Re: [Qemu-devel] Regression caused by d2f39ad "exec.c: Ensure right alignment also for file backed ram"

[Igor Mammedov <imammedo@redhat.com>]

On Fri, 21 Oct 2016 16:39:08 +0800
Haozhong Zhang <haozhong.zhang@intel.com> wrote:

> On 10/21/2016 04:28 PM, Haozhong Zhang wrote:
> > Hi,
> >
> > Commit d2f39ad "exec.c: Ensure right alignment also for file backed
> > ram" added an additional alignment requirement besides to the previous
> > page size for the size of the backend file. On x86, the alignment
> > requirement for the size of the backend file is changed from 4KB in
> > QEMU 2.6 to 2MB in QEMU 2.7.
> >
> > This change immediately breaks the usage of "-object
> > memory-backend-file,...,size=$SIZE"  
> 
> I forgot the frontend part, the broken usage should be
>     -object memory-backend-file,id=mem1,mem-path=$FILE,size=$SIZE
>     -device nvdimm,id=nv1,memdev=mem1
> where $SIZE is multiple of 4KB but not 2MB.
> 
> It works on QEMU 2.6, but fails on 2.7 when pc_dimm_memory_plug() tries
> to plug the nvdimm device.
it affect's not only nvdimm but also pc-dimm, 

qemu -m 1G,slots=10,maxmem=10G \
  -object memory-backend-file,id=mem1,mem-path=/tmp,size=0x3e2000
  -device pc-dimm,memdev=mem1
qemu: -device pc-dimm,memdev=mem1: backend memory size must be multiple of 0x200000

As for nvdimms d2f39ad commit changes auto-generated GPA
when 'addr' property is not provided

pre bug:
mhp_pc_dimm_assigned_address 0x100000000
mhp_pc_dimm_assigned_slot 0x0
mhp_pc_dimm_assigned_address 0x1003e2000
mhp_pc_dimm_assigned_slot 0x1

vs post bug:
mhp_pc_dimm_assigned_address 0x100000000
mhp_pc_dimm_assigned_slot 0
mhp_pc_dimm_assigned_address 0x100200000
mhp_pc_dimm_assigned_slot 1

or if 'addr' is provided, qemu will fail with

"address must be aligned to 0x200000 bytes"

so migration from qemu-2.6 and older builds to 2.7 and later isn't possible due to
inability to start target with 2.6-page_based alignment.

> 
> Haozhong
> 
> > on x86, where $SIZE is multiple of 4KB but not 2MB. It works on QEMU 2.6.
> > Could this be considered as a regression?
> >
> > The commit message shows it's for s390. I'm wondering whether the same
> > regression happens on s390 and ARM. If not, could I fix the regression
> > on x86 by keeping using the original alignment only on x86, e.g.
> >
> > modified   exec.c
> > @@ -1254,7 +1254,11 @@ static void *file_ram_alloc(RAMBlock *block,
> >     }
> >
> >     block->page_size = qemu_fd_getpagesize(fd);
> > +#if defined(__x86_64__) || defined(__i386__)
> > +    block->mr->align = block->page_size;
> > +#else
> >     block->mr->align = MAX(block->page_size, QEMU_VMALLOC_ALIGN);
> > +#endif
> >
> >     if (memory < block->page_size) {
> >         error_setg(errp, "memory size 0x" RAM_ADDR_FMT " must be equal
> > to "
> >
> >
> > Thanks,
> > Haozhong
> >  
>