RE: [PATCH for-4.17 1/2] docs: Document the minimal requirement of static heap

[Henry Wang <Henry.Wang@xxxxxxx>]

Hi Julien,

> -----Original Message-----
> From: Julien Grall <julien@xxxxxxx>
> >>> +Users should be mindful that the static heap should at least satisfy the
> >>> +allocation of the P2M maps for all guests. Currently, the minimal
> >> requirement
> >>> +of per-domain P2M pages pool is in-sync with function
> >>> +libxl__get_required_paging_memory() (for xl-created domUs) and
> >>> +domain_p2m_pages() (for dom0less domUs), that is, 1MB per vCPU,
> plus
> >> 4KiB per
> >>> +MiB of RAM for the P2M map, and plus 512KiB to cover extended
> regions.
> >>
> >> I think this wording is OK if the feature is a tech preview. However, if
> >> this is security supported, we need to provide some more details about
> >> the size.
> >>
> >> In particular, this doesn't tell a user how they can find the size that
> >> would fit them. Can this be decided with a formula?
> > My feeling of the formula would be:
> >
> > Total heap size needed per guest =  1MB * num_guest_vcpu +
> >      4KB * guest_ram_size_in_mb + 512KB +
> >      the memory allocated from heap by xzalloc/xzalloc_array for
>       various uses
> >      for example alloc_domain_struct(), d->shared_info, evtchn_bucket, etc.
> 
> There are also some pages allocated using alloc_{xen,dom}heap_pages().
> We also need to take into account runtime allocation done by some
> hypercalls (I can't remember which one) or subsystem like OPTee.
> 
> In addition to that, you also have memory for the system. E.g
> frametables, Xen page-tables, various driver allocations...
> 
> >
> > Is this formula somehow make sense to you? I think we need to have a
> > rough estimation of the last part (boot time allocation) though.
> 
> That's going to be hard. It will vary depending on your system and this
> could change in the future as we add more features. For instance, I
> expect the PCI passthrough will need some memory to keep track of all
> the devices.
> 
> I am worry the formula will become complex. Ideally we need to have a
> very simple formula. If that's not possible, then we need to provide a
> way for the user to estimate it at runtime (like what I suggested before).

I agree, I think the simple formula can only be achieved is we have an
estimation of the worst case scenario of those scattered memory usages.
I remember I once had a try so let me try to find the results back that time...

I am also very interested in the method that you proposed to provide a
mechanism for users to get the system memory allocation at runtime. But
IIUC this needs some work in another series. Could you please confirm if I
am understanding correctly? Or probably Xen has some mechanisms that
I am likely unaware? Thanks!

Kind regards,
Henry

> 
> Cheers,
> 
> --
> Julien Grall