Re: [PATCH V4 7/9] vhost: do not use RCU to synchronize MMU notifier with worker

From: Jason Wang
Date: Thu Aug 08 2019 - 08:55:04 EST

Next message: Sudeep Holla: "Re: [PATCH 6/6] PSCI: cpuidle: Refactor CPU suspend power_state parameter handling"
Previous message: Julia Lawall: "drivers/clk/bcm/clk-bcm2835.c:2144:1-13: WARNING: Use devm_platform_ioremap_resource for cprman -> regs (fwd)"
Next in thread: Jason Wang: "Re: [PATCH V4 7/9] vhost: do not use RCU to synchronize MMU notifier with worker"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On 2019/8/7 äå10:02, Jason Wang wrote:
>
> On 2019/8/7 äå8:07, Jason Gunthorpe wrote:
>> On Wed, Aug 07, 2019 at 03:06:15AM -0400, Jason Wang wrote:
>>> We used to use RCU to synchronize MMU notifier with worker. This leads
>>> calling synchronize_rcu() in invalidate_range_start(). But on a busy
>>> system, there would be many factors that may slow down the
>>> synchronize_rcu() which makes it unsuitable to be called in MMU
>>> notifier.
>>>
>>> So this patch switches use seqlock counter to track whether or not the
>>> map was used. The counter was increased when vq try to start or finish
>>> uses the map. This means, when it was even, we're sure there's no
>>> readers and MMU notifier is synchronized. When it was odd, it means
>>> there's a reader we need to wait it to be even again then we are
>>> synchronized. Consider the read critical section is pretty small the
>>> synchronization should be done very fast.
>>>
>>> Reported-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
>>> Fixes: 7f466032dc9e ("vhost: access vq metadata through kernel
>>> virtual address")
>>> Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
>>> Â drivers/vhost/vhost.c | 141
>>> ++++++++++++++++++++++++++----------------
>>> Â drivers/vhost/vhost.h |ÂÂ 7 ++-
>>> Â 2 files changed, 90 insertions(+), 58 deletions(-)
>>>
>>> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
>>> index cfc11f9ed9c9..57bfbb60d960 100644
>>> +++ b/drivers/vhost/vhost.c
>>> @@ -324,17 +324,16 @@ static void vhost_uninit_vq_maps(struct
>>> vhost_virtqueue *vq)
>>> Â ÂÂÂÂÂ spin_lock(&vq->mmu_lock);
>>> ÂÂÂÂÂ for (i = 0; i < VHOST_NUM_ADDRS; i++) {
>>> -ÂÂÂÂÂÂÂ map[i] = rcu_dereference_protected(vq->maps[i],
>>> -ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ lockdep_is_held(&vq->mmu_lock));
>>> +ÂÂÂÂÂÂÂ map[i] = vq->maps[i];
>>> ÂÂÂÂÂÂÂÂÂ if (map[i]) {
>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂ vhost_set_map_dirty(vq, map[i], i);
>>> -ÂÂÂÂÂÂÂÂÂÂÂ rcu_assign_pointer(vq->maps[i], NULL);
>>> +ÂÂÂÂÂÂÂÂÂÂÂ vq->maps[i] = NULL;
>>> ÂÂÂÂÂÂÂÂÂ }
>>> ÂÂÂÂÂ }
>>> ÂÂÂÂÂ spin_unlock(&vq->mmu_lock);
>>> Â -ÂÂÂ /* No need for synchronize_rcu() or kfree_rcu() since we are
>>> -ÂÂÂÂ * serialized with memory accessors (e.g vq mutex held).
>>> +ÂÂÂ /* No need for synchronization since we are serialized with
>>> +ÂÂÂÂ * memory accessors (e.g vq mutex held).
>>> ÂÂÂÂÂÂ */
>>> Â ÂÂÂÂÂ for (i = 0; i < VHOST_NUM_ADDRS; i++)
>>> @@ -362,6 +361,40 @@ static bool vhost_map_range_overlap(struct
>>> vhost_uaddr *uaddr,
>>> ÂÂÂÂÂ return !(end < uaddr->uaddr || start > uaddr->uaddr - 1 +
>>> uaddr->size);
>>> Â }
>>> Â +static void inline vhost_vq_access_map_begin(struct
>>> vhost_virtqueue *vq)
>>> +{
>>> +ÂÂÂ write_seqcount_begin(&vq->seq);
>>> +}
>>> +
>>> +static void inline vhost_vq_access_map_end(struct vhost_virtqueue *vq)
>>> +{
>>> +ÂÂÂ write_seqcount_end(&vq->seq);
>>> +}
>> The write side of a seqlock only provides write barriers. Access to
>>
>> ÂÂÂÂmap = vq->maps[VHOST_ADDR_USED];
>>
>> Still needs a read side barrier, and then I think this will be no
>> better than a normal spinlock.
>>
>> It also doesn't seem like this algorithm even needs a seqlock, as this
>> is just a one bit flag
>
>
> Right, so then I tend to use spinlock first for correctness.
>
>
>>
>> atomic_set_bit(using map)
>> smp_mb__after_atomic()
>> .. maps [...]
>> atomic_clear_bit(using map)
>>
>>
>> map = NULL;
>> smp_mb__before_atomic();
>> while (atomic_read_bit(using map))
>> ÂÂÂ relax()
>>
>> Again, not clear this could be faster than a spinlock when the
>> barriers are correct...
>

I've done some benchmark[1] on x86, and yes it looks even slower. It
looks to me the atomic stuffs is not necessary, so in order to compare
it better with spinlock. I tweak it a little bit through
smp_load_acquire()/store_releaes() + mb() like:

static struct vhost_map *vhost_vq_access_map_begin(struct
vhost_virtqueue
*vq,ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ unsigned int
type)ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

{ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ
++vq->counter;ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ /* Ensure map was read after incresing the counter.
PairedÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂ * with smp_mb() in
vhost_vq_sync_access().ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂ
*/ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ
smp_mb();ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ return
vq->maps[type];ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

}ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

static void inline vhost_vq_access_map_end(struct vhost_virtqueue
*vq)ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

{ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ /* Ensure all memory access through map was done
beforeÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂ * reducing the counter. Paired with smp_load_acquire()
inÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂ * vhost_vq_sync_access()
*/ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ smp_store_release(&vq->counter,
--vq->counter);ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

}ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

static void inline vhost_vq_sync_access(struct vhost_virtqueue
*vq)ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

{ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ /* Ensure new map value is visible before checking counter.
*/ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ
smp_mb();ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ /* Ensure map was freed after reading counter value,
pairedÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂ * with smp_store_release() in
vhost_vq_access_map_end().ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂ
*/ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ while (smp_load_acquire(&vq->counter))
{ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ if
(need_resched())ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ
schedule();ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

ÂÂÂÂÂÂÂ
}ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ

}ÂÂÂÂÂÂÂÂÂ

And the result is something like:

ÂÂÂÂÂÂÂÂ base | direct + atomic bitops | direct + spinlock() | direct +
counter + smp_mb() | direct + RCUÂÂÂÂ |
SMAP onÂ | 5.0Mpps | 5.0MppsÂÂÂÂ (+0%)ÂÂÂÂÂ | 5.7MppsÂ ÂÂÂ (+14%)ÂÂÂ Â |
5.9MppsÂ (+18%)ÂÂÂ ÂÂÂÂÂÂÂ | 6.2MppsÂ (+24%)Â |
SMAP off | 7.0Mpps | 7.0MppsÂÂÂÂ (+0%)ÂÂÂÂÂ | 7.0MppsÂÂ (+0%)ÂÂÂÂ |
7.5MppsÂ (+7%)ÂÂÂ ÂÂÂÂÂÂÂ | 8.2MppsÂ (+17%)Â |

base: normal copy_to_user()/copy_from_user() path.
direct + atomic bitops: using direct mapping but synchronize through
atomic bitops like you suggested above
direct + spinlock(): using direct mapping but synchronize through spinlocks
direct + counter + smp_mb(): using direct mapping but synchronize
through counter + smp_mb()
direct + RCU: using direct mapping and synchronize through RCU (buggy
and need to be addressed by this series)

So smp_mb() + counter is fastest way. And spinlock can still show some
improvement (+14%) in the case of SMAP, but no the case when SMAP is off.

I don't have any objection to convertÂ to spinlock() but just want to
know if any case that the above smp_mb() + counter looks good to you?

Thanks

>
> Yes, for next release we may want to use the idea from Michael like to
> mitigate the impact of mb.
>
> https://lwn.net/Articles/775871/
>
> Thanks
>
>
>>
>> Jason

Attachment: pEpkey.asc
Description: application/pgp-keys

Next message: Sudeep Holla: "Re: [PATCH 6/6] PSCI: cpuidle: Refactor CPU suspend power_state parameter handling"
Previous message: Julia Lawall: "drivers/clk/bcm/clk-bcm2835.c:2144:1-13: WARNING: Use devm_platform_ioremap_resource for cprman -> regs (fwd)"
Next in thread: Jason Wang: "Re: [PATCH V4 7/9] vhost: do not use RCU to synchronize MMU notifier with worker"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]