Re: VMA's not getting merged

[Jarkko Sakkinen <jarkko@kernel.org>]

On Mon, May 16, 2022 at 09:45:56PM +0300, Jarkko Sakkinen wrote:
> On Mon, May 16, 2022 at 07:13:42AM -0700, Dave Hansen wrote:
> > On 5/14/22 18:20, Jarkko Sakkinen wrote:
> > > Looks like VMA's do not get merged properly:
> > ...
> > > 7f8f0800b000-7f8f08014000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f08014000-7f8f08025000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f08025000-7f8f08046000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f08046000-7f8f0804a000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f0804a000-7f8f0804b000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f0804b000-7f8f0804c000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f0804c000-7f8f0804d000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > > 7f8f0804d000-7f8f0804e000 rw-s 00000000 00:05 84                         /dev/sgx_enclave
> > 
> > I remember some issue with VM_IO VMAs not getting merged, but I don't
> > remember how we fixed or addressed it.
> > 
> > Seems like something we can and should fix, though.
> 
> Access pattern here is series of small adjacent mmaps (brk()).
> 
> During upstreaming vma_close callback was eliminated to allow VMA's getting
> merged. I'll try to trace the condition in the call path triggering this.
> 
> BR, Jarkko

I wrote a small bpftrace script:

#include <linux/fs.h>

k:vma_merge
/strncmp(str(((struct vm_area_struct *)arg1)->vm_file->f_path.dentry->d_name.name), "sgx_enclave", 11) == 0/
{
	@merge[tid] = true;
}

kr:vma_merge /@merge[tid] && retval > 0/
{
	@merged = count();
	delete(@merge[tid]);
}

kr:vma_merge /@merge[tid]/
{
	@unmerged = count();
	delete(@merge[tid]);
}

When running with the same payload, I get:

@merged: 0
@unmerged: 209

So it gets down to vma_merge but each time the function returns NULL.

BR, Jarkko