Update the inflated memory in the mm core on change.
Signed-off-by: Alexander Atanasov <[email protected]>
---
drivers/misc/vmw_balloon.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/misc/vmw_balloon.c b/drivers/misc/vmw_balloon.c
index 91d4d2a285c5..3bfd845898f5 100644
--- a/drivers/misc/vmw_balloon.c
+++ b/drivers/misc/vmw_balloon.c
@@ -1507,6 +1507,7 @@ static void vmballoon_work(struct work_struct *work)
queue_delayed_work(system_freezable_wq,
dwork, round_jiffies_relative(HZ));
+ balloon_set_inflated_free(atomic64_read(&b->size) << 2);
}
/**
--
2.31.1
On Oct 19, 2022, at 12:56 PM, Alexander Atanasov <[email protected]> wrote:
> Update the inflated memory in the mm core on change.
>
> Signed-off-by: Alexander Atanasov <[email protected]>
> ---
> drivers/misc/vmw_balloon.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/drivers/misc/vmw_balloon.c b/drivers/misc/vmw_balloon.c
> index 91d4d2a285c5..3bfd845898f5 100644
> --- a/drivers/misc/vmw_balloon.c
> +++ b/drivers/misc/vmw_balloon.c
> @@ -1507,6 +1507,7 @@ static void vmballoon_work(struct work_struct *work)
> queue_delayed_work(system_freezable_wq,
> dwork, round_jiffies_relative(HZ));
>
> + balloon_set_inflated_free(atomic64_read(&b->size) << 2);
> }
I don’t like it in general (I think that something like that should go into
some common infra).
But more concretely there are at least 2 problems here. First, queueing the
work should come last. Second, there are other places that change the
balloon size (e.g., vmballoon_reset()), which are not handled by this patch.
If you added calls to balloon_set_inflated_free() from these places, you can
get races while calling balloon_set_inflated_free() and the last value that
would be latched would be the wrong one. I don’t see anything in the logic
or comments that clarify how something like that should be resolved.
On Oct 21, 2022, at 10:25 AM, Alexander Atanasov <[email protected]> wrote:
>
> Ok,I will move it before the enqueue call.
> But are you sure about this the reset?
> vmballoon_reset(...) is called only from vmballoon_work(...) which does
> the update ? what i am missing?
My bad. But when the module is unloaded, vmballoon_pop() is called.
On 21.10.22 9:50, Nadav Amit wrote:
> On Oct 19, 2022, at 12:56 PM, Alexander Atanasov <[email protected]> wrote:
>
>> Update the inflated memory in the mm core on change.
>>
>> Signed-off-by: Alexander Atanasov <[email protected]>
>> ---
>> drivers/misc/vmw_balloon.c | 1 +
>> 1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/misc/vmw_balloon.c b/drivers/misc/vmw_balloon.c
>> index 91d4d2a285c5..3bfd845898f5 100644
>> --- a/drivers/misc/vmw_balloon.c
>> +++ b/drivers/misc/vmw_balloon.c
>> @@ -1507,6 +1507,7 @@ static void vmballoon_work(struct work_struct *work)
>> queue_delayed_work(system_freezable_wq,
>> dwork, round_jiffies_relative(HZ));
>>
>> + balloon_set_inflated_free(atomic64_read(&b->size) << 2);
>> }
>
> I don’t like it in general (I think that something like that should go into
> some common infra).
My goal is to create that infra, sure there is a place for improvement.
I think of it as a commit point so plugging it into something existing
does not fit or at least i don't see how.
> But more concretely there are at least 2 problems here. First, queueing the
> work should come last. Second, there are other places that change the
> balloon size (e.g., vmballoon_reset()), which are not handled by this patch.
>
> If you added calls to balloon_set_inflated_free() from these places, you can
> get races while calling balloon_set_inflated_free() and the last value that
> would be latched would be the wrong one. I don’t see anything in the logic
> or comments that clarify how something like that should be resolved.
>
Ok,I will move it before the enqueue call.
But are you sure about this the reset?
vmballoon_reset(...) is called only from vmballoon_work(...) which does
the update ? what i am missing?
--
Regards,
Alexander Atanasov
On 21.10.22 10:31, Nadav Amit wrote:
> On Oct 21, 2022, at 10:25 AM, Alexander Atanasov <[email protected]> wrote:
>
>>
>> Ok,I will move it before the enqueue call.
>> But are you sure about this the reset?
>> vmballoon_reset(...) is called only from vmballoon_work(...) which does
>> the update ? what i am missing?
>
> My bad. But when the module is unloaded, vmballoon_pop() is called.
Yes, i missed the unload - i will just set it to zero there.
--
Regards,
Alexander Atanasov