There is an ordering issue with spin_unlock_wait() on powerpc, because
the spin_lock primitive is an ACQUIRE and an ACQUIRE is only ordering
the load part of the operation with memory operations following it.
Therefore the following event sequence can happen:
CPU 1 CPU 2 CPU 3
================== ==================== ==============
spin_unlock(&lock);
spin_lock(&lock):
r1 = *lock; // r1 == 0;
o = object; o = READ_ONCE(object); // reordered here
object = NULL;
smp_mb();
spin_unlock_wait(&lock);
*lock = 1;
smp_mb();
o->dead = true; < o = READ_ONCE(object); > // reordered upwards
if (o) // true
BUG_ON(o->dead); // true!!
To fix this, we add a "nop" ll/sc loop in arch_spin_unlock_wait() on
ppc, the "nop" ll/sc loop reads the lock
value and writes it back atomically, in this way it will synchronize the
view of the lock on CPU1 with that on CPU2. Therefore in the scenario
above, either CPU2 will fail to get the lock at first or CPU1 will see
the lock acquired by CPU2, both cases will eliminate this bug. This is a
similar idea as what Will Deacon did for ARM64 in:
d86b8da04d ("arm64: spinlock: serialise spin_unlock_wait against concurrent lockers")
Furthermore, if the "nop" ll/sc figures out the lock is locked, we
actually don't need to do the "nop" ll/sc trick again, we can just do a
normal load+check loop for the lock to be released, because in that
case, spin_unlock_wait() is called when someone is holding the lock, and
the store part of the "nop" ll/sc happens before the lock release of the
current lock holder:
"nop" ll/sc -> spin_unlock()
and the lock release happens before the next lock acquisition:
spin_unlock() -> spin_lock() <next holder>
which means the "nop" ll/sc happens before the next lock acquisition:
"nop" ll/sc -> spin_unlock() -> spin_lock() <next holder>
With a smp_mb() preceding spin_unlock_wait(), the store of object is
guaranteed to be observed by the next lock holder:
STORE -> smp_mb() -> "nop" ll/sc
-> spin_unlock() -> spin_lock() <next holder>
This patch therefore fixes the issue and also cleans the
arch_spin_unlock_wait() a little bit by removing superfluous memory
barriers in loops and consolidating the implementations for PPC32 and
PPC64 into one.
Suggested-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Reviewed-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
[mpe: Inline the "nop" ll/sc loop and set EH=0, munge change log]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
RCU is the only thing that uses smp_mb__after_unlock_lock(), and is
likely the only thing that ever will use it, so this commit makes this
macro private to RCU.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: "linux-arch@vger.kernel.org" <linux-arch@vger.kernel.org>
The kernel defines the function spin_is_locked(), which can be used to
check if a spinlock is currently locked.
Using spin_is_locked() on a lock you don't hold is obviously racy. That
is, even though you may observe that the lock is unlocked, it may become
locked at any time.
There is (at least) one exception to that, which is if two locks are
used as a pair, and the holder of each checks the status of the other
before doing any update.
Assuming *A and *B are two locks, and *COUNTER is a shared non-atomic
value:
The first CPU does:
spin_lock(*A)
if spin_is_locked(*B)
# nothing
else
smp_mb()
LOAD r = *COUNTER
r++
STORE *COUNTER = r
spin_unlock(*A)
And the second CPU does:
spin_lock(*B)
if spin_is_locked(*A)
# nothing
else
smp_mb()
LOAD r = *COUNTER
r++
STORE *COUNTER = r
spin_unlock(*B)
Although this is a strange locking construct, it should work.
It seems to be understood, but not documented, that spin_is_locked() is
not a memory barrier, so in the examples above and below the caller
inserts its own memory barrier before acting on the result of
spin_is_locked().
For now we assume spin_is_locked() is implemented as below, and we break
it out in our examples:
bool spin_is_locked(*LOCK) {
LOAD l = *LOCK
return l.locked
}
Our intuition is that there should be no problem even if the two code
sequences run simultaneously such as:
CPU 0 CPU 1
==================================================
spin_lock(*A) spin_lock(*B)
LOAD b = *B LOAD a = *A
if b.locked # true if a.locked # true
# nothing # nothing
spin_unlock(*A) spin_unlock(*B)
If one CPU gets the lock before the other then it will do the update and
the other CPU will back off:
CPU 0 CPU 1
==================================================
spin_lock(*A)
LOAD b = *B
spin_lock(*B)
if b.locked # false LOAD a = *A
else if a.locked # true
smp_mb() # nothing
LOAD r1 = *COUNTER spin_unlock(*B)
r1++
STORE *COUNTER = r1
spin_unlock(*A)
However in reality spin_lock() itself is not indivisible. On powerpc we
implement it as a load-and-reserve and store-conditional.
Ignoring the retry logic for the lost reservation case, it boils down to:
spin_lock(*LOCK) {
LOAD l = *LOCK
l.locked = true
STORE *LOCK = l
ACQUIRE_BARRIER
}
The ACQUIRE_BARRIER is required to give spin_lock() ACQUIRE semantics as
defined in memory-barriers.txt:
This acts as a one-way permeable barrier. It guarantees that all
memory operations after the ACQUIRE operation will appear to happen
after the ACQUIRE operation with respect to the other components of
the system.
On modern powerpc systems we use lwsync for ACQUIRE_BARRIER. lwsync is
also know as "lightweight sync", or "sync 1".
As described in Power ISA v2.07 section B.2.1.1, in this scenario the
lwsync is not the barrier itself. It instead causes the LOAD of *LOCK to
act as the barrier, preventing any loads or stores in the locked region
from occurring prior to the load of *LOCK.
Whether this behaviour is in accordance with the definition of ACQUIRE
semantics in memory-barriers.txt is open to discussion, we may switch to
a different barrier in future.
What this means in practice is that the following can occur:
CPU 0 CPU 1
==================================================
LOAD a = *A LOAD b = *B
a.locked = true b.locked = true
LOAD b = *B LOAD a = *A
STORE *A = a STORE *B = b
if b.locked # false if a.locked # false
else else
smp_mb() smp_mb()
LOAD r1 = *COUNTER LOAD r2 = *COUNTER
r1++ r2++
STORE *COUNTER = r1
STORE *COUNTER = r2 # Lost update
spin_unlock(*A) spin_unlock(*B)
That is, the load of *B can occur prior to the store that makes *A
visibly locked. And similarly for CPU 1. The result is both CPUs hold
their lock and believe the other lock is unlocked.
The easiest fix for this is to add a full memory barrier to the start of
spin_is_locked(), so adding to our previous definition would give us:
bool spin_is_locked(*LOCK) {
smp_mb()
LOAD l = *LOCK
return l.locked
}
The new barrier orders the store to the lock we are locking vs the load
of the other lock:
CPU 0 CPU 1
==================================================
LOAD a = *A LOAD b = *B
a.locked = true b.locked = true
STORE *A = a STORE *B = b
smp_mb() smp_mb()
LOAD b = *B LOAD a = *A
if b.locked # true if a.locked # true
# nothing # nothing
spin_unlock(*A) spin_unlock(*B)
Although the above example is theoretical, there is code similar to this
example in sem_lock() in ipc/sem.c. This commit in addition to the next
commit appears to be a fix for crashes we are seeing in that code where
we believe this race happens in practice.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Pull powerpc updates from Ben Herrenschmidt:
"So here's my next branch for powerpc. A bit late as I was on vacation
last week. It's mostly the same stuff that was in next already, I
just added two patches today which are the wiring up of lockref for
powerpc, which for some reason fell through the cracks last time and
is trivial.
The highlights are, in addition to a bunch of bug fixes:
- Reworked Machine Check handling on kernels running without a
hypervisor (or acting as a hypervisor). Provides hooks to handle
some errors in real mode such as TLB errors, handle SLB errors,
etc...
- Support for retrieving memory error information from the service
processor on IBM servers running without a hypervisor and routing
them to the memory poison infrastructure.
- _PAGE_NUMA support on server processors
- 32-bit BookE relocatable kernel support
- FSL e6500 hardware tablewalk support
- A bunch of new/revived board support
- FSL e6500 deeper idle states and altivec powerdown support
You'll notice a generic mm change here, it has been acked by the
relevant authorities and is a pre-req for our _PAGE_NUMA support"
* 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/benh/powerpc: (121 commits)
powerpc: Implement arch_spin_is_locked() using arch_spin_value_unlocked()
powerpc: Add support for the optimised lockref implementation
powerpc/powernv: Call OPAL sync before kexec'ing
powerpc/eeh: Escalate error on non-existing PE
powerpc/eeh: Handle multiple EEH errors
powerpc: Fix transactional FP/VMX/VSX unavailable handlers
powerpc: Don't corrupt transactional state when using FP/VMX in kernel
powerpc: Reclaim two unused thread_info flag bits
powerpc: Fix races with irq_work
Move precessing of MCE queued event out from syscall exit path.
pseries/cpuidle: Remove redundant call to ppc64_runlatch_off() in cpu idle routines
powerpc: Make add_system_ram_resources() __init
powerpc: add SATA_MV to ppc64_defconfig
powerpc/powernv: Increase candidate fw image size
powerpc: Add debug checks to catch invalid cpu-to-node mappings
powerpc: Fix the setup of CPU-to-Node mappings during CPU online
powerpc/iommu: Don't detach device without IOMMU group
powerpc/eeh: Hotplug improvement
powerpc/eeh: Call opal_pci_reinit() on powernv for restoring config space
powerpc/eeh: Add restore_config operation
...
At a glance these are just the inverse of each other. The one subtlety
is that arch_spin_value_unlocked() takes the lock by value, rather than
as a pointer, which is important for the lockref code.
On the other hand arch_spin_is_locked() doesn't really care, so
implement it in terms of arch_spin_value_unlocked().
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This commit adds the architecture support required to enable the
optimised implementation of lockrefs.
That's as simple as defining arch_spin_value_unlocked() and selecting
the Kconfig option.
We also define cmpxchg64_relaxed(), because the lockref code does not
need the cmpxchg to have barrier semantics.
Using Linus' test case[1] on one system I see a 4x improvement for the
basic enablement, and a further 1.3x for cmpxchg64_relaxed(), for a
total of 5.3x vs the baseline.
On another system I see more like 2x improvement.
[1]: http://marc.info/?l=linux-fsdevel&m=137782380714721&w=4
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The powerpc lock acquisition sequence is as follows:
lwarx; cmpwi; bne; stwcx.; lwsync;
Lock release is as follows:
lwsync; stw;
If CPU 0 does a store (say, x=1) then a lock release, and CPU 1
does a lock acquisition then a load (say, r1=y), then there is
no guarantee of a full memory barrier between the store to 'x'
and the load from 'y'. To see this, suppose that CPUs 0 and 1
are hardware threads in the same core that share a store buffer,
and that CPU 2 is in some other core, and that CPU 2 does the
following:
y = 1; sync; r2 = x;
If 'x' and 'y' are both initially zero, then the lock
acquisition and release sequences above can result in r1 and r2
both being equal to zero, which could not happen if unlock+lock
was a full barrier.
This commit therefore makes powerpc's
smp_mb__after_unlock_lock() be a full barrier.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: linuxppc-dev@lists.ozlabs.org
Cc: <linux-arch@vger.kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/1386799151-2219-8-git-send-email-paulmck@linux.vnet.ibm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Our ppc64 spinlocks and rwlocks use a trick where a lock token and
the paca index are placed in the lock with a single store. Since we
are using two u16s they need adjusting for little endian.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Although the shared_proc field in the lppaca works today, it is
not architected. A shared processor partition will always have a non
zero yield_count so use that instead. Create a wrapper so users
don't have to know about the details.
In order for older kernels to continue to work on KVM we need
to set the shared_proc bit. While here, remove the ugly bitfield.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Use local_paca directly in macro SHARED_PROCESSOR, as all processors
have the same value for the field shared_proc, so we don't need care
racy here.
Reported-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
For performance reasons we are about to change ISYNC_ON_SMP to sometimes be
lwsync. Now that the macro name doesn't make sense, change it and LWSYNC_ON_SMP
to better explain what the barriers are doing.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Recent versions of the PowerPC architecture added a hint bit to the larx
instructions to differentiate between an atomic operation and a lock operation:
> 0 Other programs might attempt to modify the word in storage addressed by EA
> even if the subsequent Store Conditional succeeds.
>
> 1 Other programs will not attempt to modify the word in storage addressed by
> EA until the program that has acquired the lock performs a subsequent store
> releasing the lock.
To avoid a binutils dependency this patch create macros for the extended lwarx
format and uses it in the spinlock code. To test this change I used a simple
test case that acquires and releases a global pthread mutex:
pthread_mutex_lock(&mutex);
pthread_mutex_unlock(&mutex);
On a 32 core POWER6, running 32 test threads we spend almost all our time in
the futex spinlock code:
94.37% perf [kernel] [k] ._raw_spin_lock
|
|--99.95%-- ._raw_spin_lock
| |
| |--63.29%-- .futex_wake
| |
| |--36.64%-- .futex_wait_setup
Which is a good test for this patch. The results (in lock/unlock operations per
second) are:
before: 1538203 ops/sec
after: 2189219 ops/sec
An improvement of 42%
A 32 core POWER7 improves even more:
before: 1279529 ops/sec
after: 2282076 ops/sec
An improvement of 78%
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Name space cleanup for rwlock functions. No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: linux-arch@vger.kernel.org
Not strictly necessary for -rt as -rt does not have non sleeping
rwlocks, but it's odd to not have a consistent naming convention.
No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: linux-arch@vger.kernel.org
Name space cleanup. No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: linux-arch@vger.kernel.org
The raw_spin* namespace was taken by lockdep for the architecture
specific implementations. raw_spin_* would be the ideal name space for
the spinlocks which are not converted to sleeping locks in preempt-rt.
Linus suggested to convert the raw_ to arch_ locks and cleanup the
name space instead of using an artifical name like core_spin,
atomic_spin or whatever
No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: linux-arch@vger.kernel.org
Needed to avoid namespace conflicts when the common code
function bodies of _spin_try_lock() etc. are moved to a header
file where the function name would be __spin_try_lock().
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Horst Hartmann <horsth@linux.vnet.ibm.com>
Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David Miller <davem@davemloft.net>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Roman Zippel <zippel@linux-m68k.org>
Cc: <linux-arch@vger.kernel.org>
LKML-Reference: <20090831124415.918799705@de.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Pass the original flags to rwlock arch-code, so that it can re-enable
interrupts if implemented for that architecture.
Initially, make __raw_read_lock_flags and __raw_write_lock_flags stubs
which just do the same thing as non-flags variants.
Signed-off-by: Petr Tesarik <ptesarik@suse.cz>
Signed-off-by: Robin Holt <holt@sgi.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: <linux-arch@vger.kernel.org>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We have several instances of inline assembly code that use the addic
or addic. instructions, but don't include XER in the list of clobbers.
The addic and addic. instructions affect the carry bit, which is in
the XER register.
This adds "xer" to the list of clobbers for those inline asm
statements that use addic or addic. and didn't already have it.
Signed-off-by: Paul Mackerras <paulus@samba.org>
from include/asm-powerpc. This is the result of a
mkdir arch/powerpc/include/asm
git mv include/asm-powerpc/* arch/powerpc/include/asm
Followed by a few documentation/comment fixups and a couple of places
where <asm-powepc/...> was being used explicitly. Of the latter only
one was outside the arch code and it is a driver only built for powerpc.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Pan Xinhui