7 years agoALSA: hdsp: Fix wrong boolean ctl value accesses
Takashi Iwai [Mon, 29 Feb 2016 13:26:43 +0000 (14:26 +0100)]
ALSA: hdsp: Fix wrong boolean ctl value accesses

commit eab3c4db193f5fcccf70e884de9a922ca2c63d80 upstream.

snd-hdsp driver accesses enum item values (int) instead of boolean
values (long) wrongly for some ctl elements.  This patch fixes them.

Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: hdspm: Fix wrong boolean ctl value accesses
Takashi Iwai [Mon, 29 Feb 2016 13:25:16 +0000 (14:25 +0100)]
ALSA: hdspm: Fix wrong boolean ctl value accesses

commit 537e48136295c5860a92138c5ea3959b9542868b upstream.

snd-hdspm driver accesses enum item values (int) instead of boolean
values (long) wrongly for some ctl elements.  This patch fixes them.

Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: drop change to snd_hdspm_put_system_sample_rate()]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: timer: Fix broken compat timer user status ioctl
Takashi Iwai [Sun, 28 Feb 2016 10:36:14 +0000 (11:36 +0100)]
ALSA: timer: Fix broken compat timer user status ioctl

commit 3a72494ac2a3bd229db941d51e7efe2f6ccd947b upstream.

The timer user status compat ioctl returned the bogus struct used for
64bit architectures instead of the 32bit one.  This patch addresses
it to return the proper struct.

Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agohpfs: don't truncate the file when delete fails
Mikulas Patocka [Thu, 25 Feb 2016 17:17:38 +0000 (18:17 +0100)]
hpfs: don't truncate the file when delete fails

commit b6853f78e763d42c7a158d8de3549c9827c604ab upstream.

The delete opration can allocate additional space on the HPFS filesystem
due to btree split. The HPFS driver checks in advance if there is
available space, so that it won't corrupt the btree if we run out of space
during splitting.

If there is not enough available space, the HPFS driver attempted to
truncate the file, but this results in a deadlock since the commit
7dd29d8d865efdb00c0542a5d2c87af8c52ea6c7 ("HPFS: Introduce a global mutex
and lock it on every callback from VFS").

This patch removes the code that tries to truncate the file and -ENOSPC is
returned instead. If the user hits -ENOSPC on delete, he should try to
delete other files (that are stored in a leaf btree node), so that the
delete operation will make some space for deleting the file stored in
non-leaf btree node.

Reported-by: Al Viro <>
Signed-off-by: Mikulas Patocka <>
Signed-off-by: Al Viro <>
[bwh: Backported to 3.2: deleted code is slightly different]
Signed-off-by: Ben Hutchings <>
7 years agomm: thp: fix SMP race condition between THP page fault and MADV_DONTNEED
Andrea Arcangeli [Fri, 26 Feb 2016 23:19:28 +0000 (15:19 -0800)]
mm: thp: fix SMP race condition between THP page fault and MADV_DONTNEED

commit ad33bb04b2a6cee6c1f99fabb15cddbf93ff0433 upstream.

pmd_trans_unstable()/pmd_none_or_trans_huge_or_clear_bad() were
introduced to locklessy (but atomically) detect when a pmd is a regular
(stable) pmd or when the pmd is unstable and can infinitely transition
from pmd_none() and pmd_trans_huge() from under us, while only holding
the mmap_sem for reading (for writing not).

While holding the mmap_sem only for reading, MADV_DONTNEED can run from
under us and so before we can assume the pmd to be a regular stable pmd
we need to compare it against pmd_none() and pmd_trans_huge() in an
atomic way, with pmd_trans_unstable().  The old pmd_trans_huge() left a
tiny window for a race.

Useful applications are unlikely to notice the difference as doing
MADV_DONTNEED concurrently with a page fault would lead to undefined

[ tidy up comment grammar/layout]
Signed-off-by: Andrea Arcangeli <>
Reported-by: Kirill A. Shutemov <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoipr: Fix regression when loading firmware
Gabriel Krisman Bertazi [Thu, 25 Feb 2016 16:54:20 +0000 (13:54 -0300)]
ipr: Fix regression when loading firmware

commit 21b81716c6bff24cda52dc75588455f879ddbfe9 upstream.

Commit d63c7dd5bcb9 ("ipr: Fix out-of-bounds null overwrite") removed
the end of line handling when storing the update_fw sysfs attribute.
This changed the userpace API because it started refusing writes
terminated by a line feed, which broke the update tools we already have.

This patch re-adds that handling, so both a write terminated by a line
feed or not can make it through with the update.

Fixes: d63c7dd5bcb9 ("ipr: Fix out-of-bounds null overwrite")
Signed-off-by: Gabriel Krisman Bertazi <>
Cc: Insu Yun <>
Acked-by: Brian King <>
Signed-off-by: Martin K. Petersen <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoipr: Fix out-of-bounds null overwrite
Insu Yun [Wed, 6 Jan 2016 17:44:01 +0000 (12:44 -0500)]
ipr: Fix out-of-bounds null overwrite

commit d63c7dd5bcb9441af0526d370c43a65ca2c980d9 upstream.

Return value of snprintf is not bound by size value, 2nd argument.
Return value is number of printed chars, can be larger than 2nd
argument.  Therefore, it can write null byte out of bounds ofbuffer.
Since snprintf puts null, it does not need to put additional null byte.

Signed-off-by: Insu Yun <>
Reviewed-by: Shane Seymour <>
Signed-off-by: Martin K. Petersen <>
Signed-off-by: Ben Hutchings <>
7 years agolibata: Align ata_device's id on a cacheline
Harvey Hunt [Wed, 24 Feb 2016 15:16:43 +0000 (15:16 +0000)]
libata: Align ata_device's id on a cacheline

commit 4ee34ea3a12396f35b26d90a094c75db95080baa upstream.

The id buffer in ata_device is a DMA target, but it isn't explicitly
cacheline aligned. Due to this, adjacent fields can be overwritten with
stale data from memory on non coherent architectures. As a result, the
kernel is sometimes unable to communicate with an ATA device.

Fix this by ensuring that the id buffer is cacheline aligned.

This issue is similar to that fixed by Commit 84bda12af31f
("libata: align ap->sector_buf").

Signed-off-by: Harvey Hunt <>
Signed-off-by: Tejun Heo <>
Signed-off-by: Ben Hutchings <>
7 years agoFix directory hardlinks from deleted directories
David Woodhouse [Mon, 1 Feb 2016 14:04:46 +0000 (14:04 +0000)]
Fix directory hardlinks from deleted directories

commit be629c62a603e5935f8177fd8a19e014100a259e upstream.

When a directory is deleted, we don't take too much care about killing off
all the dirents that belong to it — on the basis that on remount, the scan
will conclude that the directory is dead anyway.

This doesn't work though, when the deleted directory contained a child
directory which was moved *out*. In the early stages of the fs build
we can then end up with an apparent hard link, with the child directory
appearing both in its true location, and as a child of the original
directory which are this stage of the mount process we don't *yet* know
is defunct.

To resolve this, take out the early special-casing of the "directories
shall not have hard links" rule in jffs2_build_inode_pass1(), and let the
normal nlink processing happen for directories as well as other inodes.

Then later in the build process we can set ic->pino_nlink to the parent
inode#, as is required for directories during normal operaton, instead
of the nlink. And complain only *then* about hard links which are still
in evidence even after killing off all the unreachable paths.

Reported-by: Liu Song <>
Signed-off-by: David Woodhouse <>
Signed-off-by: Ben Hutchings <>
7 years agojffs2: Fix page lock / f->sem deadlock
David Woodhouse [Mon, 1 Feb 2016 12:37:20 +0000 (12:37 +0000)]
jffs2: Fix page lock / f->sem deadlock

commit 49e91e7079febe59a20ca885a87dd1c54240d0f1 upstream.

With this fix, all code paths should now be obtaining the page lock before

Reported-by: Szabó Tamás <>
Tested-by: Thomas Betker <>
Signed-off-by: David Woodhouse <>
Signed-off-by: Ben Hutchings <>
7 years agoRevert "jffs2: Fix lock acquisition order bug in jffs2_write_begin"
Thomas Betker [Tue, 10 Nov 2015 21:18:15 +0000 (22:18 +0100)]
Revert "jffs2: Fix lock acquisition order bug in jffs2_write_begin"

commit 157078f64b8a9cd7011b6b900b2f2498df850748 upstream.

This reverts commit 5ffd3412ae55
("jffs2: Fix lock acquisition order bug in jffs2_write_begin").

The commit modified jffs2_write_begin() to remove a deadlock with
jffs2_garbage_collect_live(), but this introduced new deadlocks found
by multiple users. page_lock() actually has to be called before
mutex_lock(&c->alloc_sem) or mutex_lock(&f->sem) because
jffs2_write_end() and jffs2_readpage() are called with the page locked,
and they acquire c->alloc_sem and f->sem, resp.

In other words, the lock order in jffs2_write_begin() was correct, and
it is the jffs2_garbage_collect_live() path that has to be changed.

Revert the commit to get rid of the new deadlocks, and to clear the way
for a better fix of the original deadlock.

Reported-by: Deng Chao <>
Reported-by: Ming Liu <>
Reported-by: wangzaiwei <>
Signed-off-by: Thomas Betker <>
Signed-off-by: David Woodhouse <>
[bwh: Backported to 3.2: use D1(printk(KERN_DEBUG ...)) instead of
 jffs2_dbg(1, ...)]
Signed-off-by: Ben Hutchings <>
7 years agoKVM: async_pf: do not warn on page allocation failures
Christian Borntraeger [Fri, 19 Feb 2016 12:11:46 +0000 (13:11 +0100)]
KVM: async_pf: do not warn on page allocation failures

commit d7444794a02ff655eda87e3cc54e86b940e7736f upstream.

In async_pf we try to allocate with NOWAIT to get an element quickly
or fail. This code also handle failures gracefully. Lets silence
potential page allocation failures under load.

qemu-system-s39: page allocation failure: order:0,mode:0x2200000
Call Trace:
([<00000000001146b8>] show_trace+0xf8/0x148)
[<000000000011476a>] show_stack+0x62/0xe8
[<00000000004a36b8>] dump_stack+0x70/0x98
[<0000000000272c3a>] warn_alloc_failed+0xd2/0x148
[<000000000027709e>] __alloc_pages_nodemask+0x94e/0xb38
[<00000000002cd36a>] new_slab+0x382/0x400
[<00000000002cf7ac>] ___slab_alloc.constprop.30+0x2dc/0x378
[<00000000002d03d0>] kmem_cache_alloc+0x160/0x1d0
[<0000000000133db4>] kvm_setup_async_pf+0x6c/0x198
[<000000000013dee8>] kvm_arch_vcpu_ioctl_run+0xd48/0xd58
[<000000000012fcaa>] kvm_vcpu_ioctl+0x372/0x690
[<00000000002f66f6>] do_vfs_ioctl+0x3be/0x510
[<00000000002f68ec>] SyS_ioctl+0xa4/0xb8
[<0000000000781c5e>] system_call+0xd6/0x264
[<000003ffa24fa06a>] 0x3ffa24fa06a

Signed-off-by: Christian Borntraeger <>
Reviewed-by: Dominik Dingel <>
Signed-off-by: Paolo Bonzini <>
Signed-off-by: Ben Hutchings <>
7 years agosunrpc/cache: fix off-by-one in qword_get()
Stefan Hajnoczi [Thu, 18 Feb 2016 18:55:54 +0000 (18:55 +0000)]
sunrpc/cache: fix off-by-one in qword_get()

commit b7052cd7bcf3c1478796e93e3dff2b44c9e82943 upstream.

The qword_get() function NUL-terminates its output buffer.  If the input
string is in hex format \xXXXX... and the same length as the output
buffer, there is an off-by-one:

  int qword_get(char **bpp, char *dest, int bufsize)
      while (len < bufsize) {
          *dest++ = (h << 4) | l;
      *dest = '\0';
      return len;

This patch ensures the NUL terminator doesn't fall outside the output

Signed-off-by: Stefan Hajnoczi <>
Signed-off-by: J. Bruce Fields <>
Signed-off-by: Ben Hutchings <>
7 years agomac80211: minstrel_ht: set default tx aggregation timeout to 0
Felix Fietkau [Thu, 18 Feb 2016 18:49:18 +0000 (19:49 +0100)]
mac80211: minstrel_ht: set default tx aggregation timeout to 0

commit 7a36b930e6ed4702c866dc74a5ad07318a57c688 upstream.

The value 5000 was put here with the addition of the timeout field to
ieee80211_start_tx_ba_session. It was originally added in mac80211 to
save resources for drivers like iwlwifi, which only supports a limited
number of concurrent aggregation sessions.

Since iwlwifi does not use minstrel_ht and other drivers don't need
this, 0 is a better default - especially since there have been
recent reports of aggregation setup related issues reproduced with
ath9k. This should improve stability without causing any adverse

Acked-by: Avery Pennarun <>
Signed-off-by: Felix Fietkau <>
Signed-off-by: Johannes Berg <>
Signed-off-by: Ben Hutchings <>
7 years agocan: ems_usb: Fix possible tx overflow
Gerhard Uttenthaler [Tue, 22 Dec 2015 16:29:16 +0000 (17:29 +0100)]
can: ems_usb: Fix possible tx overflow

commit 90cfde46586d2286488d8ed636929e936c0c9ab2 upstream.

This patch fixes the problem that more CAN messages could be sent to the
interface as could be send on the CAN bus. This was more likely for slow baud
rates. The sleeping _start_xmit was woken up in the _write_bulk_callback. Under
heavy TX load this produced another bulk transfer without checking the
free_slots variable and hence caused the overflow in the interface.

Signed-off-by: Gerhard Uttenthaler <>
Signed-off-by: Marc Kleine-Budde <>
Signed-off-by: Ben Hutchings <>
7 years agokernel/resource.c: fix muxed resource handling in __request_region()
Simon Guinot [Wed, 9 Sep 2015 22:15:18 +0000 (00:15 +0200)]
kernel/resource.c: fix muxed resource handling in __request_region()

commit 59ceeaaf355fa0fb16558ef7c24413c804932ada upstream.

In __request_region, if a conflict with a BUSY and MUXED resource is
detected, then the caller goes to sleep and waits for the resource to be
released.  A pointer on the conflicting resource is kept.  At wake-up
this pointer is used as a parent to retry to request the region.

A first problem is that this pointer might well be invalid (if for
example the conflicting resource have already been freed).  Another
problem is that the next call to __request_region() fails to detect a
remaining conflict.  The previously conflicting resource is passed as a
parameter and __request_region() will look for a conflict among the
children of this resource and not at the resource itself.  It is likely
to succeed anyway, even if there is still a conflict.

Instead, the parent of the conflicting resource should be passed to

As a fix, this patch doesn't update the parent resource pointer in the
case we have to wait for a muxed region right after.

Reported-and-tested-by: Vincent Pelletier <>
Signed-off-by: Simon Guinot <>
Tested-by: Vincent Donnefort <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
7 years agoext4: fix bh->b_state corruption
Jan Kara [Fri, 19 Feb 2016 05:18:25 +0000 (00:18 -0500)]
ext4: fix bh->b_state corruption

commit ed8ad83808f009ade97ebbf6519bc3a97fefbc0c upstream.

ext4 can update bh->b_state non-atomically in _ext4_get_block() and
ext4_da_get_block_prep(). Usually this is fine since bh is just a
temporary storage for mapping information on stack but in some cases it
can be fully living bh attached to a page. In such case non-atomic
update of bh->b_state can race with an atomic update which then gets
lost. Usually when we are mapping bh and thus updating bh->b_state
non-atomically, nobody else touches the bh and so things work out fine
but there is one case to especially worry about: ext4_finish_bio() uses
BH_Uptodate_Lock on the first bh in the page to synchronize handling of
PageWriteback state. So when blocksize < pagesize, we can be atomically
modifying bh->b_state of a buffer that actually isn't under IO and thus
can race e.g. with delalloc trying to map that buffer. The result is
that we can mistakenly set / clear BH_Uptodate_Lock bit resulting in the
corruption of PageWriteback state or missed unlock of BH_Uptodate_Lock.

Fix the problem by always updating bh->b_state bits atomically.

Reported-by: Nikolay Borisov <>
Signed-off-by: Jan Kara <>
Signed-off-by: Theodore Ts'o <>
[bwh: Backported to 3.2:
 - Adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoAdding Intel Lewisburg device IDs for SATA
Alexandra Yates [Thu, 18 Feb 2016 03:36:20 +0000 (19:36 -0800)]
Adding Intel Lewisburg device IDs for SATA

commit f5bdd66c705484b4bc77eb914be15c1b7881fae7 upstream.

This patch complements the list of device IDs previously
added for lewisburg sata.

Signed-off-by: Alexandra Yates <>
Signed-off-by: Tejun Heo <>
Signed-off-by: Ben Hutchings <>
7 years agoUSB: option: add "4G LTE usb-modem U901"
Bjørn Mork [Fri, 12 Feb 2016 15:40:00 +0000 (16:40 +0100)]
USB: option: add "4G LTE usb-modem U901"

commit d061c1caa31d4d9792cfe48a2c6b309a0e01ef46 upstream.

Thomas reports:

T:  Bus=01 Lev=01 Prnt=01 Port=03 Cnt=01 Dev#=  4 Spd=480 MxCh= 0
D:  Ver= 2.00 Cls=00(>ifc ) Sub=00 Prot=00 MxPS=64 #Cfgs=  1
P:  Vendor=05c6 ProdID=6001 Rev=00.00
S:  Manufacturer=USB Modem
S:  Product=USB Modem
S:  SerialNumber=1234567890ABCDEF
C:  #Ifs= 5 Cfg#= 1 Atr=e0 MxPwr=500mA
I:  If#= 0 Alt= 0 #EPs= 2 Cls=ff(vend.) Sub=ff Prot=ff Driver=option
I:  If#= 1 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=ff Prot=ff Driver=option
I:  If#= 2 Alt= 0 #EPs= 2 Cls=ff(vend.) Sub=ff Prot=ff Driver=option
I:  If#= 3 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=ff Prot=ff Driver=qmi_wwan
I:  If#= 4 Alt= 0 #EPs= 2 Cls=08(stor.) Sub=06 Prot=50 Driver=usb-storage

Reported-by: Thomas Schäfer <>
Signed-off-by: Bjørn Mork <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
7 years agoUSB: cp210x: add IDs for GE B650V3 and B850V3 boards
Ken Lin [Mon, 1 Feb 2016 19:57:25 +0000 (14:57 -0500)]
USB: cp210x: add IDs for GE B650V3 and B850V3 boards

commit 6627ae19385283b89356a199d7f03c75ba35fb29 upstream.

Add USB ID for cp2104/5 devices on GE B650v3 and B850v3 boards.

Signed-off-by: Ken Lin <>
Signed-off-by: Akshay Bhat <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
7 years agoUSB: option: add support for SIM7100E
Andrey Skvortsov [Thu, 28 Jan 2016 21:07:30 +0000 (00:07 +0300)]
USB: option: add support for SIM7100E

commit 3158a8d416f4e1b79dcc867d67cb50013140772c upstream.

$ lsusb:
Bus 001 Device 101: ID 1e0e:9001 Qualcomm / Option

$ usb-devices:
T:  Bus=01 Lev=02 Prnt=02 Port=00 Cnt=01 Dev#=101 Spd=480  MxCh= 0
D:  Ver= 2.00 Cls=00(>ifc ) Sub=00 Prot=00 MxPS=64 #Cfgs=  2
P:  Vendor=1e0e ProdID=9001 Rev= 2.32
S:  Manufacturer=SimTech, Incorporated
S:  Product=SimTech, Incorporated
S:  SerialNumber=0123456789ABCDEF
C:* #Ifs= 7 Cfg#= 1 Atr=80 MxPwr=500mA
I:* If#= 0 Alt= 0 #EPs= 2 Cls=ff(vend.) Sub=ff Prot=ff Driver=option
I:* If#= 1 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=00 Prot=00 Driver=option
I:* If#= 2 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=00 Prot=00 Driver=option
I:* If#= 3 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=00 Prot=00 Driver=option
I:* If#= 4 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=00 Prot=00 Driver=option
I:* If#= 5 Alt= 0 #EPs= 3 Cls=ff(vend.) Sub=ff Prot=ff Driver=qmi_wwan
I:* If#= 6 Alt= 0 #EPs= 2 Cls=ff(vend.) Sub=42 Prot=01 Driver=(none)

The last interface (6) is used for Android Composite ADB interface.

Serial port layout:
2: AT
4: audio

Signed-off-by: Andrey Skvortsov <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
7 years agonet/mlx4_en: Count HW buffer overrun only once
Amir Vadai [Wed, 17 Feb 2016 15:24:22 +0000 (17:24 +0200)]
net/mlx4_en: Count HW buffer overrun only once

commit 281e8b2fdf8e4ef366b899453cae50e09b577ada upstream.

RdropOvflw counts overrun of HW buffer, therefore should
be used for rx_fifo_errors only.

Currently RdropOvflw counter is mistakenly also set into
rx_missed_errors and rx_over_errors too, which makes the
device total dropped packets accounting to show wrong results.

Fix that. Use it for rx_fifo_errors only.

Fixes: c27a02cd94d6 ('mlx4_en: Add driver for Mellanox ConnectX 10GbE NIC')
Signed-off-by: Amir Vadai <>
Signed-off-by: Eugenia Emantayev <>
Signed-off-by: Or Gerlitz <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
7 years agousb: dwc3: Fix assignment of EP transfer resources
John Youn [Wed, 17 Feb 2016 04:10:53 +0000 (20:10 -0800)]
usb: dwc3: Fix assignment of EP transfer resources

commit c450960187f45d4260db87c7dd4fc0bceb5565d8 upstream.

The assignement of EP transfer resources was not handled properly in the
dwc3 driver. Commit aebda6187181 ("usb: dwc3: Reset the transfer
resource index on SET_INTERFACE") previously fixed one aspect of this
where resources may be exhausted with multiple calls to SET_INTERFACE.
However, it introduced an issue where composite devices with multiple
interfaces can be assigned the same transfer resources for different
endpoints. This patch solves both issues.

The assignment of transfer resources cannot perfectly follow the data
book due to the fact that the controller driver does not have all
knowledge of the configuration in advance. It is given this information
piecemeal by the composite gadget framework after every
SET_CONFIGURATION and SET_INTERFACE. Trying to follow the databook
programming model in this scenario can cause errors. For two reasons:

1) The databook says to do DEPSTARTCFG for every SET_CONFIGURATION and
SET_INTERFACE (8.1.5). This is incorrect in the scenario of multiple

2) The databook does not mention doing more DEPXFERCFG for new endpoint
on alt setting (8.1.6).

The following simplified method is used instead:

All hardware endpoints can be assigned a transfer resource and this
setting will stay persistent until either a core reset or hibernation.
So whenever we do a DEPSTARTCFG(0) we can go ahead and do DEPXFERCFG for
every hardware endpoint as well. We are guaranteed that there are as
many transfer resources as endpoints.

This patch triggers off of the calling dwc3_gadget_start_config() for
EP0-out, which always happens first, and which should only happen in one
of the above conditions.

Fixes: aebda6187181 ("usb: dwc3: Reset the transfer resource index on SET_INTERFACE")
Reported-by: Ravi Babu <>
Signed-off-by: John Youn <>
Signed-off-by: Felipe Balbi <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agox86/uaccess/64: Handle the caching of 4-byte nocache copies properly in __copy_user_n...
Toshi Kani [Thu, 11 Feb 2016 21:24:17 +0000 (14:24 -0700)]
x86/uaccess/64: Handle the caching of 4-byte nocache copies properly in __copy_user_nocache()

commit a82eee7424525e34e98d821dd059ce14560a1e35 upstream.

Data corruption issues were observed in tests which initiated
a system crash/reset while accessing BTT devices.  This problem
is reproducible.

The BTT driver calls pmem_rw_bytes() to update data in pmem
devices.  This interface calls __copy_user_nocache(), which
uses non-temporal stores so that the stores to pmem are

__copy_user_nocache() uses non-temporal stores when a request
size is 8 bytes or larger (and is aligned by 8 bytes).  The
BTT driver updates the BTT map table, which entry size is
4 bytes.  Therefore, updates to the map table entries remain
cached, and are not written to pmem after a crash.

Change __copy_user_nocache() to use non-temporal store when
a request size is 4 bytes.  The change extends the current
byte-copy path for a less-than-8-bytes request, and does not
add any overhead to the regular path.

Reported-and-tested-by: Micah Parrish <>
Reported-and-tested-by: Brian Boylston <>
Signed-off-by: Toshi Kani <>
Cc: Andrew Morton <>
Cc: Andy Lutomirski <>
Cc: Borislav Petkov <>
Cc: Borislav Petkov <>
Cc: Brian Gerst <>
Cc: Dan Williams <>
Cc: Denys Vlasenko <>
Cc: H. Peter Anvin <>
Cc: Linus Torvalds <>
Cc: Luis R. Rodriguez <>
Cc: Peter Zijlstra <>
Cc: Ross Zwisler <>
Cc: Thomas Gleixner <>
Cc: Toshi Kani <>
Cc: Vishal Verma <>
[ Small readability edits. ]
Signed-off-by: Ingo Molnar <>
[bwh: Backported to 3.2: aadjust filename, context]
Signed-off-by: Ben Hutchings <>
7 years agox86/uaccess/64: Make the __copy_user_nocache() assembly code more readable
Toshi Kani [Thu, 11 Feb 2016 21:24:16 +0000 (14:24 -0700)]
x86/uaccess/64: Make the __copy_user_nocache() assembly code more readable

commit ee9737c924706aaa72c2ead93e3ad5644681dc1c upstream.

Add comments to __copy_user_nocache() to clarify its procedures
and alignment requirements.

Also change numeric branch target labels to named local labels.

No code changed:


    text    data     bss     dec     hex filename
    1239       0       0    1239     4d7 copy_user_64.o.before
    1239       0       0    1239     4d7 copy_user_64.o.after

    58bed94c2db98c1ca9a2d46d0680aaae  copy_user_64.o.before.asm
    58bed94c2db98c1ca9a2d46d0680aaae  copy_user_64.o.after.asm

Signed-off-by: Toshi Kani <>
Cc: Andrew Morton <>
Cc: Andy Lutomirski <>
Cc: Borislav Petkov <>
Cc: Borislav Petkov <>
Cc: Brian Gerst <>
Cc: Denys Vlasenko <>
Cc: H. Peter Anvin <>
Cc: Linus Torvalds <>
Cc: Luis R. Rodriguez <>
Cc: Peter Zijlstra <>
Cc: Thomas Gleixner <>
Cc: Toshi Kani <>
[ Small readability edits and added object file comparison. ]
Signed-off-by: Ingo Molnar <>
[bwh: Backported to 3.2: aadjust filename, context]
Signed-off-by: Ben Hutchings <>
7 years agox86, extable: Remove open-coded exception table entries in arch/x86/lib/copy_user_noc...
H. Peter Anvin [Fri, 20 Apr 2012 19:19:51 +0000 (12:19 -0700)]
x86, extable: Remove open-coded exception table entries in arch/x86/lib/copy_user_nocache_64.S

commit 0d8559feafbc9dc5a2c17ba42aea7de824b18308 upstream.

Remove open-coded exception table entries in arch/x86/lib/copy_user_nocache_64.S,
and replace them with _ASM_EXTABLE() macros; this will allow us to
change the format and type of the exception table entries.

Signed-off-by: H. Peter Anvin <>
Cc: David Daney <>
Signed-off-by: Ben Hutchings <>
7 years agoaf_unix: Guard against other == sk in unix_dgram_sendmsg
Rainer Weikusat [Thu, 11 Feb 2016 19:37:27 +0000 (19:37 +0000)]
af_unix: Guard against other == sk in unix_dgram_sendmsg

commit a5527dda344fff0514b7989ef7a755729769daa1 upstream.

The unix_dgram_sendmsg routine use the following test

if (unlikely(unix_peer(other) != sk && unix_recvq_full(other))) {

to determine if sk and other are in an n:1 association (either
established via connect or by using sendto to send messages to an
unrelated socket identified by address). This isn't correct as the
specified address could have been bound to the sending socket itself or
because this socket could have been connected to itself by the time of
the unix_peer_get but disconnected before the unix_state_lock(other). In
both cases, the if-block would be entered despite other == sk which
might either block the sender unintentionally or lead to trying to unlock
the same spin lock twice for a non-blocking send. Add a other != sk
check to guard against this.

Fixes: 7d267278a9ec ("unix: avoid use-after-free in ep_remove_wait_queue")
Reported-By: Philipp Hahn <>
Signed-off-by: Rainer Weikusat <>
Tested-by: Philipp Hahn <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
7 years agoaf_unix: Don't set err in unix_stream_read_generic unless there was an error
Rainer Weikusat [Mon, 8 Feb 2016 18:47:19 +0000 (18:47 +0000)]
af_unix: Don't set err in unix_stream_read_generic unless there was an error

commit 1b92ee3d03af6643df395300ba7748f19ecdb0c5 upstream.

The present unix_stream_read_generic contains various code sequences of
the form

if (<test>)
goto out;

This has the unfortunate side effect of possibly causing the error code
to bleed through to the final

return copied ? : err;

and then to be wrongly returned if no data was copied because the caller
didn't supply a data buffer, as demonstrated by the program available at

Change it such that err is only set if an error condition was detected.

Fixes: 3822b5c2fc62 ("af_unix: Revert 'lock_interruptible' in stream receive code")
Reported-by: Joseph Salisbury <>
Signed-off-by: Rainer Weikusat <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: seq: Fix double port list deletion
Takashi Iwai [Tue, 16 Feb 2016 13:15:59 +0000 (14:15 +0100)]
ALSA: seq: Fix double port list deletion

commit 13d5e5d4725c64ec06040d636832e78453f477b7 upstream.

The commit [7f0973e973cd: ALSA: seq: Fix lockdep warnings due to
double mutex locks] split the management of two linked lists (source
and destination) into two individual calls for avoiding the AB/BA
deadlock.  However, this may leave the possible double deletion of one
of two lists when the counterpart is being deleted concurrently.
It ends up with a list corruption, as revealed by syzkaller fuzzer.

This patch fixes it by checking the list emptiness and skipping the
deletion and the following process.

Fixes: 7f0973e973cd ('ALSA: seq: Fix lockdep warnings due to 'double mutex locks)
Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agotracing: Fix freak link error caused by branch tracer
Arnd Bergmann [Fri, 12 Feb 2016 21:26:42 +0000 (22:26 +0100)]
tracing: Fix freak link error caused by branch tracer

commit b33c8ff4431a343561e2319f17c14286f2aa52e2 upstream.

In my randconfig tests, I came across a bug that involves several

* gcc-4.9 through at least 5.3
* CONFIG_GCOV_PROFILE_ALL enabling -fprofile-arcs for all files
* CONFIG_PROFILE_ALL_BRANCHES overriding every if()
* The optimized implementation of do_div() that tries to
  replace a library call with an division by multiplication
* code in drivers/media/dvb-frontends/zl10353.c doing

        u32 adc_clock = 450560; /* 45.056 MHz */
        if (state->config.adc_clock)
                adc_clock = state->config.adc_clock;
        do_div(value, adc_clock);

In this case, gcc fails to determine whether the divisor
in do_div() is __builtin_constant_p(). In particular, it
concludes that __builtin_constant_p(adc_clock) is false, while
__builtin_constant_p(!!adc_clock) is true.

That in turn throws off the logic in do_div() that also uses
__builtin_constant_p(), and instead of picking either the
constant- optimized division, and the code in ilog2() that uses
__builtin_constant_p() to figure out whether it knows the answer at
compile time. The result is a link error from failing to find
multiple symbols that should never have been called based on
the __builtin_constant_p():

dvb-frontends/zl10353.c:138: undefined reference to `____ilog2_NaN'
dvb-frontends/zl10353.c:138: undefined reference to `__aeabi_uldivmod'
ERROR: "____ilog2_NaN" [drivers/media/dvb-frontends/zl10353.ko] undefined!
ERROR: "__aeabi_uldivmod" [drivers/media/dvb-frontends/zl10353.ko] undefined!

This patch avoids the problem by changing __trace_if() to check
whether the condition is known at compile-time to be nonzero, rather
than checking whether it is actually a constant.

I see this one link error in roughly one out of 1600 randconfig builds
on ARM, and the patch fixes all known instances.

Acked-by: Nicolas Pitre <>
Fixes: ab3c9c686e22 ("branch tracer, intel-iommu: fix build with CONFIG_BRANCH_TRACER=y")
Signed-off-by: Arnd Bergmann <>
Signed-off-by: Steven Rostedt <>
Signed-off-by: Ben Hutchings <>
7 years agotracepoints: Do not trace when cpu is offline
Steven Rostedt (Red Hat) [Mon, 15 Feb 2016 17:36:14 +0000 (12:36 -0500)]
tracepoints: Do not trace when cpu is offline

commit f37755490fe9bf76f6ba1d8c6591745d3574a6a6 upstream.

The tracepoint infrastructure uses RCU sched protection to enable and
disable tracepoints safely. There are some instances where tracepoints are
used in infrastructure code (like kfree()) that get called after a CPU is
going offline, and perhaps when it is coming back online but hasn't been
registered yet.

This can probuce the following warning:

 [ INFO: suspicious RCU usage. ]
 4.4.0-00006-g0fe53e8-dirty #34 Tainted: G S
 include/trace/events/kmem.h:141 suspicious rcu_dereference_check() usage!

 other info that might help us debug this:

 RCU used illegally from offline CPU!  rcu_scheduler_active = 1, debug_locks = 1
 no locks held by swapper/8/0.

 stack backtrace:
  CPU: 8 PID: 0 Comm: swapper/8 Tainted: G S              4.4.0-00006-g0fe53e8-dirty #34
  Call Trace:
  [c0000005b76c78d0] [c0000000008b9540] .dump_stack+0x98/0xd4 (unreliable)
  [c0000005b76c7950] [c00000000010c898] .lockdep_rcu_suspicious+0x108/0x170
  [c0000005b76c79e0] [c00000000029adc0] .kfree+0x390/0x440
  [c0000005b76c7a80] [c000000000055f74] .destroy_context+0x44/0x100
  [c0000005b76c7b00] [c0000000000934a0] .__mmdrop+0x60/0x150
  [c0000005b76c7b90] [c0000000000e3ff0] .idle_task_exit+0x130/0x140
  [c0000005b76c7c20] [c000000000075804] .pseries_mach_cpu_die+0x64/0x310
  [c0000005b76c7cd0] [c000000000043e7c] .cpu_die+0x3c/0x60
  [c0000005b76c7d40] [c0000000000188d8] .arch_cpu_idle_dead+0x28/0x40
  [c0000005b76c7db0] [c000000000101e6c] .cpu_startup_entry+0x50c/0x560
  [c0000005b76c7ed0] [c000000000043bd8] .start_secondary+0x328/0x360
  [c0000005b76c7f90] [c000000000008a6c] start_secondary_prolog+0x10/0x14

This warning is not a false positive either. RCU is not protecting code that
is being executed while the CPU is offline.

Instead of playing "whack-a-mole(TM)" and adding conditional statements to
the tracepoints we find that are used in this instance, simply add a
cpu_online() test to the tracepoint code where the tracepoint will be
ignored if the CPU is offline.

Use of raw_smp_processor_id() is fine, as there should never be a case where
the tracepoint code goes from running on a CPU that is online and suddenly
gets migrated to a CPU that is offline.

Reported-by: Denis Kirjanov <>
Fixes: 97e1c18e8d17b ("tracing: Kernel Tracepoints")
Signed-off-by: Steven Rostedt <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: seq: Fix leak of pool buffer at concurrent writes
Takashi Iwai [Mon, 15 Feb 2016 15:20:24 +0000 (16:20 +0100)]
ALSA: seq: Fix leak of pool buffer at concurrent writes

commit d99a36f4728fcbcc501b78447f625bdcce15b842 upstream.

When multiple concurrent writes happen on the ALSA sequencer device
right after the open, it may try to allocate vmalloc buffer for each
write and leak some of them.  It's because the presence check and the
assignment of the buffer is done outside the spinlock for the pool.

The fix is to move the check and the assignment into the spinlock.

(The current implementation is suboptimal, as there can be multiple
 unnecessary vmallocs because the allocation is done before the check
 in the spinlock.  But the pool size is already checked beforehand, so
 this isn't a big problem; that is, the only possible path is the
 multiple writes before any pool assignment, and practically seen, the
 current coverage should be "good enough".)

The issue was triggered by syzkaller fuzzer.

Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoxen/pcifront: Fix mysterious crashes when NUMA locality information was extracted.
Konrad Rzeszutek Wilk [Thu, 11 Feb 2016 21:10:26 +0000 (16:10 -0500)]
xen/pcifront: Fix mysterious crashes when NUMA locality information was extracted.

commit 4d8c8bd6f2062c9988817183a91fe2e623c8aa5e upstream.

Occasionaly PV guests would crash with:

pciback 0000:00:00.1: Xen PCI mapped GSI0 to IRQ16
BUG: unable to handle kernel paging request at 0000000d1a8c0be0
.. snip..
  <ffffffff8139ce1b>] find_next_bit+0xb/0x10
  [<ffffffff81387f22>] cpumask_next_and+0x22/0x40
  [<ffffffff813c1ef8>] pci_device_probe+0xb8/0x120
  [<ffffffff81529097>] ? driver_sysfs_add+0x77/0xa0
  [<ffffffff815293e4>] driver_probe_device+0x1a4/0x2d0
  [<ffffffff813c1ddd>] ? pci_match_device+0xdd/0x110
  [<ffffffff81529657>] __device_attach_driver+0xa7/0xb0
  [<ffffffff815295b0>] ? __driver_attach+0xa0/0xa0
  [<ffffffff81527622>] bus_for_each_drv+0x62/0x90
  [<ffffffff8152978d>] __device_attach+0xbd/0x110
  [<ffffffff815297fb>] device_attach+0xb/0x10
  [<ffffffff813b75ac>] pci_bus_add_device+0x3c/0x70
  [<ffffffff813b7618>] pci_bus_add_devices+0x38/0x80
  [<ffffffff813dc34e>] pcifront_scan_root+0x13e/0x1a0
  [<ffffffff817a0692>] pcifront_backend_changed+0x262/0x60b
  [<ffffffff814644c6>] ? xenbus_gather+0xd6/0x160
  [<ffffffff8120900f>] ? put_object+0x2f/0x50
  [<ffffffff81465c1d>] xenbus_otherend_changed+0x9d/0xa0
  [<ffffffff814678ee>] backend_changed+0xe/0x10
  [<ffffffff81463a28>] xenwatch_thread+0xc8/0x190
  [<ffffffff810f22f0>] ? woken_wake_function+0x10/0x10

which was the result of two things:

When we call pci_scan_root_bus we would pass in 'sd' (sysdata)
pointer which was an 'pcifront_sd' structure. However in the
pci_device_add it expects that the 'sd' is 'struct sysdata' and
sets the dev->node to what is in sd->node (offset 4):

set_dev_node(&dev->dev, pcibus_to_node(bus));

 __pcibus_to_node(const struct pci_bus *bus)
        const struct pci_sysdata *sd = bus->sysdata;

        return sd->node;

However our structure was pcifront_sd which had nothing at that

struct pcifront_sd {
        int                        domain;    /*     0     4 */
        /* XXX 4 bytes hole, try to pack */
        struct pcifront_device *   pdev;      /*     8     8 */

That is an hole - filled with garbage as we used kmalloc instead of
kzalloc (the second problem).

This patch fixes the issue by:
 1) Use kzalloc to initialize to a well known state.
 2) Put 'struct pci_sysdata' at the start of 'pcifront_sd'. That
    way access to the 'node' will access the right offset.

Signed-off-by: Konrad Rzeszutek Wilk <>
Reviewed-by: Boris Ostrovsky <>
Signed-off-by: David Vrabel <>
Signed-off-by: Ben Hutchings <>
7 years agoxen/pciback: Save the number of MSI-X entries to be copied later.
Konrad Rzeszutek Wilk [Thu, 11 Feb 2016 21:10:24 +0000 (16:10 -0500)]
xen/pciback: Save the number of MSI-X entries to be copied later.

commit d159457b84395927b5a52adb72f748dd089ad5e5 upstream.

Commit 8135cf8b092723dbfcc611fe6fdcb3a36c9951c5 (xen/pciback: Save
xen_pci_op commands before processing it) broke enabling MSI-X because
it would never copy the resulting vectors into the response.  The
number of vectors requested was being overwritten by the return value
(typically zero for success).

Save the number of vectors before processing the op, so the correct
number of vectors are copied afterwards.

Signed-off-by: Konrad Rzeszutek Wilk <>
Reviewed-by: Jan Beulich <>
Signed-off-by: David Vrabel <>
Signed-off-by: Ben Hutchings <>
7 years agoxen/pciback: Check PF instead of VF for PCI_COMMAND_MEMORY
Konrad Rzeszutek Wilk [Thu, 11 Feb 2016 21:10:23 +0000 (16:10 -0500)]
xen/pciback: Check PF instead of VF for PCI_COMMAND_MEMORY

commit 8d47065f7d1980dde52abb874b301054f3013602 upstream.

Commit 408fb0e5aa7fda0059db282ff58c3b2a4278baa0 (xen/pciback: Don't
allow MSI-X ops if PCI_COMMAND_MEMORY is not set) prevented enabling
MSI-X on passed-through virtual functions, because it checked the VF
for PCI_COMMAND_MEMORY but this is not a valid bit for VFs.

Instead, check the physical function for PCI_COMMAND_MEMORY.

Signed-off-by: Konrad Rzeszutek Wilk <>
Reviewed-by: Jan Beulich <>
Signed-off-by: David Vrabel <>
Signed-off-by: Ben Hutchings <>
7 years agolibata: fix HDIO_GET_32BIT ioctl
Arnd Bergmann [Thu, 11 Feb 2016 13:16:27 +0000 (14:16 +0100)]
libata: fix HDIO_GET_32BIT ioctl

commit 287e6611ab1eac76c2c5ebf6e345e04c80ca9c61 upstream.

As reported by Soohoon Lee, the HDIO_GET_32BIT ioctl does not
work correctly in compat mode with libata.

I have investigated the issue further and found multiple problems
that all appeared with the same commit that originally introduced
HDIO_GET_32BIT handling in libata back in linux-2.6.8 and presumably
also linux-2.4, as the code uses "copy_to_user(arg, &val, 1)" to copy
a 'long' variable containing either 0 or 1 to user space.

The problems with this are:

* On big-endian machines, this will always write a zero because it
  stores the wrong byte into user space.

* In compat mode, the upper three bytes of the variable are updated
  by the compat_hdio_ioctl() function, but they now contain
  uninitialized stack data.

* The hdparm tool calling this ioctl uses a 'static long' variable
  to store the result. This means at least the upper bytes are
  initialized to zero, but calling another ioctl like HDIO_GET_MULTCOUNT
  would fill them with data that remains stale when the low byte
  is overwritten. Fortunately libata doesn't implement any of the
  affected ioctl commands, so this would only happen when we query
  both an IDE and an ATA device in the same command such as
  "hdparm -N -c /dev/hda /dev/sda"

* The libata code for unknown reasons started using ATA_IOC_GET_IO32
  and ATA_IOC_SET_IO32 as aliases for HDIO_GET_32BIT and HDIO_SET_32BIT,
  while the ioctl commands that were added later use the normal
  HDIO_* names. This is harmless but rather confusing.

This addresses all four issues by changing the code to use put_user()
on an 'unsigned long' variable in HDIO_GET_32BIT, like the IDE subsystem
does, and by clarifying the names of the ioctl commands.

Signed-off-by: Arnd Bergmann <>
Reported-by: Soohoon Lee <>
Tested-by: Soohoon Lee <>
Signed-off-by: Tejun Heo <>
Signed-off-by: Ben Hutchings <>
7 years agos390/dasd: fix refcount for PAV reassignment
Stefan Haberland [Tue, 15 Dec 2015 09:45:05 +0000 (10:45 +0100)]
s390/dasd: fix refcount for PAV reassignment

commit 9d862ababb609439c5d6987f6d3ddd09e703aa0b upstream.

Add refcount to the DASD device when a summary unit check worker is
scheduled. This prevents that the device is set offline with worker
in place.

Signed-off-by: Stefan Haberland <>
Signed-off-by: Martin Schwidefsky <>
Signed-off-by: Ben Hutchings <>
7 years agos390/dasd: prevent incorrect length error under z/VM after PAV changes
Stefan Haberland [Tue, 15 Dec 2015 09:16:43 +0000 (10:16 +0100)]
s390/dasd: prevent incorrect length error under z/VM after PAV changes

commit 020bf042e5b397479c1174081b935d0ff15d1a64 upstream.

The channel checks the specified length and the provided amount of
data for CCWs and provides an incorrect length error if the size does
not match. Under z/VM with simulation activated the length may get
changed. Having the suppress length indication bit set is stated as
good CCW coding practice and avoids errors under z/VM.

Signed-off-by: Stefan Haberland <>
Signed-off-by: Martin Schwidefsky <>
Signed-off-by: Ben Hutchings <>
7 years agocifs: fix erroneous return value
Anton Protopopov [Wed, 10 Feb 2016 17:50:21 +0000 (12:50 -0500)]
cifs: fix erroneous return value

commit 4b550af519854421dfec9f7732cdddeb057134b2 upstream.

The setup_ntlmv2_rsp() function may return positive value ENOMEM instead
of -ENOMEM in case of kmalloc failure.

Signed-off-by: Anton Protopopov <>
Signed-off-by: Steve French <>
Signed-off-by: Ben Hutchings <>
7 years agodrm/i915: fix error path in intel_setup_gmbus()
Rasmus Villemoes [Tue, 9 Feb 2016 20:11:13 +0000 (21:11 +0100)]
drm/i915: fix error path in intel_setup_gmbus()

commit ed3f9fd1e865975ceefdb2a43b453e090b1fd787 upstream.

This fails to undo the setup for pin==0; moreover, something
interesting happens if the setup failed already at pin==0.

Signed-off-by: Rasmus Villemoes <>
Fixes: f899fc64cda8 ("drm/i915: use GMBUS to manage i2c links")
Signed-off-by: Jani Nikula <>
(cherry picked from commit 2417c8c03f508841b85bf61acc91836b7b0e2560)
Signed-off-by: Jani Nikula <>
[bwh: Backported to 3.2: index variable is i, not pin]
Signed-off-by: Ben Hutchings <>
7 years agonfs: fix nfs_size_to_loff_t
Christoph Hellwig [Mon, 8 Feb 2016 20:11:50 +0000 (21:11 +0100)]
nfs: fix nfs_size_to_loff_t

commit 50ab8ec74a153eb30db26529088bc57dd700b24c upstream.

See http: //
Content-Transfer-Encoding: 8bit
Mime-Version: 1.0

We support OFFSET_MAX just fine, so don't round down below it.  Also
switch to using min_t to make the helper more readable.

Signed-off-by: Christoph Hellwig <>
Fixes: 433c92379d9c ("NFS: Clean up nfs_size_to_loff_t()")
Signed-off-by: Trond Myklebust <>
Signed-off-by: Ben Hutchings <>
7 years agomac80211: fix use of uninitialised values in RX aggregation
Chris Bainbridge [Wed, 27 Jan 2016 15:46:18 +0000 (15:46 +0000)]
mac80211: fix use of uninitialised values in RX aggregation

commit f39ea2690bd61efec97622c48323f40ed6e16317 upstream.

Use kzalloc instead of kmalloc for struct tid_ampdu_rx to
initialize the "removed" field (all others are initialized
manually). That fixes:

UBSAN: Undefined behaviour in net/mac80211/rx.c:932:29
load of value 2 is not a valid value for type '_Bool'
CPU: 3 PID: 1134 Comm: kworker/u16:7 Not tainted 4.5.0-rc1+ #265
Workqueue: phy0 rt2x00usb_work_rxdone
 0000000000000004 ffff880254a7ba50 ffffffff8181d866 0000000000000007
 ffff880254a7ba78 ffff880254a7ba68 ffffffff8188422d ffffffff8379b500
 ffff880254a7bab8 ffffffff81884747 0000000000000202 0000000348620032
Call Trace:
 [<ffffffff8181d866>] dump_stack+0x45/0x5f
 [<ffffffff8188422d>] ubsan_epilogue+0xd/0x40
 [<ffffffff81884747>] __ubsan_handle_load_invalid_value+0x67/0x70
 [<ffffffff82227b4d>] ieee80211_sta_reorder_release.isra.16+0x5ed/0x730
 [<ffffffff8222ca14>] ieee80211_prepare_and_rx_handle+0xd04/0x1c00
 [<ffffffff8222db03>] __ieee80211_rx_handle_packet+0x1f3/0x750
 [<ffffffff8222e4a7>] ieee80211_rx_napi+0x447/0x990

While at it, convert to use sizeof(*tid_agg_rx) instead.

Fixes: 788211d81bfdf ("mac80211: fix RX A-MPDU session reorder timer deletion")
Signed-off-by: Chris Bainbridge <>
[reword commit message, use sizeof(*tid_agg_rx)]
Signed-off-by: Johannes Berg <>
Signed-off-by: Ben Hutchings <>
7 years agocfg80211/wext: fix message ordering
Johannes Berg [Wed, 27 Jan 2016 12:29:34 +0000 (13:29 +0100)]
cfg80211/wext: fix message ordering

commit cb150b9d23be6ee7f3a0fff29784f1c5b5ac514d upstream.

Since cfg80211 frequently takes actions from its netdev notifier
call, wireless extensions messages could still be ordered badly
since the wext netdev notifier, since wext is built into the
kernel, runs before the cfg80211 netdev notifier. For example,
the following can happen:

5: wlan1: <BROADCAST,MULTICAST> mtu 1500 qdisc mq state DOWN group default
    link/ether 02:00:00:00:01:00 brd ff:ff:ff:ff:ff:ff

when setting the interface down causes the wext message.

To also fix this, export the wireless_nlevent_flush() function
and also call it from the cfg80211 notifier.

Signed-off-by: Johannes Berg <>
[bwh: Backported to 3.2:
 - Add default case in cfg80211_netdev_notifier_call() which bypasses
   the added wireless_nlevent_flush() (added upstream by commit
   6784c7db8d43 "cfg80211: change return value of notifier function")
 - Adjust context]
Signed-off-by: Ben Hutchings <>
7 years agowext: fix message delay/ordering
Johannes Berg [Wed, 27 Jan 2016 11:37:52 +0000 (12:37 +0100)]
wext: fix message delay/ordering

commit 8bf862739a7786ae72409220914df960a0aa80d8 upstream.

Beniamino reported that he was getting an RTM_NEWLINK message for a
given interface, after the RTM_DELLINK for it. It turns out that the
message is a wireless extensions message, which was sent because the
interface had been connected and disconnection while it was deleted
caused a wext message.

For its netlink messages, wext uses RTM_NEWLINK, but the message is
without all the regular rtnetlink attributes, so "ip monitor link"
prints just rudimentary information:

5: wlan1: <BROADCAST,MULTICAST> mtu 1500 qdisc mq state DOWN group default
    link/ether 02:00:00:00:01:00 brd ff:ff:ff:ff:ff:ff
Deleted 5: wlan1: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN group default
    link/ether 02:00:00:00:01:00 brd ff:ff:ff:ff:ff:ff
(from my hwsim reproduction)

This can cause userspace to get confused since it doesn't expect an

The reason for this is that wext schedules a worker to send out the
messages, and the scheduling delay can cause the messages to get out
to userspace in different order.

To fix this, have wext register a netdevice notifier and flush out
any pending messages when netdevice state changes. This fixes any
ordering whenever the original message wasn't sent by a notifier

Reported-by: Beniamino Galvani <>
Signed-off-by: Johannes Berg <>
Signed-off-by: Ben Hutchings <>
7 years agoiommu/vt-d: Fix 64-bit accesses to 32-bit DMAR_GSTS_REG
CQ Tang [Wed, 13 Jan 2016 21:15:03 +0000 (21:15 +0000)]
iommu/vt-d: Fix 64-bit accesses to 32-bit DMAR_GSTS_REG

commit fda3bec12d0979aae3f02ee645913d66fbc8a26e upstream.

This is a 32-bit register. Apparently harmless on real hardware, but
causing justified warnings in simulation.

Signed-off-by: CQ Tang <>
Signed-off-by: David Woodhouse <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
7 years agocrypto: {blk,giv}cipher: Set has_setkey
Ben Hutchings [Sun, 6 Mar 2016 19:52:46 +0000 (19:52 +0000)]
crypto: {blk,giv}cipher: Set has_setkey

Commit a1383cd86a06 ("crypto: skcipher - Add crypto_skcipher_has_setkey")
was incorrectly backported to the 3.2.y and 3.16.y stable branches.
We need to set ablkcipher_tfm::has_setkey in the
crypto_init_blkcipher_ops_async() and crypto_init_givcipher_ops()
functions as well as crypto_init_ablkcipher_ops().

Signed-off-by: Ben Hutchings <>
7 years agoRevert "crypto: algif_skcipher - Do not dereference ctx without socket lock"
Ben Hutchings [Sun, 6 Mar 2016 22:07:37 +0000 (22:07 +0000)]
Revert "crypto: algif_skcipher - Do not dereference ctx without socket lock"

This reverts commit c54ddfbb1b691d77c52b76ca6e13ca7082eb3b82, which
was a poorly backported version of commit
6454c2b83f719057069777132b13949e4c6b6350 upstream.  The small part I
was able to backport makes no sense by itself.

Signed-off-by: Ben Hutchings <>
7 years agoLinux 3.2.78 v3.2.78
Ben Hutchings [Sat, 27 Feb 2016 14:28:50 +0000 (14:28 +0000)]
Linux 3.2.78

7 years agosched: fix __sched_setscheduler() vs load balancing race
Mike Galbraith [Wed, 17 Feb 2016 03:02:59 +0000 (04:02 +0100)]
sched: fix __sched_setscheduler() vs load balancing race

__sched_setscheduler() may release rq->lock in pull_rt_task() as a task is
being changed rt -> fair class.  load balancing may sneak in, move the task
behind __sched_setscheduler()'s back, which explodes in switched_to_fair()
when the passed but no longer valid rq is used.  Tell can_migrate_task() to
say no if ->pi_lock is held.

@stable: Kernels that predate SCHED_DEADLINE can use this simple (and tested)
check in lieu of backport of the full 18 patch mainline treatment.

Signed-off-by: Mike Galbraith <>
[bwh: Backported to 3.2:
 - Adjust numbering in the comment
 - Adjust filename]
Signed-off-by: Ben Hutchings <>
Cc: Byungchul Park <>
Cc: Peter Zijlstra <>
Cc: Willy Tarreau <>
7 years agopipe: Fix buffer offset after partially failed read
Ben Hutchings [Sat, 13 Feb 2016 02:34:52 +0000 (02:34 +0000)]
pipe: Fix buffer offset after partially failed read

Quoting the RHEL advisory:

> It was found that the fix for CVE-2015-1805 incorrectly kept buffer
> offset and buffer length in sync on a failed atomic read, potentially
> resulting in a pipe buffer state corruption. A local, unprivileged user
> could use this flaw to crash the system or leak kernel memory to user
> space. (CVE-2016-0774, Moderate)

The same flawed fix was applied to stable branches from 2.6.32.y to
3.14.y inclusive, and I was able to reproduce the issue on 3.2.y.
We need to give pipe_iov_copy_to_user() a separate offset variable
and only update the buffer offset if it succeeds.

Signed-off-by: Ben Hutchings <>
7 years agoiw_cxgb3: Fix incorrectly returning error on success
Hariprasad S [Fri, 11 Dec 2015 08:29:17 +0000 (13:59 +0530)]
iw_cxgb3: Fix incorrectly returning error on success

commit 67f1aee6f45059fd6b0f5b0ecb2c97ad0451f6b3 upstream.

The cxgb3_*_send() functions return NET_XMIT_ values, which are
positive integers values. So don't treat positive return values
as an error.

Signed-off-by: Steve Wise <>
Signed-off-by: Hariprasad Shenai <>
Signed-off-by: Doug Ledford <>
Signed-off-by: Ben Hutchings <>
7 years agopipe: limit the per-user amount of pages allocated in pipes
Willy Tarreau [Mon, 18 Jan 2016 15:36:09 +0000 (16:36 +0100)]
pipe: limit the per-user amount of pages allocated in pipes

commit 759c01142a5d0f364a462346168a56de28a80f52 upstream.

On no-so-small systems, it is possible for a single process to cause an
OOM condition by filling large pipes with data that are never read. A
typical process filling 4000 pipes with 1 MB of data will use 4 GB of
memory. On small systems it may be tricky to set the pipe max size to
prevent this from happening.

This patch makes it possible to enforce a per-user soft limit above
which new pipes will be limited to a single page, effectively limiting
them to 4 kB each, as well as a hard limit above which no new pipes may
be created for this user. This has the effect of protecting the system
against memory abuse without hurting other users, and still allowing
pipes to work correctly though with less data at once.

The limit are controlled by two new sysctls : pipe-user-pages-soft, and
pipe-user-pages-hard. Both may be disabled by setting them to zero. The
default soft limit allows the default number of FDs per process (1024)
to create pipes of the default size (64kB), thus reaching a limit of 64MB
before starting to create only smaller pipes. With 256 processes limited
to 1024 FDs each, this results in 1024*64kB + (256*1024 - 1024) * 4kB =
1084 MB of memory allocated for a user. The hard limit is disabled by
default to avoid breaking existing applications that make intensive use
of pipes (eg: for splicing).

Reported-by: Tetsuo Handa <>
Mitigates: CVE-2013-4312 (Linux 2.0+)
Suggested-by: Linus Torvalds <>
Signed-off-by: Willy Tarreau <>
Signed-off-by: Al Viro <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agounix: correctly track in-flight fds in sending process user_struct
Hannes Frederic Sowa [Wed, 3 Feb 2016 01:11:03 +0000 (02:11 +0100)]
unix: correctly track in-flight fds in sending process user_struct

commit 415e3d3e90ce9e18727e8843ae343eda5a58fad6 upstream.

The commit referenced in the Fixes tag incorrectly accounted the number
of in-flight fds over a unix domain socket to the original opener
of the file-descriptor. This allows another process to arbitrary
deplete the original file-openers resource limit for the maximum of
open files. Instead the sending processes and its struct cred should
be credited.

To do so, we add a reference counted struct user_struct pointer to the
scm_fp_list and use it to account for the number of inflight unix fds.

Fixes: 712f4aad406bb1 ("unix: properly account for FDs passed over unix sockets")
Reported-by: David Herrmann <>
Cc: David Herrmann <>
Cc: Willy Tarreau <>
Cc: Linus Torvalds <>
Suggested-by: Linus Torvalds <>
Signed-off-by: Hannes Frederic Sowa <>
Signed-off-by: David S. Miller <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agounix: properly account for FDs passed over unix sockets
willy tarreau [Sun, 10 Jan 2016 06:54:56 +0000 (07:54 +0100)]
unix: properly account for FDs passed over unix sockets

commit 712f4aad406bb1ed67f3f98d04c044191f0ff593 upstream.

It is possible for a process to allocate and accumulate far more FDs than
the process' limit by sending them over a unix socket then closing them
to keep the process' fd count low.

This change addresses this problem by keeping track of the number of FDs
in flight per user and preventing non-privileged processes from having
more FDs in flight than their configured FD limit.

Reported-by: Tetsuo Handa <>
Mitigates: CVE-2013-4312 (Linux 2.0+)
Suggested-by: Linus Torvalds <>
Acked-by: Hannes Frederic Sowa <>
Signed-off-by: Willy Tarreau <>
Signed-off-by: David S. Miller <>
[carnil: Backported to 3.16: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: usb-audio: avoid freeing umidi object twice
Andrey Konovalov [Sat, 13 Feb 2016 08:08:06 +0000 (11:08 +0300)]
ALSA: usb-audio: avoid freeing umidi object twice

commit 07d86ca93db7e5cdf4743564d98292042ec21af7 upstream.

The 'umidi' object will be free'd on the error path by snd_usbmidi_free()
when tearing down the rawmidi interface. So we shouldn't try to free it
in snd_usbmidi_create() after having registered the rawmidi interface.

Found by KASAN.

Signed-off-by: Andrey Konovalov <>
Acked-by: Clemens Ladisch <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agobtrfs: properly set the termination value of ctx->pos in readdir
David Sterba [Fri, 13 Nov 2015 12:44:28 +0000 (13:44 +0100)]
btrfs: properly set the termination value of ctx->pos in readdir

commit bc4ef7592f657ae81b017207a1098817126ad4cb upstream.

The value of ctx->pos in the last readdir call is supposed to be set to
INT_MAX due to 32bit compatibility, unless 'pos' is intentially set to a
larger value, then it's LLONG_MAX.

There's a report from PaX SIZE_OVERFLOW plugin that "ctx->pos++"
overflows (, on a
64bit arch, where the value is 0x7fffffffffffffff ie. LLONG_MAX before
the increment.

We can get to that situation like that:

* emit all regular readdir entries
* still in the same call to readdir, bump the last pos to INT_MAX
* next call to readdir will not emit any entries, but will reach the
  bump code again, finds pos to be INT_MAX and sets it to LLONG_MAX

Normally this is not a problem, but if we call readdir again, we'll find
'pos' set to LLONG_MAX and the unconditional increment will overflow.

The report from Victor at
( with debugging
print shows that pattern:

 Overflow: e
 Overflow: 7fffffff
 Overflow: 7fffffffffffffff
 PAX: size overflow detected in function btrfs_real_readdir
   fs/btrfs/inode.c:5760 cicus.935_282 max, count: 9, decl: pos; num: 0;
   context: dir_context;
 CPU: 0 PID: 2630 Comm: polkitd Not tainted 4.2.3-grsec #1
 Hardware name: Gigabyte Technology Co., Ltd. H81ND2H/H81ND2H, BIOS F3 08/11/2015
  ffffffff81901608 0000000000000000 ffffffff819015e6 ffffc90004973d48
  ffffffff81742f0f 0000000000000007 ffffffff81901608 ffffc90004973d78
  ffffffff811cb706 0000000000000000 ffff8800d47359e0 ffffc90004973ed8
 Call Trace:
  [<ffffffff81742f0f>] dump_stack+0x4c/0x7f
  [<ffffffff811cb706>] report_size_overflow+0x36/0x40
  [<ffffffff812ef0bc>] btrfs_real_readdir+0x69c/0x6d0
  [<ffffffff811dafc8>] iterate_dir+0xa8/0x150
  [<ffffffff811e6d8d>] ? __fget_light+0x2d/0x70
  [<ffffffff811dba3a>] SyS_getdents+0xba/0x1c0
 Overflow: 1a
  [<ffffffff811db070>] ? iterate_dir+0x150/0x150
  [<ffffffff81749b69>] entry_SYSCALL_64_fastpath+0x12/0x83

The jump from 7fffffff to 7fffffffffffffff happens when new dir entries
are not yet synced and are processed from the delayed list. Then the code
could go to the bump section again even though it might not emit any new
dir entries from the delayed list.

The fix avoids entering the "bump" section again once we've finished
emitting the entries, both for synced and delayed entries.

Reported-by: Victor <>
Signed-off-by: David Sterba <>
Tested-by: Holger Hoffstätte <>
Signed-off-by: Chris Mason <>
[bwh: Backported to 3.2:
 - s/ctx->pos/filp->f_pos/
 - Adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoARM: 8519/1: ICST: try other dividends than 1
Linus Walleij [Wed, 10 Feb 2016 08:25:17 +0000 (09:25 +0100)]
ARM: 8519/1: ICST: try other dividends than 1

commit e972c37459c813190461dabfeaac228e00aae259 upstream.

Since the dawn of time the ICST code has only supported divide
by one or hang in an eternal loop. Luckily we were always dividing
by one because the reference frequency for the systems using
the ICSTs is 24MHz and the [min,max] values for the PLL input
if [10,320] MHz for ICST307 and [6,200] for ICST525, so the loop
will always terminate immediately without assigning any divisor
for the reference frequency.

But for the code to make sense, let's insert the missing i++

Reported-by: David Binderman <>
Signed-off-by: Linus Walleij <>
Signed-off-by: Russell King <>
Signed-off-by: Ben Hutchings <>
7 years agoahci: Intel DNV device IDs SATA
Alexandra Yates [Fri, 5 Feb 2016 23:27:49 +0000 (15:27 -0800)]
ahci: Intel DNV device IDs SATA

commit 342decff2b846b46fa61eb5ee40986fab79a9a32 upstream.

Adding Intel codename DNV platform device IDs for SATA.

Signed-off-by: Alexandra Yates <>
Signed-off-by: Tejun Heo <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: timer: Fix race at concurrent reads
Takashi Iwai [Mon, 8 Feb 2016 16:26:58 +0000 (17:26 +0100)]
ALSA: timer: Fix race at concurrent reads

commit 4dff5c7b7093b19c19d3a100f8a3ad87cb7cd9e7 upstream.

snd_timer_user_read() has a potential race among parallel reads, as
qhead and qused are updated outside the critical section due to
copy_to_user() calls.  Move them into the critical section, and also
sanitize the relevant code a bit.

Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: there's no check for tu->connected to fix up]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: timer: Fix race between stop and interrupt
Takashi Iwai [Tue, 9 Feb 2016 11:02:32 +0000 (12:02 +0100)]
ALSA: timer: Fix race between stop and interrupt

commit ed8b1d6d2c741ab26d60d499d7fbb7ac801f0f51 upstream.

A slave timer element also unlinks at snd_timer_stop() but it takes
only slave_active_lock.  When a slave is assigned to a master,
however, this may become a race against the master's interrupt
handling, eventually resulting in a list corruption.  The actual bug
could be seen with a syzkaller fuzzer test case in BugLink below.

As a fix, we need to take timeri->timer->lock when timer isn't NULL,
i.e. assigned to a master, while the assignment to a master itself is
protected by slave_active_lock.

Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: adjust context, indentation]
Signed-off-by: Ben Hutchings <>
7 years agosctp: translate network order to host order when users get a hmacid
Xin Long [Wed, 3 Feb 2016 15:33:30 +0000 (23:33 +0800)]
sctp: translate network order to host order when users get a hmacid

commit 7a84bd46647ff181eb2659fdc99590e6f16e501d upstream.

Commit ed5a377d87dc ("sctp: translate host order to network order when
setting a hmacid") corrected the hmacid byte-order when setting a hmacid.
but the same issue also exists on getting a hmacid.

We fix it by changing hmacids to host order when users get them with

Fixes: Commit ed5a377d87dc ("sctp: translate host order to network order when setting a hmacid")
Signed-off-by: Xin Long <>
Acked-by: Marcelo Ricardo Leitner <>
Signed-off-by: David S. Miller <>
Signed-off-by: Ben Hutchings <>
7 years agoARM: 8517/1: ICST: avoid arithmetic overflow in icst_hz()
Linus Walleij [Mon, 8 Feb 2016 08:14:37 +0000 (09:14 +0100)]
ARM: 8517/1: ICST: avoid arithmetic overflow in icst_hz()

commit 5070fb14a0154f075c8b418e5bc58a620ae85a45 upstream.

When trying to set the ICST 307 clock to 25174000 Hz I ran into
this arithmetic error: the icst_hz_to_vco() correctly figure out
DIVIDE=2, RDW=100 and VDW=99 yielding a frequency of
25174000 Hz out of the VCO. (I replicated the icst_hz() function
in a spreadsheet to verify this.)

However, when I called icst_hz() on these VCO settings it would
instead return 4122709 Hz. This causes an error in the common
clock driver for ICST as the common clock framework will call
.round_rate() on the clock which will utilize icst_hz_to_vco()
followed by icst_hz() suggesting the erroneous frequency, and
then the clock gets set to this.

The error did not manifest in the old clock framework since
this high frequency was only used by the CLCD, which calls
clk_set_rate() without first calling clk_round_rate() and since
the old clock framework would not call clk_round_rate() before
setting the frequency, the correct values propagated into
the VCO.

After some experimenting I figured out that it was due to a simple
arithmetic overflow: the divisor for 24Mhz reference frequency
as reference becomes 24000000*2*(99+8)=0x132212400 and the "1"
in bit 32 overflows and is lost.

But introducing an explicit 64-by-32 bit do_div() and casting
the divisor into (u64) we get the right frequency back, and the
right frequency gets set.

Tested on the ARM Versatile.

Cc: Pawel Moll <>
Signed-off-by: Linus Walleij <>
Signed-off-by: Russell King <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: timer: Fix wrong instance passed to slave callbacks
Takashi Iwai [Mon, 8 Feb 2016 16:36:25 +0000 (17:36 +0100)]
ALSA: timer: Fix wrong instance passed to slave callbacks

commit 117159f0b9d392fb433a7871426fad50317f06f7 upstream.

In snd_timer_notify1(), the wrong timer instance was passed for slave
ccallback function.  This leads to the access to the wrong data when
an incompatible master is handled (e.g. the master is the sequencer
timer and the slave is a user timer), as spotted by syzkaller fuzzer.

This patch fixes that wrong assignment.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: dummy: Implement timer backend switching more safely
Takashi Iwai [Tue, 2 Feb 2016 14:27:36 +0000 (15:27 +0100)]
ALSA: dummy: Implement timer backend switching more safely

commit ddce57a6f0a2d8d1bfacfa77f06043bc760403c2 upstream.

Currently the selected timer backend is referred at any moment from
the running PCM callbacks.  When the backend is switched, it's
possible to lead to inconsistency from the running backend.  This was
pointed by syzkaller fuzzer, and the commit [7ee96216c31a: ALSA:
dummy: Disable switching timer backend via sysfs] disabled the dynamic
switching for avoiding the crash.

This patch improves the handling of timer backend switching.  It keeps
the reference to the selected backend during the whole operation of an
opened stream so that it won't be changed by other streams.

Together with this change, the hrtimer parameter is reenabled as
writable now.

NOTE: this patch also turned out to fix the still remaining race.
Namely, ops was still replaced dynamically at dummy_pcm_open:

  static int dummy_pcm_open(struct snd_pcm_substream *substream)
          dummy->timer_ops = &dummy_systimer_ops;
          if (hrtimer)
                  dummy->timer_ops = &dummy_hrtimer_ops;

Since dummy->timer_ops is common among all streams, and when the
replacement happens during accesses of other streams, it may lead to a
crash.  This was actually triggered by syzkaller fuzzer and KASAN.

This patch rewrites the code not to use the ops shared by all streams
any longer, too.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoklist: fix starting point removed bug in klist iterators
James Bottomley [Wed, 13 Jan 2016 16:10:31 +0000 (08:10 -0800)]
klist: fix starting point removed bug in klist iterators

commit 00cd29b799e3449f0c68b1cc77cd4a5f95b42d17 upstream.

The starting node for a klist iteration is often passed in from
somewhere way above the klist infrastructure, meaning there's no
guarantee the node is still on the list.  We've seen this in SCSI where
we use bus_find_device() to iterate through a list of devices.  In the
face of heavy hotplug activity, the last device returned by
bus_find_device() can be removed before the next call.  This leads to

Dec  3 13:22:02 localhost kernel: WARNING: CPU: 2 PID: 28073 at include/linux/kref.h:47 klist_iter_init_node+0x3d/0x50()
Dec  3 13:22:02 localhost kernel: Modules linked in: scsi_debug x86_pkg_temp_thermal kvm_intel kvm irqbypass crc32c_intel joydev iTCO_wdt dcdbas ipmi_devintf acpi_power_meter iTCO_vendor_support ipmi_si imsghandler pcspkr wmi acpi_cpufreq tpm_tis tpm shpchp lpc_ich mfd_core nfsd nfs_acl lockd grace sunrpc tg3 ptp pps_core
Dec  3 13:22:02 localhost kernel: CPU: 2 PID: 28073 Comm: cat Not tainted 4.4.0-rc1+ #2
Dec  3 13:22:02 localhost kernel: Hardware name: Dell Inc. PowerEdge R320/08VT7V, BIOS 2.0.22 11/19/2013
Dec  3 13:22:02 localhost kernel: ffffffff81a20e77 ffff880613acfd18 ffffffff81321eef 0000000000000000
Dec  3 13:22:02 localhost kernel: ffff880613acfd50 ffffffff8107ca52 ffff88061176b198 0000000000000000
Dec  3 13:22:02 localhost kernel: ffffffff814542b0 ffff880610cfb100 ffff88061176b198 ffff880613acfd60
Dec  3 13:22:02 localhost kernel: Call Trace:
Dec  3 13:22:02 localhost kernel: [<ffffffff81321eef>] dump_stack+0x44/0x55
Dec  3 13:22:02 localhost kernel: [<ffffffff8107ca52>] warn_slowpath_common+0x82/0xc0
Dec  3 13:22:02 localhost kernel: [<ffffffff814542b0>] ? proc_scsi_show+0x20/0x20
Dec  3 13:22:02 localhost kernel: [<ffffffff8107cb4a>] warn_slowpath_null+0x1a/0x20
Dec  3 13:22:02 localhost kernel: [<ffffffff8167225d>] klist_iter_init_node+0x3d/0x50
Dec  3 13:22:02 localhost kernel: [<ffffffff81421d41>] bus_find_device+0x51/0xb0
Dec  3 13:22:02 localhost kernel: [<ffffffff814545ad>] scsi_seq_next+0x2d/0x40

And an eventual crash. It can actually occur in any hotplug system
which has a device finder and a starting device.

We can fix this globally by making sure the starting node for
klist_iter_init_node() is actually a member of the list before using it
(and by starting from the beginning if it isn't).

Reported-by: Ewan D. Milne <>
Tested-by: Ewan D. Milne <>
Signed-off-by: James Bottomley <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
7 years agocrypto: algif_skcipher - Do not dereference ctx without socket lock
Herbert Xu [Wed, 3 Feb 2016 13:39:26 +0000 (21:39 +0800)]
crypto: algif_skcipher - Do not dereference ctx without socket lock

commit 6454c2b83f719057069777132b13949e4c6b6350 upstream.

Any access to non-constant bits of the private context must be
done under the socket lock, in particular, this includes ctx->req.

This patch moves such accesses under the lock, and fetches the
tfm from the parent socket which is guaranteed to be constant,
rather than from ctx->req.

Signed-off-by: Herbert Xu <>
[bwh: Backported to 3.2:
 - Drop changes to skcipher_recvmsg_async
 - s/skcipher/ablkcipher/ in many places
 - s/skc->skcipher/skc->base/]
Signed-off-by: Ben Hutchings <>
7 years agocrypto: user - lock crypto_alg_list on alg dump
Mathias Krause [Mon, 1 Feb 2016 13:27:30 +0000 (14:27 +0100)]
crypto: user - lock crypto_alg_list on alg dump

commit 63e41ebc6630f39422d87f8a4bade1e793f37a01 upstream.

We miss to take the crypto_alg_sem semaphore when traversing the
crypto_alg_list for CRYPTO_MSG_GETALG dumps. This allows a race with
crypto_unregister_alg() removing algorithms from the list while we're
still traversing it, thereby leading to a use-after-free as show below:

[ 3482.071639] general protection fault: 0000 [#1] SMP
[ 3482.075639] Modules linked in: aes_x86_64 glue_helper lrw ablk_helper cryptd gf128mul ipv6 pcspkr serio_raw virtio_net microcode virtio_pci virtio_ring virtio sr_mod cdrom [last unloaded: aesni_intel]
[ 3482.075639] CPU: 1 PID: 11065 Comm: crconf Not tainted 4.3.4-grsec+ #126
[ 3482.075639] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.7.5-20140531_083030-gandalf 04/01/2014
[ 3482.075639] task: ffff88001cd41a40 ti: ffff88001cd422c8 task.ti: ffff88001cd422c8
[ 3482.075639] RIP: 0010:[<ffffffff93722bd3>]  [<ffffffff93722bd3>] strncpy+0x13/0x30
[ 3482.075639] RSP: 0018:ffff88001f713b60  EFLAGS: 00010202
[ 3482.075639] RAX: ffff88001f6c4430 RBX: ffff88001f6c43a0 RCX: ffff88001f6c4430
[ 3482.075639] RDX: 0000000000000040 RSI: fefefefefefeff16 RDI: ffff88001f6c4430
[ 3482.075639] RBP: ffff88001f713b60 R08: ffff88001f6c4470 R09: ffff88001f6c4480
[ 3482.075639] R10: 0000000000000002 R11: 0000000000000246 R12: ffff88001ce2aa28
[ 3482.075639] R13: ffff880000093700 R14: ffff88001f5e4bf8 R15: 0000000000003b20
[ 3482.075639] FS:  0000033826fa2700(0000) GS:ffff88001e900000(0000) knlGS:0000000000000000
[ 3482.075639] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 3482.075639] CR2: ffffffffff600400 CR3: 00000000139ec000 CR4: 00000000001606f0
[ 3482.075639] Stack:
[ 3482.075639]  ffff88001f713bd8 ffffffff936ccd00 ffff88001e5c4200 ffff880000093700
[ 3482.075639]  ffff88001f713bd0 ffffffff938ef4bf 0000000000000000 0000000000003b20
[ 3482.075639]  ffff88001f5e4bf8 ffff88001f5e4848 0000000000000000 0000000000003b20
[ 3482.075639] Call Trace:
[ 3482.075639]  [<ffffffff936ccd00>] crypto_report_alg+0xc0/0x3e0
[ 3482.075639]  [<ffffffff938ef4bf>] ? __alloc_skb+0x16f/0x300
[ 3482.075639]  [<ffffffff936cd08a>] crypto_dump_report+0x6a/0x90
[ 3482.075639]  [<ffffffff93935707>] netlink_dump+0x147/0x2e0
[ 3482.075639]  [<ffffffff93935f99>] __netlink_dump_start+0x159/0x190
[ 3482.075639]  [<ffffffff936ccb13>] crypto_user_rcv_msg+0xc3/0x130
[ 3482.075639]  [<ffffffff936cd020>] ? crypto_report_alg+0x3e0/0x3e0
[ 3482.075639]  [<ffffffff936cc4b0>] ? alg_test_crc32c+0x120/0x120
[ 3482.075639]  [<ffffffff93933145>] ? __netlink_lookup+0xd5/0x120
[ 3482.075639]  [<ffffffff936cca50>] ? crypto_add_alg+0x1d0/0x1d0
[ 3482.075639]  [<ffffffff93938141>] netlink_rcv_skb+0xe1/0x130
[ 3482.075639]  [<ffffffff936cc4f8>] crypto_netlink_rcv+0x28/0x40
[ 3482.075639]  [<ffffffff939375a8>] netlink_unicast+0x108/0x180
[ 3482.075639]  [<ffffffff93937c21>] netlink_sendmsg+0x541/0x770
[ 3482.075639]  [<ffffffff938e31e1>] sock_sendmsg+0x21/0x40
[ 3482.075639]  [<ffffffff938e4763>] SyS_sendto+0xf3/0x130
[ 3482.075639]  [<ffffffff93444203>] ? bad_area_nosemaphore+0x13/0x20
[ 3482.075639]  [<ffffffff93444470>] ? __do_page_fault+0x80/0x3a0
[ 3482.075639]  [<ffffffff939d80cb>] entry_SYSCALL_64_fastpath+0x12/0x6e
[ 3482.075639] Code: 88 4a ff 75 ed 5d 48 0f ba 2c 24 3f c3 66 66 2e 0f 1f 84 00 00 00 00 00 55 48 85 d2 48 89 f8 48 89 f9 4c 8d 04 17 48 89 e5 74 15 <0f> b6 16 80 fa 01 88 11 48 83 de ff 48 83 c1 01 4c 39 c1 75 eb
[ 3482.075639] RIP  [<ffffffff93722bd3>] strncpy+0x13/0x30

To trigger the race run the following loops simultaneously for a while:
  $ while : ; do modprobe aesni-intel; rmmod aesni-intel; done
  $ while : ; do crconf show all > /dev/null; done

Fix the race by taking the crypto_alg_sem read lock, thereby preventing
crypto_unregister_alg() from modifying the algorithm list during the

This bug has been detected by the PaX memory sanitize feature.

Signed-off-by: Mathias Krause <>
Cc: Steffen Klassert <>
Cc: PaX Team <>
Signed-off-by: Herbert Xu <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoocfs2/dlm: clear refmap bit of recovery lock while doing local recovery cleanup
xuejiufei [Fri, 5 Feb 2016 23:36:47 +0000 (15:36 -0800)]
ocfs2/dlm: clear refmap bit of recovery lock while doing local recovery cleanup

commit c95a51807b730e4681e2ecbdfd669ca52601959e upstream.

When recovery master down, dlm_do_local_recovery_cleanup() only remove
the $RECOVERY lock owned by dead node, but do not clear the refmap bit.
Which will make umount thread falling in dead loop migrating $RECOVERY
to the dead node.

Signed-off-by: xuejiufei <>
Reviewed-by: Joseph Qi <>
Cc: Mark Fasheh <>
Cc: Joel Becker <>
Cc: Junxiao Bi <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
7 years agomm, vmstat: fix wrong WQ sleep when memory reclaim doesn't make any progress
Tetsuo Handa [Fri, 5 Feb 2016 23:36:30 +0000 (15:36 -0800)]
mm, vmstat: fix wrong WQ sleep when memory reclaim doesn't make any progress

commit 564e81a57f9788b1475127012e0fd44e9049e342 upstream.

Jan Stancek has reported that system occasionally hanging after "oom01"
testcase from LTP triggers OOM.  Guessing from a result that there is a
kworker thread doing memory allocation and the values between "Node 0
Normal free:" and "Node 0 Normal:" differs when hanging, vmstat is not
up-to-date for some reason.

According to commit 373ccbe59270 ("mm, vmstat: allow WQ concurrency to
discover memory reclaim doesn't make any progress"), it meant to force
the kworker thread to take a short sleep, but it by error used
schedule_timeout(1).  We missed that schedule_timeout() in state
TASK_RUNNING doesn't do anything.

Fix it by using schedule_timeout_uninterruptible(1) which forces the
kworker thread to take a short sleep in order to make sure that vmstat
is up-to-date.

Fixes: 373ccbe59270 ("mm, vmstat: allow WQ concurrency to discover memory reclaim doesn't make any progress")
Signed-off-by: Tetsuo Handa <>
Reported-by: Jan Stancek <>
Acked-by: Michal Hocko <>
Cc: Tejun Heo <>
Cc: Cristopher Lameter <>
Cc: Joonsoo Kim <>
Cc: Arkadiusz Miskiewicz <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Ben Hutchings <>
7 years agoscsi_dh_rdac: always retry MODE SELECT on command lock violation
Hannes Reinecke [Fri, 22 Jan 2016 14:42:41 +0000 (15:42 +0100)]
scsi_dh_rdac: always retry MODE SELECT on command lock violation

commit d2d06d4fe0f2cc2df9b17fefec96e6e1a1271d91 upstream.

If MODE SELECT returns with sense '05/91/36' (command lock violation)
it should always be retried without counting the number of retries.
During an HBA upgrade or similar circumstances one might see a flood
of MODE SELECT command from various HBAs, which will easily trigger
the sense code and exceed the retry count.

Signed-off-by: Hannes Reinecke <>
Reviewed-by: Johannes Thumshirn <>
Signed-off-by: Martin K. Petersen <>
Signed-off-by: Ben Hutchings <>
7 years agosaa7134-alsa: Only frees registered sound cards
Mauro Carvalho Chehab [Thu, 4 Feb 2016 17:59:43 +0000 (15:59 -0200)]
saa7134-alsa: Only frees registered sound cards

commit ac75fe5d8fe4a0bf063be18fb29684405279e79e upstream.

That prevents this bug:
[ 2382.269496] BUG: unable to handle kernel NULL pointer dereference at 0000000000000540
[ 2382.270013] IP: [<ffffffffa01fe616>] snd_card_free+0x36/0x70 [snd]
[ 2382.270013] PGD 0
[ 2382.270013] Oops: 0002 [#1] SMP
[ 2382.270013] Modules linked in: saa7134_alsa(-) tda1004x saa7134_dvb videobuf2_dvb dvb_core tda827x tda8290 tuner saa7134 tveeprom videobuf2_dma_sg videobuf2_memops videobuf2_v4l2 videobuf2_core v4l2_common videodev media auth_rpcgss nfsv4 dns_resolver nfs lockd grace sunrpc tun bridge stp llc ebtables ip6table_filter ip6_tables nf_conntrack_ipv4 nf_defrag_ipv4 xt_conntrack nf_conntrack it87 hwmon_vid snd_hda_codec_idt snd_hda_codec_generic iTCO_wdt iTCO_vendor_support snd_hda_intel snd_hda_codec snd_hwdep snd_hda_core snd_seq pcspkr i2c_i801 snd_seq_device snd_pcm snd_timer lpc_ich snd mfd_core soundcore binfmt_misc i915 video i2c_algo_bit drm_kms_helper drm r8169 ata_generic serio_raw pata_acpi mii i2c_core [last unloaded: videobuf2_memops]
[ 2382.270013] CPU: 0 PID: 4899 Comm: rmmod Not tainted 4.5.0-rc1+ #4
[ 2382.270013] Hardware name: PCCHIPS P17G/P17G, BIOS 080012  05/14/2008
[ 2382.270013] task: ffff880039c38000 ti: ffff88003c764000 task.ti: ffff88003c764000
[ 2382.270013] RIP: 0010:[<ffffffffa01fe616>]  [<ffffffffa01fe616>] snd_card_free+0x36/0x70 [snd]
[ 2382.270013] RSP: 0018:ffff88003c767ea0  EFLAGS: 00010286
[ 2382.270013] RAX: ffff88003c767eb8 RBX: 0000000000000000 RCX: 0000000000006260
[ 2382.270013] RDX: ffffffffa020a060 RSI: ffffffffa0206de1 RDI: ffff88003c767eb0
[ 2382.270013] RBP: ffff88003c767ed8 R08: 0000000000019960 R09: ffffffff811a5412
[ 2382.270013] R10: ffffea0000d7c200 R11: 0000000000000000 R12: ffff88003c767ea8
[ 2382.270013] R13: 00007ffe760617f7 R14: 0000000000000000 R15: 0000557625d7f1e0
[ 2382.270013] FS:  00007f80bb1c0700(0000) GS:ffff88003f400000(0000) knlGS:0000000000000000
[ 2382.270013] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[ 2382.270013] CR2: 0000000000000540 CR3: 000000003c00f000 CR4: 00000000000006f0
[ 2382.270013] Stack:
[ 2382.270013]  000000003c767ed8 ffffffff00000000 ffff880000000000 ffff88003c767eb8
[ 2382.270013]  ffff88003c767eb8 ffffffffa049a890 00007ffe76060060 ffff88003c767ef0
[ 2382.270013]  ffffffffa049889d ffffffffa049a500 ffff88003c767f48 ffffffff8111079c
[ 2382.270013] Call Trace:
[ 2382.270013]  [<ffffffffa049889d>] saa7134_alsa_exit+0x1d/0x780 [saa7134_alsa]
[ 2382.270013]  [<ffffffff8111079c>] SyS_delete_module+0x19c/0x1f0
[ 2382.270013]  [<ffffffff8170fc2e>] entry_SYSCALL_64_fastpath+0x12/0x71
[ 2382.270013] Code: 20 a0 48 c7 c6 e1 6d 20 a0 48 89 e5 41 54 53 4c 8d 65 d0 48 89 fb 48 83 ec 28 c7 45 d0 00 00 00 00 49 8d 7c 24 08 e8 7a 55 ed e0 <4c> 89 a3 40 05 00 00 48 89 df e8 eb fd ff ff 85 c0 75 1a 48 8d
[ 2382.270013] RIP  [<ffffffffa01fe616>] snd_card_free+0x36/0x70 [snd]
[ 2382.270013]  RSP <ffff88003c767ea0>
[ 2382.270013] CR2: 0000000000000540

Signed-off-by: Mauro Carvalho Chehab <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: timer: Fix leftover link at closing
Takashi Iwai [Thu, 4 Feb 2016 16:06:13 +0000 (17:06 +0100)]
ALSA: timer: Fix leftover link at closing

commit 094fd3be87b0f102589e2d5c3fa5d06b7e20496d upstream.

In ALSA timer core, the active timer instance is managed in
active_list linked list.  Each element is added / removed dynamically
at timer start, stop and in timer interrupt.  The problem is that
snd_timer_interrupt() has a thinko and leaves the element in
active_list when it's the last opened element.  This eventually leads
to list corruption or use-after-free error.

This hasn't been revealed because we used to delete the list forcibly
in snd_timer_stop() in the past.  However, the recent fix avoids the
double-stop behavior (in commit [f784beb75ce8: ALSA: timer: Fix link
corruption due to double start or stop]), and this leak hits reality.

This patch fixes the link management in snd_timer_interrupt().  Now it
simply unlinks no matter which stream is.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agotda1004x: only update the frontend properties if locked
Mauro Carvalho Chehab [Wed, 3 Feb 2016 19:33:48 +0000 (17:33 -0200)]
tda1004x: only update the frontend properties if locked

commit e8beb02343e7582980c6705816cd957cf4f74c7a upstream.

The tda1004x was updating the properties cache before locking.
If the device is not locked, the data at the registers are just
random values with no real meaning.

This caused the driver to fail with libdvbv5, as such library
calls GET_PROPERTY from time to time, in order to return the
DVB stats.

Tested with a saa7134 card 78:
ASUSTeK P7131 Dual, vendor PCI ID: 1043:4862

Signed-off-by: Mauro Carvalho Chehab <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
7 years agoxhci: Fix list corruption in urb dequeue at host removal
Mathias Nyman [Tue, 26 Jan 2016 15:50:12 +0000 (17:50 +0200)]
xhci: Fix list corruption in urb dequeue at host removal

commit 5c82171167adb8e4ac77b91a42cd49fb211a81a0 upstream.

xhci driver frees data for all devices, both usb2 and and usb3 the
first time usb_remove_hcd() is called, including td_list and and xhci_ring

When usb_remove_hcd() is called a second time for the second xhci bus it
will try to dequeue all pending urbs, and touches td_list which is already
freed for that endpoint.

Reported-by: Joe Lawrence <>
Tested-by: Joe Lawrence <>
Signed-off-by: Mathias Nyman <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
7 years agousb: xhci: apply XHCI_PME_STUCK_QUIRK to Intel Broxton-M platforms
Lu Baolu [Tue, 26 Jan 2016 15:50:08 +0000 (17:50 +0200)]
usb: xhci: apply XHCI_PME_STUCK_QUIRK to Intel Broxton-M platforms

commit ccc04afb72cddbdf7c0e1c17e92886405a71b754 upstream.

Intel Broxton M was verifed to require XHCI_PME_STUCK_QUIRK quirk as well.

Signed-off-by: Lu Baolu <>
Signed-off-by: Mathias Nyman <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
7 years agoRevert "xhci: don't finish a TD if we get a short-transfer event mid TD"
Mathias Nyman [Tue, 26 Jan 2016 15:50:04 +0000 (17:50 +0200)]
Revert "xhci: don't finish a TD if we get a short-transfer event mid TD"

commit a6835090716a85f2297668ba593bd00e1051e662 upstream.

This reverts commit e210c422b6fd ("xhci: don't finish a TD if we get a
short transfer event mid TD")

Turns out that most host controllers do not follow the xHCI specs and never
send the second event for the last TRB in the TD if there was a short event

Returning the URB directly after the first short-transfer event is far
better than never returning the URB. (class drivers usually timeout
after 30sec). For the hosts that do send the second event we will go
back to treating it as misplaced event and print an error message for it.

The origial patch was sent to stable kernels and needs to be reverted from
there as well

Signed-off-by: Mathias Nyman <>
Signed-off-by: Greg Kroah-Hartman <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: seq: Fix lockdep warnings due to double mutex locks
Takashi Iwai [Wed, 3 Feb 2016 07:32:44 +0000 (08:32 +0100)]
ALSA: seq: Fix lockdep warnings due to double mutex locks

commit 7f0973e973cd74aa40747c9d38844560cd184ee8 upstream.

The port subscription code uses double mutex locks for source and
destination ports, and this may become racy once when wrongly set up.
It leads to lockdep warning splat, typically triggered by fuzzer like
syzkaller, although the actual deadlock hasn't been seen, so far.

This patch simplifies the handling by reducing to two single locks, so
that no lockdep warning will be trigger any longer.

By splitting to two actions, a still-in-progress element shall be
added in one list while handling another.  For ignoring this element,
a new check is added in deliver_to_subscribers().

Along with it, the code to add/remove the subscribers list element was
cleaned up and refactored.

Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: rawmidi: Fix race at copying & updating the position
Takashi Iwai [Wed, 3 Feb 2016 13:41:22 +0000 (14:41 +0100)]
ALSA: rawmidi: Fix race at copying & updating the position

commit 81f577542af15640cbcb6ef68baa4caa610cbbfc upstream.

The rawmidi read and write functions manage runtime stream status
such as runtime->appl_ptr and runtime->avail.  These point where to
copy the new data and how many bytes have been copied (or to be
read).  The problem is that rawmidi read/write call copy_from_user()
or copy_to_user(), and the runtime spinlock is temporarily unlocked
and relocked while copying user-space.  Since the current code
advances and updates the runtime status after the spin unlock/relock,
the copy and the update may be asynchronous, and eventually
runtime->avail might go to a negative value when many concurrent
accesses are done.  This may lead to memory corruption in the end.

For fixing this race, in this patch, the status update code is
performed in the same lock before the temporary unlock.  Also, the
spinlock is now taken more widely in snd_rawmidi_kernel_read1() for
protecting more properly during the whole operation.

Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: rawmidi: Make snd_rawmidi_transmit() race-free
Takashi Iwai [Sun, 31 Jan 2016 10:57:41 +0000 (11:57 +0100)]
ALSA: rawmidi: Make snd_rawmidi_transmit() race-free

commit 06ab30034ed9c200a570ab13c017bde248ddb2a6 upstream.

A kernel WARNING in snd_rawmidi_transmit_ack() is triggered by
syzkaller fuzzer:
  WARNING: CPU: 1 PID: 20739 at sound/core/rawmidi.c:1136
Call Trace:
 [<     inline     >] __dump_stack lib/dump_stack.c:15
 [<ffffffff82999e2d>] dump_stack+0x6f/0xa2 lib/dump_stack.c:50
 [<ffffffff81352089>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:482
 [<ffffffff813522b9>] warn_slowpath_null+0x29/0x30 kernel/panic.c:515
 [<ffffffff84f80bd5>] snd_rawmidi_transmit_ack+0x275/0x400 sound/core/rawmidi.c:1136
 [<ffffffff84fdb3c1>] snd_virmidi_output_trigger+0x4b1/0x5a0 sound/core/seq/seq_virmidi.c:163
 [<     inline     >] snd_rawmidi_output_trigger sound/core/rawmidi.c:150
 [<ffffffff84f87ed9>] snd_rawmidi_kernel_write1+0x549/0x780 sound/core/rawmidi.c:1223
 [<ffffffff84f89fd3>] snd_rawmidi_write+0x543/0xb30 sound/core/rawmidi.c:1273
 [<ffffffff817b0323>] __vfs_write+0x113/0x480 fs/read_write.c:528
 [<ffffffff817b1db7>] vfs_write+0x167/0x4a0 fs/read_write.c:577
 [<     inline     >] SYSC_write fs/read_write.c:624
 [<ffffffff817b50a1>] SyS_write+0x111/0x220 fs/read_write.c:616
 [<ffffffff86336c36>] entry_SYSCALL_64_fastpath+0x16/0x7a arch/x86/entry/entry_64.S:185

Also a similar warning is found but in another path:
Call Trace:
 [<     inline     >] __dump_stack lib/dump_stack.c:15
 [<ffffffff82be2c0d>] dump_stack+0x6f/0xa2 lib/dump_stack.c:50
 [<ffffffff81355139>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:482
 [<ffffffff81355369>] warn_slowpath_null+0x29/0x30 kernel/panic.c:515
 [<ffffffff8527e69a>] rawmidi_transmit_ack+0x24a/0x3b0 sound/core/rawmidi.c:1133
 [<ffffffff8527e851>] snd_rawmidi_transmit_ack+0x51/0x80 sound/core/rawmidi.c:1163
 [<ffffffff852d9046>] snd_virmidi_output_trigger+0x2b6/0x570 sound/core/seq/seq_virmidi.c:185
 [<     inline     >] snd_rawmidi_output_trigger sound/core/rawmidi.c:150
 [<ffffffff85285a0b>] snd_rawmidi_kernel_write1+0x4bb/0x760 sound/core/rawmidi.c:1252
 [<ffffffff85287b73>] snd_rawmidi_write+0x543/0xb30 sound/core/rawmidi.c:1302
 [<ffffffff817ba5f3>] __vfs_write+0x113/0x480 fs/read_write.c:528
 [<ffffffff817bc087>] vfs_write+0x167/0x4a0 fs/read_write.c:577
 [<     inline     >] SYSC_write fs/read_write.c:624
 [<ffffffff817bf371>] SyS_write+0x111/0x220 fs/read_write.c:616
 [<ffffffff86660276>] entry_SYSCALL_64_fastpath+0x16/0x7a arch/x86/entry/entry_64.S:185

In the former case, the reason is that virmidi has an open code
calling snd_rawmidi_transmit_ack() with the value calculated outside
the spinlock.   We may use snd_rawmidi_transmit() in a loop just for
consuming the input data, but even there, there is a race between
snd_rawmidi_transmit_peek() and snd_rawmidi_tranmit_ack().

Similarly in the latter case, it calls snd_rawmidi_transmit_peek() and
snd_rawmidi_tranmit_ack() separately without protection, so they are
racy as well.

The patch tries to address these issues by the following ways:
- Introduce the unlocked versions of snd_rawmidi_transmit_peek() and
  snd_rawmidi_transmit_ack() to be called inside the explicit lock.
- Rewrite snd_rawmidi_transmit() to be race-free (the former case).
- Make the split calls (the latter case) protected in the rawmidi spin

Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <>
7 years agolibata: fix sff host state machine locking while polling
Tejun Heo [Mon, 1 Feb 2016 16:33:21 +0000 (11:33 -0500)]
libata: fix sff host state machine locking while polling

commit 8eee1d3ed5b6fc8e14389567c9a6f53f82bb7224 upstream.

The bulk of ATA host state machine is implemented by
ata_sff_hsm_move().  The function is called from either the interrupt
handler or, if polling, a work item.  Unlike from the interrupt path,
the polling path calls the function without holding the host lock and
ata_sff_hsm_move() selectively grabs the lock.

This is completely broken.  If an IRQ triggers while polling is in
progress, the two can easily race and end up accessing the hardware
and updating state machine state at the same time.  This can put the
state machine in an illegal state and lead to a crash like the

  kernel BUG at drivers/ata/libata-sff.c:1302!
  invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN
  Modules linked in:
  CPU: 1 PID: 10679 Comm: syz-executor Not tainted 4.5.0-rc1+ #300
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
  task: ffff88002bd00000 ti: ffff88002e048000 task.ti: ffff88002e048000
  RIP: 0010:[<ffffffff83a83409>]  [<ffffffff83a83409>] ata_sff_hsm_move+0x619/0x1c60
  Call Trace:
   [<ffffffff83a84c31>] __ata_sff_port_intr+0x1e1/0x3a0 drivers/ata/libata-sff.c:1584
   [<ffffffff83a85611>] ata_bmdma_port_intr+0x71/0x400 drivers/ata/libata-sff.c:2877
   [<     inline     >] __ata_sff_interrupt drivers/ata/libata-sff.c:1629
   [<ffffffff83a85bf3>] ata_bmdma_interrupt+0x253/0x580 drivers/ata/libata-sff.c:2902
   [<ffffffff81479f98>] handle_irq_event_percpu+0x108/0x7e0 kernel/irq/handle.c:157
   [<ffffffff8147a717>] handle_irq_event+0xa7/0x140 kernel/irq/handle.c:205
   [<ffffffff81484573>] handle_edge_irq+0x1e3/0x8d0 kernel/irq/chip.c:623
   [<     inline     >] generic_handle_irq_desc include/linux/irqdesc.h:146
   [<ffffffff811a92bc>] handle_irq+0x10c/0x2a0 arch/x86/kernel/irq_64.c:78
   [<ffffffff811a7e4d>] do_IRQ+0x7d/0x1a0 arch/x86/kernel/irq.c:240
   [<ffffffff86653d4c>] common_interrupt+0x8c/0x8c arch/x86/entry/entry_64.S:520
   [<     inline     >] rcu_lock_acquire include/linux/rcupdate.h:490
   [<     inline     >] rcu_read_lock include/linux/rcupdate.h:874
   [<ffffffff8164b4a1>] filemap_map_pages+0x131/0xba0 mm/filemap.c:2145
   [<     inline     >] do_fault_around mm/memory.c:2943
   [<     inline     >] do_read_fault mm/memory.c:2962
   [<     inline     >] do_fault mm/memory.c:3133
   [<     inline     >] handle_pte_fault mm/memory.c:3308
   [<     inline     >] __handle_mm_fault mm/memory.c:3418
   [<ffffffff816efb16>] handle_mm_fault+0x2516/0x49a0 mm/memory.c:3447
   [<ffffffff8127dc16>] __do_page_fault+0x376/0x960 arch/x86/mm/fault.c:1238
   [<ffffffff8127e358>] trace_do_page_fault+0xe8/0x420 arch/x86/mm/fault.c:1331
   [<ffffffff8126f514>] do_async_page_fault+0x14/0xd0 arch/x86/kernel/kvm.c:264
   [<ffffffff86655578>] async_page_fault+0x28/0x30 arch/x86/entry/entry_64.S:986

Fix it by ensuring that the polling path is holding the host lock
before entering ata_sff_hsm_move() so that all hardware accesses and
state updates are performed under the host lock.

Signed-off-by: Tejun Heo <>
Reported-and-tested-by: Dmitry Vyukov <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: timer: Fix link corruption due to double start or stop
Takashi Iwai [Sat, 30 Jan 2016 22:09:08 +0000 (23:09 +0100)]
ALSA: timer: Fix link corruption due to double start or stop

commit f784beb75ce82f4136f8a0960d3ee872f7109e09 upstream.

Although ALSA timer code got hardening for races, it still causes
use-after-free error.  This is however rather a corrupted linked list,
not actually the concurrent accesses.  Namely, when timer start is
triggered twice, list_add_tail() is called twice, too.  This ends
up with the link corruption and triggers KASAN error.

The simplest fix would be replacing list_add_tail() with
list_move_tail(), but fundamentally it's the problem that we don't
check the double start/stop correctly.  So, the right fix here is to
add the proper checks to snd_timer_start() and snd_timer_stop() (and
their variants).

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: adjust context, indentation]
Signed-off-by: Ben Hutchings <>
7 years agoALSA: seq: Fix yet another races among ALSA timer accesses
Takashi Iwai [Sat, 30 Jan 2016 22:30:25 +0000 (23:30 +0100)]
ALSA: seq: Fix yet another races among ALSA timer accesses

commit 2cdc7b636d55cbcf42e1e6c8accd85e62d3e9ae8 upstream.

ALSA sequencer may open/close and control ALSA timer instance
dynamically either via sequencer events or direct ioctls.  These are
done mostly asynchronously, and it may call still some timer action
like snd_timer_start() while another is calling snd_timer_close().
Since the instance gets removed by snd_timer_close(), it may lead to
a use-after-free.

This patch tries to address such a race by protecting each
snd_timer_*() call via the existing spinlock and also by avoiding the
access to timer during close call.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: pcm: Fix potential deadlock in OSS emulation
Takashi Iwai [Sun, 31 Jan 2016 09:32:37 +0000 (10:32 +0100)]
ALSA: pcm: Fix potential deadlock in OSS emulation

commit b248371628aad599a48540962f6b85a21a8a0c3f upstream.

There are potential deadlocks in PCM OSS emulation code while
accessing read/write and mmap concurrently.  This comes from the
infamous mmap_sem usage in copy_from/to_user().  Namely,

   snd_pcm_oss_write() ->
     &runtime->oss.params_lock ->
        copy_to_user() ->
  mmap() ->
    &mm->mmap_sem ->
      snd_pcm_oss_mmap() ->

Since we can't avoid taking params_lock from mmap code path, use
trylock variant and aborts with -EAGAIN as a workaround of this AB/BA

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: rawmidi: Remove kernel WARNING for NULL user-space buffer check
Takashi Iwai [Mon, 1 Feb 2016 11:04:55 +0000 (12:04 +0100)]
ALSA: rawmidi: Remove kernel WARNING for NULL user-space buffer check

commit cc85f7a634cfaf9f0713c6aa06d08817424db37a upstream.

NULL user-space buffer can be passed even in a normal path, thus it's
not good to spew a kernel warning with stack trace at each time.
Just drop snd_BUG_ON() macro usage there.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: seq: Fix race at closing in virmidi driver
Takashi Iwai [Mon, 1 Feb 2016 11:06:42 +0000 (12:06 +0100)]
ALSA: seq: Fix race at closing in virmidi driver

commit 2d1b5c08366acd46c35a2e9aba5d650cb5bf5c19 upstream.

The virmidi driver has an open race at closing its assigned rawmidi
device, and this may lead to use-after-free in

Plug the hole by properly protecting the linked list deletion and
calling in the right order in snd_virmidi_input_close().

Reported-by: Dmitry Vyukov <>
Tested-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agointel_scu_ipcutil: underflow in scu_reg_access()
Dan Carpenter [Tue, 26 Jan 2016 09:24:25 +0000 (12:24 +0300)]
intel_scu_ipcutil: underflow in scu_reg_access()

commit b1d353ad3d5835b16724653b33c05124e1b5acf1 upstream.

"count" is controlled by the user and it can be negative.  Let's prevent
that by making it unsigned.  You have to have CAP_SYS_RAWIO to call this
function so the bug is not as serious as it could be.

Fixes: 5369c02d951a ('intel_scu_ipc: Utility driver for intel scu ipc')
Signed-off-by: Dan Carpenter <>
Signed-off-by: Darren Hart <>
Signed-off-by: Ben Hutchings <>
7 years agocrypto: algif_hash - wait for crypto_ahash_init() to complete
Wang, Rui Y [Wed, 27 Jan 2016 09:08:37 +0000 (17:08 +0800)]
crypto: algif_hash - wait for crypto_ahash_init() to complete

commit fe09786178f9df713a4b2dd6b93c0a722346bf5e upstream.

hash_sendmsg/sendpage() need to wait for the completion
of crypto_ahash_init() otherwise it can cause panic.

Signed-off-by: Rui Wang <>
Signed-off-by: Herbert Xu <>
Signed-off-by: Ben Hutchings <>
7 years agox86/mm/pat: Avoid truncation when converting cpa->numpages to address
Matt Fleming [Fri, 29 Jan 2016 11:36:10 +0000 (11:36 +0000)]
x86/mm/pat: Avoid truncation when converting cpa->numpages to address

commit 742563777e8da62197d6cb4b99f4027f59454735 upstream.

There are a couple of nasty truncation bugs lurking in the pageattr
code that can be triggered when mapping EFI regions, e.g. when we pass
a cpa->pgd pointer. Because cpa->numpages is a 32-bit value, shifting
left by PAGE_SHIFT will truncate the resultant address to 32-bits.

Viorel-Cătălin managed to trigger this bug on his Dell machine that
provides a ~5GB EFI region which requires 1236992 pages to be mapped.
When calling populate_pud() the end of the region gets calculated
incorrectly in the following buggy expression,

  end = start + (cpa->numpages << PAGE_SHIFT);

And only 188416 pages are mapped. Next, populate_pud() gets invoked
for a second time because of the loop in __change_page_attr_set_clr(),
only this time no pages get mapped because shifting the remaining
number of pages (1048576) by PAGE_SHIFT is zero. At which point the
loop in __change_page_attr_set_clr() spins forever because we fail to
map progress.

Hitting this bug depends very much on the virtual address we pick to
map the large region at and how many pages we map on the initial run
through the loop. This explains why this issue was only recently hit
with the introduction of commit

  a5caa209ba9c ("x86/efi: Fix boot crash by mapping EFI memmap
   entries bottom-up at runtime, instead of top-down")

It's interesting to note that safe uses of cpa->numpages do exist in
the pageattr code. If instead of shifting ->numpages we multiply by
PAGE_SIZE, no truncation occurs because PAGE_SIZE is a UL value, and
so the result is unsigned long.

To avoid surprises when users try to convert very large cpa->numpages
values to addresses, change the data type from 'int' to 'unsigned
long', thereby making it suitable for shifting by PAGE_SHIFT without
any type casting.

The alternative would be to make liberal use of casting, but that is
far more likely to cause problems in the future when someone adds more
code and fails to cast properly; this bug was difficult enough to
track down in the first place.

Reported-and-tested-by: Viorel-Cătălin Răpițeanu <>
Acked-by: Borislav Petkov <>
Cc: Sai Praneeth Prakhya <>
Signed-off-by: Matt Fleming <>
Signed-off-by: Thomas Gleixner <>
Signed-off-by: Ben Hutchings <>
7 years agodrm/vmwgfx: respect 'nomodeset'
Rob Clark [Wed, 15 Oct 2014 19:00:47 +0000 (15:00 -0400)]
drm/vmwgfx: respect 'nomodeset'

commit 96c5d076f0a5e2023ecdb44d8261f87641ee71e0 upstream.

Signed-off-by: Rob Clark <>
Reviewed-by: Thomas Hellstrom <>.
Signed-off-by: Dave Airlie <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: dummy: Disable switching timer backend via sysfs
Takashi Iwai [Thu, 28 Jan 2016 06:54:16 +0000 (07:54 +0100)]
ALSA: dummy: Disable switching timer backend via sysfs

commit 7ee96216c31aabe1eb42fb91ff50dae9fcd014b2 upstream.

ALSA dummy driver can switch the timer backend between system timer
and hrtimer via its hrtimer module option.  This can be also switched
dynamically via sysfs, but it may lead to a memory corruption when
switching is done while a PCM stream is running; the stream instance
for the newly switched timer method tries to access the memory that
was allocated by another timer method although the sizes differ.

As the simplest fix, this patch just disables the switch via sysfs by
dropping the writable bit.

Reported-by: Dmitry Vyukov <>
Signed-off-by: Takashi Iwai <>
Signed-off-by: Ben Hutchings <>
7 years agocrypto: shash - Fix has_key setting
Herbert Xu [Tue, 26 Jan 2016 16:16:37 +0000 (00:16 +0800)]
crypto: shash - Fix has_key setting

commit 00420a65fa2beb3206090ead86942484df2275f3 upstream.

The has_key logic is wrong for shash algorithms as they always
have a setkey function.  So we should instead be testing against

Fixes: a5596d633278 ("crypto: hash - Add crypto_ahash_has_setkey")
Reported-by: Stephan Mueller <>
Signed-off-by: Herbert Xu <>
Tested-by: Stephan Mueller <>
Signed-off-by: Ben Hutchings <>
7 years agotty: Fix unsafe ldisc reference via ioctl(TIOCGETD)
Peter Hurley [Mon, 11 Jan 2016 06:40:55 +0000 (22:40 -0800)]
tty: Fix unsafe ldisc reference via ioctl(TIOCGETD)

commit 5c17c861a357e9458001f021a7afa7aab9937439 upstream.

ioctl(TIOCGETD) retrieves the line discipline id directly from the
ldisc because the line discipline id (c_line) in termios is untrustworthy;
userspace may have set termios via ioctl(TCSETS*) without actually
changing the line discipline via ioctl(TIOCSETD).

However, directly accessing the current ldisc via tty->ldisc is
unsafe; the ldisc ptr dereferenced may be stale if the line discipline
is changing via ioctl(TIOCSETD) or hangup.

Wait for the line discipline reference (just like read() or write())
to retrieve the "current" line discipline id.

Signed-off-by: Peter Hurley <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Ben Hutchings <>
7 years agoSCSI: fix crashes in sd and sr runtime PM
Alan Stern [Wed, 20 Jan 2016 16:26:01 +0000 (11:26 -0500)]
SCSI: fix crashes in sd and sr runtime PM

commit 13b4389143413a1f18127c07f72c74cad5b563e8 upstream.

Runtime suspend during driver probe and removal can cause problems.
The driver's runtime_suspend or runtime_resume callbacks may invoked
before the driver has finished binding to the device or after the
driver has unbound from the device.

This problem shows up with the sd and sr drivers, and can cause disk
or CD/DVD drives to become unusable as a result.  The fix is simple.
The drivers store a pointer to the scsi_disk or scsi_cd structure as
their private device data when probing is finished, so we simply have
to be sure to clear the private data during removal and test it during
runtime suspend/resume.

This fixes <>.

Signed-off-by: Alan Stern <>
Reported-by: Paul Menzel <>
Reported-by: Erich Schubert <>
Reported-by: Alexandre Rossi <>
Tested-by: Paul Menzel <>
Tested-by: Erich Schubert <>
Signed-off-by: James Bottomley <>
[bwh: Backported to 3.2: drop changes to sr as it doesn't support runtime PM]
Signed-off-by: Ben Hutchings <>
7 years agoperf annotate browser: Fix behaviour of Shift-Tab with nothing focussed
Markus Trippelsdorf [Mon, 14 Dec 2015 15:44:03 +0000 (16:44 +0100)]
perf annotate browser: Fix behaviour of Shift-Tab with nothing focussed

commit d4913cbd05bab685e49c8174896e563b2487d054 upstream.

The issue was pointed out by gcc-6's -Wmisleading-indentation.

Signed-off-by: Markus Trippelsdorf <>
Acked-by: Ingo Molnar <>
Cc: Ben Hutchings <>
Cc: Matt Fleming <>
Cc: Peter Zijlstra <>
Fixes: c97cf42219b7 ("perf top: Live TUI Annotation")
Signed-off-by: Arnaldo Carvalho de Melo <>
[bwh: Backported to 3.2: adjust filename, context]
Signed-off-by: Ben Hutchings <>
7 years agorfkill: fix rfkill_fop_read wait_event usage
Johannes Berg [Tue, 26 Jan 2016 10:29:03 +0000 (11:29 +0100)]
rfkill: fix rfkill_fop_read wait_event usage

commit 6736fde9672ff6717ac576e9bba2fd5f3dfec822 upstream.

The code within wait_event_interruptible() is called with
!TASK_RUNNING, so mustn't call any functions that can sleep,
like mutex_lock().

Since we re-check the list_empty() in a loop after the wait,
it's safe to simply use list_empty() without locking.

This bug has existed forever, but was only discovered now
because all userspace implementations, including the default
'rfkill' tool, use poll() or select() to get a readable fd
before attempting to read.

Fixes: c64fb01627e24 ("rfkill: create useful userspace interface")
Reported-by: Dmitry Vyukov <>
Signed-off-by: Johannes Berg <>
Signed-off-by: Ben Hutchings <>
7 years agovirtio_pci: fix use after free on release
Michael S. Tsirkin [Thu, 14 Jan 2016 14:00:41 +0000 (16:00 +0200)]
virtio_pci: fix use after free on release

commit 2989be09a8a9d62a785137586ad941f916e08f83 upstream.

KASan detected a use-after-free error in virtio-pci remove code. In
virtio_pci_remove(), vp_dev is still used after being freed in
unregister_virtio_device() (in virtio_pci_release_dev() more

To fix, keep a reference until cleanup is done.

Fixes: 63bd62a08ca4 ("virtio_pci: defer kfree until release callback")
Reported-by: Jerome Marchand <>
Cc: Sasha Levin <>
Signed-off-by: Michael S. Tsirkin <>
Tested-by: Jerome Marchand <>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <>
7 years agolibata: disable forced PORTS_IMPL for >= AHCI 1.3
Tejun Heo [Fri, 15 Jan 2016 20:13:05 +0000 (15:13 -0500)]
libata: disable forced PORTS_IMPL for >= AHCI 1.3

commit 566d1827df2ef0cbe921d3d6946ac3007b1a6938 upstream.

Some early controllers incorrectly reported zero ports in PORTS_IMPL
register and the ahci driver fabricates PORTS_IMPL from the number of
ports in those cases.  This hasn't mattered but with the new nvme
controllers there are cases where zero PORTS_IMPL is valid and should
be honored.

Disable the workaround for >= AHCI 1.3.

Signed-off-by: Tejun Heo <>
Reported-by: Andy Lutomirski <>
Cc: Sergei Shtylyov <>
Signed-off-by: Ben Hutchings <>
7 years agoPCI/AER: Flush workqueue on device remove to avoid use-after-free
Sebastian Andrzej Siewior [Mon, 25 Jan 2016 16:08:00 +0000 (10:08 -0600)]
PCI/AER: Flush workqueue on device remove to avoid use-after-free

commit 4ae2182b1e3407de369f8c5d799543b7db74221b upstream.

A Root Port's AER structure (rpc) contains a queue of events.  aer_irq()
enqueues AER status information and schedules aer_isr() to dequeue and
process it.  When we remove a device, aer_remove() waits for the queue to
be empty, then frees the rpc struct.

But aer_isr() references the rpc struct after dequeueing and possibly
emptying the queue, which can cause a use-after-free error as in the
following scenario with two threads, aer_isr() on the left and a
concurrent aer_remove() on the right:

  Thread A                      Thread B
  --------                      --------
                                  wait_event(rpc->prod_idx == rpc->cons_idx)
                                  # now blocked until queue becomes empty
  aer_isr():                      # ...
    rpc->cons_idx++               # unblocked because queue is now empty
    ...                           kfree(rpc)

To prevent this problem, use flush_work() to wait until the last scheduled
instance of aer_isr() has completed before freeing the rpc struct in

I reproduced this use-after-free by flashing a device FPGA and
re-enumerating the bus to find the new device.  With SLUB debug, this
crashes with 0x6b bytes (POISON_FREE, the use-after-free magic number) in

  pcieport 0000:00:00.0: AER: Multiple Corrected error received: id=0000
  Unable to handle kernel paging request for data at address 0x27ef9e3e
  Workqueue: events aer_isr
  GPR24: dd6aa000 6b6b6b6b 605f8378 605f8360 d99b12c0 604fc674 606b1704 d99b12c0
  NIP [602f5328] pci_walk_bus+0xd4/0x104

[bhelgaas: changelog, stable tag]
Signed-off-by: Sebastian Andrzej Siewior <>
Signed-off-by: Bjorn Helgaas <>
Signed-off-by: Ben Hutchings <>
7 years agoUSB: serial: ftdi_sio: add support for Yaesu SCU-18 cable
Greg Kroah-Hartman [Wed, 20 Jan 2016 07:43:13 +0000 (23:43 -0800)]
USB: serial: ftdi_sio: add support for Yaesu SCU-18 cable

commit e03cdf22a2727c60307be6a729233edab3bfda9c upstream.

Harald Linden reports that the ftdi_sio driver works properly for the
Yaesu SCU-18 cable if the device ids are added to the driver.  So let's
add them.

Reported-by: Harald Linden <>
Signed-off-by: Greg Kroah-Hartman <>
Signed-off-by: Johan Hovold <>
Signed-off-by: Ben Hutchings <>
7 years agoALSA: seq: Degrade the error message for too many opens
Takashi Iwai [Mon, 25 Jan 2016 10:24:56 +0000 (11:24 +0100)]
ALSA: seq: Degrade the error message for too many opens

commit da10816e3d923565b470fec78a674baba794ed33 upstream.

ALSA OSS sequencer spews a kernel error message ("ALSA: seq_oss: too
many applications") when user-space tries to open more than the
limit.  This means that it can easily fill the log buffer.

Since it's merely a normal error, it's safe to suppress it via
pr_debug() instead.

Signed-off-by: Takashi Iwai <>
[bwh: Backported to 3.2: this was still using snd_printk()]
Signed-off-by: Ben Hutchings <>