[Pvfs2-developers] Hang at server shutdown

[Scott Atchley]

On Jan 18, 2007, at 7:49 PM, Scott Atchley wrote:

> I should add that I not using the kernel module on the client. I am  
> simply using pvfs2-ls, pvfs2-cp, pvfs2-rm several times in a row,  
> then ^C the server.
>
> For each command line tool, the client does a connect to the  
> server. The server cleans up the old client state and resets it. I  
> assume it is related to multiple calls to  
> bmi_method_addr_reg_callback() on the server.
>
> Scott

It seems I get a call to BMI_mx_set_info() for each time I use a  
command line app on the client. For example, on the client I call:

$ pvfs2-ls -l /mnt/pvfs2 ; pvfs2-rm /mnt/pvfs2/test ; pvfs2-cp -t  
test /mnt/pvfs2/test
$ pvfs2-ls -l /mnt/pvfs2 ; pvfs2-rm /mnt/pvfs2/test ; pvfs2-cp -t  
test /mnt/pvfs2/test
$ pvfs2-ls -l /mnt/pvfs2 ; pvfs2-rm /mnt/pvfs2/test ; pvfs2-cp -t  
test /mnt/pvfs2/test
$ pvfs2-ls -l /mnt/pvfs2 ; pvfs2-rm /mnt/pvfs2/test ; pvfs2-cp -t  
test /mnt/pvfs2/test

and then I ^C the server, I see 12 calls into BMI_mx_set_info() after  
BMX_mx_finalize() in the server's output:

[D 01/19 08:39] bmi_mx: entering BMI_mx_finalize.
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop mx://fog34:0:0 map  
0x0x81e3620 mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).
[D 01/19 08:39] bmi_mx: BMI_mx_set_info drop NULL map 0x0x81e3620  
mxmap 0x0x81e362c).

I added a flag to my peer struct so that I only call  
bmi_method_addr_reg_callback() the first time the server connects to  
the client. It now behaves properly and exits cleanly.

I thought IB was calling that each time in  
ib_tcp_server_check_new_connections(), but I must have misunderstood  
the code.

Scott