10 years agoMerge git://
Linus Torvalds [Fri, 5 Aug 2011 16:42:01 +0000 (06:42 -1000)]
Merge git://git./linux/kernel/git/davem/net

* git:// (54 commits)
  ipv6: check for IPv4 mapped addresses when connecting IPv6 sockets
  mlx4: decreasing ref count when removing mac
  net: Fix security_socket_sendmsg() bypass problem.
  net: Cap number of elements for sendmmsg
  net: sendmmsg should only return an error if no messages were sent
  ixgbe: fix PHY link setup for 82599
  ixgbe: fix __ixgbe_notify_dca() bail out code
  igb: fix WOL on second port of i350 device
  e1000e: minor re-order of #include files
  e1000e: remove unnecessary check for NULL pointer
  intel drivers: repair missing flush operations
  macb: restore wrap bit when performing underrun cleanup
  cdc_ncm: fix endianness problem.
  irda: use PCI_VENDOR_ID_*
  mlx4: Fixing Ethernet unicast packet steering
  net: fix NULL dereferences in check_peer_redir()
  bnx2x: Clear MDIO access warning during first driver load
  bnx2x: Fix BCM578xx MAC test
  bnx2x: Fix BCM54618se invalid link indication
  bnx2x: Fix BCM84833 link

10 years agoMerge git://
Linus Torvalds [Fri, 5 Aug 2011 16:41:10 +0000 (06:41 -1000)]
Merge git://git./linux/kernel/git/davem/ide

* git://
  drivers/ide/cy82c693.c: Add missing pci_dev_put
  ide: Fix irq flags madness

10 years agoipv6: check for IPv4 mapped addresses when connecting IPv6 sockets
Max Matveev [Fri, 5 Aug 2011 10:56:30 +0000 (03:56 -0700)]
ipv6: check for IPv4 mapped addresses when connecting IPv6 sockets

When support for binding to 'mapped INADDR_ANY (::ffff.' was added
in 0f8d3c7ac3693d7b6c731bf2159273a59bf70e12 the rest of the code
wasn't told so now it's possible to bind IPv6 datagram socket to
::ffff., connect it to another IPv4 address and it will all
work except for getsockhame() which does not return the local address
as expected.

To give getsockname() something to work with check for 'mapped INADDR_ANY'
when connecting and update the in-core source addresses appropriately.

Signed-off-by: Max Matveev <>
Signed-off-by: David S. Miller <>
10 years agomlx4: decreasing ref count when removing mac
Yevgeny Petrilin [Thu, 4 Aug 2011 01:05:12 +0000 (01:05 +0000)]
mlx4: decreasing ref count when removing mac

For older FW versions, when a Mac address removed from Mac table,
we should set 0 for reference count for the corresponding Mac index.
Fixes a bug where removing Mac from the table still left that entry as

Signed-off-by: Yevgeny Petrilin <>
Tested-by: Roland Dreier <>
Signed-off-by: David S. Miller <>
10 years agoMerge branch 'master' of
David S. Miller [Fri, 5 Aug 2011 10:35:00 +0000 (03:35 -0700)]
Merge branch 'master' of /linux/kernel/git/jkirsher/net

10 years agonet: Fix security_socket_sendmsg() bypass problem.
Tetsuo Handa [Thu, 4 Aug 2011 14:07:40 +0000 (14:07 +0000)]
net: Fix security_socket_sendmsg() bypass problem.

The sendmmsg() introduced by commit 228e548e "net: Add sendmmsg socket system
call" is capable of sending to multiple different destination addresses.

SMACK is using destination's address for checking sendmsg() permission.
However, security_socket_sendmsg() is called for only once even if multiple
different destination addresses are passed to sendmmsg().

Therefore, we need to call security_socket_sendmsg() for each destination
address rather than only the first destination address.

Since calling security_socket_sendmsg() every time when only single destination
address was passed to sendmmsg() is a waste of time, omit calling
security_socket_sendmsg() unless destination address of previous datagram and
that of current datagram differs.

Signed-off-by: Tetsuo Handa <>
Acked-by: Anton Blanchard <>
Cc: stable <> [3.0+]
Signed-off-by: David S. Miller <>
10 years agonet: Cap number of elements for sendmmsg
Anton Blanchard [Thu, 4 Aug 2011 14:07:39 +0000 (14:07 +0000)]
net: Cap number of elements for sendmmsg

To limit the amount of time we can spend in sendmmsg, cap the
number of elements to UIO_MAXIOV (currently 1024).

For error handling an application using sendmmsg needs to retry at
the first unsent message, so capping is simpler and requires less
application logic than returning EINVAL.

Signed-off-by: Anton Blanchard <>
Cc: stable <> [3.0+]
Signed-off-by: David S. Miller <>
10 years agonet: sendmmsg should only return an error if no messages were sent
Anton Blanchard [Thu, 4 Aug 2011 14:07:38 +0000 (14:07 +0000)]
net: sendmmsg should only return an error if no messages were sent

sendmmsg uses a similar error return strategy as recvmmsg but it
turns out to be a confusing way to communicate errors.

The current code stores the error code away and returns it on the next
sendmmsg call. This means a call with completely valid arguments could
get an error from a previous call.

Change things so we only return an error if no datagrams could be sent.
If less than the requested number of messages were sent, the application
must retry starting at the first failed one and if the problem is
persistent the error will be returned.

This matches the behaviour of other syscalls like read/write - it
is not an error if less than the requested number of elements are sent.

Signed-off-by: Anton Blanchard <>
Cc: stable <> [3.0+]
Signed-off-by: David S. Miller <>
10 years agoDo 'shm_init_ns()' in an early pure_initcall
Linus Torvalds [Fri, 5 Aug 2011 05:35:59 +0000 (19:35 -1000)]
Do 'shm_init_ns()' in an early pure_initcall

This isn't really critical any more, since other patches (commit
298507d4d2cf: "shm: optimize exit_shm()") have caused us to not actually
need to touch the rw_mutex unless there are actual shm segments
associated with the namespace, but we really should do tne shm_init_ns()
earlier than we do now.

This, together with commit 288d5abec831 ("Boot up with usermodehelper
disabled") will mean that we really do initialize the initial ipc
namespace data structure before we run any tasks.

Tested-by: Marc Zyngier <>
Signed-off-by: Linus Torvalds <>
10 years agoMerge branch 'for-linus' of git://
Linus Torvalds [Fri, 5 Aug 2011 02:44:40 +0000 (16:44 -1000)]
Merge branch 'for-linus' of git://git./linux/kernel/git/viro/vfs-2.6

* 'for-linus' of git://
  RCUify freeing acls, let check_acl() go ahead in RCU mode if acl is cached
  get rid of boilerplate switches in posix_acl.h
  fix block device fallout from ->fsync() changes

10 years agoMerge branch 'fixefi' of git://
Linus Torvalds [Fri, 5 Aug 2011 02:44:23 +0000 (16:44 -1000)]
Merge branch 'fixefi' of git://git./linux/kernel/git/aegl/linux-2.6

* 'fixefi' of git://
  efi: Fix argument types for SetVariable() for ia64

10 years agoMerge branch 'core-urgent-for-linus' of git://
Linus Torvalds [Fri, 5 Aug 2011 02:44:04 +0000 (16:44 -1000)]
Merge branch 'core-urgent-for-linus' of git://git./linux/kernel/git/tip/linux-2.6-tip

* 'core-urgent-for-linus' of git://
  slab, lockdep: Annotate the locks before using them
  lockdep: Clear whole lockdep_map on initialization
  slab, lockdep: Annotate slab -> rcu -> debug_object -> slab
  lockdep: Fix up warning
  lockdep: Fix trace_hardirqs_on_caller()
  futex: Fix regression with read only mappings

10 years agoMerge branch 'next' of git://
Linus Torvalds [Fri, 5 Aug 2011 02:43:43 +0000 (16:43 -1000)]
Merge branch 'next' of git://git./linux/kernel/git/djbw/async_tx

* 'next' of git://
  dmaengine: use DEFINE_IDR for static initialization
  ioat: fix xor_idx_to_desc
  Avoid section type conflict in dma/ioat/dma_v3.c
  ioat: Adding PCI IDs for IOAT devices on SandyBridge platforms

10 years agocpuidle: Consistent spelling of cpuidle_idle_call()
David Brown [Thu, 4 Aug 2011 16:24:31 +0000 (09:24 -0700)]
cpuidle: Consistent spelling of cpuidle_idle_call()

Commit a0bfa1373859e9d11dc92561a8667588803e42d8 mispells
cpuidle_idle_call() on ARM and SH code.  Fix this to be consistent.

Cc: Kevin Hilman <>
Cc: Paul Mundt <>
Cc: Len Brown <>
Signed-off-by: David Brown <>
[ Also done by Mark Brown - th ebug has been around forever, and was
  noticed in -next, but the idle tree never picked it up. Bad bad bad ]
Signed-off-by: Linus Torvalds <>
10 years agoefi: Fix argument types for SetVariable() for ia64
Matthew Garrett [Wed, 6 Jul 2011 20:48:49 +0000 (16:48 -0400)]
efi: Fix argument types for SetVariable() for ia64

The spec says this takes uint32 for attributes, not uintn.

Signed-off-by: Matthew Garrett <>
Signed-off-by: Tony Luck <>
10 years agoMerge branch 'devicetree/merge' of git://
Linus Torvalds [Thu, 4 Aug 2011 16:37:07 +0000 (06:37 -1000)]
Merge branch 'devicetree/merge' of git://

* 'devicetree/merge' of git://
  Revert "dt: add of_alias_scan and of_alias_get_id"
  dt: remove of_alias_get_id() reference

10 years agoMerge branch 'fixes' of git://
Linus Torvalds [Thu, 4 Aug 2011 16:36:20 +0000 (06:36 -1000)]
Merge branch 'fixes' of git://git./linux/kernel/git/jejb/parisc-2.6

* 'fixes' of git://
  [PARISC] wire up sendmmsg syscall
  [PARISC] fix return type of __atomic64_add_return
  [PARISC] Fix futex support

10 years agoMerge branch 'for-linus' of git://
Linus Torvalds [Thu, 4 Aug 2011 16:35:51 +0000 (06:35 -1000)]
Merge branch 'for-linus' of git://

* 'for-linus' of git://
  [S390] signal: use set_restore_sigmask() helper
  [S390] smp: remove pointless comments in startup_secondary()
  [S390] qdio: Use kstrtoul_from_user
  [S390] sclp_async: Use kstrtoul_from_user
  [S390] exec: remove redundant set_fs(USER_DS)
  [S390] cpu hotplug: on cpu start wait until being marked active
  [S390] signal: convert to use set_current_blocked()
  [S390] asm offsets: fix coding style
  [S390] Add support for IBM zEnterprise 114
  [S390] dasd: check if raw track access is supported
  [S390] Use diagnose 308 for system reset
  [S390] Export store_status() function
  [S390] dasd: use vmalloc for statistics input buffer
  [S390] Add PSW restart shutdown trigger
  [S390] missing return in page_table_alloc_pgste
  [S390] qdio: 2nd stage retry on SIGA-W busy conditions

10 years agoeisa/pci_eisa.c: fix BUG introduced by 005bdad7b80
Arnaud Lacombe [Thu, 4 Aug 2011 14:39:44 +0000 (10:39 -0400)]
eisa/pci_eisa.c: fix BUG introduced by 005bdad7b80

While `pci_eisa_driver' still refer `pci_eisa_init', the .probe() function
should not be called after init memory release, as pointed out by commit
74b9a297. The structure is still referenced in the drivers subsystem, and can
be accesseed through sysfs, so the modpost warning is a false positive. Mark
it as such.

In the same time, the warning referenced in 005bdad7b80 did only mention
`pci_eisa_driver', not `pci_eisa_pci_tbl', so remove its marking.

Broken-by: Arnaud Lacombe <> (in 005bdad7b80)
Reported-by: Tetsuo Handa <>
Signed-off-by: Arnaud Lacombe <>
Signed-off-by: Linus Torvalds <>
10 years agoixgbe: fix PHY link setup for 82599
Emil Tantilov [Thu, 28 Jul 2011 06:17:04 +0000 (06:17 +0000)]
ixgbe: fix PHY link setup for 82599

Fix pointer to setup_link for 82599.

This resolves some link issues when advertising modes unsupported
by the link partner.

Signed-off-by: Emil Tantilov <>
Tested-by: Phil Schmitt <>
Signed-off-by: Jeff Kirsher <>
10 years agoixgbe: fix __ixgbe_notify_dca() bail out code
Don Skidmore [Wed, 20 Jul 2011 02:27:05 +0000 (02:27 +0000)]
ixgbe: fix __ixgbe_notify_dca() bail out code

The way __ixgbe_notify_dca() was currently set up it would not be
possible to add a requester. Both cases of the IXGBE_FLAG_DCA_ENABLED
bit being on and off would lead to the function exiting for a

Signed-off-by: Don Skidmore <>
Tested-by: Phil Schmitt <>
Signed-off-by: Jeff Kirsher <>
10 years agoigb: fix WOL on second port of i350 device
Carolyn Wyborny [Thu, 7 Jul 2011 00:24:56 +0000 (00:24 +0000)]
igb: fix WOL on second port of i350 device

This patch fixes a problem where WOL would fail on second port of i350

Reported-by: Martin Wilck <>
Reported-by: Stefan Assmann<>
Signed-off-by: Carolyn Wyborny <>
Tested-by: Aaron Brown <>
Signed-off-by: Jeff Kirsher <>
10 years agoe1000e: minor re-order of #include files
Bruce Allan [Fri, 29 Jul 2011 05:52:51 +0000 (05:52 +0000)]
e1000e: minor re-order of #include files

The recent commit a6b7a407 when back-ported to the out-of-tree e1000e
driver caused a compilation error on older kernels which required a
re-ordering of the #include files.  This cosmetic patch syncs the two
drivers for easier maintainability.

Signed-off-by: Bruce Allan <>
Tested-by: Aaron Brown <>
Signed-off-by: Jeff Kirsher <>
10 years agoe1000e: remove unnecessary check for NULL pointer
Bruce Allan [Fri, 22 Jul 2011 06:21:41 +0000 (06:21 +0000)]
e1000e: remove unnecessary check for NULL pointer

The array shadow_ram is never NULL.

Signed-off-by: Bruce Allan <>
Tested-by: Aaron Brown <>
Signed-off-by: Jeff Kirsher <>
10 years agointel drivers: repair missing flush operations
Jesse Brandeburg [Wed, 20 Jul 2011 00:56:21 +0000 (00:56 +0000)]
intel drivers: repair missing flush operations

after review of all intel drivers, found several instances where
drivers had the incorrect pattern of:
memory mapped write();

which should always be:
memory mapped write();
write flush(); /* aka memory mapped read */

The reason for including the flush is that writes can be held
(posted) in PCI/PCIe bridges, but the read always has to complete
synchronously and therefore has to flush all pending writes to a
device.  If a write is held and followed by a delay, the delay
means nothing because the write may not have reached hardware
(maybe even not until the next read)

Signed-off-by: Jesse Brandeburg <>
Tested-by: Aaron Brown <>
Signed-off-by: Jeff Kirsher <>
10 years agoRevert "dt: add of_alias_scan and of_alias_get_id"
Grant Likely [Thu, 4 Aug 2011 09:27:32 +0000 (10:27 +0100)]
Revert "dt: add of_alias_scan and of_alias_get_id"

This reverts commit 750f463a749e28464151ad26938d11b07b1c43cb.

of_alias_* still needs work to be generalized for 'promtree' dt
platforms, and to no implicitly create entries for available ids.

Signed-off-by: Grant Likely <>
10 years agodt: remove of_alias_get_id() reference
Grant Likely [Thu, 4 Aug 2011 09:47:40 +0000 (10:47 +0100)]
dt: remove of_alias_get_id() reference

of_alias_get_id() is broken and being reverted.  Remove the reference
to it and replace with a single incrementing id number.

There is no risk of regression here on the imx driver since the imx
change to use of_alias_get_id() is commit 22698aa2, "serial/imx: add
device tree probe support" which is new for v3.1, and it won't get
used unless CONFIG_OF is enabled and the board is booted using a
device tree.  A single incrementing integer is sufficient for now.

Signed-off-by: Grant Likely <>
Acked-by: Shawn Guo <>
10 years agomacb: restore wrap bit when performing underrun cleanup
Tord Andersson [Wed, 3 Aug 2011 22:11:47 +0000 (22:11 +0000)]
macb: restore wrap bit when performing underrun cleanup

When TX underrun occurs, a cleanup is performed that marks all buffers
as used. As a side effect it also clears the wrap bit in the last
buffer. This patch will restore the wrap bit.

Signed-off-by: Tord Andersson <>
Acked-by: Nicolas Ferre <>
Signed-off-by: David S. Miller <>
10 years agocdc_ncm: fix endianness problem.
Giuseppe Scrivano [Wed, 3 Aug 2011 22:10:29 +0000 (22:10 +0000)]
cdc_ncm: fix endianness problem.

Fix a misusage of the struct usb_cdc_notification to pass arguments to the
usb_control_msg function.  The usb_control_msg function expects host endian
arguments but usb_cdc_notification stores these values as little endian.

Now usb_control_msg is directly invoked with host endian values.

Signed-off-by: Giuseppe Scrivano <>
Signed-off-by: David S. Miller <>
10 years agodrivers/ide/cy82c693.c: Add missing pci_dev_put
Julia Lawall [Thu, 4 Aug 2011 08:30:34 +0000 (01:30 -0700)]
drivers/ide/cy82c693.c: Add missing pci_dev_put

Pci_get_slot calls pci_dev_get, so pci_dev_put is needed before leaving the
function in the case where pci_get_slot is locally used.

The semantic match that finds this problem is as follows:

// <smpl>
local idexpression x;
expression e;

*x = pci_get_slot(...)
... when != true x == NULL
    when != pci_dev_put(x)
    when != e = x
    when != if (x != NULL) {<+... pci_dev_put(x); ...+>}
*return ...;
// </smpl>

Signed-off-by: Julia Lawall <>
Signed-off-by: David S. Miller <>
10 years agoide: Fix irq flags madness
Thomas Gleixner [Thu, 4 Aug 2011 08:29:51 +0000 (01:29 -0700)]
ide: Fix irq flags madness

commit ec1a123 (IDE: pass IRQ flags to the IDE core) introduced the
bogosity of passing unfiltered resource->flags to the irq_flags which
are used for request_irq. It results in random bits set (especially

Filter the bits proper.

Signed-off-by: Thomas Gleixner <>
Signed-off-by: David S. Miller <>
10 years agoslab, lockdep: Annotate the locks before using them
Peter Zijlstra [Thu, 28 Jul 2011 21:22:56 +0000 (23:22 +0200)]
slab, lockdep: Annotate the locks before using them

Fernando found we hit the regular OFF_SLAB 'recursion' before we
annotate the locks, cure this.

The relevant portion of the stack-trace:

> [    0.000000]  [<c085e24f>] rt_spin_lock+0x50/0x56
> [    0.000000]  [<c04fb406>] __cache_free+0x43/0xc3
> [    0.000000]  [<c04fb23f>] kmem_cache_free+0x6c/0xdc
> [    0.000000]  [<c04fb2fe>] slab_destroy+0x4f/0x53
> [    0.000000]  [<c04fb396>] free_block+0x94/0xc1
> [    0.000000]  [<c04fc551>] do_tune_cpucache+0x10b/0x2bb
> [    0.000000]  [<c04fc8dc>] enable_cpucache+0x7b/0xa7
> [    0.000000]  [<c0bd9d3c>] kmem_cache_init_late+0x1f/0x61
> [    0.000000]  [<c0bba687>] start_kernel+0x24c/0x363
> [    0.000000]  [<c0bba0ba>] i386_start_kernel+0xa9/0xaf

Reported-by: Fernando Lopez-Lezcano <nando@ccrma.Stanford.EDU>
Acked-by: Pekka Enberg <>
Signed-off-by: Peter Zijlstra <>
Signed-off-by: Ingo Molnar <>
10 years agolockdep: Clear whole lockdep_map on initialization
Tejun Heo [Thu, 14 Jul 2011 13:19:09 +0000 (15:19 +0200)]
lockdep: Clear whole lockdep_map on initialization

lockdep_init_map() only initializes parts of lockdep_map and triggers
kmemcheck warning when it is copied as a whole.  There isn't anything
to be gained by clearing selectively.  memset() the whole structure
and remove loop for ->class_cache[] clearing.


Signed-off-by: Tejun Heo <>
Reported-and-tested-by: Christian Casteyde <>
Signed-off-by: Peter Zijlstra <>
Signed-off-by: Ingo Molnar <>
10 years agoslab, lockdep: Annotate slab -> rcu -> debug_object -> slab
Peter Zijlstra [Fri, 22 Jul 2011 13:26:05 +0000 (15:26 +0200)]
slab, lockdep: Annotate slab -> rcu -> debug_object -> slab

Lockdep thinks there's lock recursion through:

    spin_lock(&l3->list_lock)  <----------------.
    free_block()                                |
      slab_destroy()                            |
call_rcu()                              |
  debug_object_activate()               |
    debug_object_init()                 |
      __debug_object_init()             |
kmem_cache_alloc()              |
  cache_alloc_refill()          |
    spin_lock(&l3->list_lock) --'

Now debug objects doesn't use SLAB_DESTROY_BY_RCU and hence there is no
actual possibility of recursing. Luckily debug objects marks it slab
with SLAB_DEBUG_OBJECTS so we can identify the thing.

Mark all SLAB_DEBUG_OBJECTS (all one!) slab caches with a special
lockdep key so that lockdep sees its a different cachep.

Also add a WARN on trying to create a SLAB_DESTROY_BY_RCU |
SLAB_DEBUG_OBJECTS cache, to avoid possible future trouble.

Reported-and-tested-by: Sebastian Siewior <>
[ fixes to the initial patch ]
Reported-by: Thomas Gleixner <>
Acked-by: Pekka Enberg <>
Signed-off-by: Peter Zijlstra <>
Signed-off-by: Ingo Molnar <>
10 years agolockdep: Fix up warning
Peter Zijlstra [Mon, 25 Jul 2011 10:09:59 +0000 (12:09 +0200)]
lockdep: Fix up warning

On Sun, 2011-07-24 at 21:06 -0400, Arnaud Lacombe wrote:

> /src/linux/linux/kernel/lockdep.c: In function 'mark_held_locks':
> /src/linux/linux/kernel/lockdep.c:2471:31: warning: comparison of
> distinct pointer types lacks a cast

The warning is harmless in this case, but the below makes it go away.

Reported-by: Arnaud Lacombe <>
Signed-off-by: Peter Zijlstra <>
Signed-off-by: Ingo Molnar <>
10 years agolockdep: Fix trace_hardirqs_on_caller()
Peter Zijlstra [Tue, 26 Jul 2011 11:13:44 +0000 (13:13 +0200)]
lockdep: Fix trace_hardirqs_on_caller()

Commit dd4e5d3ac4a ("lockdep: Fix trace_[soft,hard]irqs_[on,off]()
recursion") made a bit of a mess of the various checks and error

In particular it moved the check for !irqs_disabled() before the
spurious enable test, resulting in some warnings.

Reported-by: Arnaud Lacombe <>
Reported-by: Dave Jones <>
Reported-and-tested-by: Sergey Senozhatsky <>
Signed-off-by: Peter Zijlstra <>
Signed-off-by: Ingo Molnar <>
10 years agoBoot up with usermodehelper disabled
Linus Torvalds [Thu, 4 Aug 2011 08:03:29 +0000 (22:03 -1000)]
Boot up with usermodehelper disabled

The core device layer sends tons of uevent notifications for each device
it finds, and if the kernel has been built with a non-empty
CONFIG_UEVENT_HELPER_PATH that will make us try to execute the usermode
helper binary for all these events very early in the boot.

Not only won't the root filesystem even be mounted at that point, we
literally won't have necessarily even initialized all the process
handling data structures at that point, which causes no end of silly
problems even when the usermode helper doesn't actually succeed in

So just use our existing infrastructure to disable the usermodehelpers
to make the kernel start out with them disabled.  We enable them when
we've at least initialized stuff a bit.

Problems related to an uninitialized


reported by various people.

Reported-by: Manuel Lauss <>
Reported-by: Richard Weinberger <>
Reported-by: Marc Zyngier <>
Acked-by: Kay Sievers <>
Cc: Andrew Morton <>
Cc: Vasiliy Kulikov <>
Cc: Greg KH <>
Signed-off-by: Linus Torvalds <>
10 years agox86: don't include xen/xen.h in <asm/io.h> unless XEN is enabled
Linus Torvalds [Thu, 4 Aug 2011 08:00:38 +0000 (22:00 -1000)]
x86: don't include xen/xen.h in <asm/io.h> unless XEN is enabled

Dmitry Kasatkin reports:
  "kernel-devel package with kernel headers have no <include/xen>
   directory if XEN is disabled.  Modules which inclide asm/io.h won't

   XEN related content is behind the CONFIG_XEN flag in the io.h.  And
   <xen/xen.h> should be also behind CONFIG_XEN flag."

So move the include of <xen/xen.h> down into the section that is
conditional on CONFIG_XEN.

Reported-by: Dmitry Kasatkin <>
Signed-off-by: Linus Torvalds <>
10 years agoMerge branch 'for-linus' of git://
Linus Torvalds [Thu, 4 Aug 2011 08:00:09 +0000 (22:00 -1000)]
Merge branch 'for-linus' of git://git./linux/kernel/git/dtor/input

* 'for-linus' of git://
  Input: ad7879 - fix deficient device disable
  Input: gpio_keys - fix two typos in devicetree documentation
  Input: mma8450 - add device tree probe support
  Input: gpio_keys - return proper error code if memory allocation fails
  Input: lm8323 - add missing device_remove_file for dev_attr_time
  Input: tegra-kbc - fix computation of polling time
  Input: kxtj9 - explicitly include module.h
  Input: psmouse - hgpk.c needs module.h

10 years agoMerge branch 'idle-release' of git://
Linus Torvalds [Thu, 4 Aug 2011 07:54:15 +0000 (21:54 -1000)]
Merge branch 'idle-release' of git://git./linux/kernel/git/lenb/linux-idle-2.6

* 'idle-release' of git://
  cpuidle: stop depending on pm_idle
  x86 idle: move mwait_idle_with_hints() to where it is used
  cpuidle: replace xen access to x86 pm_idle and default_idle
  cpuidle: create bootparam ""
  mrst_pmu: driver for Intel Moorestown Power Management Unit

10 years agoMerge branch 'apei-release' of git://
Linus Torvalds [Thu, 4 Aug 2011 07:53:27 +0000 (21:53 -1000)]
Merge branch 'apei-release' of git://git./linux/kernel/git/lenb/linux-acpi-2.6

* 'apei-release' of git://
  ACPI, APEI, EINJ Param support is disabled by default
  APEI GHES: 32-bit buildfix
  ACPI: APEI build fix
  ACPI, APEI, GHES: Add hardware memory error recovery support
  HWPoison: add memory_failure_queue()
  ACPI, APEI, GHES, Error records content based throttle
  ACPI, APEI, GHES, printk support for recoverable error via NMI
  lib, Make gen_pool memory allocator lockless
  lib, Add lock-less NULL terminated single list
  ACPI, APEI, Add WHEA _OSC support
  ACPI, APEI, Add APEI bit support in generic _OSC call
  ACPI, APEI, GHES, Support disable GHES at boot time
  ACPI, APEI, GHES, Prevent GHES to be built as module
  ACPI, APEI, Use apei_exec_run_optional in APEI EINJ and ERST
  ACPI, APEI, Add apei_exec_run_optional
  ACPI, APEI, GHES, Do not ratelimit fatal error printk before panic
  ACPI, APEI, ERST, Fix erst-dbg long record reading issue
  ACPI, APEI, ERST, Prevent erst_dbg from loading if ERST is disabled

10 years agoMerge branch 'linus' into core/urgent
Ingo Molnar [Thu, 4 Aug 2011 07:09:27 +0000 (09:09 +0200)]
Merge branch 'linus' into core/urgent

10 years agodmaengine: use DEFINE_IDR for static initialization
Axel Lin [Wed, 20 Jul 2011 03:32:28 +0000 (11:32 +0800)]
dmaengine: use DEFINE_IDR for static initialization

We could use DEFINE_IDR for statically allocated idr
that allow us to save a few lines of code.

And also remove unneeded mutex_init() for dma_list_mutex, as
dma_list_mutex is initialized automatically by DEFINE_MUTEX().

Signed-off-by: Axel Lin <>
Signed-off-by: Dan Williams <>
10 years agoioat: fix xor_idx_to_desc
Dan Williams [Fri, 22 Jul 2011 21:20:46 +0000 (14:20 -0700)]
ioat: fix xor_idx_to_desc

For versions of the device that implement operation-types 0x87, 0x88
(IOAT_OP_XOR, IOAT_OP_XOR_VAL) this map determines whether a given
source is located in the base or extended descriptor.  Source addresses
6 through 8 require an extended descriptor, hence 0xe0, not 0xd0.  No
shipping hardware currently implements these operation types.

Reported-by: Evgueni Smogailov <>
Signed-off-by: Dan Williams <>
10 years agoirda: use PCI_VENDOR_ID_*
Jon Mason [Wed, 3 Aug 2011 06:42:42 +0000 (06:42 +0000)]
irda: use PCI_VENDOR_ID_*

Use PCI_VENDOR_ID_* from pci_ids.h instead of creating #define locally.

Signed-off-by: Jon Mason <>
Signed-off-by: David S. Miller <>
10 years agoMerge branch 'for-next' of git://
Linus Torvalds [Thu, 4 Aug 2011 01:12:09 +0000 (15:12 -1000)]
Merge branch 'for-next' of git://git./linux/kernel/git/nab/target-pending

* 'for-next' of git://
  tcm_fc: Handle DDP/SW fc_frame_payload_get failures in ft_recv_write_data
  target: Fix bug for transport_generic_wait_for_tasks with direct operation
  target: iscsi_target depends on NET
  target: Fix WRITE_SAME_16 lba assignment breakage
  MAINTAINERS: Add target-devel list for drivers/target/
  iscsi-target: Fix CONFIG_SMP=n and CONFIG_MODULES=n build failure
  iscsi-target: Fix snprintf usage with MAX_PORTAL_LEN
  iscsi-target: Fix uninitialized usage of cmd->pad_bytes
  iscsi-target: strlen() doesn't count the terminator
  iscsi-target: Fix NULL dereference on allocation failure

10 years agoMerge branch 'devicetree/next' of git://
Linus Torvalds [Thu, 4 Aug 2011 01:10:30 +0000 (15:10 -1000)]
Merge branch 'devicetree/next' of git://

* 'devicetree/next' of git://
  dt: add of_alias_scan and of_alias_get_id

10 years agoMerge branch 'for_linus' of git://
Linus Torvalds [Thu, 4 Aug 2011 01:09:10 +0000 (15:09 -1000)]
Merge branch 'for_linus' of git://git./linux/kernel/git/tytso/ext4

* 'for_linus' of git://
  ext4: use kzalloc in ext4_kzalloc()

10 years agoshm: optimize exit_shm()
Vasiliy Kulikov [Wed, 3 Aug 2011 18:28:26 +0000 (22:28 +0400)]
shm: optimize exit_shm()

We may optimistically check .in_use == 0 without holding the rw_mutex:
it's the common case, and if it's zero, there certainly won't be any
segments associated with us.

After taking the lock, the idr_for_each() will do the right thing, so we
could now drop the re-check inside the lock without any real cost.  But
it won't hurt.

Signed-off-by: Vasiliy Kulikov <>
Signed-off-by: Linus Torvalds <>
10 years agoshm: fix wrong tests
Vasiliy Kulikov [Wed, 3 Aug 2011 18:26:55 +0000 (22:26 +0400)]
shm: fix wrong tests

Commit 4c677e2eefdb ("shm: optimize locking and ipc_namespace getting")
introduced a copy-paste bug.  Due to the bug cycle optimizations were

Signed-off-by: Vasiliy Kulikov <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: expand "help" to explain value of TMPFS_POSIX_ACL
Robert P. J. Day [Wed, 3 Aug 2011 23:21:29 +0000 (16:21 -0700)]
tmpfs: expand "help" to explain value of TMPFS_POSIX_ACL

Expand the fs/Kconfig "help" info to clarify why it's a bad idea to
deselect the TMPFS_POSIX_ACL config variable.

Signed-off-by: Robert P. J. Day <>
Acked-by: Randy Dunlap <>
Acked-by: Hugh Dickins <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agomm: clarify the radix_tree exceptional cases
Hugh Dickins [Wed, 3 Aug 2011 23:21:28 +0000 (16:21 -0700)]
mm: clarify the radix_tree exceptional cases

Make the radix_tree exceptional cases, mostly in filemap.c, clearer.

It's hard to devise a suitable snappy name that illuminates the use by
shmem/tmpfs for swap, while keeping filemap/pagecache/radix_tree
generality.  And akpm points out that /* radix_tree_deref_retry(page) */
comments look like calls that have been commented out for unknown

Skirt the naming difficulty by rearranging these blocks to handle the
transient radix_tree_deref_retry(page) case first; then just explain the
remaining shmem/tmpfs swap case in a comment.

Signed-off-by: Hugh Dickins <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs radix_tree: locate_item to speed up swapoff
Hugh Dickins [Wed, 3 Aug 2011 23:21:27 +0000 (16:21 -0700)]
tmpfs radix_tree: locate_item to speed up swapoff

We have already acknowledged that swapoff of a tmpfs file is slower than
it was before conversion to the generic radix_tree: a little slower
there will be acceptable, if the hotter paths are faster.

But it was a shock to find swapoff of a 500MB file 20 times slower on my
laptop, taking 10 minutes; and at that rate it significantly slows down
my testing.

Now, most of that turned out to be overhead from PROVE_LOCKING and
PROVE_RCU: without those it was only 4 times slower than before; and
more realistic tests on other machines don't fare as badly.

I've tried a number of things to improve it, including tagging the swap
entries, then doing lookup by tag: I'd expected that to halve the time,
but in practice it's erratic, and often counter-productive.

The only change I've so far found to make a consistent improvement, is
to short-circuit the way we go back and forth, gang lookup packing
entries into the array supplied, then shmem scanning that array for the
target entry.  Scanning in place doubles the speed, so it's now only
twice as slow as before (or three times slower when the PROVEs are on).

So, add radix_tree_locate_item() as an expedient, once-off,
single-caller hack to do the lookup directly in place.  #ifdef it on
CONFIG_SHMEM and CONFIG_SWAP, as much to document its limited
applicability as save space in other configurations.  And, sadly,
#include sched.h for cond_resched().

Signed-off-by: Hugh Dickins <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agomm: a few small updates for radix-swap
Hugh Dickins [Wed, 3 Aug 2011 23:21:27 +0000 (16:21 -0700)]
mm: a few small updates for radix-swap

Remove PageSwapBacked (!page_is_file_cache) cases from
add_to_page_cache_locked() and add_to_page_cache_lru(): those pages now
go through shmem_add_to_page_cache().

Remove a comment on maximum tmpfs size from fsstack_copy_inode_size(),
and add a comment on swap entries to invalidate_mapping_pages().

And mincore_page() uses find_get_page() on what might be shmem or a
tmpfs file: allow for a radix_tree_exceptional_entry(), and proceed to
find_get_page() on swapper_space if so (oh, swapper_space needs #ifdef).

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: use kmemdup for short symlinks
Hugh Dickins [Wed, 3 Aug 2011 23:21:26 +0000 (16:21 -0700)]
tmpfs: use kmemdup for short symlinks

But we've not yet removed the old swp_entry_t i_direct[16] from
shmem_inode_info.  That's because it was still being shared with the
inline symlink.  Remove it now (saving 64 or 128 bytes from shmem inode
size), and use kmemdup() for short symlinks, say, those up to 128 bytes.

I wonder why mpol_free_shared_policy() is done in shmem_destroy_inode()
rather than shmem_evict_inode(), where we usually do such freeing? I
guess it doesn't matter, and I'm not into NUMA mpol testing right now.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Reviewed-by: Pekka Enberg <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: convert shmem_writepage and enable swap
Hugh Dickins [Wed, 3 Aug 2011 23:21:25 +0000 (16:21 -0700)]
tmpfs: convert shmem_writepage and enable swap

Convert shmem_writepage() to use shmem_delete_from_page_cache() to use
shmem_radix_tree_replace() to substitute swap entry for page pointer
atomically in the radix tree.

As with shmem_add_to_page_cache(), it's not entirely satisfactory to be
copying such code from delete_from_swap_cache, but again judged easier
to sell than making its other callers go through the extras.

Remove the toy implementation's shmem_put_swap() and shmem_get_swap(),
now unreferenced, and the hack to disable swap: it's now good to go.

The way things have worked out, info->lock no longer helps to guard the
shmem_swaplist: we increment swapped under shmem_swaplist_mutex only.
That global mutex exclusion between shmem_writepage() and shmem_unuse()
is not pretty, and we ought to find another way; but it's been forced on
us by recent race discoveries, not a consequence of this patchset.

And what has become of the WARN_ON_ONCE(1) free_swap_and_cache() if a
swap entry was found already present? That's no longer possible, the
(unknown) one inserting this page into filecache would hit the swap
entry occupying that slot.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: convert mem_cgroup shmem to radix-swap
Hugh Dickins [Wed, 3 Aug 2011 23:21:24 +0000 (16:21 -0700)]
tmpfs: convert mem_cgroup shmem to radix-swap

Remove mem_cgroup_shmem_charge_fallback(): it was only required when we
had to move swappage to filecache with GFP_NOWAIT.

Remove the GFP_NOWAIT special case from mem_cgroup_cache_charge(), by
moving its call out from shmem_add_to_page_cache() to two of thats three
callers.  But leave it doing mem_cgroup_uncharge_cache_page() on error:
although asymmetrical, it's easier for all 3 callers to handle.

These two changes would also be appropriate if anyone were to start
using shmem_read_mapping_page_gfp() with GFP_NOWAIT.

Remove mem_cgroup_get_shmem_target(): mc_handle_file_pte() can test
radix_tree_exceptional_entry() to get what it needs for itself.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: convert shmem_getpage_gfp to radix-swap
Hugh Dickins [Wed, 3 Aug 2011 23:21:24 +0000 (16:21 -0700)]
tmpfs: convert shmem_getpage_gfp to radix-swap

Convert shmem_getpage_gfp(), the engine-room of shmem, to expect page or
swap entry returned from radix tree by find_lock_page().

Whereas the repetitive old method proceeded mainly under info->lock,
dropping and repeating whenever one of the conditions needed was not
met, now we can proceed without it, leaving shmem_add_to_page_cache() to
check for a race.

This way there is no need to preallocate a page, no need for an early
radix_tree_preload(), no need for mem_cgroup_shmem_charge_fallback().

Move the error unwinding down to the bottom instead of repeating it
throughout.  ENOSPC handling is a little different from before: there is
no longer any race between find_lock_page() and finding swap, but we can
arrive at ENOSPC before calling shmem_recalc_inode(), which might
occasionally discover freed space.

Be stricter to check i_size before returning.  info->lock is used for
little but alloced, swapped, i_blocks updates.  Move i_blocks updates
out from under the max_blocks check, so even an unlimited size=0 mount
can show accurate du.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: convert shmem_unuse_inode to radix-swap
Hugh Dickins [Wed, 3 Aug 2011 23:21:23 +0000 (16:21 -0700)]
tmpfs: convert shmem_unuse_inode to radix-swap

Convert shmem_unuse_inode() to use a lockless gang lookup of the radix
tree, searching for matching swap.

This is somewhat slower than the old method: because of repeated radix
tree descents, because of copying entries up, but probably most because
the old method noted and skipped once a vector page was cleared of swap.
Perhaps we can devise a use of radix tree tagging to achieve that later.

shmem_add_to_page_cache() uses shmem_radix_tree_replace() to compensate
for the lockless lookup by checking that the expected entry is in place,
under lock.  It is not very satisfactory to be copying this much from
add_to_page_cache_locked(), but I think easier to sell than insisting
that every caller of add_to_page_cache*() go through the extras.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: convert shmem_truncate_range to radix-swap
Hugh Dickins [Wed, 3 Aug 2011 23:21:22 +0000 (16:21 -0700)]
tmpfs: convert shmem_truncate_range to radix-swap

Disable the toy swapping implementation in shmem_writepage() - it's hard
to support two schemes at once - and convert shmem_truncate_range() to a
lockless gang lookup of swap entries along with pages, freeing both.

Since the second loop tightens its noose until all entries of either
kind have been squeezed out (and we shall make sure that there's not an
instant when neither is visible), there is no longer a need for yet
another pass below.

shmem_radix_tree_replace() compensates for the lockless lookup by
checking that the expected entry is in place, under lock, before
replacing it.  Here it just deletes, but will be used in later patches
to substitute swap entry for page or page for swap entry.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: copy truncate_inode_pages_range
Hugh Dickins [Wed, 3 Aug 2011 23:21:21 +0000 (16:21 -0700)]
tmpfs: copy truncate_inode_pages_range

Bring truncate.c's code for truncate_inode_pages_range() inline into
shmem_truncate_range(), replacing its first call (there's a followup
call below, but leave that one, it will disappear next).

Don't play with it yet, apart from leaving out the cleancache flush, and
(importantly) the nrpages == 0 skip, and moving shmem_setattr()'s
partial page preparation into its partial page handling.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: miscellaneous trivial cleanups
Hugh Dickins [Wed, 3 Aug 2011 23:21:21 +0000 (16:21 -0700)]
tmpfs: miscellaneous trivial cleanups

While it's at its least, make a number of boring nitpicky cleanups to
shmem.c, mostly for consistency of variable naming.  Things like "swap"
instead of "entry", "pgoff_t index" instead of "unsigned long idx".

And since everything else here is prefixed "shmem_", better change
init_tmpfs() to shmem_init().

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotmpfs: demolish old swap vector support
Hugh Dickins [Wed, 3 Aug 2011 23:21:20 +0000 (16:21 -0700)]
tmpfs: demolish old swap vector support

The maximum size of a shmem/tmpfs file has been limited by the maximum
size of its triple-indirect swap vector.  With 4kB page size, maximum
filesize was just over 2TB on a 32-bit kernel, but sadly one eighth of
that on a 64-bit kernel.  (With 8kB page size, maximum filesize was just
over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
MAX_LFS_FILESIZE being then more restrictive than swap vector layout.)

It's a shame that tmpfs should be more restrictive than ramfs, and this
limitation has now been noticed.  Add another level to the swap vector?
No, it became obscure and hard to maintain, once I complicated it to
make use of highmem pages nine years ago: better choose another way.

Surely, if 2.4 had had the radix tree pagecache introduced in 2.5, then
tmpfs would never have invented its own peculiar radix tree: we would
have fitted swap entries into the common radix tree instead, in much the
same way as we fit swap entries into page tables.

And why should each file have a separate radix tree for its pages and
for its swap entries? The swap entries are required precisely where and
when the pages are not.  We want to put them together in a single radix
tree: which can then avoid much of the locking which was needed to
prevent them from being exchanged underneath us.

This also avoids the waste of memory devoted to swap vectors, first in
the shmem_inode itself, then at least two more pages once a file grew
beyond 16 data pages (pages accounted by df and du, but not by memcg).
Allocated upfront, to avoid allocation when under swapping pressure, but
pure waste when CONFIG_SWAP is not set - I have never spattered around
the ifdefs to prevent that, preferring this move to sharing the common
radix tree instead.

There are three downsides to sharing the radix tree.  One, that it binds
tmpfs more tightly to the rest of mm, either requiring knowledge of swap
entries in radix tree there, or duplication of its code here in shmem.c.
I believe that the simplications and memory savings (and probable higher
performance, not yet measured) justify that.

Two, that on HIGHMEM systems with SWAP enabled, it's the lowmem radix
nodes that cannot be freed under memory pressure - whereas before it was
the less precious highmem swap vector pages that could not be freed.
I'm hoping that 64-bit has now been accessible for long enough, that the
highmem argument has grown much less persuasive.

Three, that swapoff is slower than it used to be on tmpfs files, since
it's using a simple generic mechanism not tailored to it: I find this
noticeable, and shall want to improve, but maybe nobody else will

So...  now remove most of the old swap vector code from shmem.c.  But,
for the moment, keep the simple i_direct vector of 16 pages, with simple
accessors shmem_put_swap() and shmem_get_swap(), as a toy implementation
to help mark where swap needs to be handled in subsequent patches.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agomm: let swap use exceptional entries
Hugh Dickins [Wed, 3 Aug 2011 23:21:19 +0000 (16:21 -0700)]
mm: let swap use exceptional entries

If swap entries are to be stored along with struct page pointers in a
radix tree, they need to be distinguished as exceptional entries.

Most of the handling of swap entries in radix tree will be contained in
shmem.c, but a few functions in filemap.c's common code need to check
for their appearance: find_get_page(), find_lock_page(),
find_get_pages() and find_get_pages_contig().

So as not to slow their fast paths, tuck those checks inside the
existing checks for unlikely radix_tree_deref_slot(); except for
find_lock_page(), where it is an added test.  And make it a BUG in
find_get_pages_tag(), which is not applied to tmpfs files.

A part of the reason for eliminating shmem_readpage() earlier, was to
minimize the places where common code would need to allow for swap

The swp_entry_t known to swapfile.c must be massaged into a slightly
different form when stored in the radix tree, just as it gets massaged
into a pte_t when stored in page tables.

In an i386 kernel this limits its information (type and page offset) to
30 bits: given 32 "types" of swapfile and 4kB pagesize, that's a maximum
swapfile size of 128GB.  Which is less than the 512GB we previously
allowed with X86_PAE (where the swap entry can occupy the entire upper
32 bits of a pte_t), but not a new limitation on 32-bit without PAE; and
there's not a new limitation on 64-bit (where swap filesize is already
limited to 16TB by a 32-bit page offset).  Thirty areas of 128GB is
probably still enough swap for a 64GB 32-bit machine.

Provide swp_to_radix_entry() and radix_to_swp_entry() conversions, and
enforce filesize limit in read_swap_header(), just as for ptes.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agoradix_tree: exceptional entries and indices
Hugh Dickins [Wed, 3 Aug 2011 23:21:18 +0000 (16:21 -0700)]
radix_tree: exceptional entries and indices

A patchset to extend tmpfs to MAX_LFS_FILESIZE by abandoning its
peculiar swap vector, instead keeping a file's swap entries in the same
radix tree as its struct page pointers: thus saving memory, and
simplifying its code and locking.

This patch:

The radix_tree is used by several subsystems for different purposes.  A
major use is to store the struct page pointers of a file's pagecache for
memory management.  But what if mm wanted to store something other than
page pointers there too?

The low bit of a radix_tree entry is already used to denote an indirect
pointer, for internal use, and the unlikely radix_tree_deref_retry()

Define the next bit as denoting an exceptional entry, and supply inline
functions radix_tree_exception() to return non-0 in either unlikely
case, and radix_tree_exceptional_entry() to return non-0 in the second

If a subsystem already uses radix_tree with that bit set, no problem: it
does not affect internal workings at all, but is defined for the
convenience of those storing well-aligned pointers in the radix_tree.

The radix_tree_gang_lookups have an implicit assumption that the caller
can deduce the offset of each entry returned e.g.  by the page->index of
a struct page.  But that may not be feasible for some kinds of item to
be stored there.

radix_tree_gang_lookup_slot() allow for an optional indices argument,
output array in which to return those offsets.  The same could be added
to other radix_tree_gang_lookups, but for now keep it to the only one
for which we need it.

Signed-off-by: Hugh Dickins <>
Acked-by: Rik van Riel <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agodrivers/video/backlight/aat2870_bl.c: make it buildable as a module
Andrew Morton [Wed, 3 Aug 2011 23:21:17 +0000 (16:21 -0700)]
drivers/video/backlight/aat2870_bl.c: make it buildable as a module

i386 allmodconfig:

  drivers/built-in.o: In function `aat2870_bl_remove':
  aat2870_bl.c:(.text+0x414f9): undefined reference to `backlight_device_unregister'
  drivers/built-in.o: In function `aat2870_bl_probe':
  aat2870_bl.c:(.text+0x418fc): undefined reference to `backlight_device_register'
  aat2870_bl.c:(.text+0x41a31): undefined reference to `backlight_device_unregiste

Cc: Jin Park <>
Cc: Samuel Ortiz <>
Cc: Axel Lin <>
Cc: Richard Purdie <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agodrivers/video/backlight/aat2870_bl.c: fix setting max_current
Axel Lin [Wed, 3 Aug 2011 23:21:17 +0000 (16:21 -0700)]
drivers/video/backlight/aat2870_bl.c: fix setting max_current

 - Current implementation tests wrong value for setting

 - In the current implementation, we cannot differentiate between 2 cases:

   a) if pdata->max_current is not set , or

   b) pdata->max_current is set to AAT2870_CURRENT_0_45 (which is also 0).

   Fix it by setting AAT2870_CURRENT_0_45 to be 1 and adjust the equation in
   aat2870_brightness() accordingly.

Signed-off-by: Axel Lin <>
Cc: Richard Purdie <>
Cc: Samuel Ortiz <>
Tested-by: Jin Park <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agodrivers/video/backlight/aat2870_bl.c: fix error checking for backlight_device_register
Axel Lin [Wed, 3 Aug 2011 23:21:16 +0000 (16:21 -0700)]
drivers/video/backlight/aat2870_bl.c: fix error checking for backlight_device_register

backlight_device_register() returns ERR_PTR() on error.

Signed-off-by: Axel Lin <>
Cc: Richard Purdie <>
Cc: Jin Park <>
Cc: Samuel Ortiz <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agocris: add missing declaration of kgdb_init() and breakpoint()
WANG Cong [Wed, 3 Aug 2011 23:21:15 +0000 (16:21 -0700)]
cris: add missing declaration of kgdb_init() and breakpoint()


  arch/cris/arch-v10/kernel/irq.c:239: error: implicit declaration of function 'kgdb_init'
  arch/cris/arch-v10/kernel/irq.c:240: error: implicit declaration of function 'breakpoint'

Declare these two functions.

Reported-by: Geert Uytterhoeven <>
Signed-off-by: WANG Cong <>
Cc: Mikael Starvik <>
Cc: Jesper Nilsson <>
Cc: <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agocris: fix the prototype of sync_serial_ioctl()
WANG Cong [Wed, 3 Aug 2011 23:21:14 +0000 (16:21 -0700)]
cris: fix the prototype of sync_serial_ioctl()


  arch/cris/arch-v10/drivers/sync_serial.c:961: error: conflicting types for 'sync_serial_ioctl'

Reported-by: Geert Uytterhoeven <>
Signed-off-by: WANG Cong <>
Cc: Mikael Starvik <>
Cc: Jesper Nilsson <>
Cc: <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agocris: fix a build error in sync_serial_open()
WANG Cong [Wed, 3 Aug 2011 23:21:14 +0000 (16:21 -0700)]
cris: fix a build error in sync_serial_open()


  arch/cris/arch-v10/drivers/sync_serial.c:628: error: 'ret' undeclared (first use in this function)

'ret' should be 'err'.

Reported-by: Geert Uytterhoeven <>
Signed-off-by: WANG Cong <>
Cc: Mikael Starvik <>
Cc: Jesper Nilsson <>
Cc: <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agocris: fix a build error in kernel/fork.c
WANG Cong [Wed, 3 Aug 2011 23:21:12 +0000 (16:21 -0700)]
cris: fix a build error in kernel/fork.c

Fix this error:

  kernel/fork.c:267: error: implicit declaration of function 'alloc_thread_info_node'

This is due to renaming alloc_thread_info() to alloc_thread_info_node().

[ coding-style fixes]
Reported-by: Geert Uytterhoeven <>
Signed-off-by: WANG Cong <>
Cc: Mikael Starvik <>
Cc: Jesper Nilsson <>
Cc: <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotpm_tis: fix build when ACPI is not enabled
Randy Dunlap [Wed, 3 Aug 2011 23:21:11 +0000 (16:21 -0700)]
tpm_tis: fix build when ACPI is not enabled

Fix tpm_tis.c build when CONFIG_ACPI is not enabled by providing a stub
function.  Fixes many build errors/warnings:

  drivers/char/tpm/tpm_tis.c:89: error: dereferencing pointer to incomplete type
  drivers/char/tpm/tpm_tis.c:89: warning: type defaults to 'int' in declaration of 'type name'
  drivers/char/tpm/tpm_tis.c:89: error: request for member 'list' in something not a structure or union

Signed-off-by: Randy Dunlap <>
Cc: Leendert van Doorn <>
Cc: James Morris <>
Cc: Ingo Molnar <>
Cc: Stephen Rothwell <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agomm: page_alloc: increase __GFP_BITS_SHIFT to include __GFP_OTHER_NODE
Johannes Weiner [Wed, 3 Aug 2011 23:21:09 +0000 (16:21 -0700)]
mm: page_alloc: increase __GFP_BITS_SHIFT to include __GFP_OTHER_NODE

__GFP_OTHER_NODE is used for NUMA allocations on behalf of other nodes.
It's supposed to be passed through from the page allocator to
zone_statistics(), but it never gets there as gfp_allowed_mask is not
wide enough and masks out the flag early in the allocation path.

The result is an accounting glitch where successful NUMA allocations
by-agent are not properly attributed as local.

Increase __GFP_BITS_SHIFT so that it includes __GFP_OTHER_NODE.

Signed-off-by: Johannes Weiner <>
Acked-by: Andi Kleen <>
Reviewed-by: Minchan Kim <>
Acked-by: Mel Gorman <>
Reviewed-by: Michal Hocko <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agoramoops: update module parameters
Sergiu Iordache [Wed, 3 Aug 2011 23:21:09 +0000 (16:21 -0700)]
ramoops: update module parameters

Update the module parameters when platform data is used.  This means
that they can be read from /sys/module/ramoops/parameters in order to
parse the memory area.

Signed-off-by: Sergiu Iordache <>
Cc: Marco Stornelli <>
Cc: Seiji Aguchi <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agoDocumentation: add pointer to name_to_dev_t for root= values
Will Drewry [Wed, 3 Aug 2011 23:21:08 +0000 (16:21 -0700)]
Documentation: add pointer to name_to_dev_t for root= values

Update kernel-parameters.txt to point users to the authoritative comment
for name_to_dev_t.  In addition, updates other places where some
name_to_dev_t behavior was discussed.  All other references to root=
appear to be for explicit sample usage or just side comments when
discussing other kernel parameters.

Signed-off-by: Will Drewry <>
Cc: Kay Sievers <>
Cc: Randy Dunlap <>
Cc: Namhyung Kim <>
Cc: Trond Myklebust <>
Cc: Jens Axboe <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agofs/dcache.c: fix new kernel-doc warning
Randy Dunlap [Wed, 3 Aug 2011 23:21:07 +0000 (16:21 -0700)]
fs/dcache.c: fix new kernel-doc warning

Fix new kernel-doc warning in fs/dcache.c:

  Warning(fs/dcache.c:797): No description found for parameter 'sb'

Signed-off-by: Randy Dunlap <>
Cc: Al Viro <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agoida: simplified functions for id allocation
Rusty Russell [Wed, 3 Aug 2011 23:21:06 +0000 (16:21 -0700)]
ida: simplified functions for id allocation

The current hyper-optimized functions are overkill if you simply want to
allocate an id for a device.  Create versions which use an internal

In followup patches, numerous drivers are converted to use this

Thanks to Tejun for feedback.

Signed-off-by: Rusty Russell <>
Acked-by: Tejun Heo <>
Acked-by: Jonathan Cameron <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotaskstats: add_del_listener() should ignore !valid listeners
Oleg Nesterov [Wed, 3 Aug 2011 23:21:05 +0000 (16:21 -0700)]
taskstats: add_del_listener() should ignore !valid listeners

When send_cpu_listeners() finds the orphaned listener it marks it as
!valid and drops listeners->sem.  Before it takes this sem for writing,
s->pid can be reused and add_del_listener() can wrongly try to re-use
this entry.

Change add_del_listener() to check ->valid = T.

Signed-off-by: Oleg Nesterov <>
Reviewed-by: Vasiliy Kulikov <>
Acked-by: Balbir Singh <>
Cc: Jerome Marchand <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agotaskstats: add_del_listener() shouldn't use the wrong node
Oleg Nesterov [Wed, 3 Aug 2011 23:21:04 +0000 (16:21 -0700)]
taskstats: add_del_listener() shouldn't use the wrong node

1. Commit 26c4caea9d69 "don't allow duplicate entries in listener mode"
   changed add_del_listener(REGISTER) so that "next_cpu:" can reuse the
   listener allocated for the previous cpu, this doesn't look exactly
   right even if minor.

   Change the code to kfree() in the already-registered case, this case
   is unlikely anyway so the extra kmalloc_node() shouldn't hurt but
   looke more correct and clean.

2. use the plain list_for_each_entry() instead of _safe() to scan

3. Remove the unneeded INIT_LIST_HEAD(&s->list), we are going to

Signed-off-by: Oleg Nesterov <>
Reviewed-by: Vasiliy Kulikov <>
Cc: Balbir Singh <>
Reviewed-by: Jerome Marchand <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agortc-omap: fix initialization of control register
Daniel Glöckner [Wed, 3 Aug 2011 23:21:02 +0000 (16:21 -0700)]
rtc-omap: fix initialization of control register

As the comment explains, the intention of the code is to clear the
OMAP_RTC_CTRL_MODE_12_24 bit, but instead it only clears the
OMAP_RTC_CTRL_TEST, and OMAP_RTC_CTRL_ROUND_30S are also better off
being cleared.

Signed-off-by: Daniel Glöckner <>
Cc: Alessandro Zummo <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agofault-injection: add ability to export fault_attr in arbitrary directory
Akinobu Mita [Wed, 3 Aug 2011 23:21:01 +0000 (16:21 -0700)]
fault-injection: add ability to export fault_attr in arbitrary directory

init_fault_attr_dentries() is used to export fault_attr via debugfs.
But it can only export it in debugfs root directory.

Per Forlin is working on mmc_fail_request which adds support to inject
data errors after a completed host transfer in MMC subsystem.

The fault_attr for mmc_fail_request should be defined per mmc host and
export it in debugfs directory per mmc host like

init_fault_attr_dentries() doesn't help for mmc_fail_request.  So this
introduces fault_create_debugfs_attr() which is able to create a
directory in the arbitrary directory and replace

[ extraneous semicolon, per Randy]
Signed-off-by: Akinobu Mita <>
Tested-by: Per Forlin <>
Cc: Jens Axboe <>
Cc: Christoph Lameter <>
Cc: Pekka Enberg <>
Cc: Matt Mackall <>
Cc: Randy Dunlap <>
Cc: Stephen Rothwell <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
10 years agomlx4: Fixing Ethernet unicast packet steering
Yevgeny Petrilin [Wed, 3 Aug 2011 23:38:59 +0000 (16:38 -0700)]
mlx4: Fixing Ethernet unicast packet steering

For older FW versions, fixing the usage of per port Mac table.
For each port we must define the base QP number, which is passed
to the HW.
Setting the correct value in SET_PORT FW command to enable the steering.

Reported-by: Roland Dreier <>
Tested-by: Roland Dreier <>
Signed-off-by: Yevgeny Petrilin <>
Signed-off-by: David S. Miller <>
10 years agoMerge branch 'for-davem' of git://
David S. Miller [Wed, 3 Aug 2011 23:36:41 +0000 (16:36 -0700)]
Merge branch 'for-davem' of git://git./linux/kernel/git/linville/wireless

10 years agocpuidle: stop depending on pm_idle
Len Brown [Fri, 1 Apr 2011 23:34:59 +0000 (19:34 -0400)]
cpuidle: stop depending on pm_idle

cpuidle users should call cpuidle_call_idle() directly
rather than via (pm_idle)() function pointer.

Architecture may choose to continue using (pm_idle)(),
but cpuidle need not depend on it:


cc: Kevin Hilman <>
cc: Paul Mundt <>
Acked-by: H. Peter Anvin <>
Signed-off-by: Len Brown <>
10 years agox86 idle: move mwait_idle_with_hints() to where it is used
Len Brown [Thu, 31 Mar 2011 03:52:29 +0000 (23:52 -0400)]
x86 idle: move mwait_idle_with_hints() to where it is used

...and make it static

no functional change

Acked-by: H. Peter Anvin <>
Signed-off-by: Len Brown <>
10 years agocpuidle: replace xen access to x86 pm_idle and default_idle
Len Brown [Fri, 1 Apr 2011 22:28:35 +0000 (18:28 -0400)]
cpuidle: replace xen access to x86 pm_idle and default_idle

When a Xen Dom0 kernel boots on a hypervisor, it gets access
to the raw-hardware ACPI tables.  While it parses the idle tables
for the hypervisor's beneift, it uses HLT for its own idle.

Rather than have xen scribble on pm_idle and access default_idle,
have it simply disable_cpuidle() so acpi_idle will not load and
architecture default HLT will be used.

Tested-by: Konrad Rzeszutek Wilk <>
Acked-by: H. Peter Anvin <>
Signed-off-by: Len Brown <>
10 years agocpuidle: create bootparam ""
Len Brown [Fri, 1 Apr 2011 22:13:10 +0000 (18:13 -0400)]
cpuidle: create bootparam ""

useful for disabling cpuidle to fall back
to architecture-default idle loop

cpuidle drivers and governors will fail to register.
on x86 they'll say so:

intel_idle: intel_idle yielding to (null)
ACPI: acpi_idle yielding to (null)

Signed-off-by: Len Brown <>
10 years agomrst_pmu: driver for Intel Moorestown Power Management Unit
Len Brown [Wed, 13 Jul 2011 02:29:32 +0000 (22:29 -0400)]
mrst_pmu: driver for Intel Moorestown Power Management Unit

The Moorestown (MRST) Power Management Unit (PMU) driver
directs the SOC power states in the "Langwell" south complex (SCU).

It hooks pci_platform_pm_ops[] and thus observes all PCI ".set_state"
requests.  For devices in the SC, the pmu driver translates those
PCI requests into the appropriate commands for the SCU.

The PMU driver helps implement S0i3, a deep system idle power idle state.
Entry into S0i3 is via cpuidle, just like regular processor c-states.
S0i3 depends on pre-conditions including uni-processor, graphics off,
and certain IO devices in the SC must be off.  If those pre-conditions
are met, then the PMU allows cpuidle to enter S0i3, otherwise such requests
are demoted, either to Atom C4 or Atom C6.

This driver is based on prototype work by Bruce Flemming,
Illyas Mansoor, Rajeev D. Muralidhar, Vishwesh M. Rudramuni,
Hari Seshadri and Sujith Thomas.  The current driver also
includes contributions from H. Peter Anvin, Arjan van de Ven,
Kristen Accardi, and Yong Wang.

Thanks for additional review feedback from Alan Cox and Randy Dunlap.

Acked-by: Alan Cox <>
Acked-by: H. Peter Anvin <>
Signed-off-by: Len Brown <>
10 years agoext4: use kzalloc in ext4_kzalloc()
Mathias Krause [Wed, 3 Aug 2011 18:57:11 +0000 (14:57 -0400)]
ext4: use kzalloc in ext4_kzalloc()

Commit 9933fc0i (ext4: introduce ext4_kvmalloc(), ext4_kzalloc(), and
ext4_kvfree()) intruduced wrappers around k*alloc/vmalloc but introduced
a typo for ext4_kzalloc() by not using kzalloc() but kmalloc().

Signed-off-by: Mathias Krause <>
Signed-off-by: "Theodore Ts'o" <>
10 years agoMerge branch 'apei' into apei-release
Len Brown [Wed, 3 Aug 2011 15:30:42 +0000 (11:30 -0400)]
Merge branch 'apei' into apei-release

Some trivial conflicts due to other various merges
adding to the end of common lists sooner than this one.


Signed-off-by: Len Brown <>
10 years agoACPI, APEI, EINJ Param support is disabled by default
Huang Ying [Wed, 20 Jul 2011 08:09:29 +0000 (16:09 +0800)]
ACPI, APEI, EINJ Param support is disabled by default

EINJ parameter support is only usable for some specific BIOS.
Originally, it is expected to have no harm for BIOS does not support
it.  But now, we found it will cause issue (memory overwriting) for
some BIOS.  So param support is disabled by default and only enabled
when newly added module parameter named "param_extension" is
explicitly specified.

Signed-off-by: Huang Ying <>
Cc: Matthew Garrett <>
Acked-by: Don Zickus <>
Acked-by: Tony Luck <>
Signed-off-by: Len Brown <>
10 years agoAPEI GHES: 32-bit buildfix
Len Brown [Tue, 2 Aug 2011 22:00:21 +0000 (18:00 -0400)]
APEI GHES: 32-bit buildfix

drivers/acpi/apei/ghes.c:542: warning: integer overflow in expression
drivers/acpi/apei/ghes.c:619: warning: integer overflow in expression

ghes.c:(.text+0x46289): undefined reference to `__udivdi3'
  in function ghes_estatus_cache_add().

Reported-by: Randy Dunlap <>
Signed-off-by: Len Brown <>
10 years agoACPI: APEI build fix
Len Brown [Sat, 16 Jul 2011 22:14:21 +0000 (18:14 -0400)]
ACPI: APEI build fix

as GHES is optional...

When # CONFIG_ACPI_APEI_GHES is not set:

(.init.text+0x4c22): undefined reference to `ghes_disable'

Reported-by: Randy Dunlap <>
Acked-by: Randy Dunlap <>
Signed-off-by: Len Brown <>
10 years agoACPI, APEI, GHES: Add hardware memory error recovery support
Huang Ying [Wed, 13 Jul 2011 05:14:28 +0000 (13:14 +0800)]
ACPI, APEI, GHES: Add hardware memory error recovery support

memory_failure_queue() is called when recoverable memory errors are
notified by firmware to do the recovery work.

Signed-off-by: Huang Ying <>
Signed-off-by: Len Brown <>
10 years agoHWPoison: add memory_failure_queue()
Huang Ying [Wed, 13 Jul 2011 05:14:27 +0000 (13:14 +0800)]
HWPoison: add memory_failure_queue()

memory_failure() is the entry point for HWPoison memory error
recovery.  It must be called in process context.  But commonly
hardware memory errors are notified via MCE or NMI, so some delayed
execution mechanism must be used.  In MCE handler, a work queue + ring
buffer mechanism is used.

In addition to MCE, now APEI (ACPI Platform Error Interface) GHES
(Generic Hardware Error Source) can be used to report memory errors
too.  To add support to APEI GHES memory recovery, a mechanism similar
to that of MCE is implemented.  memory_failure_queue() is the new
entry point that can be called in IRQ context.  The next step is to
make MCE handler uses this interface too.

Signed-off-by: Huang Ying <>
Cc: Andi Kleen <>
Cc: Wu Fengguang <>
Cc: Andrew Morton <>
Signed-off-by: Len Brown <>
10 years agoACPI, APEI, GHES, Error records content based throttle
Huang Ying [Wed, 13 Jul 2011 05:14:26 +0000 (13:14 +0800)]
ACPI, APEI, GHES, Error records content based throttle

printk is used by GHES to report hardware errors.  Ratelimit is
enforced on the printk to avoid too many hardware error reports in
kernel log.  Because there may be thousands or even millions of
corrected hardware errors during system running.

Currently, a simple scheme is used.  That is, the total number of
hardware error reporting is ratelimited.  This may cause some issues
in practice.

For example, there are two kinds of hardware errors occurred in
system.  One is corrected memory error, because the fault memory
address is accessed frequently, there may be hundreds error report
per-second.  The other is corrected PCIe AER error, it will be
reported once per-second.  Because they share one ratelimit control
structure, it is highly possible that only memory error is reported.

To avoid the above issue, an error record content based throttle
algorithm is implemented in the patch.  Where after the first
successful reporting, all error records that are same are throttled for
some time, to let other kinds of error records have the opportunity to
be reported.

In above example, the memory errors will be throttled for some time,
after being printked.  Then the PCIe AER error will be printked

Signed-off-by: Huang Ying <>
Signed-off-by: Len Brown <>
10 years agoACPI, APEI, GHES, printk support for recoverable error via NMI
Huang Ying [Wed, 13 Jul 2011 05:14:25 +0000 (13:14 +0800)]
ACPI, APEI, GHES, printk support for recoverable error via NMI

Some APEI GHES recoverable errors are reported via NMI, but printk is
not safe in NMI context.

To solve the issue, a lock-less memory allocator is used to allocate
memory in NMI handler, save the error record into the allocated
memory, put the error record into a lock-less list.  On the other
hand, an irq_work is used to delay the operation from NMI context to
IRQ context.  The irq_work IRQ handler will remove nodes from
lock-less list, printk the error record and do some further processing
include recovery operation, then free the memory.

Signed-off-by: Huang Ying <>
Signed-off-by: Len Brown <>
10 years agolib, Make gen_pool memory allocator lockless
Huang Ying [Wed, 13 Jul 2011 05:14:24 +0000 (13:14 +0800)]
lib, Make gen_pool memory allocator lockless

This version of the gen_pool memory allocator supports lockless

This makes it safe to use in NMI handlers and other special
unblockable contexts that could otherwise deadlock on locks.  This is
implemented by using atomic operations and retries on any conflicts.
The disadvantage is that there may be livelocks in extreme cases.  For
better scalability, one gen_pool allocator can be used for each CPU.

The lockless operation only works if there is enough memory available.
If new memory is added to the pool a lock has to be still taken.  So
any user relying on locklessness has to ensure that sufficient memory
is preallocated.

The basic atomic operation of this allocator is cmpxchg on long.  On
architectures that don't have NMI-safe cmpxchg implementation, the
allocator can NOT be used in NMI handler.  So code uses the allocator
in NMI handler should depend on CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG.

Signed-off-by: Huang Ying <>
Reviewed-by: Andi Kleen <>
Reviewed-by: Mathieu Desnoyers <>
Cc: Andrew Morton <>
Signed-off-by: Len Brown <>
10 years agolib, Add lock-less NULL terminated single list
Huang Ying [Wed, 13 Jul 2011 05:14:23 +0000 (13:14 +0800)]
lib, Add lock-less NULL terminated single list

Cmpxchg is used to implement adding new entry to the list, deleting
all entries from the list, deleting first entry of the list and some
other operations.

Because this is a single list, so the tail can not be accessed in O(1).

If there are multiple producers and multiple consumers, llist_add can
be used in producers and llist_del_all can be used in consumers.  They
can work simultaneously without lock.  But llist_del_first can not be
used here.  Because llist_del_first depends on list->first->next does
not changed if list->first is not changed during its operation, but
llist_del_first, llist_add, llist_add (or llist_del_all, llist_add,
llist_add) sequence in another consumer may violate that.

If there are multiple producers and one consumer, llist_add can be
used in producers and llist_del_all or llist_del_first can be used in
the consumer.

This can be summarized as follow:

           |   add    | del_first |  del_all
 add       |    -     |     -     |     -
 del_first |          |     L     |     L
 del_all   |          |           |     -

Where "-" stands for no lock is needed, while "L" stands for lock is

The list entries deleted via llist_del_all can be traversed with
traversing function such as llist_for_each etc.  But the list entries
can not be traversed safely before deleted from the list.  The order
of deleted entries is from the newest to the oldest added one.  If you
want to traverse from the oldest to the newest, you must reverse the
order by yourself before traversing.

The basic atomic operation of this list is cmpxchg on long.  On
architectures that don't have NMI-safe cmpxchg implementation, the
list can NOT be used in NMI handler.  So code uses the list in NMI
handler should depend on CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG.

Signed-off-by: Huang Ying <>
Reviewed-by: Andi Kleen <>
Reviewed-by: Mathieu Desnoyers <>
Cc: Andrew Morton <>
Signed-off-by: Len Brown <>