Re: PROBLEM: Bug in __pollwait() can cause select() and poll() to hang in

Re: PROBLEM: Bug in __pollwait() can cause select() and poll() to hang in

[Manfred Spraul]

Hi Ray,

your bug description seems to be correct, but the fix is wrong:
If the allocation is for the 2nd page of wait queue heads, then 
"current->state = TASK_INTERRUPTIBLE" can lead to lost wakeups, if an fd 
that is stored in the first page gets ready during the allocation. 
Setting the state to interruptible is only permitted if a full scan of 
all file descriptors happens before calling schedule(). This is 
expensive and should be avoided.

The correct fix is current->state = TASK_RUNNING just before calling 
yield() in the rebalance code.

--
    Manfred

Re: PROBLEM: Bug in __pollwait() can cause select() and poll() tohang in

[Ray Bryant]

Manfred Spraul wrote:
> 
> Hi Ray,
> 
> your bug description seems to be correct, but the fix is wrong:
> If the allocation is for the 2nd page of wait queue heads, then
> "current->state = TASK_INTERRUPTIBLE" can lead to lost wakeups, if an fd
> that is stored in the first page gets ready during the allocation.

Hi Manfred,

Grumble.  :-) Yes, I believe you are correct.

> Setting the state to interruptible is only permitted if a full scan of
> all file descriptors happens before calling schedule(). This is
> expensive and should be avoided.
> 
> The correct fix is current->state = TASK_RUNNING just before calling
> yield() in the rebalance code.

But doesn't this have the same kind of problem?  e. g., just before
calling yield() in the rebalance code we save current->state, set it to
TASK_RUNNING, then restore current->state on return from yield().  If a
fd becomes ready after the call to yield(), and we entered
__alloc_pages() with state TASK_INTERRUPTIBLE, aren't we in exactly the
same situation as described above?

Let me think about this some more.

Thanks,
-- 
Best Regards,
Ray
-----------------------------------------------
                  Ray Bryant
512-453-9679 (work)         512-507-7807 (cell)
raybry@sgi.com             raybry@austin.rr.com
The box said: "Requires Windows 98 or better",
           so I installed Linux.
-----------------------------------------------

Re: PROBLEM: Bug in __pollwait() can cause select() and poll() tohang in

[Andrew Morton]

Ray Bryant <raybry@sgi.com> wrote:
>
>  > The correct fix is current->state = TASK_RUNNING just before calling
>  > yield() in the rebalance code.
> 
>  But doesn't this have the same kind of problem?  e. g., just before
>  calling yield() in the rebalance code we save current->state, set it to
>  TASK_RUNNING, then restore current->state on return from yield().  If a
>  fd becomes ready after the call to yield(), and we entered
>  __alloc_pages() with state TASK_INTERRUPTIBLE, aren't we in exactly the
>  same situation as described above?

No, you cannot restore the task state after having set it to TASK_RUNNING.

Just leave the state at TASK_RUNNING.  The (silly) code which called the
page allocator in state TASK_[IN]TERRUPTIBLE will just go around its wait
loop an extra time and go back to sleep.  This almost always works.