[Top] [All Lists]

Re: PROBLEM + POSS FIX: kernel stack overflow, xfs, many disks, heavy wr

To: Chris Mason <chris.mason@xxxxxxxxxx>, Dave Chinner <david@xxxxxxxxxxxxx>, John Berthels <john@xxxxxxxxx>, linux-kernel@xxxxxxxxxxxxxxx, Nick Gregory <nick@xxxxxxxxx>, Rob Sanderson <rob@xxxxxxxxx>, xfs@xxxxxxxxxxx, linux-mm@xxxxxxxxx
Subject: Re: PROBLEM + POSS FIX: kernel stack overflow, xfs, many disks, heavy write load, 8k stack, x86-64
From: John Berthels <john@xxxxxxxxx>
Date: Fri, 16 Apr 2010 14:41:52 +0100
In-reply-to: <20100409113850.GE13327@think>
References: <4BBC6719.7080304@xxxxxxxxx> <20100407140523.GJ11036@dastard> <4BBCAB57.3000106@xxxxxxxxx> <20100407234341.GK11036@dastard> <20100408030347.GM11036@dastard> <4BBDC92D.8060503@xxxxxxxxx> <4BBDEC9A.9070903@xxxxxxxxx> <20100408233837.GP11036@dastard> <20100409113850.GE13327@think>
User-agent: Thunderbird (X11/20100411)
Chris Mason wrote:
shrink_zone on my box isn't 500 bytes, but lets try the easy stuff
first.  This is against .34, if you have any trouble applying to .32,
just add the word noinline after the word static on the function

This makes shrink_zone disappear from my check_stack.pl output.
Basically I think the compiler is inlining the shrink_active_zone and
shrink_inactive_zone code into shrink_zone.

Hi Chris,

I hadn't seen the followup discussion on lkml until today, but this message:


allowed me to look at stack usage in our build environment. If I've understood correctly, it seems that a build with gcc-4.4 and gcc-4.3 have very different stack usages for shrink_zone(): 0x88 versus 0x1d8. (details below).

The reason appears to be the -fconserve-stack compilation option specified when using 4.4, since running the cmdline from mm/.vmscan.cmd with gcc-4.4 but *without* -fconserve-stack gives the same result as with 4.3.

According to the discussion when the flag was added, http://www.gossamer-threads.com/lists/linux/kernel/1131612 this flag seems to primarily affects inlining, so I double-checked the noinline patch you sent to the list and discovered that it had been incorrectly applied to the build tree. Correctly applying that patch to mm/vmscan.c (and using gcc-4.3) gives a

sub    $0x78,%rsp

line. I'm very sorry that this test or ours wasn't correct and I'm sorry for sending bad info to the list.

We're currently building a kernel with gcc-4.4 and will let you know if it blows the 8k limit or not.

Thanks for your help.



$ gcc-4.3 --version
gcc-4.3 (Ubuntu 4.3.4-5ubuntu1) 4.3.4
$ gcc-4.4 --version
gcc-4.4 (Ubuntu 4.4.1-4ubuntu9) 4.4.1

$ make CC=gcc-4.4 mm/vmscan.o
$ objdump -d mm/vmscan.o  | less +/shrink_zone
0000000000002830 <shrink_zone>:
    2830:       55                      push   %rbp
    2831:       48 89 e5                mov    %rsp,%rbp
    2834:       41 57                   push   %r15
    2836:       41 56                   push   %r14
    2838:       41 55                   push   %r13
    283a:       41 54                   push   %r12
    283c:       53                      push   %rbx
    283d:       48 81 ec 88 00 00 00    sub    $0x88,%rsp
    2844:       e8 00 00 00 00          callq  2849 <shrink_zone+0x19>
$ make clean
$ make CC=gcc-4.3 mm/vmscan.o
$ objdump -d mm/vmscan.o  | less +/shrink_zone
0000000000001ca0 <shrink_zone>:
    1ca0:       55                      push   %rbp
    1ca1:       48 89 e5                mov    %rsp,%rbp
    1ca4:       41 57                   push   %r15
    1ca6:       41 56                   push   %r14
    1ca8:       41 55                   push   %r13
    1caa:       41 54                   push   %r12
    1cac:       53                      push   %rbx
    1cad:       48 81 ec d8 01 00 00    sub    $0x1d8,%rsp
    1cb4:       e8 00 00 00 00          callq  1cb9 <shrink_zone+0x19>

<Prev in Thread] Current Thread [Next in Thread>