with
no data returned.

This can be reproduced by issuing SG_IO ioctl()s in one thread while
constantly sending signals to it.

Fixes: 35dc248 [SCSI] sg: Fix user memory corruption when SG_IO is interrupted by a signal
Signed-off-by: Johannes Thumshirn <>
Signed-off-by: Hannes Reinecke <>
Signed-off-by: Jens Axboe <>
[bwh: Backported to 3.2:
 - Adjust filename
 - Put the assignment in the existing 'else' block]
Signed-off-by: Ben Hutchings <>
6 years agosctp: Fix port hash table size computation
Neil Horman [Thu, 18 Feb 2016 21:10:57 +0000 (16:10 -0500)]
sctp: Fix port hash table size computation

[ Upstream commit d9749fb5942f51555dc9ce1ac0dbb1806960a975 ]

Dmitry Vyukov noted recently that the sctp_port_hashtable had an error in
its size computation, observing that the current method never guaranteed
that the hashsize (measured in number of entries) would be a power of two,
which the input hash function for that table requires.  The root cause of
the problem is that two values need to be computed (one, the allocation
order of the storage requries, as passed to __get_free_pages, and two the
number of entries for the hash table).  Both need to be ^2, but for
different reasons, and the existing code is simply computing one order
value, and using it as the basis for both, which is wrong (i.e. it assumes
that ((1<<order)*PAGE_SIZE)/sizeof(bucket) is still ^2 when its not).

To fix this, we change the logic slightly.  We start by computing a goal
allocation order (which is limited by the maximum size hash table we want
to support.  Then we attempt to allocate that size table, decreasing the
order until a successful allocation is made.  Then, with the resultant
successful order we compute the number of buckets that hash table supports,
which we then round down to the nearest power of two, giving us the number
of entries the table actually supports.

I've tested this locally here, using non-debug and spinlock-debug kernels,
and the number of entries in the hashtable consistently work out to be
powers of two in all cases.

Signed-off-by: Neil Horman <>
Reported-by: Dmitry Vyukov <>
CC: Dmitry Vyukov <>
CC: Vladislav Yasevich <>
CC: "David S. Miller" <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agopppoe: fix reference counting in PPPoE proxy
Guillaume Nault [Mon, 15 Feb 2016 16:01:10 +0000 (17:01 +0100)]
pppoe: fix reference counting in PPPoE proxy

[ Upstream commit 29e73269aa4d36f92b35610c25f8b01c789b0dc8 ]

Drop reference on the relay_po socket when __pppoe_xmit() succeeds.
This is already handled correctly in the error path.

Signed-off-by: Guillaume Nault <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv4: fix memory leaks in ip_cmsg_send() callers
Eric Dumazet [Thu, 4 Feb 2016 14:23:28 +0000 (06:23 -0800)]
ipv4: fix memory leaks in ip_cmsg_send() callers

[ Upstream commit 919483096bfe75dda338e98d56da91a263746a0a ]

Dmitry reported memory leaks of IP options allocated in
ip_cmsg_send() when/if this function returns an error.

Callers are responsible for the freeing.

Many thanks to Dmitry for the report and diagnostic.

Reported-by: Dmitry Vyukov <>
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 agonet/ipv6: add sysctl option accept_ra_min_hop_limit
Hangbin Liu [Thu, 30 Jul 2015 06:28:42 +0000 (14:28 +0800)]
net/ipv6: add sysctl option accept_ra_min_hop_limit

[ Upstream commit 8013d1d7eafb0589ca766db6b74026f76b7f5cb4 ]

Commit 6fd99094de2b ("ipv6: Don't reduce hop limit for an interface")
disabled accept hop limit from RA if it is smaller than the current hop
limit for security stuff. But this behavior kind of break the RFC definition.

RFC 4861, 6.3.4.  Processing Received Router Advertisements
   A Router Advertisement field (e.g., Cur Hop Limit, Reachable Time,
   and Retrans Timer) may contain a value denoting that it is
   unspecified.  In such cases, the parameter should be ignored and the
   host should continue using whatever value it is already using.

   If the received Cur Hop Limit value is non-zero, the host SHOULD set
   its CurHopLimit variable to the received value.

So add sysctl option accept_ra_min_hop_limit to let user choose the minimum
hop limit value they can accept from RA. And set default to 1 to meet RFC

Signed-off-by: Hangbin Liu <>
Acked-by: YOSHIFUJI Hideaki <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Adjust filename, context
 - Number DEVCONF enumerators explicitly to match upstream]
Signed-off-by: Ben Hutchings <>
6 years agoipv6/udp: use sticky pktinfo egress ifindex on connect()
Paolo Abeni [Fri, 29 Jan 2016 11:30:20 +0000 (12:30 +0100)]
ipv6/udp: use sticky pktinfo egress ifindex on connect()

[ Upstream commit 1cdda91871470f15e79375991bd2eddc6e86ddb1 ]

Currently, the egress interface index specified via IPV6_PKTINFO
is ignored by __ip6_datagram_connect(), so that RFC 3542 section 6.7
can be subverted when the user space application calls connect()
before sendmsg().
Fix it by initializing properly flowi6_oif in connect() before
performing the route lookup.

Signed-off-by: Paolo Abeni <>
Acked-by: Hannes Frederic Sowa <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agonet: dp83640: Fix tx timestamp overflow handling.
Manfred Rudigier [Wed, 20 Jan 2016 10:22:28 +0000 (11:22 +0100)]
net: dp83640: Fix tx timestamp overflow handling.

[ Upstream commit 81e8f2e930fe76b9814c71b9d87c30760b5eb705 ]

PHY status frames are not reliable, the PHY may not be able to send them
during heavy receive traffic. This overflow condition is signaled by the
PHY in the next status frame, but the driver did not make use of it.
Instead it always reported wrong tx timestamps to user space after an
overflow happened because it assigned newly received tx timestamps to old
packets in the queue.

This commit fixes this issue by clearing the tx timestamp queue every time
an overflow happens, so that no timestamps are delivered for overflow
packets. This way time stamping will continue correctly after an overflow.

Signed-off-by: Manfred Rudigier <>
Acked-by: Richard Cochran <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoaf_iucv: Validate socket address length in iucv_sock_bind()
Ursula Braun [Tue, 19 Jan 2016 09:41:33 +0000 (10:41 +0100)]
af_iucv: Validate socket address length in iucv_sock_bind()

[ Upstream commit 52a82e23b9f2a9e1d429c5207f8575784290d008 ]

Signed-off-by: Ursula Braun <>
Reported-by: Dmitry Vyukov <>
Reviewed-by: Evgeny Cherkashin <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agoipv6: update skb->csum when CE mark is propagated
Eric Dumazet [Fri, 15 Jan 2016 12:56:56 +0000 (04:56 -0800)]
ipv6: update skb->csum when CE mark is propagated

[ Upstream commit 34ae6a1aa0540f0f781dd265366036355fdc8930 ]

When a tunnel decapsulates the outer header, it has to comply
with RFC 6080 and eventually propagate CE mark into inner header.

It turns out IP6_ECN_set_ce() does not correctly update skb->csum
for CHECKSUM_COMPLETE packets, triggering infamous "hw csum failure"
messages and stack traces.

Signed-off-by: Eric Dumazet <>
Acked-by: Herbert Xu <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2:
 - Adjust context
 - Add skb argument to other callers of IP6_ECN_set_ce()]
Signed-off-by: Ben Hutchings <>
6 years agophonet: properly unshare skbs in phonet_rcv()
Eric Dumazet [Tue, 12 Jan 2016 16:58:00 +0000 (08:58 -0800)]
phonet: properly unshare skbs in phonet_rcv()

[ Upstream commit 7aaed57c5c2890634cfadf725173c7c68ea4cb4f ]

Ivaylo Dimitrov reported a regression caused by commit 7866a621043f
("dev: add per net_device packet type chains").

skb->dev becomes NULL and we crash in __netif_receive_skb_core().

Before above commit, different kind of bugs or corruptions could happen
without major crash.

But the root cause is that phonet_rcv() can queue skb without checking
if skb is shared or not.

Many thanks to Ivaylo Dimitrov for his help, diagnosis and tests.

Reported-by: Ivaylo Dimitrov <>
Tested-by: Ivaylo Dimitrov <>
Signed-off-by: Eric Dumazet <>
Cc: Remi Denis-Courmont <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agotcp_yeah: don't set ssthresh below 2
Neal Cardwell [Mon, 11 Jan 2016 18:42:43 +0000 (13:42 -0500)]
tcp_yeah: don't set ssthresh below 2

[ Upstream commit 83d15e70c4d8909d722c0d64747d8fb42e38a48f ]

For tcp_yeah, use an ssthresh floor of 2, the same floor used by Reno
and CUBIC, per RFC 5681 (equation 4).

tcp_yeah_ssthresh() was sometimes returning a 0 or negative ssthresh
value if the intended reduction is as big or bigger than the current
cwnd. Congestion control modules should never return a zero or
negative ssthresh. A zero ssthresh generally results in a zero cwnd,
causing the connection to stall. A negative ssthresh value will be
interpreted as a u32 and will set a target cwnd for PRR near 4

Oleksandr Natalenko reported that a system using tcp_yeah with ECN
could see a warning about a prior_cwnd of 0 in
tcp_cwnd_reduction(). Testing verified that this was due to
tcp_yeah_ssthresh() misbehaving in this way.

Reported-by: Oleksandr Natalenko <>
Signed-off-by: Neal Cardwell <>
Signed-off-by: Yuchung Cheng <>
Signed-off-by: Eric Dumazet <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
6 years agobridge: Only call /sbin/bridge-stp for the initial network namespace
Hannes Frederic Sowa [Tue, 5 Jan 2016 09:46:00 +0000 (10:46 +0100)]
bridge: Only call /sbin/bridge-stp for the initial network namespace

[ Upstream commit ff62198553e43cdffa9d539f6165d3e83f8a42bc ]

[I stole this patch from Eric Biederman. He wrote:]

> There is no defined mechanism to pass network namespace information
> into /sbin/bridge-stp therefore don't even try to invoke it except
> for bridge devices in the initial network namespace.
> It is possible for unprivileged users to cause /sbin/bridge-stp to be
> invoked for any network device name which if /sbin/bridge-stp does not
> guard against unreasonable arguments or being invoked twice on the
> same network device could cause problems.

[Hannes: changed patch using netns_eq]

Cc: Eric W. Biederman <>
Signed-off-by: Eric W. Biederman <>
Signed-off-by: Hannes Frederic Sowa <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
6 years agoconnector: bump skb->users before callback invocation
Florian Westphal [Thu, 31 Dec 2015 13:26:33 +0000 (14:26 +0100)]
connector: bump skb->users before callback invocation

[ Upstream commit 55285bf09427c5abf43ee1d54e892f352092b1f1 ]

Dmitry reports memleak with syskaller program.
Problem is that connector bumps skb usecount but might not invoke callback.

So move skb_get to where we invoke the callback.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Florian Westphal <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>