Message ID | 1308158600-7169-2-git-send-email-will.deacon@arm.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Wed, Jun 15, 2011 at 06:23:12PM +0100, Will Deacon wrote: > The l2x0_disable function attempts to writel with the l2x0_lock held. > This results in deadlock when the writel contains an outer_sync call > for the platform since the l2x0_lock is already held by the disable > function. A further problem is that disabling the L2 without flushing it > first can lead to the spin_lock operation becoming visible after the > spin_unlock, causing any subsequent L2 maintenance to deadlock. > > This patch replaces the writel with a call to writel_relaxed in the > disabling code and adds a flush before disabling in the control > register, preventing livelock from occurring. > > Signed-off-by: Will Deacon <will.deacon@arm.com> Acked-by: Catalin Marinas <catalin.marinas@arm.com>
diff --git a/arch/arm/mm/cache-l2x0.c b/arch/arm/mm/cache-l2x0.c index ef59099..44c0867 100644 --- a/arch/arm/mm/cache-l2x0.c +++ b/arch/arm/mm/cache-l2x0.c @@ -120,17 +120,22 @@ static void l2x0_cache_sync(void) spin_unlock_irqrestore(&l2x0_lock, flags); } -static void l2x0_flush_all(void) +static void __l2x0_flush_all(void) { - unsigned long flags; - - /* clean all ways */ - spin_lock_irqsave(&l2x0_lock, flags); debug_writel(0x03); writel_relaxed(l2x0_way_mask, l2x0_base + L2X0_CLEAN_INV_WAY); cache_wait_way(l2x0_base + L2X0_CLEAN_INV_WAY, l2x0_way_mask); cache_sync(); debug_writel(0x00); +} + +static void l2x0_flush_all(void) +{ + unsigned long flags; + + /* clean all ways */ + spin_lock_irqsave(&l2x0_lock, flags); + __l2x0_flush_all(); spin_unlock_irqrestore(&l2x0_lock, flags); } @@ -266,7 +271,9 @@ static void l2x0_disable(void) unsigned long flags; spin_lock_irqsave(&l2x0_lock, flags); - writel(0, l2x0_base + L2X0_CTRL); + __l2x0_flush_all(); + writel_relaxed(0, l2x0_base + L2X0_CTRL); + dsb(); spin_unlock_irqrestore(&l2x0_lock, flags); }
The l2x0_disable function attempts to writel with the l2x0_lock held. This results in deadlock when the writel contains an outer_sync call for the platform since the l2x0_lock is already held by the disable function. A further problem is that disabling the L2 without flushing it first can lead to the spin_lock operation becoming visible after the spin_unlock, causing any subsequent L2 maintenance to deadlock. This patch replaces the writel with a call to writel_relaxed in the disabling code and adds a flush before disabling in the control register, preventing livelock from occurring. Signed-off-by: Will Deacon <will.deacon@arm.com> --- arch/arm/mm/cache-l2x0.c | 19 +++++++++++++------ 1 files changed, 13 insertions(+), 6 deletions(-)