[haiku-gsoc] Re: [HCD]: Bfs bug #1
- From: Ingo Weinhold <ingo_weinhold@xxxxxx>
- To: haiku-gsoc@xxxxxxxxxxxxx
- Date: Mon, 23 Jun 2008 00:06:38 +0200
On 2008-06-22 at 14:30:00 [+0200], Salvatore Benedetto <emitrax@xxxxxxxxx>
wrote:
> Another test. Different approach, different behavior.
>
> To try and support my thesis about "starvation", I ran the test on vmware
> and left the computer, while usually I used to use it regurarly while
> running
> the test, so this time teorically all CPU would have been reserved to
> Haiku, and thus the low memory thread should have been able to free the
> vnode in 10 seconds.
BTW, what is you machine's configuration? Host system physical memory,
memory assigned to the virtual machine, size of the kernel tracing buffer.
> The test this time in fact, last 3 hours before crashing (nice nap by
> the way :-) ),
> with the same error, but the output of tar was quite different. See
> screenshot attached.
> Any comments?
Mmh, the stack trace looks like the one attached to the ticket. Not sure
what to make of the kernel tracing output. Seems neither related nor wrong.
> Anyway, I wanted to test with condition variable, but I'm not sure on
> how to proceed, despite of that
> though, I think we should hanle the error E_BUSY differently, because
> 10 seconds with our current
> scheduler, and with not I/O scheduler doesn't seem to be a good choice
> on slow system like running
> on vmware is.
You would need to remove the vnode::busy flag and add a ConditionVariable*
busy_condition instead. When NULL the vnode is not busy. Whereever a vnode
is currently marked busy, a condition variable needs to be created and
busy_condition set respectively. In get_vnode() where one waits for vnode
to become unbusy, one needs to wait for the condition variable instead of
snooze()ing (everything else remains the same). In most cases the condition
variable can be allocated on the stack, since it is unset in the same
function, but unfortunately new_vnode() also marks the vnode busy -- there
it needs to be allocated on the heap and deleted in
publish_vnode()/remove_vnode().
> Anyway, is Marcus subscribed to this mailing list?
I don't think he is.
> What does he think
> about the possibility of being
> a scsi problem?
He's probably not aware of the bug yet. Respective comments should be added
to the ticket and he should be notified.
CU, Ingo
Other related posts:
