Re: crashing socket tests

[Vladimir T. <vtz...@gm...>]

Hi Bruno,

On Sun, May 28, 2017 at 5:20 PM, Vladimir Tzankov <vtz...@gm...>
wrote:

> > The stream object has been moved due to compaction but strm_encoding has
>> > either not moved (unlikely - its heap has shrunk) or has not been
>> updated.
>>
>> I'd also guess that it has not been updated. Is the reclength of the
>> stream appropriate? Has the reclength been changed since the
>> allocate_stream
>> call?
>>
>
> Stream record looks ok but encoding object is not getting marked during GC.
>


The problem is GC mark phase.

Streams with strmtype equal to Rectype_Weakpointer or Rectype_Weakpointer
are not getting marked since SXrecord_nonweak_length() returns 0 for them!

Here is gdb output for this case:

(gdb) p /x *(Stream)0xb0000005ac0
$32 = {header = {_GCself = 0xb0000005ac0, flags = {0xb0000005ac0}},
strmtype = 0x11, strmflags = 0x41, reclength = 0x15, recxlength = 0x60,
strm_rd_by = 0x464b98,
  strm_rd_by_array = 0x464b80, strm_wr_by = 0x464b68, strm_wr_by_array =
0x464b50, strm_rd_ch = 0x462df0, strm_pk_ch = 0x45af68, strm_rd_ch_array =
0x462918,
  strm_rd_ch_last = 0x2000000000a, strm_wr_ch = 0x464b08, strm_wr_ch_array
= 0x464af0, strm_wr_ch_npnl = 0x464b08, strm_wr_ch_array_npnl = 0x464af0,
  strm_wr_ch_lpos = 0x200000000000, strm_other = 0xb0000005b38}
(gdb) p /x *(Record)0xb0000005ac0
$33 = {header = {_GCself = 0xb0000005ac0, flags = {0xb0000005ac0}}, rectype
= 0x11, recflags = 0x41, reclength = 0x6015, recdata = 0xb0000005ad0}
(gdb) p /x Rectype_Weakpointer
$34 = 0x11

And SXrecord_nonweak_length() explicitly uses Record_type() to check for
weak pointers.

This also explains why I experience this only in non-libffcall builds. With
DYNAMIC_FFI, Rectype_Weakpointer is 0x14 because of the three additional
FFI record types.