We need to use post-decrement to get the dma_map_page undone also for
i==0, and to avoid some very unpleasant behaviour if dma_map_page
failed already at i==0.
Signed-off-by: NRasmus Villemoes <linux@rasmusvillemoes.dk>
Reviewed-by: NBen Skeggs <bskeggs@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Because we record connector_mask using 1 << drm_connector_index now
the connector_mask should stay the same even when other connectors
are removed. This was not the case with MST, in that case when removing
a connector all other connectors may change their index.

This is fixed by waiting until the first get_connector_state to allocate
connector_state, and force reallocation when state is too small.

As a side effect connector arrays no longer have to be preallocated,
and can be allocated on first use which means a less allocations in
the page flip only path.

Changes since v1:
- Whitespace. (Ville)
- Call ida_remove when destroying the connector. (Ville)
- u32 alloc -> int. (Ville)

Fixes: 14de6c44 ("drm/atomic: Remove drm_atomic_connectors_for_crtc.")
Signed-off-by: NMaarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NLyude <cpaul@redhat.com>
Reviewed-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Looks like g4x hpd live status bits actually agree with the spec. At
least they do on the machine I have, and apparently on Nick Bowler's
g4x as well.

So gm45 may be the only platform where they don't agree. At least
that seems to be the case based on the (somewhat incomplete)
logs/dumps in [1], and Daniel has also tested this on his gm45
sometime in the past.

So let's change the bits to match the spec on g4x. That actually makes
the g4x bits identical to vlv/chv so we can just share the code
between those platforms, leaving gm45 as the special case.

[1] https://bugzilla.kernel.org/show_bug.cgi?id=52361

Cc: Shashank Sharma <shashank.sharma@intel.com>
Cc: Sonika Jindal <sonika.jindal@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Nick Bowler <nbowler@draconx.ca>
References: https://lists.freedesktop.org/archives/dri-devel/2016-February/100382.htmlReported-by: NNick Bowler <nbowler@draconx.ca>
Cc: stable@vger.kernel.org
Fixes: 237ed86c ("drm/i915: Check live status before reading edid")
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1455127145-20087-1-git-send-email-ville.syrjala@linux.intel.comReviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
(cherry picked from commit 0780cd36)
Signed-off-by: NJani Nikula <jani.nikula@intel.com>

Commit ca369d51 ("sd: Fix device-imposed transfer length limits")
introduced a new queue limit max_dev_sectors which limits the maximum
sectors for requests. The default value leads to small dasd requests
and therefor to a performance drop.
Set the max_dev_sectors value to the same value as the max_hw_sectors
to use the maximum available request size for DASD devices.
Signed-off-by: NStefan Haberland <sth@linux.vnet.ibm.com>
Cc: stable@vger.kernel.org # 4.4+
Signed-off-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>

The qxl_gem_prime_mmap() function returns ENOSYS instead of -ENOSYS
Signed-off-by: NAnton Protopopov <a.s.protopopov@gmail.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

In the display resume path, move the calls to drm_vblank_on()
after the point when the display engine is running again.

Since changes were made to drm_update_vblank_count() in Linux 4.4+
to emulate hw vblank counters via vblank timestamping, the function
drm_vblank_on() now needs working high precision vblank timestamping
and therefore working scanout position queries at time of call.
These don't work before the display engine gets restarted, causing
miscalculation of vblank counter increments and thereby large forward
jumps in vblank count at display resume. These jumps can cause client
hangs on resume, or desktop hangs in the case of composited desktops.

Fix this Linux 4.4 regression by reordering calls accordingly.
Signed-off-by: NMario Kleiner <mario.kleiner.de@gmail.com>
Cc: <stable@vger.kernel.org> # 4.4+
Cc: Ben Skeggs <bskeggs@redhat.com>
Cc: ville.syrjala@linux.intel.com
Cc: daniel.vetter@ffwll.ch
Cc: dri-devel@lists.freedesktop.org
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Make sure that drm_vblank_get/put() stay balanced in
case drm_vblank_get fails, by skipping the corresponding
put.
Signed-off-by: NMario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Cc: michel@daenzer.net
Cc: dri-devel@lists.freedesktop.org
Cc: alexander.deucher@amd.com
Cc: christian.koenig@amd.com
Signed-off-by: NDave Airlie <airlied@redhat.com>

drm_vblank_offdelay can have three different types of values:

< 0 is to be always treated the same as dev->vblank_disable_immediate
= 0 is to be treated as "never disable vblanks"
> 0 is to be treated as disable immediate if kms driver wants it
    that way via dev->vblank_disable_immediate. Otherwise it is
    a disable timeout in msecs.

This got broken in Linux 3.18+ for the implementation of
drm_vblank_on. If the user specified a value of zero which should
always reenable vblank irqs in this function, a kms driver could
override the users choice by setting vblank_disable_immediate
to true. This patch fixes the regression and keeps the user in
control.

v2: Only reenable vblank if there are clients left or the user
    requested to "never disable vblanks" via offdelay 0. Enabling
    vblanks even in the "delayed disable" case (offdelay > 0) was
    specifically added by Ville in commit cd19e52a
    ("drm: Kick start vblank interrupts at drm_vblank_on()"),
    but after discussion it turns out that this was done by accident.

    Citing Ville: "I think it just ended up as a mess due to changing
    some of the semantics of offdelay<0 vs. offdelay==0 vs.
    disable_immediate during the review of the series. So yeah, given
    how drm_vblank_put() works now, I'd just make this check for
    offdelay==0."
Signed-off-by: NMario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Cc: <stable@vger.kernel.org> # 3.18+
Cc: michel@daenzer.net
Cc: vbabka@suse.cz
Cc: ville.syrjala@linux.intel.com
Cc: daniel.vetter@ffwll.ch
Cc: dri-devel@lists.freedesktop.org
Cc: alexander.deucher@amd.com
Cc: christian.koenig@amd.com
Signed-off-by: NDave Airlie <airlied@redhat.com>

Changes to drm_update_vblank_count() in Linux 4.4 broke the
behaviour of the pre/post modeset functions as the new update
code doesn't deal with hw vblank counter resets inbetween calls
to drm_vblank_pre_modeset an drm_vblank_post_modeset, as it
should.

This causes mistreatment of such hw counter resets as counter
wraparound, and thereby large forward jumps of the software
vblank counter which in turn cause vblank event dispatching
and vblank waits to fail/hang --> userspace clients hang.

This symptom was reported on radeon-kms to cause a infinite
hang of KDE Plasma 5 shell's login procedure, preventing users
from logging in.

Fix this by detecting when drm_update_vblank_count() is called
inside a pre->post modeset interval. If so, clamp valid vblank
increments to the safe values 0 and 1, pretty much restoring
the update behavior of the old update code of Linux 4.3 and
earlier. Also reset the last recorded hw vblank count at call
to drm_vblank_post_modeset() to be safe against hw that after
modesetting, dpms on etc. only fires its first vblank irq after
drm_vblank_post_modeset() was already called.
Reported-by: NVlastimil Babka <vbabka@suse.cz>
Signed-off-by: NMario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Tested-by: NVlastimil Babka <vbabka@suse.cz>

Cc: <stable@vger.kernel.org> # 4.4+
Cc: michel@daenzer.net
Cc: vbabka@suse.cz
Cc: ville.syrjala@linux.intel.com
Cc: daniel.vetter@ffwll.ch
Cc: dri-devel@lists.freedesktop.org
Cc: alexander.deucher@amd.com
Cc: christian.koenig@amd.com
Signed-off-by: NDave Airlie <airlied@redhat.com>

This fixes a regression introduced by the new drm_update_vblank_count()
implementation in Linux 4.4:

Restrict the bump of the software vblank counter in drm_update_vblank_count()
to a safe maximum value of +1 whenever there is the possibility that
concurrent readers of vblank timestamps could be active at the moment,
as the current implementation of the timestamp caching and updating is
not safe against concurrent readers for calls to store_vblank() with a
bump of anything but +1. A bump != 1 would very likely return corrupted
timestamps to userspace, because the same slot in the cache could
be concurrently written by store_vblank() and read by one of those
readers in a non-atomic fashion and without the read-retry logic
detecting this collision.

Concurrent readers can exist while drm_update_vblank_count() is called
from the drm_vblank_off() or drm_vblank_on() functions or other non-vblank-
irq callers. However, all those calls are happening with the vbl_lock
locked thereby preventing a drm_vblank_get(), so the vblank refcount
can't increase while drm_update_vblank_count() is executing. Therefore
a zero vblank refcount during execution of that function signals that
is safe for arbitrary counter bumps if called from outside vblank irq,
whereas a non-zero count is not safe.

Whenever the function is called from vblank irq, we have to assume concurrent
readers could show up any time during its execution, even if the refcount
is currently zero, as vblank irqs are usually only enabled due to the
presence of readers, and because when it is called from vblank irq it
can't hold the vbl_lock to protect it from sudden bumps in vblank refcount.
Therefore also restrict bumps to +1 when the function is called from vblank
irq.

Such bumps of more than +1 can happen at other times than reenabling
vblank irqs, e.g., when regular vblank interrupts get delayed by more
than 1 frame due to long held locks, long irq off periods, realtime
preemption on RT kernels, or system management interrupts.

A better solution would be to rewrite the timestamp caching to use
full seqlocks to allow concurrent writes and reads for arbitrary
vblank counter increments.

v2: Add code comment that this is essentially a hack and should
    be replaced by a full seqlock implementation for caching of
    timestamps.
Signed-off-by: NMario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Cc: <stable@vger.kernel.org> # 4.4+
Cc: michel@daenzer.net
Cc: vbabka@suse.cz
Cc: ville.syrjala@linux.intel.com
Cc: daniel.vetter@ffwll.ch
Cc: dri-devel@lists.freedesktop.org
Cc: alexander.deucher@amd.com
Cc: christian.koenig@amd.com
Signed-off-by: NDave Airlie <airlied@redhat.com>

Otherwise if a kms driver calls into drm_vblank_off() more than once
before calling drm_vblank_on() again, the redundant calls to
vblank_disable_and_save() will call drm_update_vblank_count()
while hw vblank counters and vblank timestamping are in a undefined
state during modesets, dpms off etc.

At least with the legacy drm helpers it is not unusual to
get multiple calls to drm_vblank_off and drm_vblank_on, e.g.,
half a dozen calls to drm_vblank_off and two calls to drm_vblank_on
were observed on radeon-kms during dpms-off -> dpms-on transition.

We don't no-op calls from atomic modesetting drivers, as they
should do a proper job of tracking hw state.

Fixes large jumps of the software maintained vblank counter due to
the hardware vblank counter resetting to zero during dpms off or
modeset, e.g., if radeon-kms is modified to use drm_vblank_off/on
instead of drm_vblank_pre/post_modeset().

This fixes a regression caused by the changes made to
drm_update_vblank_count() in Linux 4.4.

v2: Don't no-op on atomic modesetting drivers, per suggestion
    of Daniel Vetter.
Signed-off-by: NMario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Cc: <stable@vger.kernel.org> # 4.4+
Cc: michel@daenzer.net
Cc: vbabka@suse.cz
Cc: ville.syrjala@linux.intel.com
Cc: alexander.deucher@amd.com
Cc: christian.koenig@amd.com
Signed-off-by: NDave Airlie <airlied@redhat.com>

This avoids integer overflows on 32bit machines when calculating
reloc_info size, as reported by Alan Cox.

Cc: stable@vger.kernel.org
Cc: gnomes@lxorguk.ukuu.org.uk
Signed-off-by: NGerd Hoffmann <kraxel@redhat.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NDave Airlie <airlied@redhat.com>

This reverts commit cfcfa086.

This causes the tiling properties to break in some unexpected ways,

Revert it for now.

Cc: stable@vger.kernel.org
Signed-off-by: NDave Airlie <airlied@redhat.com>

This gets us functional GPU reset again, like we had until a refactor
at merge time.  Tested with a little patch to stuff in a broken binner
job every 100 frames.
Signed-off-by: NEric Anholt <eric@anholt.net>

This may actually get us a feature that the closed driver didn't have:
turning off the GPU in between rendering jobs, while the V3D device is
still opened by the client.

There may be some tuning to be applied here to use autosuspend so that
we don't bounce the device's power so much, but in steady-state
GPU-bound rendering we keep the power on (since we keep multiple jobs
outstanding) and even if we power cycle on every job we can still
manage at least 680 fps.

More importantly, though, runtime PM will allow us to power off the
device to do a GPU reset.

v2: Switch #ifdef to CONFIG_PM not CONFIG_PM_SLEEP (caught by kbuild
    test robot)
Signed-off-by: NEric Anholt <eric@anholt.net>

We were tracking the "where are the head pointers pointing" globally,
so if another job reused the same BOs and execution was at the same
point as last time we checked, we'd stop and trigger a reset even
though the GPU had made progress.
Signed-off-by: NEric Anholt <eric@anholt.net>

These ioctls end up getting exposed to fairly directly to GL users,
and having normal user operations print DRM errors is obviously wrong.
The message was originally to give us some idea of what happened when
a hang occurred, but we have a DRM_INFO from reset for that.
Signed-off-by: NEric Anholt <eric@anholt.net>

This caused the wait ioctls to claim that waiting had completed when
we actually got interrupted by a signal before it was done.  Fixes
broken rendering throttling that produced serious lag in X window
dragging.
Signed-off-by: NEric Anholt <eric@anholt.net>

Fixes igt vc4_create_bo/create-bo-0 by returning -EINVAL from the
ioctl instead of -ENOMEM.
Signed-off-by: NEric Anholt <eric@anholt.net>

Apparently in hardware (as opposed to simulation), the clear colors
need to be uploaded before the render config, otherwise they won't
take effect.  Fixes igt's vc4_wait_bo/used-bo-* subtests.
Signed-off-by: NEric Anholt <eric@anholt.net>

This is ABI future-proofing if we ever want to extend the pad to mean
something.
Signed-off-by: NEric Anholt <eric@anholt.net>

We need to use post-decrement to get the pci_map_page undone also for
i==0, and to avoid some very unpleasant behaviour if pci_map_page
failed already at i==0.
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NRasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

We need to use post-decrement to get the pci_map_page undone also for
i==0, and to avoid some very unpleasant behaviour if pci_map_page
failed already at i==0.
Reviewed-by: NChristian König <christian.koenig@amd.com>
Signed-off-by: NRasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

Previously, samsung_gpio_drection_in/output function were not covered
with a spinlock.

For example, samsung_gpio_direction_output function consists of
two functions.
1. samsung_gpio_set
2. samsung_gpio_set_direction

When 2 CPUs try to control the same gpio pin heavily,
(situation like i2c control with gpio emulation)
This situation can cause below problem.

CPU 0                                   | CPU1
                                        |
samsung_gpio_direction_output           |
   samsung_gpio_set(pin A as 1)         | samsung_gpio_direction_output
                                        |    samsung_gpio_set(pin A as 0)
   samsung_gpio_set_direction           |
                                        |    samsung_gpio_set_direction

The initial value of pin A will be set as 0 while we wanted to set pin A as 1.

This patch modifies samsung_gpio_direction_in/output function
to be done in one spinlock to fix race condition.

Additionally, the new samsung_gpio_set_value was added to implement
gpio set callback(samsung_gpio_set) with spinlock using this function.

Cc: stable@vger.kernel.org
Signed-off-by: NYoungmin Nam <ym0914@gmail.com>
Acked-by: NTomasz Figa <tomasz.figa@gmail.com>
Signed-off-by: NLinus Walleij <linus.walleij@linaro.org>

Decon requires that clocks should be disabled in reverse order. Otherwise
system hangs.
Signed-off-by: NAndrzej Hajda <a.hajda@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

dma_mmap_attrs() should be called with cpu address returned by
dma_alloc_attrs(). Existing code however passed pages array base as cpu
address. This worked only by a pure luck on ARM architecture. This patch
fixes this issue.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

BIT_IRQS_ENABLED was never set because of incorrect test in
decon_vlank_enable() function, what resulted in lack of enabling vblank
support. This patch fixes this issue.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

Patch ebf3fd40 ("drm/exynos: add
pm_runtime to DECON 5433") removed some code from decon_enable()
function, but it left set_bit(BIT_SUSPENDED, &ctx->flags) call, which
was earlier called only in error path. This patch removes it, what
finally lets driver to go out of suspended state.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

This patch fixes issue introduced by commit
cf67cc9a ("drm/exynos: remove struct
exynos_drm_display"), which removed assigning of drm bridge to drm
encoder. Lack of it caused that no bridge callbacks were called on
encoder enable/disable actions.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

There is no point exposing all internal functions to global kernel name
space, so make all internals functions static.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

MIC is SoC component and important part of kms pipeline on Exynos5433,
so convert it to use component framework like other KMS/CRTC drivers.
MIC driver is already listed on KMS component driver list in Exynos DRM
core, so without this conversion, initialization of Exynos DRM core
fails on Exynos 5433 SoC.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

Drivers should use devm_clk* interface instead of of_clk* functions.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

This patch fixes compilation warnings (on 64bit architectures) and bugs
related to casting pointers through 32bit integers.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

Drivers should use %p for printing pointers instead of hardcoding them
as hexadecimal integers. This patch fixes compilation warnings on 64bit
architectures.
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

Because PLAT_SAMSUNG isn't include exynos SoCs for arm64, but
ARCH_EXYNOS can do it. And it also needs to add ARCH_S3C64XX instead of
PLAT_SAMSUNG.
Signed-off-by: NJoonyoung Shim <jy0922.shim@samsung.com>
Signed-off-by: NMarek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: NInki Dae <inki.dae@samsung.com>

According to the VT-d specification we need to clear the PPR bit in
the Page Request Status register when handling page requests, or the
hardware won't generate any more interrupts.

This wasn't actually necessary on SKL/KBL (which may well be the
subject of a hardware erratum, although it's harmless enough). But
other implementations do appear to get it right, and we only ever get
one interrupt unless we clear the PPR bit.
Reported-by: NCQ Tang <cq.tang@intel.com>
Signed-off-by: NDavid Woodhouse <David.Woodhouse@intel.com>
Cc: stable@vger.kernel.org

When assigning mpp settings from static mpp modes to mpp groups,
we do not want any groups that have no supported setting for a
specific Kirkwood variant. However, when there is at least a
single supported setting, we need to assign the number of all
settings in this mode to grp->num_settings as we are reusing
the static modes table.

Fixes: 0581b16b ("pinctrl: mvebu: complain about missing group after checking variant")
Reported-by: NAaro Koskinen <aaro.koskinen@iki.fi>
Tested-by: NAaro Koskinen <aaro.koskinen@iki.fi>
Cc: Gregory Clement <gregory.clement@free-electrons.com>
Cc: Andrew Lunn <andrew@lunn.ch>
Cc: linux-arm-kernel@lists.infradead.org
Signed-off-by: NSebastian Hesselbarth <sebastian.hesselbarth@gmail.com>
Signed-off-by: NLinus Walleij <linus.walleij@linaro.org>

Commit 5de85b9d ("PM / runtime: Re-init runtime PM states at probe
error and driver unbind") introduced pm_runtime_reinit() that is used
to reinitialize PM runtime after -EPROBE_DEFER. This allows shutting
down the device after a failed probe.

However, for drivers using pm_runtime_use_autosuspend() this can cause
a state where suspend callback is never called after -EPROBE_DEFER.
On the following device driver probe, hardware state is different from
the PM runtime state causing omap_device to produce the following
error:

omap_device_enable() called from invalid state 1

And with omap_device and omap hardware being picky for PM, this will
block any deeper idle states in hardware.

The solution is to fix the drivers to follow the PM runtime documentation:

1. For sections of code that needs the device disabled, use
   pm_runtime_put_sync_suspend() if pm_runtime_set_autosuspend() has
   been set.

2. For driver exit code, use pm_runtime_dont_use_autosuspend() before
   pm_runtime_put_sync() if pm_runtime_use_autosuspend() has been
   set.

Fixes: 5de85b9d ("PM / runtime: Re-init runtime PM states at probe
error and driver unbind")
Cc: linux-spi@vger.kernel.org
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Kevin Hilman <khilman@baylibre.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Nishanth Menon <nm@ti.com>
Cc: Rafael J. Wysocki <rafael@kernel.org>
Cc: Ulf Hansson <ulf.hansson@linaro.org>
Cc: Tero Kristo <t-kristo@ti.com>
Signed-off-by: NTony Lindgren <tony@atomide.com>
Signed-off-by: NMark Brown <broonie@kernel.org>

Fix the RC QPs send queue overhead computation to take into account
two additional segments in the WQE which are needed for registration
operations.

The ATOMIC and UMR segments can't coexist together, so chose maximum out
of them.

The commit 9e65dc37 ("IB/mlx5: Fix RC transport send queue overhead
computation") was intended to update RC transport as commit messages
states, but added the code to UC transport.

Fixes: 9e65dc37 ("IB/mlx5: Fix RC transport send queue overhead computation")
Signed-off-by: NKamal Heib <kamalh@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

A narrow window for race condition still exist between
multicast join thread and *dev_flush workers.
A kernel crash caused by prolong erratic link state changes
was observed (most likely a faulty cabling):

[167275.656270] BUG: unable to handle kernel NULL pointer dereference at
0000000000000020
[167275.665973] IP: [<ffffffffa05f8f2e>] ipoib_mcast_join+0xae/0x1d0 [ib_ipoib]
[167275.674443] PGD 0
[167275.677373] Oops: 0000 [#1] SMP
...
[167275.977530] Call Trace:
[167275.982225]  [<ffffffffa05f92f0>] ? ipoib_mcast_free+0x200/0x200 [ib_ipoib]
[167275.992024]  [<ffffffffa05fa1b7>] ipoib_mcast_join_task+0x2a7/0x490
[ib_ipoib]
[167276.002149]  [<ffffffff8109d5fb>] process_one_work+0x17b/0x470
[167276.010754]  [<ffffffff8109e3cb>] worker_thread+0x11b/0x400
[167276.019088]  [<ffffffff8109e2b0>] ? rescuer_thread+0x400/0x400
[167276.027737]  [<ffffffff810a5aef>] kthread+0xcf/0xe0
Here was a hit spot:
ipoib_mcast_join() {
..............
      rec.qkey      = priv->broadcast->mcmember.qkey;
                                       ^^^^^^^
.....
 }
Proposed patch should prevent multicast join task to continue
if link state change is detected.
Signed-off-by: NAlex Estrin <alex.estrin@intel.com>

Changes from v4:
- as suggested by Doug Ledford, optimized spinlock usage,
i.e. ipoib_mcast_join() is called with lock held.
Changes from v3:
- sync with priv->lock before flag check.
Chages from v2:
- Move check for OPER_UP flag state to mcast_join() to
ensure no event worker is in progress.
- minor style fixes.
Changes from v1:
- No need to lock again if error detected.
Signed-off-by: NDoug Ledford <dledford@redhat.com>