json_parse_number broken by compiler optimization

[Josef 'Jeff' Sipek]

On Tue, Mar 30, 2021 at 13:34:54 -0400, Josef 'Jeff' Sipek wrote:
> On Tue, Mar 30, 2021 at 17:53:27 +0200, Christian Ehrhardt wrote:
> > Hi,
> > the recent Ubuntu (re)builds uncovered an issue with dovecot 1:2.3.13+dfsg1-1
> > build log: https://launchpadlibrarian.net/529849650/buildlog_ubuntu-hirsute-amd64.dovecot_1%3A2.3.13+dfsg1-1build1_BUILDING.txt.gz
> > A coworker tried 2.3.14 but got the same result.
> > 
> > What fails is the json_parser build time test like:
> >   test-json-parser.c:161: Assert(#25) failed:
> > null_strcmp(json_output[pos].value, value) == 0
> > 
> > I was looking into that a bit more and what I found is that it is
> > dependent on the new toolchain
> > of gcc 10.2.0-1.
> 
> FWIW, I managed to reproduce it on FreeBSD with gcc 11, so the good news for
> you is that it isn't Ubuntu specific :)
> 
> I'll debug further.

The culprit seems to be LTO.  If you disable LTO, everything should work
just fine.  So, I think that'll be the "official" workaround - and a much
better one than disabling optimization completely.

Now, the big question is, is something in the test breaking or is the parser
itself somehow triggering this.

Jeff.

> 
> Thanks again for the report,
> 
> Jeff.
> 
> > 
> > Not all calls to json_parse_* fail, e.g. the first one looks all good and passes
> > I was iterating the tests using a report function defined like
> > 
> > (gdb) define repcon
> > >c
> > >p pos
> > >p json_output[pos].type
> > >p type
> > >p json_output[pos].value
> > >p value
> > >call null_strcmp(json_output[pos].value, value)
> > >end
> > 
> > The first one to be bad was:
> > Breakpoint 2, test_json_parser_success (full_size=<optimized out>) at
> > test-json-parser.c:161
> > 161 test_assert_idx(null_strcmp(json_output[pos].value, value) == 0, pos);
> > $84 = 25
> > $85 = JSON_TYPE_NUMBER
> > $86 = JSON_TYPE_NUMBER
> > $87 = 0x555555633b25 "-12.456"
> > $88 = 0x555555693110 ""
> > $89 = 45
> > 
> > Earlier and later parsing was happy, for example
> > 
> > Breakpoint 2, test_json_parser_success (full_size=<optimized out>) at
> > test-json-parser.c:161
> > 161 test_assert_idx(null_strcmp(json_output[pos].value, value) == 0, pos);
> > $90 = 27
> > $91 = JSON_TYPE_NUMBER
> > $92 = JSON_TYPE_NUMBER
> > $93 = 0x555555633b32 "12.456e9"
> > $94 = 0x555555693110 "12.456e9"
> > $95 = 0
> > (gdb)
> > 
> > 
> > We have two things we compare here.
> > 1. json_output[] which is a static define and for this value is
> >    67 »···{ JSON_TYPE_NUMBER, "-12.456" },
> > 2. the return value that json_parse_next returns.
> >    25 »···"  \"sub2\":-12.456,\n"
> >   148 »···»···»···»···ret = json_parse_next(parser, &type, &value);
> > 
> > I tried a non negative number and got success which is suspicious
> > 
> > Breakpoint 5, test_json_parser_success (full_size=<optimized out>) at
> > test-json-parser.c:164
> > 164 test_assert_idx(null_strcmp(json_output[pos].value, value) == 0, pos);
> > $122 = 25
> > $123 = JSON_TYPE_NUMBER
> > $124 = JSON_TYPE_NUMBER
> > $125 = 0x555555633b2c "12.456"
> > $126 = 0x555555693110 "12.456"
> > $127 = 0
> > (gdb)
> > 
> > 
> > Also the return value otherwise LGTM, it is recognized as a number:
> > 540 } else if ((ret = json_parse_number(parser, value_r)) >= 0) {
> > (gdb) n
> > 541 *type_r = JSON_TYPE_NUMBER;
> > 
> > So parsing worked one might think, but the value was not set.
> > 
> > (gdb) p *value_r
> > $833 = 0x555555693110 ""
> > 
> > It turns out json_parse_number is too optimized.
> > Too optimized for debugging further into json_parse_number.
> > But also optimization is causing the issue.
> > With -O0 or -O1 it works fine, even putting the lower optimization
> > level just around json_parse_number it is enough to avoid the issue.
> > 
> > 
> > Not a fix but "avoidance" for now:
> > 
> > --- a/src/lib/json-parser.c
> > +++ b/src/lib/json-parser.c
> > @@ -386,6 +386,8 @@ static int json_parse_int(struct json_pa
> >   return 1;
> >  }
> > 
> > +#pragma GCC push_options
> > +#pragma GCC optimize ("-O0")
> >  static int json_parse_number(struct json_parser *parser, const char **value_r)
> >  {
> >   int ret;
> > @@ -415,6 +417,7 @@ static int json_parse_number(struct json
> >   *value_r = str_c(parser->value);
> >   return 1;
> >  }
> > +#pragma GCC pop_options
> > 
> >  static int json_parse_atom(struct json_parser *parser, const char *atom)
> >  {
> > 
> > 
> > I was trying to see what happened and tracked each str_append_c
> > that happened inside json_parse_number.
> > A breakpoint on str_append_c and backtraces showed that we clearly do the
> > "right" parsing. From the backtraces I've seen it does (just from positions):
> > 
> > For example
> > (gdb) bt
> > #0  str_append_c (chr=<optimized out>, str=0x5555556930d0) at str.h:44
> > #1  json_parse_digits (parser=<optimized out>) at json-parser.c:366
> > => that is a digit
> > 
> > It does "- digit digit dot digit digit digit"
> > 
> > This perfectly matches the input value, but still the result stays empty:
> > (gdb) p (char*)(str->data)
> > $738 = 0x555555693110 ""
> > 
> > In comparison - when parsing the latter 12.456e9 which works I see two
> > odd things:
> > 1. the value isn't cleared despite
> >     str_truncate(parser->value, 0);
> >    at the beginning of json_parse_number
> > 2. the value (this time) is read exactly as I assumed
> > 
> > Breakpoint 17, json_parse_number (value_r=0x7fffffffe1a0,
> > parser=0x555555692ee0) at json-parser.c:394
> > 394 str_truncate(parser->value, 0);
> > Breakpoint 18, str_append_c (chr=<optimized out>, str=0x5555556930d0)
> > at str.h:44
> > 44 buffer_append_c(str, chr);
> > (gdb) p (char*)(str->data)
> > $741 = 0x555555693110 "sub3"
> > 
> > ^^ not cleared, still has "sub3" even later on
> > 
> > Breakpoint 18, str_append_c (chr=<optimized out>, str=0x5555556930d0)
> > at str.h:44
> > 44 buffer_append_c(str, chr);
> > (gdb) p (char*)(str->data)
> > $742 = 0x555555693110 "1ub3"
> > 
> > ^^ reading left to right one char at a time "over the old value",
> > which didn't update the data.
> > 
> > So it seems in json_parse_number the ordering of
> > a) str_truncate(parser->value, 0);
> > b) the parsing itself
> > is messed up by optimization.
> > 
> > We already know from the above that reducing just this function to -O1 or less
> > avoids the issue. But what is it really - I don't know?
> > 
> > Maybe you have a better idea what is going on, the issue should be
> > reproducible when building
> > dovecot with the toolchain present on the latest Ubuntu being 21.04
> > (Hirsute) right now.
> > 
> > --
> > Christian Ehrhardt
> > Staff Engineer, Ubuntu Server
> > Canonical Ltd