Re: [PATCH] softmmu/physmem: Improve guest memory allocation failure error message

[Igor Mammedov]

On Tue, 24 Aug 2021 09:37:54 +0100
"Dr. David Alan Gilbert" <dgilbert@redhat.com> wrote:

> * David Hildenbrand (david@redhat.com) wrote:
> > On 23.08.21 12:34, Philippe Mathieu-Daudé wrote:  
> > > On 8/23/21 12:24 PM, David Hildenbrand wrote:  
> > > > On 23.08.21 12:12, Philippe Mathieu-Daudé wrote:  
> > > > > On 8/23/21 11:29 AM, David Hildenbrand wrote:  
> > > > > > On 23.08.21 11:23, Peter Maydell wrote:  
> > > > > > > On Mon, 23 Aug 2021 at 09:40, David Hildenbrand <david@redhat.com>
> > > > > > > wrote:  
> > > > > > > > Not opposed to printing the size, although I doubt that it will 
> > > > > > > > really
> > > > > > > > stop similar questions/problems getting raised.  
> > > > > > > 
> > > > > > > The case that triggered this was somebody thinking
> > > > > > > -m took a byte count, so very likely that an error message
> > > > > > > saying "you tried to allocate 38TB" would have made their
> > > > > > > mistake clear in a way that just "allocation failed" did not.
> > > > > > > It also means that if a future user asks us for help then
> > > > > > > we can look at the error message and immediately tell them
> > > > > > > the problem, rather than going "hmm, what are all the possible
> > > > > > > ways that allocation might have failed" and going off down
> > > > > > > rabbitholes like VM overcommit settings...  
> > > > > > 
> > > > > > We've had similar issues recently where Linux memory overcommit 
> > > > > > handling
> > > > > > rejected the allocation -- and the user was well aware about the 
> > > > > > actual
> > > > > > size. You won't be able to catch such reports, because people don't
> > > > > > understand how Linux memory overcommit handling works or was 
> > > > > > configured.
> > > > > > 
> > > > > > "I have 3 GiB of free memory, why can't I create a 3 GiB VM". "I 
> > > > > > have 3
> > > > > > GiB of RAM, why can't I create a 3 GiB VM even if it won't make use 
> > > > > > of
> > > > > > all 3 GiB of memory".
> > > > > > 
> > > > > > Thus my comment, it will only stop very basic usage issues. And I 
> > > > > > agree
> > > > > > that looking at the error *might* help. It didn't help for the 
> > > > > > cases I
> > > > > > just described, because we need much more system information to 
> > > > > > make a
> > > > > > guess what the user error actually is.  
> > > > > 
> > > > > Is it possible to get the maximal overcommitable amount on Linux?  
> > > > 
> > > > Not reliably I think.
> > > > 
> > > > In the "always" mode, there is none.
> > > > 
> > > > In the "guess"/"estimate" mode, the kernel takes a guess (currently
> > > > implemented as checking if the mmap size <= total RAM + total SWAP).
> > > >      Committable = MemTotal + SwapTotal
> > > > 
> > > > In the "never" mode:
> > > >      Committable = CommitLimit - Committed_AS
> > > > However, the value gets further reduced for !root applications by
> > > > /proc/sys/vm/admin_reserve_kbytes.
> > > > 
> > > > Replicating these calculations in user space would be suboptimal IMHO.  
> > > 
> > > What about simply giving a hint about memory overcommit and display
> > > a link to documentation with longer description about how to check
> > > and figure out this issue?  
> > 
> > That would be highly OS-specific -- for example, there is no memory
> > overcommit under Windows. Sure, we could add a Linux specific hint,
> > indication documentation. But I'm not sure if most end users stumbling into
> > such an error+hint would be able to make sense of memory overcommit details
> > (not to mention that they know what it even is) :)
> > 
> > You can run into memory allocation issues with many applications. Let me
> > give you a simple example
> > 
> > t480s: ~  $ dd if=/dev/zero of=/dev/null ibs=100G
> > dd: memory exhausted by input buffer of size 107374182400 bytes (100 GiB)
> > 
> > So indicating the size of the failing allocation might be just good enough.
> > For the other parts it's usually just "the way the OS was configured, it
> > does not think it can allow this allocation".  
> 
> Does it also get complicated by the use of CGroup?

And if it's not complex enough, add to that NUMA node binding,
which introduces additional limitations on RAM size that
can be allocated.

> Dave
> 
> > -- 
> > Thanks,
> > 
> > David / dhildenb
> >