OSDN Git Service

uclinux-h8/linux.git
8 years agoMerge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/net...
David S. Miller [Mon, 29 Jun 2015 00:20:02 +0000 (17:20 -0700)]
Merge branch 'master' of git://git./linux/kernel/git/jkirsher/net-queue

Jeff Kirsher says:

====================
Intel Wired LAN Driver Updates 2015-06-26

This series contains fixes for igb, e1000e and i40evf.

Todd disables IPv6 extension header processing due to a hardware errata
and bumps the driver version.

Yanir provides six fixes for e1000e.  First is a fix for a locking issue
where we were not always taking the pci_bus_sem semaphore all the time
when calling pci_disable_link_state_locked(), so fix the code to only call
pci_disable_link_state_locked() when the semaphore has been acquired,
otherwise call pci_disable_link_state().  A previous fix for i219 where
the hardware prevented ULP entry caused EEE in Sx not the be enabled, so
modify the code flow that allows both ULP and EEE in Sx.  Fix an issue
when running 10/100 full duplex on i219 where CRC errors were occurring
by increasing the IPG from 8 to 0xC as per the hardware developers.
Fix a data corruption issue found on some platforms by increasing the
minimum gap between the PHY FIFO read and write pointers.  Fix i219,
which does not require the K1 workaround for LPT devices.

Mitch provides a i40evf fix for a panic when changing MTU.  Down was
requesting queue disables, but then exited immediately without waiting
for the queues to actually be disabled.  This could allow any function
called after i40evf_down() to run immediately, including i40evf_up(),
and causes a memory leak.  Fixed the issue by removing the whole
reinit_locked function which allows for the driver to handle the state
changes by requesting reset from the periodic timer.  The second fix
resolves an issue where RSS was being configured as though it is using
the maximum number of queue.  This can cause the device to drop a lot
of receive traffic, as the packets get assigned to non-functional queues.
This is resolved by only configuring RSS with the number of active queues.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/phy: tune get_phy_c45_ids to support more c45 phy
Shengzhou Liu [Fri, 26 Jun 2015 09:58:52 +0000 (17:58 +0800)]
net/phy: tune get_phy_c45_ids to support more c45 phy

As some C45 10G PHYs(e.g. Cortina CS4315/CS4340 PHY) have
zero Devices In package, current driver can't get correct
devices_in_package value by non-zero Devices In package.
so let's probe more with zero Devices In package to support
more C45 PHYs.

Signed-off-by: Shengzhou Liu <Shengzhou.Liu@freescale.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: fix lockdep splat
Eric Dumazet [Fri, 26 Jun 2015 05:32:29 +0000 (07:32 +0200)]
bnx2x: fix lockdep splat

Michel reported following lockdep splat

[   44.718117] INFO: trying to register non-static key.
[   44.723081] the code is fine but needs lockdep annotation.
[   44.728559] turning off the locking correctness validator.
[   44.734036] CPU: 8 PID: 5483 Comm: ethtool Not tainted 4.1.0
[   44.770289] Call Trace:
[   44.772741]  [<ffffffff816eb1cd>] dump_stack+0x4c/0x65
[   44.777879]  [<ffffffff8111d921>] ? console_unlock+0x1f1/0x510
[   44.783708]  [<ffffffff811121f5>] __lock_acquire+0x1d05/0x1f10
[   44.789538]  [<ffffffff8111370a>] ? mark_held_locks+0x6a/0x90
[   44.795276]  [<ffffffff81113835>] ? trace_hardirqs_on_caller+0x105/0x1d0
[   44.801967]  [<ffffffff8111390d>] ? trace_hardirqs_on+0xd/0x10
[   44.807793]  [<ffffffff811330fa>] ? hrtimer_try_to_cancel+0x4a/0x250
[   44.814142]  [<ffffffff81112ba6>] lock_acquire+0xb6/0x290
[   44.819537]  [<ffffffff810d6675>] ? flush_work+0x5/0x280
[   44.824844]  [<ffffffff810d66ad>] flush_work+0x3d/0x280
[   44.830061]  [<ffffffff810d6675>] ? flush_work+0x5/0x280
[   44.835366]  [<ffffffff816f3c43>] ? schedule_hrtimeout_range+0x13/0x20
[   44.841889]  [<ffffffff8112ec9b>] ? usleep_range+0x4b/0x50
[   44.847365]  [<ffffffff8111370a>] ? mark_held_locks+0x6a/0x90
[   44.853102]  [<ffffffff810d8585>] ? __cancel_work_timer+0x105/0x1c0
[   44.859359]  [<ffffffff81113835>] ? trace_hardirqs_on_caller+0x105/0x1d0
[   44.866045]  [<ffffffff810d851f>] __cancel_work_timer+0x9f/0x1c0
[   44.872048]  [<ffffffffa0010982>] ? bnx2x_func_stop+0x42/0x90 [bnx2x]
[   44.878481]  [<ffffffff810d8670>] cancel_work_sync+0x10/0x20
[   44.884134]  [<ffffffffa00259e5>] bnx2x_chip_cleanup+0x245/0x730 [bnx2x]
[   44.890829]  [<ffffffff8110ce02>] ? up+0x32/0x50
[   44.895439]  [<ffffffff811306b5>] ? del_timer_sync+0x5/0xd0
[   44.901005]  [<ffffffffa005596d>] bnx2x_nic_unload+0x20d/0x8e0 [bnx2x]
[   44.907527]  [<ffffffff811f1aef>] ? might_fault+0x5f/0xb0
[   44.912921]  [<ffffffffa005851c>] bnx2x_reload_if_running+0x2c/0x50 [bnx2x]
[   44.919879]  [<ffffffffa005a3c5>] bnx2x_set_ringparam+0x2b5/0x460 [bnx2x]
[   44.926664]  [<ffffffff815d498b>] dev_ethtool+0x55b/0x1c40
[   44.932148]  [<ffffffff815dfdc7>] ? rtnl_lock+0x17/0x20
[   44.937364]  [<ffffffff815e7f8b>] dev_ioctl+0x17b/0x630
[   44.942582]  [<ffffffff815abf8d>] sock_do_ioctl+0x5d/0x70
[   44.947972]  [<ffffffff815ac013>] sock_ioctl+0x73/0x280
[   44.953192]  [<ffffffff8124c1c8>] do_vfs_ioctl+0x88/0x5b0
[   44.958587]  [<ffffffff8110d0b3>] ? up_read+0x23/0x40
[   44.963631]  [<ffffffff812584cc>] ? __fget_light+0x6c/0xa0
[   44.969105]  [<ffffffff8124c781>] SyS_ioctl+0x91/0xb0
[   44.974149]  [<ffffffff816f4dd7>] system_call_fastpath+0x12/0x6f

As bnx2x_init_ptp() is only called if bp->flags contains PTP_SUPPORTED,
we also need to guard bnx2x_stop_ptp() with same condition, otherwise
ptp_task workqueue is not initialized and kernel barfs on
cancel_work_sync()

Fixes: eeed018cbfa30 ("bnx2x: Add timestamping and PTP hardware clock support")
Reported-by: Michel Lespinasse <walken@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Michal Kalderon <Michal.Kalderon@qlogic.com>
Cc: Ariel Elior <Ariel.Elior@qlogic.com>
Cc: Yuval Mintz <Yuval.Mintz@qlogic.com>
Cc: David Decotigny <decot@google.com>
Acked-by: Sony Chacko <sony.chacko@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: fec: don't access RACC register when not available
Greg Ungerer [Sat, 20 Jun 2015 05:51:57 +0000 (15:51 +1000)]
net: fec: don't access RACC register when not available

Not all silicon implementations of the Freescale FEC hardware module
have the RACC (Receive Accelerator Function) register, so we should not
be trying to access it on those that don't. Currently none of the ColdFire
based parts with a FEC have it.

Support for RACC was introduced by commit 4c09eed9 ("net: fec: Enable imx6
enet checksum acceleration"). A fix was introduced in commit d1391930
("net: fec: Fix build for MCF5272") that disables its use on the ColdFire
M5272 part, but it doesn't fix the general case of other ColdFire parts.

To fix we create a quirk flag, FEC_QUIRK_HAS_RACC, and check it before
working with the RACC register.

Signed-off-by: Greg Ungerer <gerg@uclinux.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: phy: fix phy link up when limiting speed via device tree
Mugunthan V N [Thu, 25 Jun 2015 16:51:02 +0000 (22:21 +0530)]
net: phy: fix phy link up when limiting speed via device tree

When limiting phy link speed using "max-speed" to 100mbps or less on a
giga bit phy, phy never completes auto negotiation and phy state
machine is held in PHY_AN. Fixing this issue by comparing the giga
bit advertise though phydev->supported doesn't have it but phy has
BMSR_ESTATEN set. So that auto negotiation is restarted as old and
new advertise are different and link comes up fine.

Signed-off-by: Mugunthan V N <mugunthanvnm@ti.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agodsa: fix promiscuity leak on slave dev open error
Gilad Ben-Yossef [Thu, 25 Jun 2015 13:50:13 +0000 (16:50 +0300)]
dsa: fix promiscuity leak on slave dev open error

DSA master netdev promiscuity counter was not being properly
decremented on slave device open error path.

Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com>
CC: Gilad Ben-Yossef <giladb@ezchip.com>
CC: David S. Miller <davem@davemloft.net>
CC: Florian Fainelli <f.fainelli@gmail.com>
CC: Guenter Roeck <linux@roeck-us.net>
CC: Andrew Lunn <andrew@lunn.ch>
CC: Scott Feldman <sfeldma@gmail.com>
Acked-by: Andrew Lunn <andrew@lunn.ch>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'kill_sk_protinfo'
David S. Miller [Sun, 28 Jun 2015 23:55:44 +0000 (16:55 -0700)]
Merge branch 'kill_sk_protinfo'

David Miller says:

====================
Get rid of sock->sk_protinfo.

These two patches get rid of the last remaining user of sk_protinfo
(ax25) and then really gets rid of the struct member.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: Kill sock->sk_protinfo
David Miller [Thu, 25 Jun 2015 13:19:15 +0000 (06:19 -0700)]
net: Kill sock->sk_protinfo

No more users, so it can now be removed.

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoax25: Stop using sock->sk_protinfo.
David Miller [Thu, 25 Jun 2015 13:19:07 +0000 (06:19 -0700)]
ax25: Stop using sock->sk_protinfo.

Just make a ax25_sock structure that provides the ax25_cb pointer.

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agodrivers: net: xgene: Pre-initialize ret in xgene_enet_get_resources()
Geert Uytterhoeven [Thu, 25 Jun 2015 13:13:29 +0000 (15:13 +0200)]
drivers: net: xgene: Pre-initialize ret in xgene_enet_get_resources()

If CONFIG_ACPI=n:

    drivers/net/ethernet/apm/xgene/xgene_enet_main.c: In function ‘xgene_enet_get_resources’:
    drivers/net/ethernet/apm/xgene/xgene_enet_main.c:951: warning: ‘ret’ may be used uninitialized in this function

If the driver is bound to a legacy platform device, ret will contain
arbitrary data. If it is non-zero, it will be returned to the caller as
an error code.

Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoflow_dissector: Pre-initialize ip_proto in __skb_flow_dissect()
Geert Uytterhoeven [Thu, 25 Jun 2015 13:10:32 +0000 (15:10 +0200)]
flow_dissector: Pre-initialize ip_proto in __skb_flow_dissect()

net/core/flow_dissector.c: In function ‘__skb_flow_dissect’:
net/core/flow_dissector.c:132: warning: ‘ip_proto’ may be used uninitialized in this function

Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Acked-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoxen-netfront: Remove the meaningless code
Li, Liang Z [Fri, 26 Jun 2015 23:17:26 +0000 (07:17 +0800)]
xen-netfront: Remove the meaningless code

The function netif_set_real_num_tx_queues() will return -EINVAL if
the second parameter < 1, so call this function with the second
parameter set to 0 is meaningless.

Signed-off-by: Liang Li <liang.z.li@intel.com>
Reviewed-by: David Vrabel <david.vrabel@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoipv4: fix RCU lockdep warning from linkdown changes
Andy Gospodarek [Fri, 26 Jun 2015 23:37:11 +0000 (19:37 -0400)]
ipv4: fix RCU lockdep warning from linkdown changes

The following lockdep splat was seen due to the wrong context for
grabbing in_dev.

===============================
[ INFO: suspicious RCU usage. ]
4.1.0-next-20150626-dbg-00020-g54a6d91-dirty #244 Not tainted
-------------------------------
include/linux/inetdevice.h:205 suspicious rcu_dereference_check() usage!

other info that might help us debug this:

rcu_scheduler_active = 1, debug_locks = 0
2 locks held by ip/403:
 #0:  (rtnl_mutex){+.+.+.}, at: [<ffffffff81453305>] rtnl_lock+0x17/0x19
 #1:  ((inetaddr_chain).rwsem){.+.+.+}, at: [<ffffffff8105c327>] __blocking_notifier_call_chain+0x35/0x6a

stack backtrace:
CPU: 2 PID: 403 Comm: ip Not tainted 4.1.0-next-20150626-dbg-00020-g54a6d91-dirty #244
 0000000000000001 ffff8800b189b728 ffffffff8150a542 ffffffff8107a8b3
 ffff880037bbea40 ffff8800b189b758 ffffffff8107cb74 ffff8800379dbd00
 ffff8800bec85800 ffff8800bf9e13c0 00000000000000ff ffff8800b189b7d8
Call Trace:
 [<ffffffff8150a542>] dump_stack+0x4c/0x6e
 [<ffffffff8107a8b3>] ? up+0x39/0x3e
 [<ffffffff8107cb74>] lockdep_rcu_suspicious+0xf7/0x100
 [<ffffffff814b63c3>] fib_dump_info+0x227/0x3e2
 [<ffffffff814b6624>] rtmsg_fib+0xa6/0x116
 [<ffffffff814b978f>] fib_table_insert+0x316/0x355
 [<ffffffff814b362e>] fib_magic+0xb7/0xc7
 [<ffffffff814b4803>] fib_add_ifaddr+0xb1/0x13b
 [<ffffffff814b4d09>] fib_inetaddr_event+0x36/0x90
 [<ffffffff8105c086>] notifier_call_chain+0x4c/0x71
 [<ffffffff8105c340>] __blocking_notifier_call_chain+0x4e/0x6a
 [<ffffffff8105c370>] blocking_notifier_call_chain+0x14/0x16
 [<ffffffff814a7f50>] __inet_insert_ifa+0x1a5/0x1b3
 [<ffffffff814a894d>] inet_rtm_newaddr+0x350/0x35f
 [<ffffffff81457866>] rtnetlink_rcv_msg+0x17b/0x18a
 [<ffffffff8107e7c3>] ? trace_hardirqs_on+0xd/0xf
 [<ffffffff8146965f>] ? netlink_deliver_tap+0x1cb/0x1f7
 [<ffffffff814576eb>] ? rtnl_newlink+0x72a/0x72a
...

This patch resolves that splat.

Signed-off-by: Andy Gospodarek <gospo@cumulusnetworks.com>
Reported-by: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotipc: purge backlog queue counters when broadcast link is reset
Jon Paul Maloy [Sun, 28 Jun 2015 13:44:44 +0000 (09:44 -0400)]
tipc: purge backlog queue counters when broadcast link is reset

In commit 1f66d161ab3d8b518903fa6c3f9c1f48d6919e74
("tipc: introduce starvation free send algorithm")
we introduced a counter per priority level for buffers
in the link backlog queue. We also introduced a new
function tipc_link_purge_backlog(), to reset these
counters to zero when the link is reset.

Unfortunately, we missed to call this function when
the broadcast link is reset, with the result that the
values of these counters might be permanently skewed
when new nodes are attached. This may in the worst case
lead to permananent, but spurious, broadcast link
congestion, where no broadcast packets can be sent at
all.

We fix this bug with this commit.

Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoi40evf: don't configure unused RSS queues
Mitch Williams [Tue, 23 Jun 2015 00:26:38 +0000 (17:26 -0700)]
i40evf: don't configure unused RSS queues

The driver will only configure as many queues as there are available
CPUs, up the maximum number of queues. However, it always configures
RSS as though it is using the maximum number of queues. This can cause
the device to drop a lot of RX traffic, as the packets get assigned to
nonfunctional queues.

Fix this by only configuring RSS with the number of active queues.

Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40evf: fix panic during MTU change
Mitch Williams [Fri, 19 Jun 2015 15:56:30 +0000 (08:56 -0700)]
i40evf: fix panic during MTU change

Down was requesting queue disables, but then exited immediately
without waiting for the queues to actually disable.  This could
allow any function called after i40evf_down to run immediately,
including i40evf_up, and causes a memory leak.

Removing the whole reinit_locked function is the best way
to go about this, and allows for the driver to handle the
state changes by requesting reset from the periodic timer.

Also, add a couple WARN_ONs in slow path to help us recognize
if we re-introduce this issue or missed any cases.

Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoe1000e: i219 - k1 workaround for LPT is not required for SPT
Yanir Lubetkin [Tue, 9 Jun 2015 22:16:03 +0000 (01:16 +0300)]
e1000e: i219 - k1 workaround for LPT is not required for SPT

In SPT hardware does not require this driver workaround.
Removed the conditional that caused K1 workaround execution on SPT.

Signed-off-by: Yanir Lubetkin <yanirx.lubetkin@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoe1000e: i219 - Increase minimum FIFO read/write min gap
Yanir Lubetkin [Tue, 9 Jun 2015 22:16:01 +0000 (01:16 +0300)]
e1000e: i219 - Increase minimum FIFO read/write min gap

Due to clocking changes in the Skylake platform, there was i219
data corruption. To work around this, HW team reported the need
to increase the minimum gap between the PHY FIFO read and write pointers.

Signed-off-by: Yanir Lubetkin <yanirx.lubetkin@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoe1000e: i219 - increase IPG for speed 10/100 full duplex
Yanir Lubetkin [Tue, 9 Jun 2015 22:15:57 +0000 (01:15 +0300)]
e1000e: i219 - increase IPG for speed 10/100 full duplex

In SPT/i219, there were CRC errors in speed 10/100 full duplex.
The solution given by the HW team is to increase the IPG from 8 to 0xC

Signed-off-by: Yanir Lubetkin <yanirx.lubetkin@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoe1000e: i219 - fix to enable both ULP and EEE in Sx state
Yanir Lubetkin [Tue, 9 Jun 2015 22:15:55 +0000 (01:15 +0300)]
e1000e: i219 - fix to enable both ULP and EEE in Sx state

In i219, there is a hardware bug that prevented ULP entry.
A side effect of the original software fix for this was that EEE in
Sx couldn't be enabled.
This patch implements a modified flow that allows both ULP and EEE in Sx.

Signed-off-by: Yanir Lubetkin <yanirx.lubetkin@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoe1000e: synchronization of MAC-PHY interface only on non- ME systems
Yanir Lubetkin [Tue, 9 Jun 2015 22:15:51 +0000 (01:15 +0300)]
e1000e: synchronization of MAC-PHY interface only on non- ME systems

On power up, the MAC - PHY interface needs to be set to PCIe, even if
cable is disconnected.  In ME systems, the ME handles this on exit from
Sx state. In non-ME, the driver handles it. Added a check for non-ME
system to the driver code that handles that.

Signed-off-by: Yanir Lubetkin <yanirx.lubetkin@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoe1000e: fix locking issue with e1000e_disable_aspm
Yanir Lubetkin [Tue, 9 Jun 2015 22:15:05 +0000 (01:15 +0300)]
e1000e: fix locking issue with e1000e_disable_aspm

e1000e_disable_aspm called pci_disable_link_state_locked which requires
pci_bus_sem to be held, but is also called from places where this semaphore
was not previously acquired. This patch implements two flavors of
disable_aspm, one that acquires the lock, and the other (_locked) which
should be called when the semaphore is already acquired.

Signed-off-by: Yanir Lubetkin <yanirx.lubetkin@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoigb: bump version of igb to 5.2.18
Todd Fujinaka [Fri, 17 Apr 2015 18:25:04 +0000 (11:25 -0700)]
igb: bump version of igb to 5.2.18

Bump version of igb to igb-5.2.18

Signed-off-by: Todd Fujinaka <todd.fujinaka@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoigb: disable IPv6 extension header processing
Todd Fujinaka [Fri, 17 Apr 2015 18:24:38 +0000 (11:24 -0700)]
igb: disable IPv6 extension header processing

Disable IPv6 extension header processing as per hardware errata.

Also fix copyright date.

Signed-off-by: Todd Fujinaka <todd.fujinaka@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoMerge branch 'bnx2x'
David S. Miller [Thu, 25 Jun 2015 13:30:43 +0000 (06:30 -0700)]
Merge branch 'bnx2x'

Yuval Mintz says:

====================
bnx2x: various fixes

This patch series contains several small fixes [with the possible
exception of the first 2 link fixes] for various driver flows.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Fix linearization for encapsulated packets
Yuval Mintz [Thu, 25 Jun 2015 12:19:29 +0000 (15:19 +0300)]
bnx2x: Fix linearization for encapsulated packets

Due to FW constraints, driver must make sure that transmitted SKBs will
not be too fragmented, or in the case that they are - that each 'window'
of fragments passed to the FW would contain at least an mss worth of data.

For encapsultaed packets the calculation is wrong, since it ignores the
inner headers in the calculation of the headers' length.
This could lead to a FW assertion in case of a too-fragmented encapsulated
packet.

Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Release nvram lock on error flow
Yuval Mintz [Thu, 25 Jun 2015 12:19:28 +0000 (15:19 +0300)]
bnx2x: Release nvram lock on error flow

During an error flow when trying to access the nvram the driver doesn't
release the hw lock it acquired.

Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Fix statistics gathering on link change
Ariel Elior [Thu, 25 Jun 2015 12:19:27 +0000 (15:19 +0300)]
bnx2x: Fix statistics gathering on link change

Since driver statistics flow access MACs and those might reset during
link re-configurations, when we're about to change link properties we
have to make sure that statistics are not operational.
Statisics would be re-enabled [i.e., gathering of statistics would
re-commence] once physical link is achieved again.

Since driver employs a link-flap avoidance scheme, there are scenarios
where driver will receive no indication that the new link is up, and
as a result the statistics would not be re-enabled.

Preventing LFA from working in such cases would guarantee that we'll
always receive such indications and thus will fix statistics gathering.

Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Fix self-test for 20g devices
Yuval Mintz [Thu, 25 Jun 2015 12:19:26 +0000 (15:19 +0300)]
bnx2x: Fix self-test for 20g devices

20g-capable devices are not configured properly for self-test, using
10g as their speed which cause the link indication to remain down and
fail the internal loopback test.

Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Fix VF MAC removal
Shahed Shaikh [Thu, 25 Jun 2015 12:19:25 +0000 (15:19 +0300)]
bnx2x: Fix VF MAC removal

There's a bug in today's driver where VF requests to add/remove MAC filters
always reach the Hypervisor as add requests.
This prevents the VF from changing its MAC address, as it cannot remove the
previously configured MAC and runs out of MAC credits.

Signed-off-by: Shahed Shaikh <Shahed.Shaikh@qlogic.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Don't notify about scratchpad parities
Manish Chopra [Thu, 25 Jun 2015 12:19:24 +0000 (15:19 +0300)]
bnx2x: Don't notify about scratchpad parities

The scratchpad is a shared block between all functions of a given device.
Due to HW limitations, we can't properly close its parity notifications
to all functions on legal flows.
E.g., it's possible that while taking a register dump from one function
a parity error would be triggered on other functions.

Today driver doesn't consider this parity as a 'real' parity unless its
being accompanied by additional indications [which would happen in a real
parity scenario]; But it does print notifications for such events in the
system logs.

This eliminates such prints - in case of real parities driver would have
additional indications; But if this is the only signal user will not even
see a parity being logged in the system.

Signed-off-by: Manish Chopra <Manish.Chopra@qlogic.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Prevent false warning when accessing MACs
Yuval Mintz [Thu, 25 Jun 2015 12:19:23 +0000 (15:19 +0300)]
bnx2x: Prevent false warning when accessing MACs

Each time a flow finishes reads from the classification shadow
configuration in the driver, that flow would check for pending commands
and pass them to FW if possible.
In case there's already a completion pending command, I.e., a ramrod
that has been sent to the FW and is yet to be completed while said flow
tries to configure the pending command we would get a false error message
in logs [and panic if SOE was used for driver compilation] since the
command could not have been completed.

This prevents said print [and panic]; The pending command will be sent by
the time the completion of the current sent command would arrive.

Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Correct speed from baseT into KR.
Yuval Mintz [Thu, 25 Jun 2015 12:19:22 +0000 (15:19 +0300)]
bnx2x: Correct speed from baseT into KR.

ethtool shows KR supported/advertised speeds incorrectly as baseT
in cases the board is in fact KR-base.

Signed-off-by: Yaniv Rosner <Yaniv.Rosner@qlogic.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnx2x: Correct asymmetric flow-control
Yuval Mintz [Thu, 25 Jun 2015 12:19:21 +0000 (15:19 +0300)]
bnx2x: Correct asymmetric flow-control

This fixes several issues relating to asymmetric configuration:
 1. When user requests to disable TX, the local-device needs to
    advertise both PAUSE and ASM_DIR, but to avoid transmitting pause
    frames. In the 578xx, it would ignore the TX disable.

 2. When user advertises RX-only, ASM_DIR was advertised instead of
    PAUSE/ASM_DIR.

 3. When changing mode, the advertised PAUSE/ASM_DIR was not cleared
    before setting new one, so disabling RX or TX had no impact on the
    'advertised' as appeared in the 'ethtool -a' output.

Signed-off-by: Yaniv Rosner <Yaniv.Rosner@qlogic.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: sched: flower fix typo
Jamal Hadi Salim [Thu, 25 Jun 2015 10:55:27 +0000 (06:55 -0400)]
net: sched: flower fix typo

Fix typo in the validation rules for flower's attributes

Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoenic: use atomic_t instead of spin_lock in busy poll
Govindarajulu Varadarajan [Thu, 25 Jun 2015 10:32:04 +0000 (16:02 +0530)]
enic: use atomic_t instead of spin_lock in busy poll

We use spinlock to access a single flag. We can avoid spin_locks by using
atomic variable and atomic_cmpxchg(). Use atomic_cmpxchg to set the flag
for idle to poll. And a simple atomic_set to unlock (set idle from poll).

In napi poll, if gro is enabled, we call napi_gro_receive() to deliver the
packets. Before we call napi_complete(), i.e while re-polling, if low
latency busy poll is called, we use netif_receive_skb() to deliver the packets.
At this point if there are some skb's held in GRO, busy poll could deliver the
packets out of order. So we call napi_gro_flush() to flush skbs before we
move the napi poll to idle.

Signed-off-by: Govindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/phy: Add Vitesse 8641 phy ID
Shaohui Xie [Thu, 25 Jun 2015 05:34:27 +0000 (13:34 +0800)]
net/phy: Add Vitesse 8641 phy ID

Vitesse VSC8641 is compatible with Vitesse 82xx

Signed-off-by: Shaohui Xie <Shaohui.Xie@freescale.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/fsl: remove dependency FSL_SOC for Gianfar
Alison Wang [Thu, 25 Jun 2015 03:34:38 +0000 (11:34 +0800)]
net/fsl: remove dependency FSL_SOC for Gianfar

CONFIG_GIANFAR is not depended on FSL_SOC, it
can be built on non-PPC platforms.

Signed-off-by: Alison Wang <alison.wang@freescale.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agocavium/liquidio: fix some error handling in lio_set_phys_id()
Dan Carpenter [Wed, 24 Jun 2015 14:47:02 +0000 (17:47 +0300)]
cavium/liquidio: fix some error handling in lio_set_phys_id()

There was a missing assignment so the "if (ret)" on the next line is
never true.

Fixes: f21fb3ed364b ('Add support of Cavium Liquidio ethernet adapters')
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agorenesas: missing unlock on error path
Dan Carpenter [Wed, 24 Jun 2015 14:32:54 +0000 (17:32 +0300)]
renesas: missing unlock on error path

We need to unlock before returning here.

Fixes: a0d2f20650e8 ('Renesas Ethernet AVB PTP clock driver')
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'mlx4'
David S. Miller [Thu, 25 Jun 2015 09:06:33 +0000 (02:06 -0700)]
Merge branch 'mlx4'

Or Gerlitz says:

====================
mlx4 driver fixes, June 24, 2015

Some fixes that we made recently, all need to go into stable.

patch #1 "net/mlx4_en: Release TX QP when destroying TX ring" and patch #3
"Fix wrong csum complete report when rxvlan offload is disabled" to >= 3.19

patch #2 "Wake TX queues only when there's enough room" addressing a bug
which is there from day one... should go to whatever kernels it's still applicable

patch #4 "mlx4: Disable HA for SRIOV PF RoCE devices" to >= 4.0

The patches are marked with net but are made against net-next,
as the net tree still doesn't contain all the net-next bits.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlx4: Disable HA for SRIOV PF RoCE devices
Or Gerlitz [Thu, 25 Jun 2015 08:29:44 +0000 (11:29 +0300)]
mlx4: Disable HA for SRIOV PF RoCE devices

When in HA mode, the driver exposes an IB (RoCE) device instance with only
one port. Under SRIOV, the existing implementation doesn't go well with
the PF RoCE driver's role of Special QPs Para-Virtualization, etc.

As such, disable HA for the mlx4 PF RoCE device in SRIOV mode.

Fixes: a57500903093 ('IB/mlx4: Add port aggregation support')
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/mlx4_en: Fix wrong csum complete report when rxvlan offload is disabled
Ido Shamay [Thu, 25 Jun 2015 08:29:43 +0000 (11:29 +0300)]
net/mlx4_en: Fix wrong csum complete report when rxvlan offload is disabled

The check_csum() function relied on hwtstamp_rx_filter to know if rxvlan
offload is disabled. This is wrong since rxvlan offload can be switched
on/off regardless of hwtstamp_rx_filter.

Also moved check_csum to query CQE information to identify VLAN packets
and removed the check of IP packets, since it has been validated before.

Fixes: f8c6455bb04b ('net/mlx4_en: Extend checksum offloading by CHECKSUM COMPLETE')
Signed-off-by: Ido Shamay <idos@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/mlx4_en: Wake TX queues only when there's enough room
Ido Shamay [Thu, 25 Jun 2015 08:29:42 +0000 (11:29 +0300)]
net/mlx4_en: Wake TX queues only when there's enough room

Indication of a single completed packet, marked by txbbs_skipped
being bigger then zero, in not enough in order to wake up a
stopped TX queue. The completed packet may contain a single TXBB,
while next packet to be sent (after the wake up) may have multiple
TXBBs (LSO/TSO packets for example), causing overflow in queue followed
by WQE corruption and TX queue timeout.
Instead, wake the stopped queue only when there's enough room for the
worst case (maximum sized WQE) packet that we should need to handle after
the queue is opened again.

Also created an helper routine - mlx4_en_is_tx_ring_full, which checks
if the current TX ring is full or not. It provides better code readability
and removes code duplication.

Signed-off-by: Ido Shamay <idos@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet/mlx4_en: Release TX QP when destroying TX ring
Eran Ben Elisha [Thu, 25 Jun 2015 08:29:41 +0000 (11:29 +0300)]
net/mlx4_en: Release TX QP when destroying TX ring

TX ring QP wasn't released at mlx4_en_destroy_tx_ring. Instead, the code
used the deprecated base_tx_qpn field. Move TX QP release to
mlx4_en_destroy_tx_ring and remove the base_tx_qpn field.

Fixes: ddae0349fdb7 ('net/mlx4: Change QP allocation scheme')
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'akpm' (patches from Andrew)
Linus Torvalds [Thu, 25 Jun 2015 03:47:21 +0000 (20:47 -0700)]
Merge branch 'akpm' (patches from Andrew)

Merge first patchbomb from Andrew Morton:

 - a few misc things

 - ocfs2 udpates

 - kernel/watchdog.c feature work (took ages to get right)

 - most of MM.  A few tricky bits are held up and probably won't make 4.2.

* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (91 commits)
  mm: kmemleak_alloc_percpu() should follow the gfp from per_alloc()
  mm, thp: respect MPOL_PREFERRED policy with non-local node
  tmpfs: truncate prealloc blocks past i_size
  mm/memory hotplug: print the last vmemmap region at the end of hot add memory
  mm/mmap.c: optimization of do_mmap_pgoff function
  mm: kmemleak: optimise kmemleak_lock acquiring during kmemleak_scan
  mm: kmemleak: avoid deadlock on the kmemleak object insertion error path
  mm: kmemleak: do not acquire scan_mutex in kmemleak_do_cleanup()
  mm: kmemleak: fix delete_object_*() race when called on the same memory block
  mm: kmemleak: allow safe memory scanning during kmemleak disabling
  memcg: convert mem_cgroup->under_oom from atomic_t to int
  memcg: remove unused mem_cgroup->oom_wakeups
  frontswap: allow multiple backends
  x86, mirror: x86 enabling - find mirrored memory ranges
  mm/memblock: allocate boot time data structures from mirrored memory
  mm/memblock: add extra "flags" to memblock to allow selection of memory based on attribute
  mm: do not ignore mapping_gfp_mask in page cache allocation paths
  mm/cma.c: fix typos in comments
  mm/oom_kill.c: print points as unsigned int
  mm/hugetlb: handle races in alloc_huge_page and hugetlb_reserve_pages
  ...

8 years agoMerge tag 'please-pull-pstore' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl...
Linus Torvalds [Thu, 25 Jun 2015 03:42:21 +0000 (20:42 -0700)]
Merge tag 'please-pull-pstore' of git://git./linux/kernel/git/aegl/linux

Pull pstore updates from Tony Luck:
 "Miscellaneous pstore improvements"

* tag 'please-pull-pstore' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux:
  ramoops: make it possible to change mem_type param.
  pstore/ram: verify ramoops header before saving record
  fs/pstore: Optimization function ramoops_init_przs
  fs/pstore: update the backend parameter in pstore module
  pstore: do not use message compression without lock

8 years agoMerge tag 'for-f2fs-4.2' of git://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk...
Linus Torvalds [Thu, 25 Jun 2015 03:38:29 +0000 (20:38 -0700)]
Merge tag 'for-f2fs-4.2' of git://git./linux/kernel/git/jaegeuk/f2fs

Pull f2fs updates from Jaegeuk Kim:
 "New features:
   - per-file encryption (e.g., ext4)
   - FALLOC_FL_ZERO_RANGE
   - FALLOC_FL_COLLAPSE_RANGE
   - RENAME_WHITEOUT

  Major enhancement/fixes:
   - recovery broken superblocks
   - enhance f2fs_trim_fs with a discard_map
   - fix a race condition on dentry block allocation
   - fix a deadlock during summary operation
   - fix a missing fiemap result

  .. and many minor bug fixes and clean-ups were done"

* tag 'for-f2fs-4.2' of git://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs: (83 commits)
  f2fs: do not trim preallocated blocks when truncating after i_size
  f2fs crypto: add alloc_bounce_page
  f2fs crypto: fix to handle errors likewise ext4
  f2fs: drop the volatile_write flag only
  f2fs: skip committing valid superblock
  f2fs: setting discard option in parse_options()
  f2fs: fix to return exact trimmed size
  f2fs: support FALLOC_FL_INSERT_RANGE
  f2fs: hide common code in f2fs_replace_block
  f2fs: disable the discard option when device doesn't support
  f2fs crypto: remove alloc_page for bounce_page
  f2fs: fix a deadlock for summary page lock vs. sentry_lock
  f2fs crypto: clean up error handling in f2fs_fname_setup_filename
  f2fs crypto: avoid f2fs_inherit_context for symlink
  f2fs crypto: do not set encryption policy for non-directory by ioctl
  f2fs crypto: allow setting encryption policy once
  f2fs crypto: check context consistent for rename2
  f2fs: avoid duplicated code by reusing f2fs_read_end_io
  f2fs crypto: use per-inode tfm structure
  f2fs: recovering broken superblock during mount
  ...

8 years agoMerge branch 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs
Linus Torvalds [Thu, 25 Jun 2015 03:07:10 +0000 (20:07 -0700)]
Merge branch 'for_linus' of git://git./linux/kernel/git/jack/linux-fs

Pull UDF fixes and cleanups from Jan Kara:
 "The contains some small fixes and improvements in error handling for
  UDF.

  Bundled is also one ext3 coding style fix and a fix in quota
  documentation"

* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs:
  udf: fix udf_load_pvoldesc()
  udf: remove double err declaration in udf_file_write_iter()
  UDF: support NFSv2 export
  fs: ext3: super: fixed a space coding style issue
  quota: Update documentation
  udf: Return error from udf_find_entry()
  udf: Make udf_get_filename() return error instead of 0 length file name
  udf: bug on exotic flag in udf_get_filename()
  udf: improve error management in udf_CS0toNLS()
  udf: improve error management in udf_CS0toUTF8()
  udf: unicode: update function name in comments
  udf: remove unnecessary test in udf_build_ustr_exact()
  udf: Return -ENOMEM when allocation fails in udf_get_filename()

8 years agoMerge tag 'docs-for-linus' of git://git.lwn.net/linux-2.6
Linus Torvalds [Thu, 25 Jun 2015 03:01:36 +0000 (20:01 -0700)]
Merge tag 'docs-for-linus' of git://git.lwn.net/linux-2.6

Pull documentation updates from Jonathan Corbet:
 "The main thing here is Ingo's big subdirectory documenting feature
  support for each architecture.  Beyond that, it's the usual pile of
  fixes, tweaks, and small additions"

* tag 'docs-for-linus' of git://git.lwn.net/linux-2.6: (79 commits)
  doc:md: fix typo in md.txt.
  Documentation/mic/mpssd: don't build x86 userspace when cross compiling
  Documentation/prctl: don't build tsc tests when cross compiling
  Documentation/vDSO: don't build tests when cross compiling
  Doc:ABI/testing: Fix typo in sysfs-bus-fcoe
  Doc: Docbook: Change wikipedia's URL from http to https in scsi.tmpl
  Doc: Change wikipedia's URL from http to https
  Documentation/kernel-parameters: add missing pciserial to the earlyprintk
  Doc:pps: Fix typo in pps.txt
  kbuild : Fix documentation of INSTALL_HDR_PATH
  Documentation: filesystems: updated struct file_operations documentation in vfs.txt
  kbuild: edit explanation of clean-files variable
  Doc: ja_JP: Fix typo in HOWTO
  Move freefall program from Documentation/ to tools/
  Documentation: ARM: EXYNOS: Describe boot loaders interface
  Doc:nfc: Fix typo in nfc-hci.txt
  vfs: Minor documentation fix
  Doc: networking: txtimestamp: fix printf format warning
  Documentation, intel_pstate: Improve legacy mode internal governors description
  Documentation: extend use case for EXPORT_SYMBOL_GPL()
  ...

8 years agoMerge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input
Linus Torvalds [Thu, 25 Jun 2015 02:56:58 +0000 (19:56 -0700)]
Merge branch 'for-linus' of git://git./linux/kernel/git/dtor/input

Pull input subsystem updates from Dmitry Torokhov:
 "Thanks to Samuel Thibault input device (keyboard) LEDs are no longer
  hardwired within the input core but use LED subsystem and so allow use
  of different triggers; Hans de Goede did a large update for the ALPS
  touchpad driver; we have new TI drv2665 haptics driver and DA9063
  OnKey driver, and host of other drivers got various fixes"

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input: (55 commits)
  Input: pixcir_i2c_ts - fix receive error
  MAINTAINERS: remove non existent input mt git tree
  Input: improve usage of gpiod API
  tty/vt/keyboard: define LED triggers for VT keyboard lock states
  tty/vt/keyboard: define LED triggers for VT LED states
  Input: export LEDs as class devices in sysfs
  Input: cyttsp4 - use swap() in cyttsp4_get_touch()
  Input: goodix - do not explicitly set evbits in input device
  Input: goodix - export id and version read from device
  Input: goodix - fix variable length array warning
  Input: goodix - fix alignment issues
  Input: add OnKey driver for DA9063 MFD part
  Input: elan_i2c - add product IDs FW names
  Input: elan_i2c - add support for multi IC type and iap format
  Input: focaltech - report finger width to userspace
  tty: remove platform_sysrq_reset_seq
  Input: synaptics_i2c - use proper boolean values
  Input: psmouse - use true instead of 1 for boolean values
  Input: cyapa - fix a few typos in comments
  Input: stmpe-ts - enforce device tree only mode
  ...

8 years agoMerge tag 'edac_for_4.2_2' of git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp
Linus Torvalds [Thu, 25 Jun 2015 02:52:06 +0000 (19:52 -0700)]
Merge tag 'edac_for_4.2_2' of git://git./linux/kernel/git/bp/bp

Pull EDAC updates from Borislav Petkov:

 - New APM X-Gene SoC EDAC driver (Loc Ho)

 - AMD error injection module improvements (Aravind Gopalakrishnan)

 - Altera Arria 10 support (Thor Thayer)

 - misc fixes and cleanups all over the place

* tag 'edac_for_4.2_2' of git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp: (28 commits)
  EDAC: Update Documentation/edac.txt
  EDAC: Fix typos in Documentation/edac.txt
  EDAC, mce_amd_inj: Set MISCV on injection
  EDAC, mce_amd_inj: Move bit preparations before the injection
  EDAC, mce_amd_inj: Cleanup and simplify README
  EDAC, altera: Do not allow suspend when EDAC is enabled
  EDAC, mce_amd_inj: Make inj_type static
  arm: socfpga: dts: Add Arria10 SDRAM EDAC DTS support
  EDAC, altera: Add Arria10 EDAC support
  EDAC, altera: Refactor for Altera CycloneV SoC
  EDAC, altera: Generalize driver to use DT Memory size
  EDAC, mce_amd_inj: Add README file
  EDAC, mce_amd_inj: Add individual permissions field to dfs_node
  EDAC, mce_amd_inj: Modify flags attribute to use string arguments
  EDAC, mce_amd_inj: Read out number of MCE banks from the hardware
  EDAC, mce_amd_inj: Use MCE_INJECT_GET macro for bank node too
  EDAC, xgene: Fix cpuid abuse
  EDAC, mpc85xx: Extend error address to 64 bit
  EDAC, mpc8xxx: Adapt for FSL SoC
  EDAC, edac_stub: Drop arch-specific include
  ...

8 years agoMerge tag 'pinctrl-v4.2-1' of git://git.kernel.org/pub/scm/linux/kernel/git/linusw...
Linus Torvalds [Thu, 25 Jun 2015 02:21:02 +0000 (19:21 -0700)]
Merge tag 'pinctrl-v4.2-1' of git://git./linux/kernel/git/linusw/linux-pinctrl

Pull pin control updates from Linus Walleij:
 "Here is the bulk of pin control changes for the v4.2 series: Quite a
  lot of new SoC subdrivers and two new main drivers this time, apart
  from that business as usual.

  Details:

  Core functionality:
   - Enable exclusive pin ownership: it is possible to flag a pin
     controller so that GPIO and other functions cannot use a single pin
     simultaneously.

  New drivers:
   - NXP LPC18xx System Control Unit pin controller
   - Imagination Pistachio SoC pin controller

  New subdrivers:
   - Freescale i.MX7d SoC
   - Intel Sunrisepoint-H PCH
   - Renesas PFC R8A7793
   - Renesas PFC R8A7794
   - Mediatek MT6397, MT8127
   - SiRF Atlas 7
   - Allwinner A33
   - Qualcomm MSM8660
   - Marvell Armada 395
   - Rockchip RK3368

  Cleanups:
   - A big cleanup of the Marvell MVEBU driver rectifying it to
     correspond to reality
   - Drop platform device probing from the SH PFC driver, we are now a
     DT only shop for SuperH
   - Drop obsolte multi-platform check for SH PFC
   - Various janitorial: constification, grammar etc

  Improvements:
   - The AT91 GPIO portions now supports the set_multiple() feature
   - Split out SPI pins on the Xilinx Zynq
   - Support DTs without specific function nodes in the i.MX driver"

* tag 'pinctrl-v4.2-1' of git://git.kernel.org/pub/scm/linux/kernel/git/linusw/linux-pinctrl: (99 commits)
  pinctrl: rockchip: add support for the rk3368
  pinctrl: rockchip: generalize perpin driver-strength setting
  pinctrl: sh-pfc: r8a7794: add SDHI pin groups
  pinctrl: sh-pfc: r8a7794: add MMCIF pin groups
  pinctrl: sh-pfc: add R8A7794 PFC support
  pinctrl: make pinctrl_register() return proper error code
  pinctrl: mvebu: armada-39x: add support for Armada 395 variant
  pinctrl: mvebu: armada-39x: add missing SATA functions
  pinctrl: mvebu: armada-39x: add missing PCIe functions
  pinctrl: mvebu: armada-38x: add ptp functions
  pinctrl: mvebu: armada-38x: add ua1 functions
  pinctrl: mvebu: armada-38x: add nand functions
  pinctrl: mvebu: armada-38x: add sata functions
  pinctrl: mvebu: armada-xp: add dram functions
  pinctrl: mvebu: armada-xp: add nand rb function
  pinctrl: mvebu: armada-xp: add spi1 function
  pinctrl: mvebu: armada-39x: normalize ref clock naming
  pinctrl: mvebu: armada-xp: rename spi to spi0
  pinctrl: mvebu: armada-370: align spi1 clock pin naming
  pinctrl: mvebu: armada-370: align VDD cpu-pd pin naming with datasheet
  ...

8 years agoMerge tag 'backlight-for-linus-4.2' of git://git.kernel.org/pub/scm/linux/kernel...
Linus Torvalds [Thu, 25 Jun 2015 01:57:00 +0000 (18:57 -0700)]
Merge tag 'backlight-for-linus-4.2' of git://git./linux/kernel/git/lee/backlight

Pull backlight updates from Lee Jones:
 "Changes to existing drivers:

   - supply MODULE_DEVICE_TABLE() to ensure probing
   - constify struct; da9052_bl
   - enable compile test; lcd_l4f00242t03, lcd_lms283fg05, backlight_gpio
   - suspend/resume bugfix; lp855x_bl
   - devm_gpiod_get_optional() API fixup; pwm_bl
   - error handling fixup; backlight"

* tag 'backlight-for-linus-4.2' of git://git.kernel.org/pub/scm/linux/kernel/git/lee/backlight:
  backlight: Change the return type of backlight_update_status() to int
  backlight: pwm_bl: Simplify usage of devm_gpiod_get_optional
  backlight: lp855x: Don't clear level on suspend/blank
  backlight: Allow compile test of GPIO consumers if !GPIOLIB
  video: backlight: da9052: Constify platform_device_id
  gpio-backlight: Discover driver during boot time

8 years agomm: kmemleak_alloc_percpu() should follow the gfp from per_alloc()
Larry Finger [Wed, 24 Jun 2015 23:58:51 +0000 (16:58 -0700)]
mm: kmemleak_alloc_percpu() should follow the gfp from per_alloc()

Beginning at commit d52d3997f843 ("ipv6: Create percpu rt6_info"), the
following INFO splat is logged:

  ===============================
  [ INFO: suspicious RCU usage. ]
  4.1.0-rc7-next-20150612 #1 Not tainted
  -------------------------------
  kernel/sched/core.c:7318 Illegal context switch in RCU-bh read-side critical section!
  other info that might help us debug this:
  rcu_scheduler_active = 1, debug_locks = 0
   3 locks held by systemd/1:
   #0:  (rtnl_mutex){+.+.+.}, at: [<ffffffff815f0c8f>] rtnetlink_rcv+0x1f/0x40
   #1:  (rcu_read_lock_bh){......}, at: [<ffffffff816a34e2>] ipv6_add_addr+0x62/0x540
   #2:  (addrconf_hash_lock){+...+.}, at: [<ffffffff816a3604>] ipv6_add_addr+0x184/0x540
  stack backtrace:
  CPU: 0 PID: 1 Comm: systemd Not tainted 4.1.0-rc7-next-20150612 #1
  Hardware name: TOSHIBA TECRA A50-A/TECRA A50-A, BIOS Version 4.20   04/17/2014
  Call Trace:
    dump_stack+0x4c/0x6e
    lockdep_rcu_suspicious+0xe7/0x120
    ___might_sleep+0x1d5/0x1f0
    __might_sleep+0x4d/0x90
    kmem_cache_alloc+0x47/0x250
    create_object+0x39/0x2e0
    kmemleak_alloc_percpu+0x61/0xe0
    pcpu_alloc+0x370/0x630

Additional backtrace lines are truncated.  In addition, the above splat
is followed by several "BUG: sleeping function called from invalid
context at mm/slub.c:1268" outputs.  As suggested by Martin KaFai Lau,
these are the clue to the fix.  Routine kmemleak_alloc_percpu() always
uses GFP_KERNEL for its allocations, whereas it should follow the gfp
from its callers.

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Reviewed-by: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Larry Finger <Larry.Finger@lwfinger.net>
Cc: Martin KaFai Lau <kafai@fb.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: <stable@vger.kernel.org> [3.18+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, thp: respect MPOL_PREFERRED policy with non-local node
Vlastimil Babka [Wed, 24 Jun 2015 23:58:48 +0000 (16:58 -0700)]
mm, thp: respect MPOL_PREFERRED policy with non-local node

Since commit 077fcf116c8c ("mm/thp: allocate transparent hugepages on
local node"), we handle THP allocations on page fault in a special way -
for non-interleave memory policies, the allocation is only attempted on
the node local to the current CPU, if the policy's nodemask allows the
node.

This is motivated by the assumption that THP benefits cannot offset the
cost of remote accesses, so it's better to fallback to base pages on the
local node (which might still be available, while huge pages are not due
to fragmentation) than to allocate huge pages on a remote node.

The nodemask check prevents us from violating e.g.  MPOL_BIND policies
where the local node is not among the allowed nodes.  However, the
current implementation can still give surprising results for the
MPOL_PREFERRED policy when the preferred node is different than the
current CPU's local node.

In such case we should honor the preferred node and not use the local
node, which is what this patch does.  If hugepage allocation on the
preferred node fails, we fall back to base pages and don't try other
nodes, with the same motivation as is done for the local node hugepage
allocations.  The patch also moves the MPOL_INTERLEAVE check around to
simplify the hugepage specific test.

The difference can be demonstrated using in-tree transhuge-stress test
on the following 2-node machine where half memory on one node was
occupied to show the difference.

> numactl --hardware
available: 2 nodes (0-1)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 24 25 26 27 28 29 30 31 32 33 34 35
node 0 size: 7878 MB
node 0 free: 3623 MB
node 1 cpus: 12 13 14 15 16 17 18 19 20 21 22 23 36 37 38 39 40 41 42 43 44 45 46 47
node 1 size: 8045 MB
node 1 free: 7818 MB
node distances:
node   0   1
  0:  10  21
  1:  21  10

Before the patch:
> numactl -p0 -C0 ./transhuge-stress
transhuge-stress: 2.197 s/loop, 0.276 ms/page,   7249.168 MiB/s 7962 succeed,    0 failed, 1786 different pages

> numactl -p0 -C12 ./transhuge-stress
transhuge-stress: 2.962 s/loop, 0.372 ms/page,   5376.172 MiB/s 7962 succeed,    0 failed, 3873 different pages

Number of successful THP allocations corresponds to free memory on node 0 in
the first case and node 1 in the second case, i.e. -p parameter is ignored and
cpu binding "wins".

After the patch:
> numactl -p0 -C0 ./transhuge-stress
transhuge-stress: 2.183 s/loop, 0.274 ms/page,   7295.516 MiB/s 7962 succeed,    0 failed, 1760 different pages

> numactl -p0 -C12 ./transhuge-stress
transhuge-stress: 2.878 s/loop, 0.361 ms/page,   5533.638 MiB/s 7962 succeed,    0 failed, 1750 different pages

> numactl -p1 -C0 ./transhuge-stress
transhuge-stress: 4.628 s/loop, 0.581 ms/page,   3440.893 MiB/s 7962 succeed,    0 failed, 3918 different pages

The -p parameter is respected regardless of cpu binding.

> numactl -C0 ./transhuge-stress
transhuge-stress: 2.202 s/loop, 0.277 ms/page,   7230.003 MiB/s 7962 succeed,    0 failed, 1750 different pages

> numactl -C12 ./transhuge-stress
transhuge-stress: 3.020 s/loop, 0.379 ms/page,   5273.324 MiB/s 7962 succeed,    0 failed, 3916 different pages

Without -p parameter, hugepage restriction to CPU-local node works as before.

Fixes: 077fcf116c8c ("mm/thp: allocate transparent hugepages on local node")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: <stable@vger.kernel.org> [4.0+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agotmpfs: truncate prealloc blocks past i_size
Josef Bacik [Wed, 24 Jun 2015 23:58:45 +0000 (16:58 -0700)]
tmpfs: truncate prealloc blocks past i_size

One of the rocksdb people noticed that when you do something like this

    fallocate(fd, FALLOC_FL_KEEP_SIZE, 0, 10M)
    pwrite(fd, buf, 5M, 0)
    ftruncate(5M)

on tmpfs, the file would still take up 10M: which led to super fun
issues because we were getting ENOSPC before we thought we should be
getting ENOSPC.  This patch fixes the problem, and mirrors what all the
other fs'es do (and was agreed to be the correct behaviour at LSF).

I tested it locally to make sure it worked properly with the following

    xfs_io -f -c "falloc -k 0 10M" -c "pwrite 0 5M" -c "truncate 5M" file

Without the patch we have "Blocks: 20480", with the patch we have the
correct value of "Blocks: 10240".

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memory hotplug: print the last vmemmap region at the end of hot add memory
Zhu Guihua [Wed, 24 Jun 2015 23:58:42 +0000 (16:58 -0700)]
mm/memory hotplug: print the last vmemmap region at the end of hot add memory

When hot add two nodes continuously, we found the vmemmap region info is
a bit messed.  The last region of node 2 is printed when node 3 hot
added, like the following:

  Initmem setup node 2 [mem 0x0000000000000000-0xffffffffffffffff]
   On node 2 totalpages: 0
   Built 2 zonelists in Node order, mobility grouping on.  Total pages: 16090539
   Policy zone: Normal
   init_memory_mapping: [mem 0x40000000000-0x407ffffffff]
    [mem 0x40000000000-0x407ffffffff] page 1G
    [ffffea1000000000-ffffea10001fffff] PMD -> [ffff8a077d800000-ffff8a077d9fffff] on node 2
    [ffffea1000200000-ffffea10003fffff] PMD -> [ffff8a077de00000-ffff8a077dffffff] on node 2
  ...
    [ffffea101f600000-ffffea101f9fffff] PMD -> [ffff8a074ac00000-ffff8a074affffff] on node 2
    [ffffea101fa00000-ffffea101fdfffff] PMD -> [ffff8a074a800000-ffff8a074abfffff] on node 2
  Initmem setup node 3 [mem 0x0000000000000000-0xffffffffffffffff]
   On node 3 totalpages: 0
   Built 3 zonelists in Node order, mobility grouping on.  Total pages: 16090539
   Policy zone: Normal
   init_memory_mapping: [mem 0x60000000000-0x607ffffffff]
    [mem 0x60000000000-0x607ffffffff] page 1G
    [ffffea101fe00000-ffffea101fffffff] PMD -> [ffff8a074a400000-ffff8a074a5fffff] on node 2 <=== node 2 ???
    [ffffea1800000000-ffffea18001fffff] PMD -> [ffff8a074a600000-ffff8a074a7fffff] on node 3
    [ffffea1800200000-ffffea18005fffff] PMD -> [ffff8a074a000000-ffff8a074a3fffff] on node 3
    [ffffea1800600000-ffffea18009fffff] PMD -> [ffff8a0749c00000-ffff8a0749ffffff] on node 3
  ...

The cause is the last region was missed at the and of hot add memory,
and p_start, p_end, node_start were not reset, so when hot add memory to
a new node, it will consider they are not contiguous blocks and print
the previous one.  So we print the last vmemmap region at the end of hot
add memory to avoid the confusion.

Signed-off-by: Zhu Guihua <zhugh.fnst@cn.fujitsu.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/mmap.c: optimization of do_mmap_pgoff function
Piotr Kwapulinski [Wed, 24 Jun 2015 23:58:39 +0000 (16:58 -0700)]
mm/mmap.c: optimization of do_mmap_pgoff function

The simple check for zero length memory mapping may be performed
earlier.  So that in case of zero length memory mapping some unnecessary
code is not executed at all.  It does not make the code less readable
and saves some CPU cycles.

Signed-off-by: Piotr Kwapulinski <kwapulinski.piotr@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: kmemleak: optimise kmemleak_lock acquiring during kmemleak_scan
Catalin Marinas [Wed, 24 Jun 2015 23:58:37 +0000 (16:58 -0700)]
mm: kmemleak: optimise kmemleak_lock acquiring during kmemleak_scan

The kmemleak memory scanning uses finer grained object->lock spinlocks
primarily to avoid races with the memory block freeing.  However, the
pointer lookup in the rb tree requires the kmemleak_lock to be held.
This is currently done in the find_and_get_object() function for each
pointer-like location read during scanning.  While this allows a low
latency on kmemleak_*() callbacks on other CPUs, the memory scanning is
slower.

This patch moves the kmemleak_lock outside the scan_block() loop,
acquiring/releasing it only once per scanned memory block.  The
allow_resched logic is moved outside scan_block() and a new
scan_large_block() function is implemented which splits large blocks in
MAX_SCAN_SIZE chunks with cond_resched() calls in-between.  A redundant
(object->flags & OBJECT_NO_SCAN) check is also removed from
scan_object().

With this patch, the kmemleak scanning performance is significantly
improved: at least 50% with lock debugging disabled and over an order of
magnitude with lock proving enabled (on an arm64 system).

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: kmemleak: avoid deadlock on the kmemleak object insertion error path
Catalin Marinas [Wed, 24 Jun 2015 23:58:34 +0000 (16:58 -0700)]
mm: kmemleak: avoid deadlock on the kmemleak object insertion error path

While very unlikely (usually kmemleak or sl*b bug), the create_object()
function in mm/kmemleak.c may fail to insert a newly allocated object into
the rb tree.  When this happens, kmemleak disables itself and prints
additional information about the object already found in the rb tree.
Such printing is done with the parent->lock acquired, however the
kmemleak_lock is already held.  This is a potential race with the scanning
thread which acquires object->lock and kmemleak_lock in a

This patch removes the locking around the 'parent' object information
printing.  Such object cannot be freed or removed from object_tree_root
and object_list since kmemleak_lock is already held.  There is a very
small risk that some of the object data is being modified on another CPU
but the only downside is inconsistent information printing.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: kmemleak: do not acquire scan_mutex in kmemleak_do_cleanup()
Catalin Marinas [Wed, 24 Jun 2015 23:58:31 +0000 (16:58 -0700)]
mm: kmemleak: do not acquire scan_mutex in kmemleak_do_cleanup()

The kmemleak_do_cleanup() work thread already waits for the kmemleak_scan
thread to finish via kthread_stop().  Waiting in kthread_stop() while
scan_mutex is held may lead to deadlock if kmemleak_scan_thread() also
waits to acquire for scan_mutex.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: kmemleak: fix delete_object_*() race when called on the same memory block
Catalin Marinas [Wed, 24 Jun 2015 23:58:29 +0000 (16:58 -0700)]
mm: kmemleak: fix delete_object_*() race when called on the same memory block

Calling delete_object_*() on the same pointer is not a standard use case
(unless there is a bug in the code calling kmemleak_free()).  However,
during kmemleak disabling (error or user triggered via /sys), there is a
potential race between kmemleak_free() calls on a CPU and
__kmemleak_do_cleanup() on a different CPU.

The current delete_object_*() implementation first performs a look-up
holding kmemleak_lock, increments the object->use_count and then
re-acquires kmemleak_lock to remove the object from object_tree_root and
object_list.

This patch simplifies the delete_object_*() mechanism to both look up
and remove an object from the object_tree_root and object_list
atomically (guarded by kmemleak_lock).  This allows safe concurrent
calls to delete_object_*() on the same pointer without additional
locking for synchronising the kmemleak_free_enabled flag.

A side effect is a slight improvement in the delete_object_*() performance
by avoiding acquiring kmemleak_lock twice and incrementing/decrementing
object->use_count.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: kmemleak: allow safe memory scanning during kmemleak disabling
Catalin Marinas [Wed, 24 Jun 2015 23:58:26 +0000 (16:58 -0700)]
mm: kmemleak: allow safe memory scanning during kmemleak disabling

The kmemleak scanning thread can run for minutes.  Callbacks like
kmemleak_free() are allowed during this time, the race being taken care
of by the object->lock spinlock.  Such lock also prevents a memory block
from being freed or unmapped while it is being scanned by blocking the
kmemleak_free() -> ...  -> __delete_object() function until the lock is
released in scan_object().

When a kmemleak error occurs (e.g.  it fails to allocate its metadata),
kmemleak_enabled is set and __delete_object() is no longer called on
freed objects.  If kmemleak_scan is running at the same time,
kmemleak_free() no longer waits for the object scanning to complete,
allowing the corresponding memory block to be freed or unmapped (in the
case of vfree()).  This leads to kmemleak_scan potentially triggering a
page fault.

This patch separates the kmemleak_free() enabling/disabling from the
overall kmemleak_enabled nob so that we can defer the disabling of the
object freeing tracking until the scanning thread completed.  The
kmemleak_free_part() is deliberately ignored by this patch since this is
only called during boot before the scanning thread started.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Vignesh Radhakrishnan <vigneshr@codeaurora.org>
Tested-by: Vignesh Radhakrishnan <vigneshr@codeaurora.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomemcg: convert mem_cgroup->under_oom from atomic_t to int
Tejun Heo [Wed, 24 Jun 2015 23:58:23 +0000 (16:58 -0700)]
memcg: convert mem_cgroup->under_oom from atomic_t to int

memcg->under_oom tracks whether the memcg is under OOM conditions and is
an atomic_t counter managed with mem_cgroup_[un]mark_under_oom().  While
atomic_t appears to be simple synchronization-wise, when used as a
synchronization construct like here, it's trickier and more error-prone
due to weak memory ordering rules, especially around atomic_read(), and
false sense of security.

For example, both non-trivial read sites of memcg->under_oom are a bit
problematic although not being actually broken.

* mem_cgroup_oom_register_event()

  It isn't explicit what guarantees the memory ordering between event
  addition and memcg->under_oom check.  This isn't broken only because
  memcg_oom_lock is used for both event list and memcg->oom_lock.

* memcg_oom_recover()

  The lockless test doesn't have any explanation why this would be
  safe.

mem_cgroup_[un]mark_under_oom() are very cold paths and there's no point
in avoiding locking memcg_oom_lock there.  This patch converts
memcg->under_oom from atomic_t to int, puts their modifications under
memcg_oom_lock and documents why the lockless test in
memcg_oom_recover() is safe.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomemcg: remove unused mem_cgroup->oom_wakeups
Tejun Heo [Wed, 24 Jun 2015 23:58:21 +0000 (16:58 -0700)]
memcg: remove unused mem_cgroup->oom_wakeups

Since commit 4942642080ea ("mm: memcg: handle non-error OOM situations
more gracefully"), nobody uses mem_cgroup->oom_wakeups.  Remove it.

While at it, also fold memcg_wakeup_oom() into memcg_oom_recover() which
is its only user.  This cleanup was suggested by Michal.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agofrontswap: allow multiple backends
Dan Streetman [Wed, 24 Jun 2015 23:58:18 +0000 (16:58 -0700)]
frontswap: allow multiple backends

Change frontswap single pointer to a singly linked list of frontswap
implementations.  Update Xen tmem implementation as register no longer
returns anything.

Frontswap only keeps track of a single implementation; any
implementation that registers second (or later) will replace the
previously registered implementation, and gets a pointer to the previous
implementation that the new implementation is expected to pass all
frontswap functions to if it can't handle the function itself.  However
that method doesn't really make much sense, as passing that work on to
every implementation adds unnecessary work to implementations; instead,
frontswap should simply keep a list of all registered implementations
and try each implementation for any function.  Most importantly, neither
of the two currently existing frontswap implementations in the kernel
actually do anything with any previous frontswap implementation that
they replace when registering.

This allows frontswap to successfully manage multiple implementations by
keeping a list of them all.

Signed-off-by: Dan Streetman <ddstreet@ieee.org>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agox86, mirror: x86 enabling - find mirrored memory ranges
Tony Luck [Wed, 24 Jun 2015 23:58:15 +0000 (16:58 -0700)]
x86, mirror: x86 enabling - find mirrored memory ranges

UEFI GetMemoryMap() uses a new attribute bit to mark mirrored memory
address ranges.  See UEFI 2.5 spec pages 157-158:

  http://www.uefi.org/sites/default/files/resources/UEFI%202_5.pdf

On EFI enabled systems scan the memory map and tell memblock about any
mirrored ranges.

Signed-off-by: Tony Luck <tony.luck@intel.com>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Hanjun Guo <guohanjun@huawei.com>
Cc: Xiexiuqi <xiexiuqi@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memblock: allocate boot time data structures from mirrored memory
Tony Luck [Wed, 24 Jun 2015 23:58:12 +0000 (16:58 -0700)]
mm/memblock: allocate boot time data structures from mirrored memory

Try to allocate all boot time kernel data structures from mirrored
memory.

If we run out of mirrored memory print warnings, but fall back to using
non-mirrored memory to make sure that we still boot.

By number of bytes, most of what we allocate at boot time is the page
structures.  64 bytes per 4K page on x86_64 ...  or about 1.5% of total
system memory.  For workloads where the bulk of memory is allocated to
applications this may represent a useful improvement to system
availability since 1.5% of total memory might be a third of the memory
allocated to the kernel.

Signed-off-by: Tony Luck <tony.luck@intel.com>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Hanjun Guo <guohanjun@huawei.com>
Cc: Xiexiuqi <xiexiuqi@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memblock: add extra "flags" to memblock to allow selection of memory based on...
Tony Luck [Wed, 24 Jun 2015 23:58:09 +0000 (16:58 -0700)]
mm/memblock: add extra "flags" to memblock to allow selection of memory based on attribute

Some high end Intel Xeon systems report uncorrectable memory errors as a
recoverable machine check.  Linux has included code for some time to
process these and just signal the affected processes (or even recover
completely if the error was in a read only page that can be replaced by
reading from disk).

But we have no recovery path for errors encountered during kernel code
execution.  Except for some very specific cases were are unlikely to ever
be able to recover.

Enter memory mirroring. Actually 3rd generation of memory mirroing.

Gen1: All memory is mirrored
Pro: No s/w enabling - h/w just gets good data from other side of the
     mirror
Con: Halves effective memory capacity available to OS/applications

Gen2: Partial memory mirror - just mirror memory begind some memory controllers
Pro: Keep more of the capacity
Con: Nightmare to enable. Have to choose between allocating from
     mirrored memory for safety vs. NUMA local memory for performance

Gen3: Address range partial memory mirror - some mirror on each memory
      controller
Pro: Can tune the amount of mirror and keep NUMA performance
Con: I have to write memory management code to implement

The current plan is just to use mirrored memory for kernel allocations.
This has been broken into two phases:

1) This patch series - find the mirrored memory, use it for boot time
   allocations

2) Wade into mm/page_alloc.c and define a ZONE_MIRROR to pick up the
   unused mirrored memory from mm/memblock.c and only give it out to
   select kernel allocations (this is still being scoped because
   page_alloc.c is scary).

This patch (of 3):

Add extra "flags" to memblock to allow selection of memory based on
attribute.  No functional changes

Signed-off-by: Tony Luck <tony.luck@intel.com>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Hanjun Guo <guohanjun@huawei.com>
Cc: Xiexiuqi <xiexiuqi@huawei.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Yinghai Lu <yinghai@kernel.org>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: do not ignore mapping_gfp_mask in page cache allocation paths
Michal Hocko [Wed, 24 Jun 2015 23:58:06 +0000 (16:58 -0700)]
mm: do not ignore mapping_gfp_mask in page cache allocation paths

page_cache_read, do_generic_file_read, __generic_file_splice_read and
__ntfs_grab_cache_pages currently ignore mapping_gfp_mask when calling
add_to_page_cache_lru which might cause recursion into fs down in the
direct reclaim path if the mapping really relies on GFP_NOFS semantic.

This doesn't seem to be the case now because page_cache_read (page fault
path) doesn't seem to suffer from the reclaim recursion issues and
do_generic_file_read and __generic_file_splice_read also shouldn't be
called under fs locks which would deadlock in the reclaim path.  Anyway it
is better to obey mapping gfp mask and prevent from later breakage.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Neil Brown <neilb@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Anton Altaparmakov <anton@tuxera.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/cma.c: fix typos in comments
Shailendra Verma [Wed, 24 Jun 2015 23:58:03 +0000 (16:58 -0700)]
mm/cma.c: fix typos in comments

Signed-off-by: Shailendra Verma <shailendra.capricorn@gmail.com>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/oom_kill.c: print points as unsigned int
Wang Long [Wed, 24 Jun 2015 23:58:01 +0000 (16:58 -0700)]
mm/oom_kill.c: print points as unsigned int

In oom_kill_process(), the variable 'points' is unsigned int.  Print it as
such.

Signed-off-by: Wang Long <long.wanglong@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/hugetlb: handle races in alloc_huge_page and hugetlb_reserve_pages
Mike Kravetz [Wed, 24 Jun 2015 23:57:58 +0000 (16:57 -0700)]
mm/hugetlb: handle races in alloc_huge_page and hugetlb_reserve_pages

alloc_huge_page and hugetlb_reserve_pages use region_chg to calculate the
number of pages which will be added to the reserve map.  Subpool and
global reserve counts are adjusted based on the output of region_chg.
Before the pages are actually added to the reserve map, these routines
could race and add fewer pages than expected.  If this happens, the
subpool and global reserve counts are not correct.

Compare the number of pages actually added (region_add) to those expected
to added (region_chg).  If fewer pages are actually added, this indicates
a race and adjust counters accordingly.

Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/hugetlb: compute/return the number of regions added by region_add()
Mike Kravetz [Wed, 24 Jun 2015 23:57:55 +0000 (16:57 -0700)]
mm/hugetlb: compute/return the number of regions added by region_add()

Modify region_add() to keep track of regions(pages) added to the reserve
map and return this value.  The return value can be compared to the return
value of region_chg() to determine if the map was modified between calls.

Make vma_commit_reservation() also pass along the return value of
region_add().  In the normal case, we want vma_commit_reservation to
return the same value as the preceding call to vma_needs_reservation.
Create a common __vma_reservation_common routine to help keep the special
case return values in sync

Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/hugetlb: document the reserve map/region tracking routines
Mike Kravetz [Wed, 24 Jun 2015 23:57:52 +0000 (16:57 -0700)]
mm/hugetlb: document the reserve map/region tracking routines

While working on hugetlbfs fallocate support, I noticed the following race
in the existing code.  It is unlikely that this race is hit very often in
the current code.  However, if more functionality to add and remove pages
to hugetlbfs mappings (such as fallocate) is added the likelihood of
hitting this race will increase.

alloc_huge_page and hugetlb_reserve_pages use information from the reserve
map to determine if there are enough available huge pages to complete the
operation, as well as adjust global reserve and subpool usage counts.  The
order of operations is as follows:

- call region_chg() to determine the expected change based on reserve map
- determine if enough resources are available for this operation
- adjust global counts based on the expected change
- call region_add() to update the reserve map

The issue is that reserve map could change between the call to region_chg
and region_add.  In this case, the counters which were adjusted based on
the output of region_chg will not be correct.

In order to hit this race today, there must be an existing shared hugetlb
mmap created with the MAP_NORESERVE flag.  A page fault to allocate a huge
page via this mapping must occur at the same another task is mapping the
same region without the MAP_NORESERVE flag.

The patch set does not prevent the race from happening.  Rather, it adds
simple functionality to detect when the race has occurred.  If a race is
detected, then the incorrect counts are adjusted.

Review comments pointed out the need for documentation of the existing
region/reserve map routines.  This patch set also adds documentation in
this area.

This patch (of 3):

This is a documentation only patch and does not modify any code.
Descriptions of the routines used for reserve map/region tracking are
added.

Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoDocumentation/vm/unevictable-lru.txt: clarify MAP_LOCKED behavior
Michal Hocko [Wed, 24 Jun 2015 23:57:50 +0000 (16:57 -0700)]
Documentation/vm/unevictable-lru.txt: clarify MAP_LOCKED behavior

There is a very subtle difference between mmap()+mlock() vs
mmap(MAP_LOCKED) semantic.  The former one fails if the population of the
area fails while the later one doesn't.  This basically means that
mmap(MAPLOCKED) areas might see major fault after mmap syscall returns
which is not the case for mlock.  mmap man page has already been altered
but Documentation/vm/unevictable-lru.txt deserves a clarification as well.

Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: nommu: refactor debug and warning prints
Leon Romanovsky [Wed, 24 Jun 2015 23:57:47 +0000 (16:57 -0700)]
mm: nommu: refactor debug and warning prints

kenter/kleave/kdebug are wrapper macros to print functions flow and debug
information.  This set was written before pr_devel() was introduced, so it
was controlled by "#if 0" construction.  It is questionable if anyone is
using them [1] now.

This patch removes these macros, converts numerous printk(KERN_WARNING,
...) to use general pr_warn(...) and removes debug print line from
validate_mmap_request() function.

Signed-off-by: Leon Romanovsky <leon@leon.nu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: clarify that the function operates on hugepage pte
Aneesh Kumar K.V [Wed, 24 Jun 2015 23:57:44 +0000 (16:57 -0700)]
mm: clarify that the function operates on hugepage pte

We have confusing functions to clear pmd, pmd_clear_* and pmd_clear.  Add
_huge_ to pmdp_clear functions so that we are clear that they operate on
hugepage pte.

We don't bother about other functions like pmdp_set_wrprotect,
pmdp_clear_flush_young, because they operate on PTE bits and hence
indicate they are operating on hugepage ptes

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agopowerpc/mm: use generic version of pmdp_clear_flush()
Aneesh Kumar K.V [Wed, 24 Jun 2015 23:57:42 +0000 (16:57 -0700)]
powerpc/mm: use generic version of pmdp_clear_flush()

Also move the pmd_trans_huge check to generic code.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/thp: split out pmd collapse flush into separate functions
Aneesh Kumar K.V [Wed, 24 Jun 2015 23:57:39 +0000 (16:57 -0700)]
mm/thp: split out pmd collapse flush into separate functions

Architectures like ppc64 [1] need to do special things while clearing pmd
before a collapse.  For them this operation is largely different from a
normal hugepage pte clear.  Hence add a separate function to clear pmd
before collapse.  After this patch pmdp_* functions operate only on
hugepage pte, and not on regular pmd_t values pointing to page table.

[1] ppc64 needs to invalidate all the normal page pte mappings we already
have inserted in the hardware hash page table.  But before doing that we
need to make sure there are no parallel hash page table insert going on.
So we need to do a kick_all_cpus_sync() before flushing the older hash
table entries.  By moving this to a separate function we capture these
details and mention how it is different from a hugepage pte clear.

This patch is a cleanup and only does code movement for clarity.  There
should not be any change in functionality.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agotracing: add trace event for memory-failure
Xie XiuQi [Wed, 24 Jun 2015 23:57:36 +0000 (16:57 -0700)]
tracing: add trace event for memory-failure

RAS user space tools like rasdaemon which base on trace event, could
receive mce error event, but no memory recovery result event.  So, I want
to add this event to make this scenario complete.

This patch add a event at ras group for memory-failure.

The output like below:
#  tracer: nop
#
#  entries-in-buffer/entries-written: 2/2   #P:24
#
#                               _-----=> irqs-off
#                              / _----=> need-resched
#                             | / _---=> hardirq/softirq
#                             || / _--=> preempt-depth
#                             ||| /     delay
#            TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION
#               | |       |   ||||       |         |
       mce-inject-13150 [001] ....   277.019359: memory_failure_event: pfn 0x19869: recovery action for free buddy page: Delayed

[xiexiuqi@huawei.com: fix build error]
Signed-off-by: Xie XiuQi <xiexiuqi@huawei.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Chen Gong <gong.chen@linux.intel.com>
Cc: Jim Davis <jim.epost@gmail.com>
Signed-off-by: Xie XiuQi <xiexiuqi@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomemory-failure: change type of action_result's param 3 to enum
Xie XiuQi [Wed, 24 Jun 2015 23:57:33 +0000 (16:57 -0700)]
memory-failure: change type of action_result's param 3 to enum

Change type of action_result's param 3 to enum for type consistency,
and rename mf_outcome to mf_result for clearly.

Signed-off-by: Xie XiuQi <xiexiuqi@huawei.com>
Acked-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Chen Gong <gong.chen@linux.intel.com>
Cc: Jim Davis <jim.epost@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomemory-failure: export page_type and action result
Xie XiuQi [Wed, 24 Jun 2015 23:57:30 +0000 (16:57 -0700)]
memory-failure: export page_type and action result

Export 'outcome' and 'action_page_type' to mm.h, so we could use
this emnus outside.

This patch is preparation for adding trace events for memory-failure
recovery action.

Signed-off-by: Xie XiuQi <xiexiuqi@huawei.com>
Acked-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Chen Gong <gong.chen@linux.intel.com>
Cc: Jim Davis <jim.epost@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm, memcg: Try charging a page before setting page up to date
Mel Gorman [Wed, 24 Jun 2015 23:57:27 +0000 (16:57 -0700)]
mm, memcg: Try charging a page before setting page up to date

Historically memcg overhead was high even if memcg was unused.  This has
improved a lot but it still showed up in a profile summary as being a
problem.

/usr/src/linux-4.0-vanilla/mm/memcontrol.c                           6.6441   395842
  mem_cgroup_try_charge                                                        2.950%   175781
  __mem_cgroup_count_vm_event                                                  1.431%    85239
  mem_cgroup_page_lruvec                                                       0.456%    27156
  mem_cgroup_commit_charge                                                     0.392%    23342
  uncharge_list                                                                0.323%    19256
  mem_cgroup_update_lru_size                                                   0.278%    16538
  memcg_check_events                                                           0.216%    12858
  mem_cgroup_charge_statistics.isra.22                                         0.188%    11172
  try_charge                                                                   0.150%     8928
  commit_charge                                                                0.141%     8388
  get_mem_cgroup_from_mm                                                       0.121%     7184

That is showing that 6.64% of system CPU cycles were in memcontrol.c and
dominated by mem_cgroup_try_charge.  The annotation shows that the bulk
of the cost was checking PageSwapCache which is expected to be cache hot
but is very expensive.  The problem appears to be that __SetPageUptodate
is called just before the check which is a write barrier.  It is
required to make sure struct page and page data is written before the
PTE is updated and the data visible to userspace.  memcg charging does
not require or need the barrier but gets unfairly hit with the cost so
this patch attempts the charging before the barrier.  Aside from the
accidental cost to memcg there is the added benefit that the barrier is
avoided if the page cannot be charged.  When applied the relevant
profile summary is as follows.

/usr/src/linux-4.0-chargefirst-v2r1/mm/memcontrol.c                  3.7907   223277
  __mem_cgroup_count_vm_event                                                  1.143%    67312
  mem_cgroup_page_lruvec                                                       0.465%    27403
  mem_cgroup_commit_charge                                                     0.381%    22452
  uncharge_list                                                                0.332%    19543
  mem_cgroup_update_lru_size                                                   0.284%    16704
  get_mem_cgroup_from_mm                                                       0.271%    15952
  mem_cgroup_try_charge                                                        0.237%    13982
  memcg_check_events                                                           0.222%    13058
  mem_cgroup_charge_statistics.isra.22                                         0.185%    10920
  commit_charge                                                                0.140%     8235
  try_charge                                                                   0.131%     7716

That brings the overhead down to 3.79% and leaves the memcg fault
accounting to the root cgroup but it's an improvement.  The difference
in headline performance of the page fault microbench is marginal as
memcg is such a small component of it.

pft faults
                                       4.0.0                  4.0.0
                                     vanilla            chargefirst
Hmean    faults/cpu-1 1443258.1051 (  0.00%) 1509075.7561 (  4.56%)
Hmean    faults/cpu-3 1340385.9270 (  0.00%) 1339160.7113 ( -0.09%)
Hmean    faults/cpu-5  875599.0222 (  0.00%)  874174.1255 ( -0.16%)
Hmean    faults/cpu-7  601146.6726 (  0.00%)  601370.9977 (  0.04%)
Hmean    faults/cpu-8  510728.2754 (  0.00%)  510598.8214 ( -0.03%)
Hmean    faults/sec-1 1432084.7845 (  0.00%) 1497935.5274 (  4.60%)
Hmean    faults/sec-3 3943818.1437 (  0.00%) 3941920.1520 ( -0.05%)
Hmean    faults/sec-5 3877573.5867 (  0.00%) 3869385.7553 ( -0.21%)
Hmean    faults/sec-7 3991832.0418 (  0.00%) 3992181.4189 (  0.01%)
Hmean    faults/sec-8 3987189.8167 (  0.00%) 3986452.2204 ( -0.02%)

It's only visible at single threaded. The overhead is there for higher
threads but other factors dominate.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agohugetlb: do not account hugetlb pages as NR_FILE_PAGES
Michal Hocko [Wed, 24 Jun 2015 23:57:24 +0000 (16:57 -0700)]
hugetlb: do not account hugetlb pages as NR_FILE_PAGES

hugetlb pages uses add_to_page_cache to track shared mappings.  This is
OK from the data structure point of view but it is less so from the
NR_FILE_PAGES accounting:

- huge pages are accounted as 4k which is clearly wrong
- this counter is used as the amount of the reclaimable page
  cache which is incorrect as well because hugetlb pages are
  special and not reclaimable
- the counter is then exported to userspace via /proc/meminfo
  (in Cached:), /proc/vmstat and /proc/zoneinfo as
  nr_file_pages which is confusing at least:
  Cached:          8883504 kB
  HugePages_Free:     8348
  ...
  Cached:          8916048 kB
  HugePages_Free:      156
  ...
  thats 8192 huge pages allocated which is ~16G accounted as 32M

There are usually not that many huge pages in the system for this to
make any visible difference e.g.  by fooling __vm_enough_memory or
zone_pagecache_reclaimable.

Fix this by special casing huge pages in both __delete_from_page_cache
and __add_to_page_cache_locked.  replace_page_cache_page is currently
only used by fuse and that shouldn't touch hugetlb pages AFAICS but it
is more robust to check for special casing there as well.

Hugetlb pages shouldn't get to any other paths where we do accounting:
- migration - we have a special handling via
  hugetlbfs_migrate_page
- shmem - doesn't handle hugetlb pages directly even for
  SHM_HUGETLB resp. MAP_HUGETLB
- swapcache - hugetlb is not swapable

This has a user visible effect but I believe it is reasonable because the
previously exported number is simply bogus.

An alternative would be to account hugetlb pages with their real size and
treat them similar to shmem.  But this has some drawbacks.

First we would have to special case in kernel users of NR_FILE_PAGES and
considering how hugetlb is special we would have to do it everywhere.  We
do not want Cached exported by /proc/meminfo to include it because the
value would be even more misleading.

__vm_enough_memory and zone_pagecache_reclaimable would have to do the
same thing because those pages are simply not reclaimable.  The correction
is even not trivial because we would have to consider all active hugetlb
page sizes properly.  Users of the counter outside of the kernel would
have to do the same.

So the question is why to account something that needs to be basically
excluded for each reasonable usage.  This doesn't make much sense to me.

It seems that this has been broken since hugetlb was introduced but I
haven't checked the whole history.

[akpm@linux-foundation.org: tweak comments]
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Mel Gorman <mgorman@suse.de>
Tested-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: page_alloc: inline should_alloc_retry()
Johannes Weiner [Wed, 24 Jun 2015 23:57:21 +0000 (16:57 -0700)]
mm: page_alloc: inline should_alloc_retry()

The should_alloc_retry() function was meant to encapsulate retry
conditions of the allocator slowpath, but there are still checks
remaining in the main function, and much of how the retrying is
performed also depends on the OOM killer progress.  The physical
separation of those conditions make the code hard to follow.

Inline the should_alloc_retry() checks.  Notes:

- The __GFP_NOFAIL check is already done in __alloc_pages_may_oom(),
  replace it with looping on OOM killer progress

- The pm_suspended_storage() check is meant to skip the OOM killer
  when reclaim has no IO available, move to __alloc_pages_may_oom()

- The order <= PAGE_ALLOC_COSTLY order is re-united with its original
  counterpart of checking whether reclaim actually made any progress

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: oom_kill: simplify OOM killer locking
Johannes Weiner [Wed, 24 Jun 2015 23:57:19 +0000 (16:57 -0700)]
mm: oom_kill: simplify OOM killer locking

The zonelist locking and the oom_sem are two overlapping locks that are
used to serialize global OOM killing against different things.

The historical zonelist locking serializes OOM kills from allocations with
overlapping zonelists against each other to prevent killing more tasks
than necessary in the same memory domain.  Only when neither tasklists nor
zonelists from two concurrent OOM kills overlap (tasks in separate memcgs
bound to separate nodes) are OOM kills allowed to execute in parallel.

The younger oom_sem is a read-write lock to serialize OOM killing against
the PM code trying to disable the OOM killer altogether.

However, the OOM killer is a fairly cold error path, there is really no
reason to optimize for highly performant and concurrent OOM kills.  And
the oom_sem is just flat-out redundant.

Replace both locking schemes with a single global mutex serializing OOM
kills regardless of context.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: oom_kill: remove unnecessary locking in exit_oom_victim()
Johannes Weiner [Wed, 24 Jun 2015 23:57:16 +0000 (16:57 -0700)]
mm: oom_kill: remove unnecessary locking in exit_oom_victim()

Disabling the OOM killer needs to exclude allocators from entering, not
existing victims from exiting.

Right now the only waiter is suspend code, which achieves quiescence by
disabling the OOM killer.  But later on we want to add waits that hold
the lock instead to stop new victims from showing up.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: oom_kill: generalize OOM progress waitqueue
Johannes Weiner [Wed, 24 Jun 2015 23:57:13 +0000 (16:57 -0700)]
mm: oom_kill: generalize OOM progress waitqueue

It turns out that the mechanism to wait for exiting OOM victims is less
generic than it looks: it won't issue wakeups unless the OOM killer is
disabled.

The reason this check was added was the thought that, since only the OOM
disabling code would wait on this queue, wakeup operations could be
saved when that specific consumer is known to be absent.

However, this is quite the handgrenade.  Later attempts to reuse the
waitqueue for other purposes will lead to completely unexpected bugs and
the failure mode will appear seemingly illogical.  Generally, providers
shouldn't make unnecessary assumptions about consumers.

This could have been replaced with waitqueue_active(), but it only saves
a few instructions in one of the coldest paths in the kernel.  Simply
remove it.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: oom_kill: switch test-and-clear of known TIF_MEMDIE to clear
Johannes Weiner [Wed, 24 Jun 2015 23:57:10 +0000 (16:57 -0700)]
mm: oom_kill: switch test-and-clear of known TIF_MEMDIE to clear

exit_oom_victim() already knows that TIF_MEMDIE is set, and nobody else
can clear it concurrently.  Use clear_thread_flag() directly.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: oom_kill: clean up victim marking and exiting interfaces
Johannes Weiner [Wed, 24 Jun 2015 23:57:07 +0000 (16:57 -0700)]
mm: oom_kill: clean up victim marking and exiting interfaces

Rename unmark_oom_victim() to exit_oom_victim().  Marking and unmarking
are related in functionality, but the interface is not symmetrical at
all: one is an internal OOM killer function used during the killing, the
other is for an OOM victim to signal its own death on exit later on.
This has locking implications, see follow-up changes.

While at it, rename mark_tsk_oom_victim() to mark_oom_victim(), which
is easier on the eye.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: oom_kill: remove unnecessary locking in oom_enable()
Johannes Weiner [Wed, 24 Jun 2015 23:57:04 +0000 (16:57 -0700)]
mm: oom_kill: remove unnecessary locking in oom_enable()

Setting oom_killer_disabled to false is atomic, there is no need for
further synchronization with ongoing allocations trying to OOM-kill.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memory hotplug: init the zone's size when calculating node totalpages
Gu Zheng [Wed, 24 Jun 2015 23:57:02 +0000 (16:57 -0700)]
mm/memory hotplug: init the zone's size when calculating node totalpages

Init the zone's size when calculating node totalpages to avoid duplicated
operations in free_area_init_core().

Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/hugetlb: introduce minimum hugepage order
Naoya Horiguchi [Wed, 24 Jun 2015 23:56:59 +0000 (16:56 -0700)]
mm/hugetlb: introduce minimum hugepage order

Currently the initial value of order in dissolve_free_huge_page is 64 or
32, which leads to the following warning in static checker:

  mm/hugetlb.c:1203 dissolve_free_huge_pages()
  warn: potential right shift more than type allows '9,18,64'

This is a potential risk of infinite loop, because 1 << order (== 0) is used
in for-loop like this:

  for (pfn =3D start_pfn; pfn < end_pfn; pfn +=3D 1 << order)
      ...

So this patch fixes it by using global minimum_order calculated at boot time.

    text    data     bss     dec     hex filename
   28313     469   84236  113018   1b97a mm/hugetlb.o
   28256     473   84236  112965   1b945 mm/hugetlb.o (patched)

Fixes: c8721bbbdd36 ("mm: memory-hotplug: enable memory hotplug to handle hugepage")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agormap: fix theoretical race between do_wp_page and shrink_active_list
Vladimir Davydov [Wed, 24 Jun 2015 23:56:56 +0000 (16:56 -0700)]
rmap: fix theoretical race between do_wp_page and shrink_active_list

As noted by Paul the compiler is free to store a temporary result in a
variable on stack, heap or global unless it is explicitly marked as
volatile, see:

  http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4455.html#sample-optimizations

This can result in a race between do_wp_page() and shrink_active_list()
as follows.

In do_wp_page() we can call page_move_anon_rmap(), which sets
page->mapping as follows:

  anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
  page->mapping = (struct address_space *) anon_vma;

The page in question may be on an LRU list, because nowhere in
do_wp_page() we remove it from the list, neither do we take any LRU
related locks.  Although the page is locked, shrink_active_list() can
still call page_referenced() on it concurrently, because the latter does
not require an anonymous page to be locked:

  CPU0                          CPU1
  ----                          ----
  do_wp_page                    shrink_active_list
   lock_page                     page_referenced
                                  PageAnon->yes, so skip trylock_page
   page_move_anon_rmap
    page->mapping = anon_vma
                                  rmap_walk
                                   PageAnon->no
                                   rmap_walk_file
                                    BUG
    page->mapping += PAGE_MAPPING_ANON

This patch fixes this race by explicitly forbidding the compiler to split
page->mapping store in page_move_anon_rmap() with the aid of WRITE_ONCE.

[akpm@linux-foundation.org: tweak comment, per Minchan]
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memory-failure: me_huge_page() does nothing for thp
Naoya Horiguchi [Wed, 24 Jun 2015 23:56:53 +0000 (16:56 -0700)]
mm/memory-failure: me_huge_page() does nothing for thp

memory_failure() is supposed not to handle thp itself, but to split it.
But if something were wrong and page_action() were called on thp,
me_huge_page() (action routine for hugepages) should be better to take
no action, rather than to take wrong action prepared for hugetlb (which
triggers BUG_ON().)

This change is for potential problems, but makes sense to me because thp
is an actively developing feature and this code path can be open in the
future.

Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm: soft-offline: don't free target page in successful page migration
Naoya Horiguchi [Wed, 24 Jun 2015 23:56:50 +0000 (16:56 -0700)]
mm: soft-offline: don't free target page in successful page migration

Stress testing showed that soft offline events for a process iterating
"mmap-pagefault-munmap" loop can trigger
VM_BUG_ON(PAGE_FLAGS_CHECK_AT_PREP) in __free_one_page():

  Soft offlining page 0x70fe1 at 0x70100008d000
  Soft offlining page 0x705fb at 0x70300008d000
  page:ffffea0001c3f840 count:0 mapcount:0 mapping:          (null) index:0x2
  flags: 0x1fffff80800000(hwpoison)
  page dumped because: VM_BUG_ON_PAGE(page->flags & ((1 << 25) - 1))
  ------------[ cut here ]------------
  kernel BUG at /src/linux-dev/mm/page_alloc.c:585!
  invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC
  Modules linked in: cfg80211 rfkill crc32c_intel microcode ppdev parport_pc pcspkr serio_raw virtio_balloon parport i2c_piix4 virtio_blk virtio_net ata_generic pata_acpi floppy
  CPU: 3 PID: 1779 Comm: test_base_madv_ Not tainted 4.0.0-v4.0-150511-1451-00009-g82360a3730e6 #139
  RIP: free_pcppages_bulk+0x52a/0x6f0
  Call Trace:
    drain_pages_zone+0x3d/0x50
    drain_local_pages+0x1d/0x30
    on_each_cpu_mask+0x46/0x80
    drain_all_pages+0x14b/0x1e0
    soft_offline_page+0x432/0x6e0
    SyS_madvise+0x73c/0x780
    system_call_fastpath+0x12/0x17
  Code: ff 89 45 b4 48 8b 45 c0 48 83 b8 a8 00 00 00 00 0f 85 e3 fb ff ff 0f 1f 00 0f 0b 48 8b 7d 90 48 c7 c6 e8 95 a6 81 e8 e6 32 02 00 <0f> 0b 8b 45 cc 49 89 47 30 41 8b 47 18 83 f8 ff 0f 85 10 ff ff
  RIP  [<ffffffff811a806a>] free_pcppages_bulk+0x52a/0x6f0
   RSP <ffff88007a117d28>
  ---[ end trace 53926436e76d1f35 ]---

When soft offline successfully migrates page, the source page is supposed
to be freed.  But there is a race condition where a source page looks
isolated (i.e.  the refcount is 0 and the PageHWPoison is set) but
somewhat linked to pcplist.  Then another soft offline event calls
drain_all_pages() and tries to free such hwpoisoned page, which is
forbidden.

This odd page state seems to happen due to the race between put_page() in
putback_lru_page() and __pagevec_lru_add_fn().  But I don't want to play
with tweaking drain code as done in commit 9ab3b598d2df "mm: hwpoison:
drop lru_add_drain_all() in __soft_offline_page()", or to change page
freeing code for this soft offline's purpose.

Instead, let's think about the difference between hard offline and soft
offline.  There is an interesting difference in how to isolate the in-use
page between these, that is, hard offline marks PageHWPoison of the target
page at first, and doesn't free it by keeping its refcount 1.  OTOH, soft
offline tries to free the target page then marks PageHWPoison.  This
difference might be the source of complexity and result in bugs like the
above.  So making soft offline isolate with keeping refcount can be a
solution for this problem.

We can pass to page migration code the "reason" which shows the caller, so
let's use this more to avoid calling putback_lru_page() when called from
soft offline, which effectively does the isolation for soft offline.  With
this change, target pages of soft offline never be reused without changing
migratetype, so this patch also removes the related code.

Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memory-failure: introduce get_hwpoison_page() for consistent refcount handling
Naoya Horiguchi [Wed, 24 Jun 2015 23:56:48 +0000 (16:56 -0700)]
mm/memory-failure: introduce get_hwpoison_page() for consistent refcount handling

memory_failure() can run in 2 different mode (specified by
MF_COUNT_INCREASED) in page refcount perspective.  When
MF_COUNT_INCREASED is set, memory_failure() assumes that the caller
takes a refcount of the target page.  And if cleared, memory_failure()
takes it in it's own.

In current code, however, refcounting is done differently in each caller.
For example, madvise_hwpoison() uses get_user_pages_fast() and
hwpoison_inject() uses get_page_unless_zero().  So this inconsistent
refcounting causes refcount failure especially for thp tail pages.
Typical user visible effects are like memory leak or
VM_BUG_ON_PAGE(!page_count(page)) in isolate_lru_page().

To fix this refcounting issue, this patch introduces get_hwpoison_page()
to handle thp tail pages in the same manner for each caller of hwpoison
code.

memory_failure() might fail to split thp and in such case it returns
without completing page isolation.  This is not good because PageHWPoison
on the thp is still set and there's no easy way to unpoison such thps.  So
this patch try to roll back any action to the thp in "non anonymous thp"
case and "thp split failed" case, expecting an MCE(SRAR) generated by
later access afterward will properly free such thps.

[akpm@linux-foundation.org: fix CONFIG_HWPOISON_INJECT=m]
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agomm/memory-failure: split thp earlier in memory error handling
Naoya Horiguchi [Wed, 24 Jun 2015 23:56:45 +0000 (16:56 -0700)]
mm/memory-failure: split thp earlier in memory error handling

memory_failure() doesn't handle thp itself at this time and need to split
it before doing isolation.  Currently thp is split in the middle of
hwpoison_user_mappings(), but there're corner cases where memory_failure()
wrongly tries to handle thp without splitting.

1) "non anonymous" thp, which is not a normal operating mode of thp,
   but a memory error could hit a thp before anon_vma is initialized.  In
   such case, split_huge_page() fails and me_huge_page() (intended for
   hugetlb) is called for thp, which triggers BUG_ON in page_hstate().

2) !PageLRU case, where hwpoison_user_mappings() returns with
   SWAP_SUCCESS and the result is the same as case 1.

memory_failure() can't avoid splitting, so let's split it more earlier,
which also reduces code which are prepared for both of normal page and
thp.

Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>