xfs
[Top] [All Lists]

TAKE 821124 - Deadlock in extent conversion

To: lord@xxxxxxx
Subject: TAKE 821124 - Deadlock in extent conversion
From: pv@xxxxxxxxxxxxxxxxxxxxxx (lord@xxxxxxx)
Date: Fri, 13 Apr 2001 09:12:26 -0700 (PDT)
Cc: linux-xfs@xxxxxxxxxxx
Reply-to: sgi.bugs.xfs@xxxxxxxxxxxxxxxxx
Sender: owner-linux-xfs@xxxxxxxxxxx
 Submitter : ananth                   *Status : closed                      
 Assigned Engineer : lord             *Fixed By : lord                      
*Fixed By Domain : sgi.com            *Closed Date : 04/13/01               
 Priority : 2                         *Modified Date : 04/13/01             
*Modified User : lord                 *Modified User Domain : sgi.com       
*Fix Description :
From: steve lord <lord@xxxxxxx> (TAKE)
Date: Apr 13 2001 09:12:26AM
[pvnews version: 1.71]
----------------------------

Fix some more places to avoid memory deadlock with the xfs transaction
system, this includes the allocation of the pagebuf header and the
kiobuf structures.

Date:  Fri Apr 13 08:52:59 PDT 2001
Workarea:  jen.americas.sgi.com:/src/lord/xfs-base

The following file(s) were checked into:
  bonnie.engr.sgi.com:/isms/slinx/2.4.x-xfs


Modid:  2.4.x-xfs:slinx:92465a
linux/kernel/ksyms.c - 1.84
        - change names of exported kiobuf allocation functions.

linux/fs/iobuf.c - 1.18
        - replace kiobuf allocation functions with versions which take a
          gfp_mask as an extra parameter. Names are changed to _ versions
          and the original functions become inlines in iobuf.h and use
          the default mask of SLAB_KERNEL.

linux/include/linux/iobuf.h - 1.12
        - replace kiobuf allocation functions with versions which take a
          gfp_mask as an extra parameter. Names are changed to _ versions
          and the original functions become inlines in iobuf.h and use  
          the default mask of SLAB_KERNEL.

linux/fs/pagebuf/page_buf.c - 1.76
        - use GFP_PAGE_IO in memory allocations from within transactions.
Description :
This problem seen with XFS 1.0 bits on a 64 M, 2CPU system with
1 scsi disk. Tests being run in parallel:

while true; do holes -l 300000000 -b 4099   -i 7 -c 7 /xfs/holes; date; done

and AIM 7, run to crossover.

It took about 3-4 hours to hit this problem, by then
AIM had gone to about 170 processes. In the end,
console became unresponsive ... looking at time-stamps

.....


==========================
ADDITIONAL INFORMATION (REASSIGN)
From: lord@xxxxxxx (BugWorks)
Date: Apr 12 2001 12:22:14PM
==========================

I have a sneaking suspicion we lost a bit of code in the mod
merge, I will go take a look.

<Prev in Thread] Current Thread [Next in Thread>
  • TAKE 821124 - Deadlock in extent conversion, lord@xxxxxxx <=