Re: [PATCH v2] hw/riscv: virt: bugfix the memory-backend-file command is invalid

[Igor Mammedov]

On Mon, 18 Oct 2021 10:17:45 +0800
Bin Meng <bmeng.cn@gmail.com> wrote:

> Hi Igor,
> 
> On Fri, Oct 15, 2021 at 8:59 PM Igor Mammedov <imammedo@redhat.com> wrote:
> >
> > On Fri, 15 Oct 2021 17:25:01 +0800
> > Bin Meng <bmeng.cn@gmail.com> wrote:
> >  
> > > On Fri, Oct 15, 2021 at 4:52 PM limingwang (A) <limingwang@huawei.com> 
> > > wrote:  
> > > >
> > > >
> > > > On Wed, Oct 13, 2021 at 22:41 PM Bin Meng <bin.meng@windriver.com> 
> > > > wrote:  
> > > > >
> > > > > On Tue, Oct 12, 2021 at 9:46 AM MingWang Li <limingwang@huawei.com> 
> > > > > wrote:  
> > > > > >
> > > > > > From: Mingwang Li <limingwang@huawei.com>
> > > > > >
> > > > > > When I start the VM with the following command:
> > > > > > $ ./qemu-system-riscv64 -M virt,accel=kvm -m 4096M -cpu host 
> > > > > > -nographic \
> > > > > >     -name guest=riscv-guset \
> > > > > >     -smp 2 \
> > > > > >     -bios none \
> > > > > >     -kernel ./Image \
> > > > > >     -drive file=./guest.img,format=raw,id=hd0 \
> > > > > >     -device virtio-blk-device,drive=hd0 \
> > > > > >     -append "root=/dev/vda rw console=ttyS0 earlycon=sbi" \
> > > > > >     -object  
> > > > > memory-backend-file,id=mem,size=4096M,mem-path=/dev/hugepages,share=on
> > > > >  \  
> > > > > >     -numa node,memdev=mem -mem-prealloc \
> > > > > >     -chardev socket,id=char0,path=/mnt/vhost-net0 \
> > > > > >     -netdev type=vhost-user,id=mynet1,chardev=char0,vhostforce \
> > > > > >     -device
> > > > > > virtio-net-pci,mac=52:54:00:00:00:01,netdev=mynet1,mrg_rxbuf=on,csum=o
> > > > > > n,guest_csum=on,guest_ecn=on \
> > > > > >
> > > > > > Then, QEMU displays the following error information:
> > > > > > qemu-system-riscv64: Failed initializing vhost-user memory map,
> > > > > > consider using -object memory-backend-file share=on  
> > > > >
> > > > > I see your command line parameters already contain "-object 
> > > > > memory-backend-file
> > > > > share=on", so this error message is not accurate.  
> > > >
> > > > QEMU uses this command to alloc fd in the 
> > > > "memory_region_init_ram_from_file" function
> > > > and assigns the value of fd to mr->ram_block-fd. If the QEMU uses the 
> > > > default memory to
> > > > initialize the system, the QEMU cannot obtain the fd in the 
> > > > "vhost_user_mem_section_filter"
> > > > function when initializing the vhost-user. As a result, an error is 
> > > > reported in the "vhost_user_fill_set_mem_table_msg"
> > > > function.
> > > >
> > > > Because of the above bug, even if "-object memory-backend-file 
> > > > share=on" is added to the command line,
> > > > the QEMU still reports an error.  
> > >
> > > Yes, what I meant is that QEMU should not report such inaccurate
> > > messages because of some random codes elsewhere.
> > >
> > > With current message, it suggested user use "-object
> > > memory-backend-file share=on" in the command line, but it is already
> > > used. So this is a false alarm. The "bug" is somewhere else.  
> >
> > bug is in using memory_region_init_ram(),
> > which can't possibly handle vhost-user, and can't work as expected with
> > '-numa node,memdev' options.
> > Before main ram infrastructure was converted to memdev,
> > one should have used memory_region_allocate_system_memory() for
> > allocating main RAM, so numa usecase was broken from the start.
> > Later it old API was dropped in favor of more flexible/generic
> > MachineState::ram approach (see commits 68a86dc15ccd..f0530f14c7c35d).  
> 
> Thanks for the detailed pointers.
> 
> I wonder if it is possible to make the error message to be clearer, so
> instead of having
> 
>     "qemu-system-riscv64: Failed initializing vhost-user memory map,
> consider using -object memory-backend-file share=on"
> 
> can we do:
> 
>     "qemu-system-riscv64: Failed initializing vhost-user memory map,
> considering using MachineState::ram instead of manually initializing
> RAM memory region."
> 
> which is more straightforward?

It would only make sense in context of this thread and
give a hint to a developer how to fix bug in machine code
but won't give a clue to the end user what's wrong with
configuration.

Maybe following would be better:
 "Failed initializing vhost-user memory map, vhost requires shared system 
memory."
 "See 'System Emulation::Device Emulation::Emulated Devices::vhost-user back 
ends'
  chapter in QEMU manual."

the chapter gives an example how to correctly configure vhost-user
(albeit using old style)

 
> >
> >
> > Modulo commit message, patch looks good to me and does what
> > every machine should do. (I though that I've converted every
> > existing to generalized MachineState::ram but it looks like
> > riscv was missed).  
> 
> Indeed all riscv boards are doing the same thing.
> 
> >
> > So we can model commit message after bd457782b3b0a,
> > and also add that the patch fixes broken -numa node,memdev case,
> > which never properly worked. It also opens possibility to
> > use vhost-user/virtiosf with main RAM if main RAM is
> > provided explicitly via machine.memory-backend option
> > with shared memory backend.
> >
> > Btw: is there other riscv machines that allocate RAM directly?
> > (if yes, those should be fixed as well, a patch per machine)
> >  
> 
> I will see if I can get some patches to fix other riscv machines.
> 
> Regards,
> Bin
>