Re: [PATCH] vhost-user: add separate memslot counter for vhost-user

[Michael S. Tsirkin]

On Tue, Oct 13, 2020 at 08:58:59PM -0400, Raphael Norwitz wrote:
> On Tue, Oct 6, 2020 at 5:48 AM Igor Mammedov <imammedo@redhat.com> wrote:
> >
> > On Mon, 28 Sep 2020 21:17:31 +0800
> > Jiajun Chen <chenjiajun8@huawei.com> wrote:
> >
> > > Used_memslots is equal to dev->mem->nregions now, it is true for
> > > vhost kernel, but not for vhost user, which uses the memory regions
> > > that have file descriptor. In fact, not all of the memory regions
> > > have file descriptor.
> > > It is usefully in some scenarios, e.g. used_memslots is 8, and only
> > > 5 memory slots can be used by vhost user, it is failed to hot plug
> > > a new memory RAM because vhost_has_free_slot just returned false,
> > > but we can hot plug it safely in fact.
> >
> > I had an impression that all guest RAM has to be shared with vhost,
> > so combination of anon and fd based RAM couldn't work.
> > Am I wrong?
> 
> I'm not sure about the kernel backend, but I've tested adding anon
> memory to a VM with a vhost-user-scsi device and it works (eventually
> the VM crashed, but I could see the guest recognized the anon RAM).
> The vhost-user code is designed to work with both. I'm not sure I see
> a use case, but if there is one, this would be a valid issue. Maybe
> Jiajun or Jianjay can elaborate.

Hmm does not vhost-user skip all regions that do not have an fd?


        mr = vhost_user_get_mr_data(reg->userspace_addr, &offset, &fd);
        if (fd > 0) {
            if (track_ramblocks) {
                assert(*fd_num < VHOST_MEMORY_BASELINE_NREGIONS);
                trace_vhost_user_set_mem_table_withfd(*fd_num, mr->name,
                                                      reg->memory_size,
                                                      reg->guest_phys_addr,
                                                      reg->userspace_addr,
                                                      offset);
                u->region_rb_offset[i] = offset;
                u->region_rb[i] = mr->ram_block;
            } else if (*fd_num == VHOST_MEMORY_BASELINE_NREGIONS) {
                error_report("Failed preparing vhost-user memory table msg");
                return -1;
            }
            vhost_user_fill_msg_region(&region_buffer, reg, offset);
            msg->payload.memory.regions[*fd_num] = region_buffer;
            fds[(*fd_num)++] = fd;
        } else if (track_ramblocks) {
            u->region_rb_offset[i] = 0;
            u->region_rb[i] = NULL;
        }



In your test, is it possible that you were lucky and guest did not send
any data from anon memory to the device?



> >
> > >
> > > --
> > > ChangeList:
> > > v3:
> > > -make used_memslots a member of struct vhost_dev instead of a global 
> > > static value
> > it's global resource, so why?
> 
> I suggested it because I thought it made the code a little cleaner.
> I'm not opposed to changing it back, or having it stored at the
> vhost_user level.