[PVFS2-developers] current cvs weirdness

[Weikuan Yu]

This problem starts at least since Friday. I came across the same 
problem over the weekend with an updated tree on Friday.

When doing the following command with one server, it gives SIGSEGV.

pvfs2-cp -t /usr/lib/libc.a /mnt/pvfs2/libc.a

16 entry LIO is passed down to BMI, the arrary of vec:len is 
26143:{232313:-232313} (repeat 7 times}:232313.  memory addr is the 
same for all 16 entries. However, if using two IO-servers, the pvfs2-cp 
completes. But pvfs2-ls then shows the file /mnt/pvfs2/libc.a has a 
huge size, something like (1<<64 - 26143 - 232313).

Not sure if this is of any help,

Weikuan

On Mar 14, 2005, at 5:45 PM, Phil Carns wrote:

> Well, I don't know what the underlying problem is, but fixing a 
> superficial bug lets me avoid it :)
>
> In pvfs2-cp it checks "if (strip_size) {" at some point.  It should 
> really check if strip_size > 0, because the parse_args function sets 
> it to -1.
>
> Fixing this prevents the distribution setting code from being 
> triggered in the general case, which solves my problem.
>
> However, with this fix in place in my tree I still see the error if I 
> try to use the -s option to pvfs2-cp.  I don't know if this is showing 
> a bug in pvfs2-cp or in the distribution code, though.
>
> -Phil
>
> Phil Carns wrote:
>> Is anyone besides me seeing anything strange with the current cvs 
>> snapshot?  This fails for me on a single server/client setup:
>> ./pvfs2-cp /etc/hosts /mnt/pvfs2/foo
>> The error msg printing is broken in the case that I am hitting, so 
>> the output isn't very helpful.  However, looking in gdb I see that 
>> bmi has been asked to send 16 buffers in a list operation generated 
>> by a flow, and the writev call gets a EINVAL.  I'm pretty sure that 
>> copying this small file should only require one buffer.
>> Oddly enough, everything looks fine when I run pvfs2-cp through 
>> valgrind, and I get no error.  So.. maybe this is some subtle timing 
>> or memory thing that doesn't show up all the time.  It hits me every 
>> time right now though :)
>> I don't know where this problem came from yet.  Just randomly backing 
>> up a bit it looks like cvs from 6 weeks ago is fine, but maybe cvs 
>> from 3 weeks ago is not?
>> I'll continue digging (or maybe just narrow down the working cvs 
>> versions until I find the suspect commit), but I just thought I would 
>> hit the list in case anyone has ideas...
>> -Phil
>> _______________________________________________
>> PVFS2-developers mailing list
>> PVFS2-developers at beowulf-underground.org
>> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
>
> _______________________________________________
> PVFS2-developers mailing list
> PVFS2-developers at beowulf-underground.org
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
>