Re: [Xen-devel] [PATCH v2 12/18] xen: setup Xen specific data for PVH

[Juergen Gross <jgross@suse.com>]

On 30/10/2018 12:23, Roger Pau Monné wrote:
> On Mon, Oct 29, 2018 at 03:19:34PM +0100, Juergen Gross wrote:
>> On 29/10/2018 13:57, Roger Pau Monné wrote:
>>> On Fri, Oct 19, 2018 at 06:39:50PM +0200, Juergen Gross wrote:
>>>> On 19/10/2018 18:10, Roger Pau Monné wrote:
>>>>> On Tue, Oct 09, 2018 at 01:03:11PM +0200, Juergen Gross wrote:
>>>>>> Initialize the needed Xen specific data. This is:
>>>>>>
>>>>>> - the Xen start of day page containing the console and Xenstore ring
>>>>>>   page PFN and event channel
>>>>>> - the grant table
>>>>>> - the shared info page
>>>>>>
>>>>>> Set the RSDP address for the guest from the start_info page passed
>>>>>> as boot parameter.
>>>>>>
>>>>>> Signed-off-by: Juergen Gross <jgross@suse.com>
>>>>>> ---
>>>>>>  grub-core/kern/i386/xen/pvh.c | 107 ++++++++++++++++++++++++++++++++++++++++++
>>>>>>  1 file changed, 107 insertions(+)
>>>>>>
>>>>>> diff --git a/grub-core/kern/i386/xen/pvh.c b/grub-core/kern/i386/xen/pvh.c
>>>>>> index b4933b454..93ed68245 100644
>>>>>> --- a/grub-core/kern/i386/xen/pvh.c
>>>>>> +++ b/grub-core/kern/i386/xen/pvh.c
>>>>>> @@ -24,6 +24,7 @@
>>>>>>  #include <grub/xen.h>
>>>>>>  #include <grub/i386/linux.h>
>>>>>>  #include <grub/machine/kernel.h>
>>>>>> +#include <xen/hvm/params.h>
>>>>>>  #include <xen/memory.h>
>>>>>>  
>>>>>>  struct xen_machine_mmap_entry
>>>>>> @@ -39,6 +40,7 @@ static struct { char _entry[32]; } hypercall_page[128]
>>>>>>    __attribute__ ((aligned (GRUB_XEN_PAGE_SIZE)));
>>>>>>  
>>>>>>  static grub_uint32_t xen_cpuid_base;
>>>>>> +static struct start_info grub_xen_start_page;
>>>>>>  static struct xen_machine_mmap_entry map[128];
>>>>>>  static unsigned int nr_map_entries;
>>>>>>  
>>>>>> @@ -104,6 +106,36 @@ grub_xen_hypercall (grub_uint32_t callno, grub_uint32_t a0,
>>>>>>    return __res;
>>>>>>  }
>>>>>>  
>>>>>> +static grub_uint32_t
>>>>>> +grub_xen_get_param (int idx)
>>>>>> +{
>>>>>> +  struct xen_hvm_param xhv;
>>>>>> +  int r;
>>>>>> +
>>>>>> +  xhv.domid = DOMID_SELF;
>>>>>> +  xhv.index = idx;
>>>>>> +  r = grub_xen_hypercall (__HYPERVISOR_hvm_op, HVMOP_get_param,
>>>>>> +			  (grub_uint32_t) (&xhv), 0, 0, 0, 0);
>>>>>> +  if (r < 0)
>>>>>> +    grub_xen_early_halt ();
>>>>>> +  return xhv.value;
>>>>>> +}
>>>>>> +
>>>>>> +static void *
>>>>>> +grub_xen_add_physmap (unsigned int space, void *addr)
>>>>>> +{
>>>>>> +  struct xen_add_to_physmap xatp;
>>>>>> +
>>>>>> +  xatp.domid = DOMID_SELF;
>>>>>> +  xatp.idx = 0;
>>>>>> +  xatp.space = space;
>>>>>> +  xatp.gpfn = (grub_addr_t) addr >> GRUB_XEN_LOG_PAGE_SIZE;
>>>>>> +  if (grub_xen_hypercall (__HYPERVISOR_memory_op, XENMEM_add_to_physmap,
>>>>>> +			  (grub_uint32_t) (&xatp), 0, 0, 0, 0))
>>>>>> +    grub_xen_early_halt ();
>>>>>> +  return addr;
>>>>>> +}
>>>>>> +
>>>>>>  static void
>>>>>>  grub_xen_sort_mmap (void)
>>>>>>  {
>>>>>> @@ -190,12 +222,87 @@ grub_xen_get_mmap (void)
>>>>>>    grub_xen_sort_mmap ();
>>>>>>  }
>>>>>>  
>>>>>> +static grub_uint64_t
>>>>>> +grub_xen_find_page (grub_uint64_t start)
>>>>>> +{
>>>>>> +  unsigned int i, j;
>>>>>> +  grub_uint64_t last = start;
>>>>>> +
>>>>>> +  /* Try to find a e820 map hole below 4G. */
>>>>>
>>>>> Doing this is kind of dangerous, what if you end up placing something
>>>>> on top of an MMIO region (either emulated or from a real passthrough
>>>>> device)?
>>>>
>>>> Shouldn't those be marked as "Reserved" in the memory map?
>>>
>>> I don't think BARs are guaranteed to be in areas marked as reserved in
>>> the memory map. Unless you also scan for PCI devices and make sure
>>> there's no device with a BAR in the area you are attempting to
>>> populate I think the above is not safe.
>>
>> So either I can add scanning the PCI bus (which wouldn't be too
>> hard IMO), or we require Xen tools to add memory map entries with
>> "Reserved" attribute for passed-through PCI device's MMIO-areas
>> (we can still do that as PCI pass-through for PVH isn't possible
>> yet AFAIK).
> 
> Ideally (and that's kind of far away from what we are now), I would
> like to have an interface to request Xen for a range of gfns available
> to map stuff in them (grants/foreign mappings/shared page...). That
> interface could be a simple hypercall that would return such range, or
> an hypercall that could be used to fetch something akin to an
> 'extended memory map' with specific Xen information (like such
> regions).
> 
> In both cases this requires Xen having a clearer picture of the p2m,
> because any of the above solutions cannot rely on scanning the p2m
> table in order to figure out what's where.
> 
> So the only change I would request is that if you use a RAM page you
> update the memory map stored in Xen to match the new layout, by using
> the XENMEM_set_memory_map hypercall.

Okay, this is simple.

Nevertheless I'm adding the already finished patch for scanning the
PCI devices to find MMIO areas.


Juergen