Re: 2.6.33-rc8 breaks UML with Restrict initial stack space expansion to rlimit

["Américo Wang" <xiyou.wangcong@gmail.com>]

On Mon, Feb 15, 2010 at 03:59:26PM +0900, KOSAKI Motohiro wrote:
>> 
>> 
>> In message <20100214164023.GA2726@jm.kir.nu> you wrote:
>> > It looks like the commit 803bf5ec259941936262d10ecc84511b76a20921
>> > (fs/exec.c: restrict initial stack space expansion to rlimit) broke my
>> > user mode Linux setup by somehow preventing system setup from running
>> > properly (or killing some processes that try to mount things, etc.).
>> > This commit turned up as the reason based on git bisect and reverting it
>> > fixes my UML test setup (Ubuntu 9.10 on both host and in UML and AMD64
>> > arch for both). I have no idea what exactly would be the main cause for
>> > this issue, but this looks like a somewhat unfortunately timed
>> > regression in 2.6.33-rc8.
>> > 
>> > The failed run shows like this (with current linux-2.6.git):
>> > 
>> > ...
>> > EXT3-fs (ubda): mounted filesystem with writeback data mode
>> > VFS: Mounted root (ext3 filesystem) readonly on device 98:0.
>> > IRQ 3/console-write: IRQF_DISABLED is not guaranteed on shared IRQs
>> > IRQ 2/console: IRQF_DISABLED is not guaranteed on shared IRQs
>> > IRQ 10/winch: IRQF_DISABLED is not guaranteed on shared IRQs
>> > IRQ 10/winch: IRQF_DISABLED is not guaranteed on shared IRQs
>> > mountall: mount /sys/kernel/debug [218] killed by KILL signal
>> > mountall: Filesystem could not be mounted: /sys/kernel/debug
>> > mountall: mount /dev [219] killed by KILL signal
>> > mountall: Filesystem could not be mounted: /dev
>> > mountall: mount /tmp [220] killed by KILL signal
>> > mountall: Filesystem could not be mounted: /tmp
>> > mountall: mount /var/lock [222] killed by KILL signal
>> > mountall: Filesystem could not be mounted: /var/lock
>> > ...
>> > 
>> > 
>> > With 803bf5ec reverted, UML comes up and the output looks like this:
>> > 
>> > ...
>> > EXT3-fs (ubda): mounted filesystem with writeback data mode
>> > VFS: Mounted root (ext3 filesystem) readonly on device 98:0.
>> > IRQ 3/console-write: IRQF_DISABLED is not guaranteed on shared IRQs
>> > IRQ 2/console: IRQF_DISABLED is not guaranteed on shared IRQs
>> > IRQ 10/winch: IRQF_DISABLED is not guaranteed on shared IRQs
>> > IRQ 10/winch: IRQF_DISABLED is not guaranteed on shared IRQs
>> > init: procps main process (226) terminated with status 255
>> > fsck from util-linux-ng 2.16
>> > ...
>> 
>> Jouni,
>> 
>> I can reproduce this now.  
>> 
>> We got the logic wrong in one of the cleanups and hence we aren't
>> actually changing the stack reservation ever, when we intended on
>> allocating up to 20 new pages.  
>> 
>> The:
>> 	rlim_stack = min(rlim_stack, stack_size);
>> always chooses stack_size hence we end up not changing the stack at all.
>> This seems to cause fatal problems on UML, but is obviously not what was
>> intended for archs as well.  
>> 
>> The following works for me on PPC64 64k and 4k pages and UML on x86_64. 
>> 
>> Let me know if it fixes it for you also.
>> 
>> Mikey
>> 
>> 
>> exec/fs: fix initial stack reservation
>> 
>> 803bf5ec259941936262d10ecc84511b76a20921 (fs/exec.c: restrict initial
>> stack space expansion to rlimit) attempts to limit the initial stack to
>> 20*PAGE_SIZE.  Unfortunately, in also attempting ensure the stack is not
>> reduced in size, we ended up not changing the stack at all.  
>> 
>> This caused a regression in UML resulting in most guest processes to be
>> killed. 
>> 
>> Signed-off-by: Michael Neuling <mikey@neuling.org>
>> cc: <stable@kernel.org>
>> 
>> diff --git a/fs/exec.c b/fs/exec.c
>> index e95c692..e0e7b3c 100644
>> --- a/fs/exec.c
>> +++ b/fs/exec.c
>> @@ -637,15 +637,16 @@ int setup_arg_pages(struct linux_binprm *bprm,
>>  	 * will align it up.
>>  	 */
>>  	rlim_stack = rlimit(RLIMIT_STACK) & PAGE_MASK;
>> -	rlim_stack = min(rlim_stack, stack_size);
>>  #ifdef CONFIG_STACK_GROWSUP
>>  	if (stack_size + stack_expand > rlim_stack)
>> -		stack_base = vma->vm_start + rlim_stack;
>> +		/*  Expand only to rlimit, making sure not to shrink it */
>> +		stack_base = vma->vm_start + max(rlim_stack,stack_size);
>>  	else
>>  		stack_base = vma->vm_end + stack_expand;
>>  #else
>>  	if (stack_size + stack_expand > rlim_stack)
>> -		stack_base = vma->vm_end - rlim_stack;
>> +		/*  Expand only to rlimit, making sure not to shrink it */
>> +		stack_base = vma->vm_end - max(rlim_stack,stack_size);
>>  	else
>>  		stack_base = vma->vm_start - stack_expand;
>>  #endif
>
>-	rlim_stack = min(rlim_stack, stack_size);
>+	/*  Expand only to rlimit, making sure not to shrink it */
>+	rlim_stack = max(rlim_stack, stack_size);
>
>is better fix?
>

Odd. If this is the right fix, 'stack_size" will be able to exceed
stack rlimit, then Michael's previous rlimit patch will be useless.
Am I missing something?