Re: [PATCH 6/7] migration: Fix arrays of pointers in JSON writer

[Peter Xu <peterx@redhat.com>]

On Wed, Jan 08, 2025 at 10:52:30AM -0300, Fabiano Rosas wrote:
> Peter Xu <peterx@redhat.com> writes:
> 
> > On Tue, Jan 07, 2025 at 04:50:24PM -0300, Fabiano Rosas wrote:
> >> Currently, if an array of pointers contains a NULL pointer, that
> >> pointer will be encoded as '0' in the stream. Since the JSON writer
> >> doesn't define a "pointer" type, that '0' will now be an uint64, which
> >> is different from the original type being pointed to, e.g. struct.
> >> 
> >> That mixed-type array shouldn't be compressed, otherwise data is lost
> >> as the code currently makes the whole array have the type of the first
> >> element.
> >> 
> >> While we could disable the array compression when a NULL pointer is
> >> found, the JSON part of the stream still makes part of downtime, so we
> >> should avoid writing unecessary bytes to it.
> >> 
> >> Keep the array compression in place, but break the array into several
> >> type-contiguous pieces if NULL and non-NULL pointers are mixed.
> >
> > Could I request for a sample JSON dump for an example array in the commit
> > log?  This whole solution looks working but is tricky.  A sample could help
> > people understand (e.g. showing the same "name" being dumped multiple
> > times..).
> 
> {"name": "s390_css", "instance_id": 0, "vmsd_name": "s390_css",
>  "version": 1, "fields": [
>    ...,
>    {"name": "css", "array_len": 254, "type": "uint8", "size": 1},
>    {"name": "css", "type": "struct", "struct": {
>     "vmsd_name": "s390_css_img", "version": 1, "fields": [{"name":
>     "chpids", "array_len": 256, "type": "struct", "struct": {"vmsd_name":
>     "s390_chp_info", "version": 1, "fields": [{"name": "in_use", "type":
>     "uint8", "size": 1}, {"name": "type", "type": "uint8", "size": 1},
>     {"name": "is_virtual", "type": "uint8", "size": 1}]}, "size": 3}]},
>     "size": 768},
>    {"name": "css", "type": "uint8", "size": 1},
>    ...
> ]}

Yes something like this would work, thanks.  We could even omit most of the
struct details but only show the important ones:

  {"name": "s390_css", "instance_id": 0, "vmsd_name": "s390_css",
   "version": 1, "fields": [
     ...,
     {"name": "css", "array_len": 254, "type": "uint8", "size": 1},
     {"name": "css", "type": "struct", "struct": {"vmsd_name": "s390_css_img", ... }, "size": 768},
     {"name": "css", "type": "uint8", "size": 1},
     ...
  ]}

> 
> >
> > Side note: I tried to dump a very basic VM's JSON out to disk, it scares me
> > on the size:
> >
> > $ ls -lhS JSON.out 
> > -rw-r--r--. 1 peterx peterx 106K Jan  7 17:18 JSON.out
> >
> > That's a simplest VM with all default stuff, mostly nothing complex.. I may
> > really need to measure how the JSON debug strings affect migration function
> > or perf at some point..
> >
> 
> Agreed.
> 
> >> 
> >> Signed-off-by: Fabiano Rosas <farosas@suse.de>
> >> ---
> >>  migration/vmstate.c          | 33 ++++++++++++++++++++++++++++++++-
> >>  scripts/analyze-migration.py |  9 ++++++++-
> >>  2 files changed, 40 insertions(+), 2 deletions(-)
> >> 
> >> diff --git a/migration/vmstate.c b/migration/vmstate.c
> >> index 52704c822c..a79ccf3875 100644
> >> --- a/migration/vmstate.c
> >> +++ b/migration/vmstate.c
> >> @@ -425,15 +425,19 @@ int vmstate_save_state_v(QEMUFile *f, const VMStateDescription *vmsd,
> >>              int size = vmstate_size(opaque, field);
> >>              uint64_t old_offset, written_bytes;
> >>              JSONWriter *vmdesc_loop = vmdesc;
> >> +            bool is_prev_null = false;
> >>  
> >>              trace_vmstate_save_state_loop(vmsd->name, field->name, n_elems);
> >>              if (field->flags & VMS_POINTER) {
> >>                  first_elem = *(void **)first_elem;
> >>                  assert(first_elem || !n_elems || !size);
> >>              }
> >> +
> >>              for (i = 0; i < n_elems; i++) {
> >>                  void *curr_elem = first_elem + size * i;
> >>                  const VMStateField *inner_field;
> >> +                bool is_null;
> >> +                int max_elems = n_elems - i;
> >>  
> >>                  old_offset = qemu_file_transferred(f);
> >>                  if (field->flags & VMS_ARRAY_OF_POINTER) {
> >> @@ -448,12 +452,39 @@ int vmstate_save_state_v(QEMUFile *f, const VMStateDescription *vmsd,
> >>                       * not follow.
> >>                       */
> >>                      inner_field = vmsd_create_fake_nullptr_field(field);
> >> +                    is_null = true;
> >>                  } else {
> >>                      inner_field = field;
> >> +                    is_null = false;
> >> +                }
> >> +
> >> +                /*
> >> +                 * Due to the fake nullptr handling above, if there's mixed
> >> +                 * null/non-null data, it doesn't make sense to emit a
> >> +                 * compressed array representation spanning the entire array
> >> +                 * because the field types will be different (e.g. struct
> >> +                 * vs. uint64_t). Search ahead for the next null/non-null
> >> +                 * element and start a new compressed array if found.
> >> +                 */
> >> +                if (field->flags & VMS_ARRAY_OF_POINTER &&
> >> +                    is_null != is_prev_null) {
> >> +
> >> +                    is_prev_null = is_null;
> >> +                    vmdesc_loop = vmdesc;
> >> +
> >> +                    for (int j = i + 1; j < n_elems; j++) {
> >> +                        void *elem = *(void **)(first_elem + size * j);
> >> +                        bool elem_is_null = !elem && size;
> >> +
> >> +                        if (is_null != elem_is_null) {
> >> +                            max_elems = j - i;
> >> +                            break;
> >> +                        }
> >> +                    }
> >>                  }
> >>  
> >>                  vmsd_desc_field_start(vmsd, vmdesc_loop, inner_field,
> >> -                                      i, n_elems);
> >> +                                      i, max_elems);
> >>  
> >>                  if (inner_field->flags & VMS_STRUCT) {
> >>                      ret = vmstate_save_state(f, inner_field->vmsd,
> >> diff --git a/scripts/analyze-migration.py b/scripts/analyze-migration.py
> >> index 4836920ddc..9138e91a11 100755
> >> --- a/scripts/analyze-migration.py
> >> +++ b/scripts/analyze-migration.py
> >> @@ -497,7 +497,14 @@ def read(self):
> >>                      raise Exception("internal index of data field unmatched (%d/%d)" % (len(a), int(field['index'])))
> >>                  a.append(field['data'])
> 
> There's actually a bug here, the code above does:
> 
>   if len(a) != int(field['index']):
>       raise Exception()
> 
> Which only works with this patch because the compressed array happens to
> come first.

I think it will work no matter how it's ordered after your patch?  IOW I'd
hope it'll keep working if the 1st is a nullptr:

     {"name": "css", "type": "uint8", "size": 1},
     {"name": "css", "type": "struct", "struct": {"vmsd_name": "s390_css_img", ... }, "size": 768},
     {"name": "css", "array_len": 254, "type": "uint8", "size": 1},

Because IIUC the python script will parse each of the lines above into a
VMSD field.

> 
> >>              else:
> >> -                self.data[field['name']] = field['data']
> >> +                # There could be multiple entries for the same field
> >> +                # name, e.g. when a compressed array was broken in
> >> +                # more than one piece.
> >> +                if (field['name'] in self.data and
> >> +                    type(self.data[field['name']]) == list):
> >> +                    self.data[field['name']].append(field['data'])
> >> +                else:
> >> +                    self.data[field['name']] = field['data']
> >
> > Do we realy need these script changes?  I thought VMSDFieldStruct always
> > breaks array_len field into "index" based anyway?
> >
> >         new_fields = []
> >         for field in self.desc['struct']['fields']:
> >             if not 'array_len' in field:
> >                 new_fields.append(field)
> >                 continue
> >             array_len = field.pop('array_len')
> >             field['index'] = 0
> >             new_fields.append(field)
> >             for i in range(1, array_len):
> >                 c = field.copy()
> >                 c['index'] = i
> >                 new_fields.append(c)
> >
> >         self.desc['struct']['fields'] = new_fields
> 
> This code is about decompressing the array, it doesn't handle multiple
> entries with the same name. See the JSON I posted up there.
> 
> This makes the single:
> 
>   {"name": "css", "array_len": 254, "type": "uint8", "size": 1},
> 
> become multiple:
> 
>   {"name": "css", "index": 0, "type": "uint8", "size": 1},
>   {"name": "css", "index": 1, "type": "uint8", "size": 1},
>   ...
>   {"name": "css", "index": 253, "type": "uint8", "size": 1},

Correct.

I think that means for each of the break-down entries there'll be an
"index" if it's an array.  What you changed above is the case where "index"
is not available, which is processing the non-array entry.  Why does that
need change?  What happens if you run this without the python part you
changed in this patch?

-- 
Peter Xu