* Daniel P. Berrangé (berrange(a)redhat.com) wrote: > The QEMU QMP service is based on JSON which is nice because that is a > widely supported "standard" data format..... > > ....except QEMU's implementation (and indeed most impls) are not strictly > standards compliant. > > Specifically the problem is around representing 64-bit integers, whether > signed or unsigned. > > The JSON standard declares that largest integer is 2^53-1 and the > likewise the smallest is -(2^53-1): > > http://www.ecma-international.org/ecma-262/6.0/index.html#sec-number.max_... > > A crazy limit inherited from its javascript origins IIUC. Ewwww.

> QEMU, and indeed many applications, want to handle 64-bit integers. > The C JSON library impls have traditionally mapped integers to the > data type 'long long int' which gives a min/max of -(2^63) / 2^63-1. > > QEMU however /really/ needs 64-bit unsigned integers, ie a max 2^64-1. > > Libvirt has historically used the YAJL library which uses 'long long int' > and thus can't officially go beyond 2^63-1 values. Fortunately it lets > libvirt get at the raw json string, so libvirt can re-parse the value > to get an 'unsigned long long'. > > We recently tried to switch to Jansson because YAJL has a dead upstream > for many years and countless unanswered bugs & patches. Unfortunately we > forgot about this need for 2^64-1 max, and Jansson also uses 'long long int' > and raises a fatal parse error for unsigned 64-bit values above 2^63-1. It > also provides no backdoor for libvirt todo its own integer parsing. Thus > we had to abort our switch to jansson as it broke parsing QEMU's JSON: > > https://bugzilla.redhat.com/show_bug.cgi?id=1614569 > > Other JSON libraries we've investigated have similar problems. I imagine > the same may well be true of non-C based JOSN impls, though I've not > investigated in any detail. > > Essentially libvirt is stuck with either using the dead YAJL library > forever, or writing its own JSON parser (most likely copying QEMU's > JSON code into libvirt's git). > > This feels like a very unappealing situation to be in as not being > able to use a JSON library of our choice is loosing one of the key > benefits of using a standard data format. > > Thus I'd like to see a solution to this to allow QMP to be reliably > consumed by any JSON library that exists. > > I can think of some options: > > 1. Encode unsigned 64-bit integers as signed 64-bit integers. > > This follows the example that most C libraries map JSON ints > to 'long long int'. This is still relying on undefined > behaviour as apps don't need to support > 2^53-1. > > Apps would need to cast back to 'unsigned long long' for > those QMP fields they know are supposed to be unsigned. > > > 2. Encode all 64-bit integers as a pair of 32-bit integers. > > This is fully compliant with the JSON spec as each half > is fully within the declared limits. App has to split or > assemble the 2 pieces from/to a signed/unsigned 64-bit > int as needed. > > > 3. Encode all 64-bit integers as strings > > The application has todo all parsing/formatting client > side. > > > None of these changes are backwards compatible, so I doubt we could make > the change transparently in QMP. Instead we would have to have a > QMP greeting message capability where the client can request enablement > of the enhanced integer handling. > > Any of the three options above would likely work for libvirt, but I > would have a slight preference for either 2 or 3, so that we become > 100% standards compliant. My preference would be 3 with the strings defined as being %x lower case hex formated with a 0x prefix and no longer than 18 characters ("0x" + 16 nybbles). Zero padding allowed but not required. It's readable and unambiguous when dealing with addresses; I don't want to have to start decoding (2) by hand when debugging.

> The Golang JSON parser decodes JSON numbers to float64 by default so > will have this precision limitation too, though at least they provide > a backdoor for custom parsing from the original serialized representation. > >> > QEMU, and indeed many applications, want to handle 64-bit integers. >> > The C JSON library impls have traditionally mapped integers to the >> > data type 'long long int' which gives a min/max of -(2^63) / 2^63-1. >> > >> > QEMU however /really/ needs 64-bit unsigned integers, ie a max 2^64-1. Correct. Support for integers 2^63..2^64-1 is relatively recent: commit 2bc7cfea095 (v2.10, 2017). Since we really needed these, the QObject input visitor silently casts negative integers to uint64_t. It still does for backward compatibility. Commit 5923f85fb82 (right after 2bc7cfea095) explains The input visitor will cast i64 input to u64 for compatibility reasons (existing json QMP client already use negative i64 for large u64, and expect an implicit cast in qemu). Note: before the patch, uint64_t values above INT64_MAX are sent over json QMP as negative values, e.g. UINT64_MAX is sent as -1. After the patch, they are sent unmodified. Clearly a bug fix, but we have to consider compatibility issues anyway. libvirt should cope fine, because its parsing of unsigned integers accepts negative values modulo 2^64. There's hope that other clients will, too.

>> > Libvirt has historically used the YAJL library which uses 'long long int' >> > and thus can't officially go beyond 2^63-1 values. Fortunately it lets >> > libvirt get at the raw json string, so libvirt can re-parse the value >> > to get an 'unsigned long long'. >> > >> > We recently tried to switch to Jansson because YAJL has a dead upstream >> > for many years and countless unanswered bugs & patches. Unfortunately we >> > forgot about this need for 2^64-1 max, and Jansson also uses 'long long int' >> > and raises a fatal parse error for unsigned 64-bit values above 2^63-1. It >> > also provides no backdoor for libvirt todo its own integer parsing. Thus >> > we had to abort our switch to jansson as it broke parsing QEMU's JSON: >> > >> > https://bugzilla.redhat.com/show_bug.cgi?id=1614569 >> > >> > Other JSON libraries we've investigated have similar problems. I imagine >> > the same may well be true of non-C based JOSN impls, though I've not >> > investigated in any detail. >> > >> > Essentially libvirt is stuck with either using the dead YAJL library >> > forever, or writing its own JSON parser (most likely copying QEMU's >> > JSON code into libvirt's git). >> > >> > This feels like a very unappealing situation to be in as not being >> > able to use a JSON library of our choice is loosing one of the key >> > benefits of using a standard data format. >> > >> > Thus I'd like to see a solution to this to allow QMP to be reliably >> > consumed by any JSON library that exists. JSON is terrible at interoperability, so good luck with that. If you reduce your order to "the commonly used JSON libraries we know", we can talk.

>> > I can think of some options: >> > >> > 1. Encode unsigned 64-bit integers as signed 64-bit integers. >> > >> > This follows the example that most C libraries map JSON ints >> > to 'long long int'. This is still relying on undefined >> > behaviour as apps don't need to support > 2^53-1. >> > >> > Apps would need to cast back to 'unsigned long long' for >> > those QMP fields they know are supposed to be unsigned. Ugly. It's also what we did until v2.10, August 2017. QMP's input direction still does it, for backward compatibility. >> > >> > >> > 2. Encode all 64-bit integers as a pair of 32-bit integers. >> > >> > This is fully compliant with the JSON spec as each half >> > is fully within the declared limits. App has to split or >> > assemble the 2 pieces from/to a signed/unsigned 64-bit >> > int as needed. Differently ugly. >> > >> > >> > 3. Encode all 64-bit integers as strings >> > >> > The application has todo all parsing/formatting client >> > side. Yet another ugly. >> > >> > >> > None of these changes are backwards compatible, so I doubt we could make >> > the change transparently in QMP. Instead we would have to have a >> > QMP greeting message capability where the client can request enablement >> > of the enhanced integer handling. We might be able to do option 1 without capability negotiation. v2.10's change from option 1 to what we have now produced zero complaints. On the other hand, we made that change for a reason, so we may want a "send large integers as negative integers" capability regardless. >> > >> > Any of the three options above would likely work for libvirt, but I >> > would have a slight preference for either 2 or 3, so that we become >> > 100% standards compliant. There's no such thing. You mean "we maximize interoperability with common implementations of JSON".

Let's talk implementation for a bit. Encoding and decoding integers in funny ways should be fairly easy in the QObject visitors. The generated QMP marshallers all use them. Trouble is a few commands still bypass the generated marshallers, and mess with the QObject themselves: * query-qmp-schema: minor hack explained in qmp_query_qmp_schema()'s comment. Should be harmless. * netdev_add: not QAPIfied. Eric's patches to QAPIfy it got stuck because they reject some abuses like passing numbers and bools as strings. * device_add: not QAPIfied. We're not sure QAPIfication is feasible. netdev_add and device_add both use qemu_opts_from_qdict(). Perhaps we could hack that to mirror what the QObject visitor do. Else, we might have to do it in the JSON parser. Should be possible, but I'd rather not. >> My preference would be 3 with the strings defined as being >> %x lower case hex formated with a 0x prefix and no longer than 18 characters >> ("0x" + 16 nybbles). Zero padding allowed but not required. >> It's readable and unambiguous when dealing with addresses; I don't want >> to have to start decoding (2) by hand when debugging. > > Yep, that's a good point about readability. QMP sending all integers in decimal is inconvenient for some values, such as addresses. QMP sending all (large) integers in hexadecimal would be inconvenient for other values. Let's keep it simple & stupid. If you want sophistication, JSON is the wrong choice. Option 1 feels simplest.

Option 2 feels ugliest. Less simple, more interoperable than option 1.

Option 3 is like option 2, just not quite as ugly.

Can we agree to eliminate option 2 from the race?

On 5/7/19 4:39 AM, Daniel P. Berrangé wrote: >> JSON is terrible at interoperability, so good luck with that. >> >> If you reduce your order to "the commonly used JSON libraries we know", >> we can talk. > > I don't particularly want us to rely on semantics of small known set > of JSON libs. I really do want us to do something that is capable of > working with any JSON impl that exists in any programming language. > > My suggested option 2 & 3 at least would manage that I believe, as > any credible JSON impl will be able to represent 32-bit integers > or strings without loosing data. > > Option 1 would not cope as some impls can't even cope with > signed 64-bit ints. > >>>>> I can think of some options: >>>>> >>>>> 1. Encode unsigned 64-bit integers as signed 64-bit integers. >>>>> >>>>> This follows the example that most C libraries map JSON ints >>>>> to 'long long int'. This is still relying on undefined >>>>> behaviour as apps don't need to support > 2^53-1. >>>>> >>>>> Apps would need to cast back to 'unsigned long long' for >>>>> those QMP fields they know are supposed to be unsigned. >> >> Ugly. It's also what we did until v2.10, August 2017. QMP's input >> direction still does it, for backward compatibility. Having qemu accept signed ints in place of large unsigned values is easy enough. But you are right that it loses precision when doubles are involved on the receiving end, and we cross the 2^53 barrier. >> >>>>> >>>>> >>>>> 2. Encode all 64-bit integers as a pair of 32-bit integers. >>>>> >>>>> This is fully compliant with the JSON spec as each half >>>>> is fully within the declared limits. App has to split or >>>>> assemble the 2 pieces from/to a signed/unsigned 64-bit >>>>> int as needed. >> >> Differently ugly. Particularly ugly as we turn 1<<55 from: "value":36028797018963968 into "value":[8388608,0] and now both qemu and the client end have to agree that an array of two integers is a valid replacement for any larger 64-bit quantity (presumably, we'd always accept the array form even for small integer values, but only produce the array form for large values). And while it manages just fine for uint64_t values, what rules would you place on int64_t values? That the resulting 2-integer array is combined with the first number as a 2's-complement signed value, and the second being a 32-bit unsigned value?

>>>>> >>>>> >>>>> 3. Encode all 64-bit integers as strings >>>>> >>>>> The application has todo all parsing/formatting client >>>>> side. >> >> Yet another ugly. But less so than option 2. "value":36028797018963968 vs. "value":"36028797018963968" is at least tolerable.

>>>>> None of these changes are backwards compatible, so I doubt we could make >>>>> the change transparently in QMP. Instead we would have to have a >>>>> QMP greeting message capability where the client can request enablement >>>>> of the enhanced integer handling. >> >> We might be able to do option 1 without capability negotiation. v2.10's >> change from option 1 to what we have now produced zero complaints. >> >> On the other hand, we made that change for a reason, so we may want a >> "send large integers as negative integers" capability regardless. >> >>>>> >>>>> Any of the three options above would likely work for libvirt, but I >>>>> would have a slight preference for either 2 or 3, so that we become >>>>> 100% standards compliant. If we're going to negotiate something, I'd lean towards option 3 (anywhere the introspection states that we accept 'int64' or similar, it is also appropriate to send a string value in its place). We'd also have to decide if we want to allow "0xabcd", or strictly insist on 43981, when stringizing an integer. And while qemu should accept a string or a number on input, we'd still have to decide/document whether it's response to the client capability negotiation is to output a string always, or only for values larger than the 2^53 threshold.

>> >> There's no such thing. You mean "we maximize interoperability with >> common implementations of JSON". > > s/common/any/ > >> Let's talk implementation for a bit. >> >> Encoding and decoding integers in funny ways should be fairly easy in >> the QObject visitors. The generated QMP marshallers all use them. >> Trouble is a few commands still bypass the generated marshallers, and >> mess with the QObject themselves: >> >> * query-qmp-schema: minor hack explained in qmp_query_qmp_schema()'s >> comment. Should be harmless. >> >> * netdev_add: not QAPIfied. Eric's patches to QAPIfy it got stuck >> because they reject some abuses like passing numbers and bools as >> strings. >> >> * device_add: not QAPIfied. We're not sure QAPIfication is feasible. >> >> netdev_add and device_add both use qemu_opts_from_qdict(). Perhaps we >> could hack that to mirror what the QObject visitor do. >> >> Else, we might have to do it in the JSON parser. Should be possible, >> but I'd rather not. >> >>>> My preference would be 3 with the strings defined as being >>>> %x lower case hex formated with a 0x prefix and no longer than 18 characters >>>> ("0x" + 16 nybbles). Zero padding allowed but not required. >>>> It's readable and unambiguous when dealing with addresses; I don't want >>>> to have to start decoding (2) by hand when debugging. >>> >>> Yep, that's a good point about readability. >> >> QMP sending all integers in decimal is inconvenient for some values, >> such as addresses. QMP sending all (large) integers in hexadecimal >> would be inconvenient for other values. >> >> Let's keep it simple & stupid. If you want sophistication, JSON is the >> wrong choice. JSON requires decimal-only, but I'm okay if we state that when negotiating the alternative representation, that we output hex-only. (JSON5 adds hex support among other things, but it is not an RFC standard, and even fewer libraries exist that parse JSON5 in addition to straight JSON). >> >> >> Option 1 feels simplest. > > But will still fail with any JSON impl that uses double precision floating > point for integers as it will loose precision. > >> Option 2 feels ugliest. Less simple, more interoperable than option 1. > > If we assume any JSON impl can do 32-bit integers without loss of > precision, then I think we can say it is guaranteed portable, but > it is certainly horrible / ugly. > >> Option 3 is like option 2, just not quite as ugly. > > I think option 3 can be guaranteed to be loss-less with /any/ JSON impl > that exists, since you're delegating all string -> int conversion to > the application code taking the JSON parser/formatter out of the equation. > > This is close to the approach libvirt takes with YAJL parser today. YAJL > parses as a int64 and we then ignore its result, and re-parse the string > again in libvirt as uint64. When generating json we format as uint64 > in libvirt and ignore YAJLs formatting for int64. > >> Can we agree to eliminate option 2 from the race? > > I'm fine with eliminating option 2. Same here.

> I guess I'd have a preference for option 3 given that it has better > interoperability Likewise - if we're going to bother with a capability that changes output and allows the input validators to accept more forms, I'd prefer a string form with correct sign over a negative integer that depends on 64-bit 2's-complement arithmetic to intepret correctly.

On Tue, May 07, 2019 at 10:47:06AM +0200, Markus Armbruster wrote: > > The Golang JSON parser decodes JSON numbers to float64 by default so > > will have this precision limitation too, though at least they provide > > a backdoor for custom parsing from the original serialized representation. > > > >> > QEMU, and indeed many applications, want to handle 64-bit integers. > >> > The C JSON library impls have traditionally mapped integers to the > >> > data type 'long long int' which gives a min/max of -(2^63) / 2^63-1. > >> > > >> > QEMU however /really/ needs 64-bit unsigned integers, ie a max 2^64-1. > > Correct. > > Support for integers 2^63..2^64-1 is relatively recent: commit > 2bc7cfea095 (v2.10, 2017). > > Since we really needed these, the QObject input visitor silently casts > negative integers to uint64_t. It still does for backward > compatibility. Commit 5923f85fb82 (right after 2bc7cfea095) explains > > The input visitor will cast i64 input to u64 for compatibility > reasons (existing json QMP client already use negative i64 for large > u64, and expect an implicit cast in qemu). > > Note: before the patch, uint64_t values above INT64_MAX are sent over > json QMP as negative values, e.g. UINT64_MAX is sent as -1. After the > patch, they are sent unmodified. Clearly a bug fix, but we have to > consider compatibility issues anyway. libvirt should cope fine, > because its parsing of unsigned integers accepts negative values > modulo 2^64. There's hope that other clients will, too. So QEMU reading stuff sent by libvirt in a back compatible manner is ok. The problem was specifically when a QEMU reply sent back UINT64_MAX value as a positive number.

> >> > Libvirt has historically used the YAJL library which uses 'long long int' > >> > and thus can't officially go beyond 2^63-1 values. Fortunately it lets > >> > libvirt get at the raw json string, so libvirt can re-parse the value > >> > to get an 'unsigned long long'. > >> > > >> > We recently tried to switch to Jansson because YAJL has a dead upstream > >> > for many years and countless unanswered bugs & patches. Unfortunately we > >> > forgot about this need for 2^64-1 max, and Jansson also uses 'long long int' > >> > and raises a fatal parse error for unsigned 64-bit values above 2^63-1. It > >> > also provides no backdoor for libvirt todo its own integer parsing. Thus > >> > we had to abort our switch to jansson as it broke parsing QEMU's JSON: > >> > > >> > https://bugzilla.redhat.com/show_bug.cgi?id=1614569 > >> > > >> > Other JSON libraries we've investigated have similar problems. I imagine > >> > the same may well be true of non-C based JOSN impls, though I've not > >> > investigated in any detail. > >> > > >> > Essentially libvirt is stuck with either using the dead YAJL library > >> > forever, or writing its own JSON parser (most likely copying QEMU's > >> > JSON code into libvirt's git). > >> > > >> > This feels like a very unappealing situation to be in as not being > >> > able to use a JSON library of our choice is loosing one of the key > >> > benefits of using a standard data format. > >> > > >> > Thus I'd like to see a solution to this to allow QMP to be reliably > >> > consumed by any JSON library that exists. > > JSON is terrible at interoperability, so good luck with that. > > If you reduce your order to "the commonly used JSON libraries we know", > we can talk. I don't particularly want us to rely on semantics of small known set of JSON libs. I really do want us to do something that is capable of working with any JSON impl that exists in any programming language.

My suggested option 2 & 3 at least would manage that I believe, as any credible JSON impl will be able to represent 32-bit integers or strings without loosing data.

Option 1 would not cope as some impls can't even cope with signed 64-bit ints. > >> > I can think of some options: > >> > > >> > 1. Encode unsigned 64-bit integers as signed 64-bit integers. > >> > > >> > This follows the example that most C libraries map JSON ints > >> > to 'long long int'. This is still relying on undefined > >> > behaviour as apps don't need to support > 2^53-1. > >> > > >> > Apps would need to cast back to 'unsigned long long' for > >> > those QMP fields they know are supposed to be unsigned. > > Ugly. It's also what we did until v2.10, August 2017. QMP's input > direction still does it, for backward compatibility. > > >> > > >> > > >> > 2. Encode all 64-bit integers as a pair of 32-bit integers. > >> > > >> > This is fully compliant with the JSON spec as each half > >> > is fully within the declared limits. App has to split or > >> > assemble the 2 pieces from/to a signed/unsigned 64-bit > >> > int as needed. > > Differently ugly. > > >> > > >> > > >> > 3. Encode all 64-bit integers as strings > >> > > >> > The application has todo all parsing/formatting client > >> > side. > > Yet another ugly. > > >> > > >> > > >> > None of these changes are backwards compatible, so I doubt we could make > >> > the change transparently in QMP. Instead we would have to have a > >> > QMP greeting message capability where the client can request enablement > >> > of the enhanced integer handling. > > We might be able to do option 1 without capability negotiation. v2.10's > change from option 1 to what we have now produced zero complaints. > > On the other hand, we made that change for a reason, so we may want a > "send large integers as negative integers" capability regardless. > > >> > > >> > Any of the three options above would likely work for libvirt, but I > >> > would have a slight preference for either 2 or 3, so that we become > >> > 100% standards compliant. > > There's no such thing. You mean "we maximize interoperability with > common implementations of JSON". s/common/any/

> Let's talk implementation for a bit. > > Encoding and decoding integers in funny ways should be fairly easy in > the QObject visitors. The generated QMP marshallers all use them. > Trouble is a few commands still bypass the generated marshallers, and > mess with the QObject themselves: > > * query-qmp-schema: minor hack explained in qmp_query_qmp_schema()'s > comment. Should be harmless. > > * netdev_add: not QAPIfied. Eric's patches to QAPIfy it got stuck > because they reject some abuses like passing numbers and bools as > strings. > > * device_add: not QAPIfied. We're not sure QAPIfication is feasible. > > netdev_add and device_add both use qemu_opts_from_qdict(). Perhaps we > could hack that to mirror what the QObject visitor do. > > Else, we might have to do it in the JSON parser. Should be possible, > but I'd rather not. > > >> My preference would be 3 with the strings defined as being > >> %x lower case hex formated with a 0x prefix and no longer than 18 characters > >> ("0x" + 16 nybbles). Zero padding allowed but not required. > >> It's readable and unambiguous when dealing with addresses; I don't want > >> to have to start decoding (2) by hand when debugging. > > > > Yep, that's a good point about readability. > > QMP sending all integers in decimal is inconvenient for some values, > such as addresses. QMP sending all (large) integers in hexadecimal > would be inconvenient for other values. > > Let's keep it simple & stupid. If you want sophistication, JSON is the > wrong choice. > > > Option 1 feels simplest. But will still fail with any JSON impl that uses double precision floating point for integers as it will loose precision. > Option 2 feels ugliest. Less simple, more interoperable than option 1. If we assume any JSON impl can do 32-bit integers without loss of precision, then I think we can say it is guaranteed portable, but it is certainly horrible / ugly. > Option 3 is like option 2, just not quite as ugly. I think option 3 can be guaranteed to be loss-less with /any/ JSON impl that exists, since you're delegating all string -> int conversion to the application code taking the JSON parser/formatter out of the equation.

This is close to the approach libvirt takes with YAJL parser today. YAJL parses as a int64 and we then ignore its result, and re-parse the string again in libvirt as uint64. When generating json we format as uint64 in libvirt and ignore YAJLs formatting for int64. > Can we agree to eliminate option 2 from the race? I'm fine with eliminating option 2.

I guess I'd have a preference for option 3 given that it has better interoperability

On Wed, May 08, 2019 at 02:44:07PM +0200, Markus Armbruster wrote: > Daniel P. Berrangé <berrange(a)redhat.com&gt; writes: > > > On Tue, May 07, 2019 at 10:47:06AM +0200, Markus Armbruster wrote: > > > >> >> > I can think of some options: > >> >> > > >> >> > 1. Encode unsigned 64-bit integers as signed 64-bit integers. > >> >> > > >> >> > This follows the example that most C libraries map JSON ints > >> >> > to 'long long int'. This is still relying on undefined > >> >> > behaviour as apps don't need to support > 2^53-1. > >> >> > > >> >> > Apps would need to cast back to 'unsigned long long' for > >> >> > those QMP fields they know are supposed to be unsigned. > >> > >> Ugly. It's also what we did until v2.10, August 2017. QMP's input > >> direction still does it, for backward compatibility. > >> > >> >> > > >> >> > > >> >> > 2. Encode all 64-bit integers as a pair of 32-bit integers. > >> >> > > >> >> > This is fully compliant with the JSON spec as each half > >> >> > is fully within the declared limits. App has to split or > >> >> > assemble the 2 pieces from/to a signed/unsigned 64-bit > >> >> > int as needed. > >> > >> Differently ugly. > >> > >> >> > > >> >> > > >> >> > 3. Encode all 64-bit integers as strings > >> >> > > >> >> > The application has todo all parsing/formatting client > >> >> > side. > >> > >> Yet another ugly. > >> > >> >> > > >> >> > > >> >> > None of these changes are backwards compatible, so I doubt we could make > >> >> > the change transparently in QMP. Instead we would have to have a > >> >> > QMP greeting message capability where the client can request enablement > >> >> > of the enhanced integer handling. > >> > >> We might be able to do option 1 without capability negotiation. v2.10's > >> change from option 1 to what we have now produced zero complaints. > >> > >> On the other hand, we made that change for a reason, so we may want a > >> "send large integers as negative integers" capability regardless. > >> > >> >> > > >> >> > Any of the three options above would likely work for libvirt, but I > >> >> > would have a slight preference for either 2 or 3, so that we become > >> >> > 100% standards compliant. > >> > >> There's no such thing. You mean "we maximize interoperability with > >> common implementations of JSON". > > > > s/common/any/ > > info: error correction applied, future applications will be silent ;-P > > >> Let's talk implementation for a bit. > >> > >> Encoding and decoding integers in funny ways should be fairly easy in > >> the QObject visitors. The generated QMP marshallers all use them. > >> Trouble is a few commands still bypass the generated marshallers, and > >> mess with the QObject themselves: > >> > >> * query-qmp-schema: minor hack explained in qmp_query_qmp_schema()'s > >> comment. Should be harmless. > >> > >> * netdev_add: not QAPIfied. Eric's patches to QAPIfy it got stuck > >> because they reject some abuses like passing numbers and bools as > >> strings. > >> > >> * device_add: not QAPIfied. We're not sure QAPIfication is feasible. > >> > >> netdev_add and device_add both use qemu_opts_from_qdict(). Perhaps we > >> could hack that to mirror what the QObject visitor do. > >> > >> Else, we might have to do it in the JSON parser. Should be possible, > >> but I'd rather not. > >> > >> >> My preference would be 3 with the strings defined as being > >> >> %x lower case hex formated with a 0x prefix and no longer than 18 characters > >> >> ("0x" + 16 nybbles). Zero padding allowed but not required. > >> >> It's readable and unambiguous when dealing with addresses; I don't want > >> >> to have to start decoding (2) by hand when debugging. > >> > > >> > Yep, that's a good point about readability. > >> > >> QMP sending all integers in decimal is inconvenient for some values, > >> such as addresses. QMP sending all (large) integers in hexadecimal > >> would be inconvenient for other values. > >> > >> Let's keep it simple & stupid. If you want sophistication, JSON is the > >> wrong choice. > >> > >> > >> Option 1 feels simplest. > > > > But will still fail with any JSON impl that uses double precision floating > > point for integers as it will loose precision. > > > >> Option 2 feels ugliest. Less simple, more interoperable than option 1. > > > > If we assume any JSON impl can do 32-bit integers without loss of > > precision, then I think we can say it is guaranteed portable, but > > it is certainly horrible / ugly. > > > >> Option 3 is like option 2, just not quite as ugly. > > > > I think option 3 can be guaranteed to be loss-less with /any/ JSON impl > > that exists, since you're delegating all string -> int conversion to > > the application code taking the JSON parser/formatter out of the equation. > > Double-checking: do you propose to encode *all* numbers as strings, or > just certain "problematic" numbers? > > If the latter, I guess your idea of "problematic" is "not representable > exactly as double precision floating-point". We have a few options 1. Use string format for values > 2^53-1, int format below that 2. Use string format for all fields which are 64-bit ints whether signed or unsigned 3. Use string format for all fields which are integers, even 32-bit ones I would probably suggest option 2. It would make the QEMU impl quite easy IIUC, we we'd just change the QAPI visitor's impl for the int64 and uint64 fields to use string format (when the right capability is negotiated by QMP). I include 3 only for completeness - I don't think there's a hugely compelling reason to mess with 32-bit ints.

Option 1 is the bare minimum needed to ensure precision, but to me it feels a bit dirty to say a given field will have different encoding depending on the value. If apps need to deal with string encoding, they might as well just use it for all values in a given field.

> > I guess I'd have a preference for option 3 given that it has better > > interoperability > > If we add a QMP capability for interoperability with JSON > implementations that set limits on range and precision that are > incompatible with the ones QMP sets, one could argue we effectively pay > the price for option 3, so we should take it for its benefits. > > Option 1 without a QMP capability merely reverts the change we made in > 2.10. We can do that if we think it's sufficient. > > You expressed a strong preference for maximizing interoperability (via > option 3). Acknowledged. However, I care a lot more about issues we > know we have than about issues somebody might have. > > You mentioned the libvirt's switch to Jansson you had to abort due to > QMP sending numbers Jansson refuses to parse. That's the kind of > non-hypothetical issue that can make me mess with the QMP language. > > You wrote Jansson "raises a fatal parse error for unsigned 64-bit values > above 2^63-1". Does that mean it rejects 9223372036854775808, but > accepts 9223372036854775808.0 (with loss of precision)? If it sees a '.' in the number, then it call strtod() and checks for the overflow conditions. If it doesn't see a '.' in the number then it calls strtoll and checks for the overflow conditions. So to answer you question, yes, it looks like it will reject 9223372036854775808 and accept 9223372036854775808.0 with loss of precision. Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|

On Mon, May 13, 2019 at 01:29:34PM +0100, Dr. David Alan Gilbert wrote: > * Daniel P. Berrangé (berrange(a)redhat.com) wrote: > > On Wed, May 08, 2019 at 02:44:07PM +0200, Markus Armbruster wrote: > > > Daniel P. Berrangé <berrange(a)redhat.com&gt; writes: > > > > > > > On Tue, May 07, 2019 at 10:47:06AM +0200, Markus Armbruster wrote: > > > > > > > >> >> > I can think of some options: > > > >> >> > > > > >> >> > 1. Encode unsigned 64-bit integers as signed 64-bit integers. > > > >> >> > > > > >> >> > This follows the example that most C libraries map JSON ints > > > >> >> > to 'long long int'. This is still relying on undefined > > > >> >> > behaviour as apps don't need to support > 2^53-1. > > > >> >> > > > > >> >> > Apps would need to cast back to 'unsigned long long' for > > > >> >> > those QMP fields they know are supposed to be unsigned. > > > >> > > > >> Ugly. It's also what we did until v2.10, August 2017. QMP's input > > > >> direction still does it, for backward compatibility. > > > >> > > > >> >> > > > > >> >> > > > > >> >> > 2. Encode all 64-bit integers as a pair of 32-bit integers. > > > >> >> > > > > >> >> > This is fully compliant with the JSON spec as each half > > > >> >> > is fully within the declared limits. App has to split or > > > >> >> > assemble the 2 pieces from/to a signed/unsigned 64-bit > > > >> >> > int as needed. > > > >> > > > >> Differently ugly. > > > >> > > > >> >> > > > > >> >> > > > > >> >> > 3. Encode all 64-bit integers as strings > > > >> >> > > > > >> >> > The application has todo all parsing/formatting client > > > >> >> > side. > > > >> > > > >> Yet another ugly. > > > >> > > > >> >> > > > > >> >> > > > > >> >> > None of these changes are backwards compatible, so I doubt we could make > > > >> >> > the change transparently in QMP. Instead we would have to have a > > > >> >> > QMP greeting message capability where the client can request enablement > > > >> >> > of the enhanced integer handling. > > > >> > > > >> We might be able to do option 1 without capability negotiation. v2.10's > > > >> change from option 1 to what we have now produced zero complaints. > > > >> > > > >> On the other hand, we made that change for a reason, so we may want a > > > >> "send large integers as negative integers" capability regardless. > > > >> > > > >> >> > > > > >> >> > Any of the three options above would likely work for libvirt, but I > > > >> >> > would have a slight preference for either 2 or 3, so that we become > > > >> >> > 100% standards compliant. > > > >> > > > >> There's no such thing. You mean "we maximize interoperability with > > > >> common implementations of JSON". > > > > > > > > s/common/any/ > > > > > > info: error correction applied, future applications will be silent ;-P > > > > > > >> Let's talk implementation for a bit. > > > >> > > > >> Encoding and decoding integers in funny ways should be fairly easy in > > > >> the QObject visitors. The generated QMP marshallers all use them. > > > >> Trouble is a few commands still bypass the generated marshallers, and > > > >> mess with the QObject themselves: > > > >> > > > >> * query-qmp-schema: minor hack explained in qmp_query_qmp_schema()'s > > > >> comment. Should be harmless. > > > >> > > > >> * netdev_add: not QAPIfied. Eric's patches to QAPIfy it got stuck > > > >> because they reject some abuses like passing numbers and bools as > > > >> strings. > > > >> > > > >> * device_add: not QAPIfied. We're not sure QAPIfication is feasible. > > > >> > > > >> netdev_add and device_add both use qemu_opts_from_qdict(). Perhaps we > > > >> could hack that to mirror what the QObject visitor do. > > > >> > > > >> Else, we might have to do it in the JSON parser. Should be possible, > > > >> but I'd rather not. > > > >> > > > >> >> My preference would be 3 with the strings defined as being > > > >> >> %x lower case hex formated with a 0x prefix and no longer than 18 characters > > > >> >> ("0x" + 16 nybbles). Zero padding allowed but not required. > > > >> >> It's readable and unambiguous when dealing with addresses; I don't want > > > >> >> to have to start decoding (2) by hand when debugging. > > > >> > > > > >> > Yep, that's a good point about readability. > > > >> > > > >> QMP sending all integers in decimal is inconvenient for some values, > > > >> such as addresses. QMP sending all (large) integers in hexadecimal > > > >> would be inconvenient for other values. > > > >> > > > >> Let's keep it simple & stupid. If you want sophistication, JSON is the > > > >> wrong choice. > > > >> > > > >> > > > >> Option 1 feels simplest. > > > > > > > > But will still fail with any JSON impl that uses double precision floating > > > > point for integers as it will loose precision. > > > > > > > >> Option 2 feels ugliest. Less simple, more interoperable than option 1. > > > > > > > > If we assume any JSON impl can do 32-bit integers without loss of > > > > precision, then I think we can say it is guaranteed portable, but > > > > it is certainly horrible / ugly. > > > > > > > >> Option 3 is like option 2, just not quite as ugly. > > > > > > > > I think option 3 can be guaranteed to be loss-less with /any/ JSON impl > > > > that exists, since you're delegating all string -> int conversion to > > > > the application code taking the JSON parser/formatter out of the equation. > > > > > > Double-checking: do you propose to encode *all* numbers as strings, or > > > just certain "problematic" numbers? > > > > > > If the latter, I guess your idea of "problematic" is "not representable > > > exactly as double precision floating-point". > > > > We have a few options > > > > 1. Use string format for values > 2^53-1, int format below that > > 2. Use string format for all fields which are 64-bit ints whether > > signed or unsigned > > 3. Use string format for all fields which are integers, even 32-bit > > ones > > > > I would probably suggest option 2. It would make the QEMU impl quite > > easy IIUC, we we'd just change the QAPI visitor's impl for the int64 > > and uint64 fields to use string format (when the right capability is > > negotiated by QMP). > > > > I include 3 only for completeness - I don't think there's a hugely > > compelling reason to mess with 32-bit ints. > > What about when the size is architecture dependent? The QAPI visitor for 'int' uses an 'int64_t' parameters, so I think that will want to be string encoded, as if it was a 64-bit int, even if built on a 32-bit platform.

> > > Option 1 is the bare minimum needed to ensure precision, but to me > > it feels a bit dirty to say a given field will have different encoding > > depending on the value. If apps need to deal with string encoding, they > > might as well just use it for all values in a given field. > > Yeh, 1 is horrid; it's too easy to miss a case which forgot to handle > the 2^53-1 because we hadn't forced a large value down that check.

Daniel P. Berrangé <berrange(a)redhat.com&gt; writes: > On Wed, May 08, 2019 at 02:44:07PM +0200, Markus Armbruster wrote: [...] >> Double-checking: do you propose to encode *all* numbers as strings, or >> just certain "problematic" numbers? >> >> If the latter, I guess your idea of "problematic" is "not representable >> exactly as double precision floating-point". > > We have a few options > > 1. Use string format for values > 2^53-1, int format below that > 2. Use string format for all fields which are 64-bit ints whether > signed or unsigned > 3. Use string format for all fields which are integers, even 32-bit > ones > > I would probably suggest option 2. It would make the QEMU impl quite > easy IIUC, we we'd just change the QAPI visitor's impl for the int64 > and uint64 fields to use string format (when the right capability is > negotiated by QMP). > > I include 3 only for completeness - I don't think there's a hugely > compelling reason to mess with 32-bit ints. Agree. > Option 1 is the bare minimum needed to ensure precision, but to me > it feels a bit dirty to say a given field will have different encoding > depending on the value. If apps need to deal with string encoding, they > might as well just use it for all values in a given field. I guess that depends on what this interoperability capability does for QMP *input*. For *output*, QEMU has to encode a number either as JSON number or as JSON string For *input*, QEMU could accept either. Or it could accept only the encoding it produces on output. Got a preference?

On 5/13/19 8:53 AM, Markus Armbruster wrote: >> We have a few options >> >> 1. Use string format for values > 2^53-1, int format below that >> 2. Use string format for all fields which are 64-bit ints whether >> signed or unsigned >> 3. Use string format for all fields which are integers, even 32-bit >> ones >> >> I would probably suggest option 2. It would make the QEMU impl quite >> easy IIUC, we we'd just change the QAPI visitor's impl for the int64 >> and uint64 fields to use string format (when the right capability is >> negotiated by QMP). >> >> I include 3 only for completeness - I don't think there's a hugely >> compelling reason to mess with 32-bit ints. > > Agree. Other than if we ever change the type of a QMP integer. Right now, if we widen from 'int32' to 'int' (aka 'int64'), it is invisible to clients; but once we start stringizing 64-bit numbers (at client request) but NOT 32-bit numbers, then changing a type from 32 to 64 bits (or the converse) becomes an API change to clients. Introspection will at least let a client know which form to expect, but it does mean we have to be more aware of typing issues going forward.

>> Option 1 is the bare minimum needed to ensure precision, but to me >> it feels a bit dirty to say a given field will have different encoding >> depending on the value. If apps need to deal with string encoding, they >> might as well just use it for all values in a given field. > > I guess that depends on what this interoperability capability does for > QMP *input*. "Be liberal in what you accept, strict in what you produce" - that argues we should accept both forms on input (it's easy enough to ALWAYS permit a string in place of an integer, and to take an in-range integer even when we would in turn output it as a string).

On Tue, May 14, 2019 at 08:02:49AM +0200, Markus Armbruster wrote: > Eric Blake <eblake(a)redhat.com&gt; writes: > > > On 5/13/19 8:53 AM, Markus Armbruster wrote: > > > >>> We have a few options > >>> > >>> 1. Use string format for values > 2^53-1, int format below that > >>> 2. Use string format for all fields which are 64-bit ints whether > >>> signed or unsigned > >>> 3. Use string format for all fields which are integers, even 32-bit > >>> ones > >>> > >>> I would probably suggest option 2. It would make the QEMU impl quite > >>> easy IIUC, we we'd just change the QAPI visitor's impl for the int64 > >>> and uint64 fields to use string format (when the right capability is > >>> negotiated by QMP). > >>> > >>> I include 3 only for completeness - I don't think there's a hugely > >>> compelling reason to mess with 32-bit ints. > >> > >> Agree. > > > > Other than if we ever change the type of a QMP integer. Right now, if we > > widen from 'int32' to 'int' (aka 'int64'), it is invisible to clients; > > but once we start stringizing 64-bit numbers (at client request) but NOT > > 32-bit numbers, then changing a type from 32 to 64 bits (or the > > converse) becomes an API change to clients. Introspection will at least > > let a client know which form to expect, but it does mean we have to be > > more aware of typing issues going forward. > > Thank you so much for helping my old synapses finally fire! Option 2 is > not what we thought it is. Let me explain. > > Introspection reports *all* QAPI integer types as "int". This is > deliberate. > > So, when the client that negotiated the interoperability capability sees > "int", it has to accept *both* integer encodings: JSON number and JSON > string. > > The difference between option 1 and option 2 for the client is that > option 2 will use only one encoding. But the client must not rely on > that! Another QEMU version may well use the other encoding (because we > narrowed or widened the QAPI integer type in the QAPI schema). > > Elsewhere in this thread, David pointed out that option 1 complicates > testing QEMU: full coverage requires passing both a small number (to > cover JSON number encoding) and a large number (to cover JSON string > encoding), to which I replied that there are very few places to test. > > Option 2 complicates testing clients: full coverage requires testing > with both a version of QEMU (or a mock-up) that uses wide integers > (encoded as JSON string) and narrow integers (encoded as JSON number). > Impractical. > > >>> Option 1 is the bare minimum needed to ensure precision, but to me > >>> it feels a bit dirty to say a given field will have different encoding > >>> depending on the value. If apps need to deal with string encoding, they > >>> might as well just use it for all values in a given field. > >> > >> I guess that depends on what this interoperability capability does for > >> QMP *input*. > > > > "Be liberal in what you accept, strict in what you produce" - that > > argues we should accept both forms on input (it's easy enough to ALWAYS > > permit a string in place of an integer, and to take an in-range integer > > even when we would in turn output it as a string). > > With option 2, QEMU *has* to be liberal in what it accepts, because the > client cannot deduce from introspection whether the integer is wide or > narrow. > > [...] > > Daniel, you wrote you'd probably suggest option 2. Would you like to > reconsider? Based on the above, let me try & summarize what we need behaviour to be: - Integer mode (current default): - QEMU & clients MUST format integer fields as numbers regardless of size - QEMU & clients MUST parse number format for any integer fields - String mode: - QEMU & clients MUST format integer fields as strings if their value can not fit in a 32-bit integer. - QEMU & clients MAY format integer fields as strings even if their value can fit in 32-bit integer - QEMU & client MUST parse both string and number format for any integer fields. Unless I'm missing something, this should ensure we don't loose precision, can always parse large numbers, and can internally change QEMU precision from int8/16/32 upto full int64 without breaking clients.

Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|

On Tue, May 14, 2019 at 10:37:55AM +0100, Dr. David Alan Gilbert wrote: > * Daniel P. Berrangé (berrange(a)redhat.com) wrote: > > On Tue, May 14, 2019 at 08:02:49AM +0200, Markus Armbruster wrote: > > > Eric Blake <eblake(a)redhat.com&gt; writes:

> > Unless I'm missing something, this should ensure we don't loose precision, > > can always parse large numbers, and can internally change QEMU precision > > from int8/16/32 upto full int64 without breaking clients. > > But we could be stricter and simpler in string mode: > > - QEMU & clients MUST format integer fields as strings, always > - QEMU & clients MUST parse only strings for integer fields. > > That's (3) above, but also meets your requirements. Yep, given that we don't actually expose the int8/int16/int32/int64 distinction via the QMP introspection data, that would be fine too. Its basically saying we'll never use JSON's number format.

We've discussed possible solutions. Is anyone working or intending to work on patches?