Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20190523143508.25387-3-armbru@redhat.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The assert checking if the value of lexer->state in next_state(),
which is used as an index to the 'json_lexer' array, incorrectly
checks for an index value less than or equal to ARRAY_SIZE(json_lexer).
Fix assert so that it just checks for an index less than the array size.
Signed-off-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <1553169472-25325-1-git-send-email-liam.merwick@oracle.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Li Qiang <liq3ea@gmail.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Commit 8bca4613 added support for %% in json strings when interpolating,
but in doing so broke handling of % when not interpolating.
When parse_string() is fed a string token containing '%', it skips the
'%' regardless of ctxt->ap, i.e. even it's not interpolating. If the
'%' is the string's last character, it fails an assertion. Else, it
"merely" swallows the '%'.
Fix parse_string() to handle '%' specially only when interpolating.
To gauge the bug's impact, let's review non-interpolating users of this
parser, i.e. code passing NULL context to json_message_parser_init():
* tests/check-qjson.c, tests/test-qobject-input-visitor.c,
tests/test-visitor-serialization.c
Plenty of tests, but we still failed to cover the buggy case.
* monitor.c: QMP input
* qga/main.c: QGA input
* qobject_from_json():
- qobject-input-visitor.c: JSON command line option arguments of
-display and -blockdev
Reproducer: -blockdev '{"%"}'
- block.c: JSON pseudo-filenames starting with "json:"
Reproducer: https://bugzilla.redhat.com/show_bug.cgi?id=1668244#c3
- block/rbd.c: JSON key pairs
Pseudo-filenames starting with "rbd:".
Command line, QMP and QGA input are trusted.
Filenames are trusted when they come from command line, QMP or HMP.
They are untrusted when they come from from image file headers.
Example: QCOW2 backing file name. Note that this is *not* the security
boundary between host and guest. It's the boundary between host and an
image file from an untrusted source.
Neither failing an assertion nor skipping a character in a filename of
your choice looks exploitable. Note that we don't support compiling
with NDEBUG.
Fixes: 8bca4613e6
Cc: qemu-stable@nongnu.org
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Message-Id: <20190102140535.11512-1-cfergeau@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Tested-by: Richard W.M. Jones <rjones@redhat.com>
[Commit message extended to discuss impact]
Signed-off-by: Markus Armbruster <armbru@redhat.com>
The lexer ignores whitespace like this:
on whitespace on non-ws spontaneously
IN_START --> IN_WHITESPACE --> JSON_SKIP --> IN_START
^ |
\__/ on whitespace
This accumulates a whitespace token in state IN_WHITESPACE, only to
throw it away on the transition via JSON_SKIP to the start state.
Wasteful. Go from IN_START to IN_START on whitespace directly,
dropping the whitespace character.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180831075841.13363-7-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180831075841.13363-6-armbru@redhat.com>
When the lexer chokes on an input character, it consumes the
character, emits a JSON error token, and enters its start state. This
can lead to suboptimal error recovery. For instance, input
0123 ,
produces the tokens
JSON_ERROR 01
JSON_INTEGER 23
JSON_COMMA ,
Make the lexer skip characters after a lexical error until a
structural character ('[', ']', '{', '}', ':', ','), an ASCII control
character, or '\xFE', or '\xFF'.
Note that we must not skip ASCII control characters, '\xFE', '\xFF',
because those are documented to force the JSON parser into known-good
state, by docs/interop/qmp-spec.txt.
The lexer now produces
JSON_ERROR 01
JSON_COMMA ,
Update qmp-test for the nicer error recovery: QMP now reports just one
error for input %p instead of two. Also drop the newline after %p; it
was needed to tease out the second error.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180831075841.13363-5-armbru@redhat.com>
[Conflict with commit ebb4d82d88 resolved]
The lexer uses macro TERMINAL_NEEDED_LOOKAHEAD() to decide whether a
state transition consumes the input character. It returns true when
the state transition is defined with the TERMINAL() macro. To detect
that, it checks whether input '\0' would have resulted in the same
state transition, and the new state is not IN_ERROR.
Why does that even work? For all states, the new state on input '\0'
is either IN_ERROR or defined with TERMINAL(). If the state
transition equals the one we'd get for input '\0', it goes to IN_ERROR
or to the argument of TERMINAL(). We never use TERMINAL(IN_ERROR),
because it makes no sense. Thus, if it doesn't go to IN_ERROR, it
must be defined with TERMINAL().
Since this isn't quite confusing enough, we negate the result to get
@char_consumed, and ignore it when @flush is true.
Instead of deriving the lookahead bit from the state transition, make
it explicit. This is easier to understand, and a bit more flexible,
too.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180831075841.13363-4-armbru@redhat.com>
When the lexer isn't in its start state at the end of input, it's
working on a token. To flush it out, it needs to transit to its start
state on "end of input" lookahead.
There are two ways to the start state, depending on the current state:
* If the lexer is in a TERMINAL(JSON_FOO) state, it can emit a
JSON_FOO token.
* Else, it can go to IN_ERROR state, and emit a JSON_ERROR token.
There are complications, however:
* The transition to IN_ERROR state consumes the input character and
adds it to the JSON_ERROR token. The latter is inappropriate for
the "end of input" character, so we suppress that. See also recent
commit a2ec6be72b "json: Fix lexer to include the bad character in
JSON_ERROR token".
* The transition to a TERMINAL(JSON_FOO) state doesn't consume the
input character. In that case, the lexer normally loops until it is
consumed. We have to suppress that for the "end of input" input
character. If we didn't, the lexer would consume it by entering
IN_ERROR state, emitting a bogus JSON_ERROR token. We fixed that in
commit bd3924a33a.
However, simply breaking the loop this way assumes that the lexer
needs exactly one state transition to reach its start state. That
assumption is correct now, but it's unclean, and I'll soon break it.
Clean up: instead of breaking the loop after one iteration, break it
after it reached the start state.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180831075841.13363-3-armbru@redhat.com>
The lexer fails to end a valid token when the lookahead character is
beyond '\x7F'. For instance, input
true\xC2\xA2
produces the tokens
JSON_ERROR true\xC2
JSON_ERROR \xA2
This should be
JSON_KEYWORD true
JSON_ERROR \xC2
JSON_ERROR \xA2
instead.
The culprit is
#define TERMINAL(state) [0 ... 0x7F] = (state)
It leaves [0x80..0xFF] zero, i.e. IN_ERROR. Has always been broken.
Fix it to initialize the complete array.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180831075841.13363-2-armbru@redhat.com>
The previous commit makes JSON strings containing '%' awkward to
express in templates: you'd have to mask the '%' with an Unicode
escape \u0025. No template currently contains such JSON strings.
Support the printf conversion specification %% in JSON strings as a
convenience anyway, because it's trivially easy to do.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-58-armbru@redhat.com>
The JSON parser optionally supports interpolation. This is used to
build QObjects by parsing string templates. The templates are C
literals, so parse errors (such as invalid interpolation
specifications) are actually programming errors. Consequently, the
functions providing parsing with interpolation
(qobject_from_jsonf_nofail(), qobject_from_vjsonf_nofail(),
qdict_from_jsonf_nofail(), qdict_from_vjsonf_nofail()) pass
&error_abort to the parser.
However, there's another, more dangerous kind of programming error:
since we use va_arg() to get the value to interpolate, behavior is
undefined when the variable argument isn't consistent with the
interpolation specification.
The same problem exists with printf()-like functions, and the solution
is to have the compiler check consistency. This is what
GCC_FMT_ATTR() is about.
To enable this type checking for interpolation as well, we carefully
chose our interpolation specifications to match printf conversion
specifications, and decorate functions parsing templates with
GCC_FMT_ATTR().
Note that this only protects against undefined behavior due to type
errors. It can't protect against use of invalid interpolation
specifications that happen to be valid printf conversion
specifications.
However, there's still a gaping hole in the type checking: GCC
recognizes '%' as start of printf conversion specification anywhere in
the template, but the parser recognizes it only outside JSON strings.
For instance, if someone were to pass a "{ '%s': %d }" template, GCC
would require a char * and an int argument, but the parser would
va_arg() only an int argument, resulting in undefined behavior.
Avoid undefined behavior by catching the programming error at run
time: have the parser recognize and reject '%' in JSON strings.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-57-armbru@redhat.com>
The recursive descent parser passes along a pointer to
JSONParserContext. It additionally passes a pointer to interpolation
state (a va_alist *) as needed to reach its consumer
parse_interpolation().
Stuffing the latter pointer into JSONParserContext saves us the
trouble of passing it along, so do that.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-56-armbru@redhat.com>
The JSON parser has three public headers, json-lexer.h, json-parser.h,
json-streamer.h. They all contain stuff that is of no interest
outside qobject/json-*.c.
Collect the public interface in include/qapi/qmp/json-parser.h, and
everything else in qobject/json-parser-int.h.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-54-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-53-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-52-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-51-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-50-armbru@redhat.com>
Token count and size limits exist to guard against excessive heap
usage. We check them only after we created the token on the heap.
That's assigning a cowboy to the barn to lasso the horse after it has
bolted. Close the barn door instead: check before we create the
token.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-49-armbru@redhat.com>
The last case where qobject_from_json() & friends return null without
setting an error is empty or blank input. Callers:
* block.c's parse_json_protocol() reports "Could not parse the JSON
options". It's marked as a work-around, because it also covered
actual bugs, but they got fixed in the previous few commits.
* qobject_input_visitor_new_str() reports "JSON parse error". Also
marked as work-around. The recent fixes have made this unreachable,
because it currently gets called only for input starting with '{'.
* check-qjson.c's empty_input() and blank_input() demonstrate the
behavior.
* The other callers are not affected since they only pass input with
exactly one JSON value or, in the case of negative tests, one error.
Fail with "Expecting a JSON value" instead of returning null, and
simplify callers.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-48-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-47-armbru@redhat.com>
json_message_process_token() accumulates tokens until it got the
sequence of tokens that comprise a single JSON value (it counts curly
braces and square brackets to decide). It feeds those token sequences
to json_parser_parse(). If a non-empty sequence of tokens remains at
the end of the parse, it's silently ignored. check-qjson.c cases
unterminated_array(), unterminated_array_comma(), unterminated_dict(),
unterminated_dict_comma() demonstrate this bug.
Fix as follows. Introduce a JSON_END_OF_INPUT token. When the
streamer receives it, it feeds the accumulated tokens to
json_parser_parse().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-46-armbru@redhat.com>
json-parser.c carefully reports end of input like this:
token = parser_context_pop_token(ctxt);
if (token == NULL) {
parse_error(ctxt, NULL, "premature EOI");
goto out;
}
Except parser_context_pop_token() can't return null, it fails its
assertion instead. Same for parser_context_peek_token(). Broken in
commit 65c0f1e955, and faithfully preserved in commit 95385fe9ac.
Only a latent bug, because the streamer throws away any input that
could trigger it.
Drop the assertions, so we can fix the streamer in the next commit.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-45-armbru@redhat.com>
qobject_from_json() & friends use the consume_json() callback to
receive either a value or an error from the parser.
When they are fed a string that contains more than either one JSON
value or one JSON syntax error, consume_json() gets called multiple
times.
When the last call receives a value, qobject_from_json() returns that
value. Any other values are leaked.
When any call receives an error, qobject_from_json() sets the first
error received. Any other errors are thrown away.
When values follow errors, qobject_from_json() returns both a value
and sets an error. That's bad. Impact:
* block.c's parse_json_protocol() ignores and leaks the value. It's
used to to parse pseudo-filenames starting with "json:". The
pseudo-filenames can come from the user or from image meta-data such
as a QCOW2 image's backing file name.
* vl.c's parse_display_qapi() ignores and leaks the error. It's used
to parse the argument of command line option -display.
* vl.c's main() case QEMU_OPTION_blockdev ignores the error and leaves
it in @err. main() will then pass a pointer to a non-null Error *
to net_init_clients(), which is forbidden. It can lead to assertion
failure or other misbehavior.
* check-qjson.c's multiple_values() demonstrates the badness.
* The other callers are not affected since they only pass strings with
exactly one JSON value or, in the case of negative tests, one
error.
The impact on the _nofail() functions is relatively harmless. They
abort when any call receives an error. Else they return the last
value, and leak the others, if any.
Fix consume_json() as follows. On the first call, save value and
error as before. On subsequent calls, if any, don't save them. If
the first call saved a value, the next call, if any, replaces the
value by an "Expecting at most one JSON value" error. Take care not
to leak values or errors that aren't saved.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-44-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-43-armbru@redhat.com>
Support for %I64d got added in commit 2c0d4b36e7 "json: fix PRId64 on
Win32". We had to hard-code I64d because we used the lexer's finite
state machine to check interpolations. No more, so clean this up.
Additional conversion specifications would be easy enough to implement
when needed.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-42-armbru@redhat.com>
Both lexer and parser reject invalid interpolation specifications.
The parser's check is useless.
The lexer ends the token right after the first bad character. This
tends to lead to suboptimal error reporting. For instance, input
[ %04d ]
produces the tokens
JSON_LSQUARE [
JSON_ERROR %0
JSON_INTEGER 4
JSON_KEYWORD d
JSON_RSQUARE ]
The parser then yields an error, an object and two more errors:
error: Invalid JSON syntax
object: 4
error: JSON parse error, invalid keyword
error: JSON parse error, expecting value
Dumb down the lexer to accept [A-Za-z0-9]*. The parser's check is now
used. Emit a proper error there.
The lexer now produces
JSON_LSQUARE [
JSON_INTERP %04d
JSON_RSQUARE ]
and the parser reports just
JSON parse error, invalid interpolation '%04d'
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-41-armbru@redhat.com>
The callback to consume JSON values takes QObject *json, Error *err.
If both are null, the callback is supposed to make up an error by
itself. This sucks.
qjson.c's consume_json() neglects to do so, which makes
qobject_from_json() null instead of failing. I consider that a bug.
The culprit is json_message_process_token(): it passes two null
pointers when it runs into a lexical error or a limit violation. Fix
it to pass a proper Error object then. Update the callbacks:
* monitor.c's handle_qmp_command(): the code to make up an error is
now dead, drop it.
* qga/main.c's process_event(): lumps the "both null" case together
with the "not a JSON object" case. The former is now gone. The
error message "Invalid JSON syntax" is misleading for the latter.
Improve it to "Input must be a JSON object".
* qobject/qjson.c's consume_json(): no update; check-qjson
demonstrates qobject_from_json() now sets an error on lexical
errors, but still doesn't on some other errors.
* tests/libqtest.c's qmp_response(): the Error object is now reliable,
so use it to improve the error message.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-40-armbru@redhat.com>
The JSON parser optionally supports interpolation. The lexer
recognizes interpolation tokens unconditionally. The parser rejects
them when interpolation is disabled, in parse_interpolation().
However, it neglects to set an error then, which can make
json_parser_parse() fail without setting an error.
Move the check for unwanted interpolation from the parser's
parse_interpolation() into the lexer's finite state machine. When
interpolation is disabled, '%' is now handled like any other
unexpected character.
The next commit will improve how such lexical errors are handled.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-39-armbru@redhat.com>
The JSON parser optionally supports interpolation. The code calls it
"escape". Awkward, because it uses the same term for escape sequences
within strings. The latter usage is consistent with RFC 8259 "The
JavaScript Object Notation (JSON) Data Interchange Format" and ISO C.
Call the former "interpolation" instead.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-38-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-37-armbru@redhat.com>
json_parser_parse() normally returns the QObject on success. Except
it returns null when its @tokens argument is null.
Its only caller json_message_process_token() passes null @tokens when
emitting a lexical error. The call is a rather opaque way to say json
= NULL then.
Simplify matters by lifting the assignment to json out of the emit
path: initialize json to null, set it to the value of
json_parser_parse() when there's no lexical error. Drop the special
case from json_parser_parse().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-36-armbru@redhat.com>
The classical way to structure parser and lexer is to have the client
call the parser to get an abstract syntax tree, the parser call the
lexer to get the next token, and the lexer call some function to get
input characters.
Another way to structure them would be to have the client feed
characters to the lexer, the lexer feed tokens to the parser, and the
parser feed abstract syntax trees to some callback provided by the
client. This way is more easily integrated into an event loop that
dispatches input characters as they arrive.
Our JSON parser is kind of between the two. The lexer feeds tokens to
a "streamer" instead of a real parser. The streamer accumulates
tokens until it got the sequence of tokens that comprise a single JSON
value (it counts curly braces and square brackets to decide). It
feeds those token sequences to a callback provided by the client. The
callback passes each token sequence to the parser, and gets back an
abstract syntax tree.
I figure it was done that way to make a straightforward recursive
descent parser possible. "Get next token" becomes "pop the first
token off the token sequence". Drawback: we need to store a complete
token sequence. Each token eats 13 + input characters + malloc
overhead bytes.
Observations:
1. This is not the only way to use recursive descent. If we replaced
"get next token" by a coroutine yield, we could do without a
streamer.
2. The lexer reports errors by passing a JSON_ERROR token to the
streamer. This communicates the offending input characters and
their location, but no more.
3. The streamer reports errors by passing a null token sequence to the
callback. The (already poor) lexical error information is thrown
away.
4. Having the callback receive a token sequence duplicates the code to
convert token sequence to abstract syntax tree in every callback.
5. Known bug: the streamer silently drops incomplete token sequences.
This commit rectifies 4. by lifting the call of the parser from the
callbacks into the streamer. Later commits will address 3. and 5.
The lifting removes a bug from qjson.c's parse_json(): it passed a
pointer to a non-null Error * in certain cases, as demonstrated by
check-qjson.c.
json_parser_parse() is now unused. It's a stupid wrapper around
json_parser_parse_err(). Drop it, and rename json_parser_parse_err()
to json_parser_parse().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-35-armbru@redhat.com>
json_lexer_init() takes the function to process a token as an
argument. It's always json_message_process_token(). Makes the code
harder to understand for no actual gain. Drop the indirection.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-34-armbru@redhat.com>
parser_context_new/free() are only used from json_parser_parse(). We
can fold the code there and avoid an allocation altogether.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Message-Id: <20180719184111.5129-9-marcandre.lureau@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20180823164025.12553-33-armbru@redhat.com>
The lexer always returns 0 when char feeding. Furthermore, none of the
caller care about the return value.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Message-Id: <20180326150916.9602-10-marcandre.lureau@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20180823164025.12553-32-armbru@redhat.com>
The JSON parser treats each half of a surrogate pair as unpaired
surrogate. Fix it to recognize surrogate pairs.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-30-armbru@redhat.com>
The JSON parser translates invalid \uXXXX to garbage instead of
rejecting it, and swallows \u0000.
Fix by using mod_utf8_encode() instead of flawed wchar_to_utf8().
Valid surrogate pairs are now differently broken: they're rejected
instead of translated to garbage. The next commit will fix them.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-29-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-28-armbru@redhat.com>
Both lexer and parser reject invalid escape sequences in strings. The
parser's check is useless.
The lexer ends the token right after the first non-well-formed byte.
This tends to lead to suboptimal error reporting. For instance, input
{"abc\@ijk": 1}
produces the tokens
JSON_LCURLY {
JSON_ERROR "abc\@
JSON_KEYWORD ijk
JSON_ERROR ": 1}\n
The parser then reports three errors
Invalid JSON syntax
JSON parse error, invalid keyword 'ijk'
Invalid JSON syntax
before it recovers at the newline.
Drop the lexer's escape sequence checking, and make it accept the same
characters after backslash it accepts elsewhere in strings. It now
produces
JSON_LCURLY {
JSON_STRING "abc\@ijk"
JSON_COLON :
JSON_INTEGER 1
JSON_RCURLY
and the parser reports just
JSON parse error, invalid escape sequence in string
While there, fix parse_string()'s inaccurate function comment.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-27-armbru@redhat.com>
Since the JSON grammer doesn't accept U+0000 anywhere, this merely
exchanges one kind of parse error for another. It's purely for
consistency with qobject_to_json(), which accepts \xC0\x80 (see commit
e2ec3f9768).
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-26-armbru@redhat.com>
Both the lexer and the parser (attempt to) validate UTF-8 in JSON
strings.
The lexer rejects bytes that can't occur in valid UTF-8: \xC0..\xC1,
\xF5..\xFF. This rejects some, but not all invalid UTF-8. It also
rejects ASCII control characters \x00..\x1F, in accordance with RFC
8259 (see recent commit "json: Reject unescaped control characters").
When the lexer rejects, it ends the token right after the first bad
byte. Good when the bad byte is a newline. Not so good when it's
something like an overlong sequence in the middle of a string. For
instance, input
{"abc\xC0\xAFijk": 1}\n
produces the tokens
JSON_LCURLY {
JSON_ERROR "abc\xC0
JSON_ERROR \xAF
JSON_KEYWORD ijk
JSON_ERROR ": 1}\n
The parser then reports four errors
Invalid JSON syntax
Invalid JSON syntax
JSON parse error, invalid keyword 'ijk'
Invalid JSON syntax
before it recovers at the newline.
The commit before previous made the parser reject invalid UTF-8
sequences. Since then, anything the lexer rejects, the parser would
reject as well. Thus, the lexer's rejecting is unnecessary for
correctness, and harmful for error reporting.
However, we want to keep rejecting ASCII control characters in the
lexer, because that produces the behavior we want for unclosed
strings.
We also need to keep rejecting \xFF in the lexer, because we
documented that as a way to reset the JSON parser
(docs/interop/qmp-spec.txt section 2.6 QGA Synchronization), which
means we can't change how we recover from this error now. I wish we
hadn't done that.
I think we should treat \xFE the same as \xFF.
Change the lexer to accept \xC0..\xC1 and \xF5..\xFD. It now rejects
only \x00..\x1F and \xFE..\xFF. Error reporting for invalid UTF-8 in
strings is much improved, except for \xFE and \xFF. For the example
above, the lexer now produces
JSON_LCURLY {
JSON_STRING "abc\xC0\xAFijk"
JSON_COLON :
JSON_INTEGER 1
JSON_RCURLY
and the parser reports just
JSON parse error, invalid UTF-8 sequence in string
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-25-armbru@redhat.com>
Quiz time! When a parser reports multiple errors, but the user gets
to see just one, which one is (on average) the least useful one?
Yes, you're right, it's the last one! You're clearly familiar with
compilers.
Which one does QEMU report?
Right again, the last one! You're clearly familiar with QEMU.
Reproducer: feeding
{"abc\xC2ijk": 1}\n
to QMP produces
{"error": {"class": "GenericError", "desc": "JSON parse error, key is not a string in object"}}
Report the first error instead. The reproducer now produces
{"error": {"class": "GenericError", "desc": "JSON parse error, invalid UTF-8 sequence in string"}}
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-24-armbru@redhat.com>
We reject bytes that can't occur in valid UTF-8 (\xC0..\xC1,
\xF5..\xFF in the lexer. That's insufficient; there's plenty of
invalid UTF-8 not containing these bytes, as demonstrated by
check-qjson:
* Malformed sequences
- Unexpected continuation bytes
- Missing continuation bytes after start bytes other than
\xC0..\xC1, \xF5..\xFD.
* Overlong sequences with start bytes other than \xC0..\xC1,
\xF5..\xFD.
* Invalid code points
Fixing this in the lexer would be bothersome. Fixing it in the parser
is straightforward, so do that.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-23-armbru@redhat.com>
Simplify loop control, and assert that the string ends with the
appropriate quote (the lexer ensures it does).
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-21-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-20-armbru@redhat.com>
Fix the lexer to reject unescaped control characters in JSON strings,
in accordance with RFC 8259 "The JavaScript Object Notation (JSON)
Data Interchange Format".
Bonus: we now recover more nicely from unclosed strings. E.g.
{"one: 1}\n{"two": 2}
now recovers cleanly after the newline, where before the lexer
remained confused until the next unpaired double quote or lexical
error.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-19-armbru@redhat.com>
json_lexer[] maps (lexer state, input character) to the new lexer
state. The input character is consumed unless the new state is
terminal and the input character doesn't belong to this token,
i.e. the state transition uses look-ahead. When this is the case,
input character '\0' would result in the same state transition.
TERMINAL_NEEDED_LOOKAHEAD() exploits this.
Except this is wrong for transitions to IN_ERROR. There, the
offending input character is in fact consumed: case IN_ERROR returns.
It isn't added to the JSON_ERROR token, though.
Fix that by making TERMINAL_NEEDED_LOOKAHEAD() return false for
transitions to IN_ERROR.
There's a slight complication. json_lexer_flush() passes input
character '\0' to flush an incomplete token. If this results in
JSON_ERROR, we'd now add the '\0' to the token. Suppress that.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180823164025.12553-18-armbru@redhat.com>
Markus Armbruster