Allow HTTP2 encoder to split headers across frames

[ladeak]

Allow the HTTP2 encoder to split headers across frames

Enable Kestrel's HTTP2 to split large HTTP headers into HEADER and CONTINUATION frames.

Description

Kestrel's HTTP2 implementation limits the max header size to the size of the frame size. If a header's size is larger than the frame size, it throws an exception: throw new HPackEncodingException(SR.net_http_hpack_encode_failure);. RFC 7540 allows the headers to be split into a HEADER frame and CONTINUATION frames.
Before the change Kestrel only used CONTINUATION frames when headers fitted fully within a frame.

This PR changes the above behavior by allowing to split even a single HTTP header into multiple frames. It uses an ArrayBufferWriter<byte> (similar to .NET runtime), to back the buffer used by the HPack encoder.

When the HPack encoder reports that a single header does not fit the available buffer, the size of the buffer is increased. Note, that the .NET runtime implementation on HttpClient writes all headers to a single buffer before pushing it onto the output, contrary to this implementation that keeps the semantics of Kestrel. It only increases the buffer when a single header fails to be written to the output, otherwise the old behavior is kept. My intention was to keep this behavior so that memory-wise it does not use more memory than the single largest header or the max frame size.
With this PR HPackHeaderWriter uses an enum to tell Http2FrameWriter to increase the buffer or not. When the buffer is too small, its size is doubled.

This behavior is also implemented for trailers. Note that in case of headers, the HEADER frame is never empty because of the response status, while this is not true for trailers. Hence there is a subtle difference when getting the buffer for the initial frame of a header vs. a trailer.

I updated existing tests asserting the previous behavior and added new tests to validate the proposed changes.

Performance

Performance-wise the change is not expected to increase throughput (given it must do more to enable this use-case) but the goal is to have the 'slow-path' is only in the case when a single header is too large. I used the existing Http2FrameWriterBenchmark to compare performance before and after:

BenchmarkDotNet=v0.13.0, OS=Windows 10.0.22631
12th Gen Intel Core i7-1255U, 1 CPU, 12 logical and 10 physical cores
.NET SDK=9.0.100-preview.4.24218.26
  [Host]     : .NET 9.0.0 (9.0.24.21901), X64 RyuJIT
  Job-KLBPPQ : .NET 9.0.0 (9.0.24.21807), X64 RyuJIT

Before changes (updated 2024. 05. 04. main):

Method	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
WriteResponseHeaders	79.64 ns	0.581 ns	0.515 ns	12,556,569.9	0.0002	-	-	32 B

Method	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
WriteResponseHeaders	77.43 ns	0.355 ns	0.315 ns	12,914,602.3	0.0002	-	-	32 B

After changes rebased on main (Updated with Validating the header length in HPackHeaderWriter)

Method	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
WriteResponseHeaders	79.42 ns	0.295 ns	0.230 ns	12,591,619.9	0.0002	-	-	32 B

Method	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
WriteResponseHeaders	80.30 ns	0.571 ns	0.506 ns	12,453,404.1	0.0002	-	-	32 B

Fixes #4722

[ladeak]

cc. @amcasey who has been in the discussions of the corresponding issue.

[amcasey]

Thanks for this! Thorough, as always.

Some notes:

1. The CI failure is unrelated. No point in re-running, since we're expecting revisions.
2. It may be a little while before I have time to review this in detail.
3. The description mentions doubling the size of a buffer, as needed - I assume there's a cap on the size of that buffer so it can't grow indefinitely and DoS the server?
4. The description (and tests) only mention spreading a header into a single CONTINUATION. I'm assuming this will work if the header is, e.g. 33 KB?
5. Is there a place we could use a simple flag to disable the new behavior and revert to the old behavior? e.g. if I wanted a NeverSplitHeaders appcontext switch, would that be straightforward to add or would be need to duplicate a bunch of code.

Some test questions (you may already have code/coverage for this - I haven't checked):

1. What happens if the header name is well-known and gets compressed? Is the 16KB limit based on the compressed or uncompressed size?
2. What happens if the header name itself is more than 16KB? (Throwing might be appropriate, but we should at least know.)
3. What happens if the header is larger than the maximum total header size? (I'm pretty sure there's an Http2ServerLimit for this.)
4. I noticed that HPACK had some handling for never-compressed literals. Does this work for those?
5. If there's a very short header before the very long header (e.g. one that compresses to one or two bytes), is that first HEADERS frame tiny?

[ladeak]

@amcasey let me reply inline:

3. The description mentions doubling the size of a buffer, as needed - I assume there's a cap on the size of that buffer so it can't grow indefinitely and DoS the server?

You are wondering about headers that are roundtripped to the client with large size (previously would fail, now it could DDoS)? I need to check Kestrel's H2 header size limits (you also mention), but there is nothing in the Http2FrameWriter in this regard.

4. The description (and tests) only mention spreading a header into a single CONTINUATION. I'm assuming this will work if the header is, e.g. 33 KB?

It can span into zero or more CONTINUATION frames.

5. Is there a place we could use a simple flag to disable the new behavior and revert to the old behavior? e.g. if I wanted a NeverSplitHeaders appcontext switch, would that be straightforward to add or would be need to duplicate a bunch of code.

There is no such place, but it could be very well built along the Kestrel's limit or an AppContext switch. Please let me know if building such a would be preference. But note, that previously "possible" use-cases still work the same as before, so the switch would only control if large headers are allowed or not -> hence a limit might be suitable option.

1. What happens if the header name is well-known and gets compressed? Is the 16KB limit based on the compressed or uncompressed size?

I did not come across compression/no-compression on this path. HPack encodes the header values into this buffer.

2. What happens if the header name itself is more than 16KB? (Throwing might be appropriate, but we should at least know.)

The header is written to a buffer, which is split into CONTINUATION frames, so it does not matter if the name or the value is being oversized.

3. What happens if the header is larger than the maximum total header size? (I'm pretty sure there's an Http2ServerLimit for this.)

MaxRequestHeaderFieldSize ? -> I need to test the behavior.

4. I noticed that HPACK had some handling for never-compressed literals. Does this work for those?

It works on anything that HPack writes to my understanding. I will double confirm.

5. If there's a very short header before the very long header (e.g. one that compresses to one or two bytes), is that first HEADERS frame tiny?

-> If the long one does not fit in the same frame, yes, the initial header will be sent in a tiny frame. This is even true for the "current" behavior.

[mgravell]

what is the behaviour here if we can't request more (currentLength == 0 && !canRequestLargerBuffer) ? I haven't pulled the code locally, but it looks like we will return MoreHeaders, which will ... keep trying? is there any way this can result in an infinite loop of not making progress?

[ladeak]

This (that it cannot request more) should be only the case for the first HEADER frame which always has status response, so no loop. We could consider completely removing this though.

[mgravell]

the side-effect of getting this wrong would be catastrophic, though; it seems that a throw (or a throw-helper) that is never hit would be a much better "things that never happen" than an infinite loop that is never hit

(my point is: if the code gets changed at some point in the future such that our expectations are no longer true: how should it manifest? in this case, it feels bad enough that a fault is preferable to an infinite loop)

[ladeak]

There is a debug assert on the call site.

[ladeak]

The canRequestLargerBuffer could dropped if we apply a do-while loop on the initial header as well, but I felt that might sacrifice perf. I will do a measurement.

[ladeak]

But if I did that, remove canRequestLargerBuffer and implement a do-while loop in WriteResponseHeadersUnsynchronized to enlarge the header, the perf decreases. Using the same Http2FrameWriterBenchmark as in the description:

Method	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
WriteResponseHeaders	80.23 ns	1.206 ns	1.128 ns	12,463,819.6	0.0002	-	-	32 B

[JamesNK]

Background: I've looked at HTTP2 headers a lot. I wrote some of the dynamic support and rewrote the writer and parser at one point.

I haven't looked through the code in detail yet. Some initial thoughts:

• Well done for jumping into HTTP2. It's complex. And Kestrel's implementation is complex.
• I like the feature. Kestrel's header limit on frame size has always felt arbitrary. We did it because it was easy and fast, not because it was the best thing to do. I recently wrote a feature that put a lot of content into a header and was mindful of this limitation.
• I know Kestrel has various limits for request headers. I don't believe there are limits on response headers (other than the max frame size). The idea is an app is responsible for sending the response headers so no need to limit it. However, we should look to see what HTTP/1.1 does in this area. If it doesn't limit what a response can do with headers, then neither should HTTP/2.
• But we should stress this and put some kind of upper limit. What happens if someone wants a 20-megabyte response header? Or a 2-gigabyte response header? Something should blow up before the server tries to double a gigabyte buffer and blows up from an excessive amount of data.
• Reading and writing response headers is extremely performance critical. Because it's stateful (the dynamic table is on the connection) it basically locks everything. Must reduce any performance overhead.

[amcasey]

@ladeak Thanks! Your responses make sense.

Regarding the appcontext switch, I regard this as a fairly risky change because it's touching such critical code (not because of anything you've done - just the circumstances). It would be good to at least think about how to make it possible to revert (as exactly as possible) to the old behavior. It may turn out to require too much duplicate code or API changes or something that makes it unacceptable, but I think we need to at least know what it would take. Open to differing opinions from @mgravell or @JamesNK.

[JamesNK]

I don't think we need a switch. If this lands in a mid .NET 9 preview, then it should go through a lot of use and testing before GA.

For example, the work David did to rewrite response writing in .NET 7(?) was much more complex and we didn't have a fallback.

[amcasey]

I don't think we need a switch. If this lands in a mid .NET 9 preview, then it should go through a lot of use and testing before GA.

For example, the work David did to rewrite response writing in .NET 7(?) was much more complex and we didn't have a fallback.

Good enough for me. I hadn't considered how well exercised this code is by TechEmpower, etc.

[amcasey]

@ladeak Did you receive the initial feedback you needed? Is this ready for a full review or are you still working on it. There's no rush - I just wondered whether the next steps were our responsibility. Thanks!

[ladeak]

@amcasey Going to come back tomorrow with some findings.

[ladeak]

Discussion about the header size limits: as I understood there is a general desire to have a limit. However, response headers mostly depend on the app, and the way app handles headers. I have not found limits for HTTP/1.1. An empty/default Kestrel ASP.NET Core webapp also allows as large headers as desired with h1. On the consumer-side I ran into limits though (H1):

.NET HttpClient has MaxResponseHeadersLength with the default of 65536 bytes.
curl on Linux - curl: (27) Rejected 140725 bytes header (max is 102400)!
curl on Windows same as on Linux - curl: (27) Out of memory
Chromium - Edge: returns ERR_RESPONSE_HEADERS_TOO_BIG at ~262000 bytes.

When I run the app with IISExpress, it seems to only returns the remainder of 64k. (header mod 65536).

@amcasey , the following questions would need further clarification:

• hardcode a "big enough" limit or to expose this on Kestrel options under the Http2 limits?
• what default value should this limit have?

[amcasey]

hardcode a "big enough" limit or to expose this on Kestrel options under the Http2 limits?
what default value should this limit have?

Since the spec doesn't give a limit, I think we need to give users a way to override whatever we decide. I'm not sure why it would be specific to http/2 though - presumably the same concern applies to http/1.1 or http/3.

My first thought for a default would be double MaxRequestHeadersTotalSize (i.e. 64 BK), but @JamesNK makes the sensible point that it's really at the server's discretion and we really just need a cap that prevents things from getting out of hand - maybe 1 MB?

If we were to decide this shouldn't get a public API, I'd want to go even higher - maybe 10 MB.

Thoughts, @JamesNK @mgravell?

[ladeak]

@amcasey I think I have not thought about http/1.1 about a limit, given it does not have currently, and setting 64KB would be breaking, wouldn't it? (I am not sure how difficult it could be to implement this for http/1.1, http/3 looks similar to h2 in code structure. But it makes sense from the point of view you describe that it could be a setting that applies to all versions of http.

A question if it is public: should it apply to a single header or to the total headers. Consumers HttpClient and Edge had a total while curl per header limit.
If it is total headers, does it include trailers?

[amcasey]

A question if it is public: should it apply to a single header or to the total headers. Consumers HttpClient and Edge had a total while curl per header limit.
If it is total headers, does it include trailers?

Because we're guarding against resource utilization rather than malformed responses, I think the limit should apply to the total size of all headers, rather than to the size of any individual header. Similarly, if we're reducing the whole detection mechanism to a single number, I would expect trailers to be included. I'm open to feedback on both.

I think I have not thought about http/1.1 about a limit, given it does not have currently, and setting 64KB would be breaking, wouldn't it?

Yes, it would. I think we generally accept breaks in service of DoS prevention, but I agree that this is a strong argument for choosing a default that is larger than we expect anyone to use in practice.

If we felt really strongly about this, I could live with adding limits to both http/2 and http/3 and not to KestrelServerLimits directly. I just don't want to end up in a state where we have three identical properties (or more, when http/4 comes out).

[ladeak]

@amcasey , I added a commit that has a new limit on KestrelServerLimits, and this is also respected by Http2FrameWriter.
I think the code ended up quite a bit complicated (enforcing the limit for written headers + headers not yet written but requiring a larger buffer, etc.).

One thing I found on the way: SETTINGS_MAX_HEADER_LIST_SIZE is an advisory setting, now with the limit might worth also respecting it. Not all clients send it (for example, HttpClient does not) - unfortunately it is not sufficient against DoS.

I would expect similar implementation would be needed on H/1.1 and H/3 (because the public setting is on Kestrel level), hence not sure if a public limit is worth all the complexity to be added.

Maybe a setting called MaxHeaderBufferSize would simplify things a lot, but I could also see why you would not want to expose such an implementation detail related setting on a public API. So that leads my thought process back to your initial suggestion: an appcontext switch for MaxHeaderBufferSize - although not sure if numbers are supported there.

[amcasey]

Thanks for the prototype and the thoughtful write-up.

@amcasey , I added a commit that has a new limit on KestrelServerLimits, and this is also respected by Http2FrameWriter. I think the code ended up quite a bit complicated (enforcing the limit for written headers + headers not yet written but requiring a larger buffer, etc.).

Assuming no perf impact, I would probably accept that level of complexity to get the extra protection. Having said that, it feels like there are ways we could reduce the complexity. What if we capped the header size before hpack? Would that let us do a single check up-front? Given how much larger we expect the limit to be than any app would reasonably use, I don't think hpack is going to be what saves people (i.e. by keeping them under the limit).

One implementation note: when you give people an option like this, it's not unusual for them to pass int.MaxValue, so you have to think very carefully about any additions and ordered comparisons you do. If we keep the checks, I think they will probably need to be hardened against that.

One thing I found on the way: SETTINGS_MAX_HEADER_LIST_SIZE is an advisory setting, now with the limit might worth also respecting it. Not all clients send it (for example, HttpClient does not) - unfortunately it is not sufficient against DoS.

It does seem like a well-behaved server ought to respect that setting. Maybe we had a reason for not already doing so? @halter73 @JamesNK?

If we were to add that functionality (possibly in a separate PR), it would be important to ensure that it uses different error text from the internal server limit so app authors know they can't control it.

I would expect similar implementation would be needed on H/1.1 and H/3 (because the public setting is on Kestrel level), hence not sure if a public limit is worth all the complexity to be added.

I would agree that a public limit should apply to all protocols. Would the H/1.1 and H/3 changes be simpler if we used a limit on the pre-compression size?

Again, any insight from @halter73 @JamesNK on why H/1.1 doesn't already have such a limit would be welcome.

Maybe a setting called MaxHeaderBufferSize would simplify things a lot, but I could also see why you would not want to expose such an implementation detail related setting on a public API.

I'm not sure I understand the suggestion. Isn't the buffer the thing we use for breaking the header into pieces? Does your proposed setting limit the size of each piece or is it a different buffer?

So that leads my thought process back to your initial suggestion: an appcontext switch for MaxHeaderBufferSize - although not sure if numbers are supported there.

It's slightly wonky. You can use Set/GetData but you get an object and then you have to check for both int and string. It's not ideal, but it's possible. I could live with making it an appcontext switch at first and eventually promoting it to a setting but, with the information I have now, I think I would still lean towards having a public setting. As always, I'm open to arguments in favor of going a different way.

There's an example here.

[ladeak]

Assuming no perf impact, I would probably accept that level of complexity to get the extra protection. Having said that, it feels like there are ways we could reduce the complexity. What if we capped the header size before hpack? Would that let us do a single check up-front? Given how much larger we expect the limit to be than any app would reasonably use, I don't think hpack is going to be what saves people (i.e. by keeping them under the limit).

I would agree that a public limit should apply to all protocols. Would the H/1.1 and H/3 changes be simpler if we used a limit on the pre-compression size?

The reason I went the way it is implemented, because the encoding is also applied by the HPack writer, but this suggestion makes sense to investigate. I will check if there is any common place for all protocols to perform this.

One implementation note: when you give people an option like this, it's not unusual for them to pass int.MaxValue, so you have to think very carefully about any additions and ordered comparisons you do. If we keep the checks, I think they will probably need to be hardened against that.

Makes sense, thank you for the reminder.

I'm not sure I understand the suggestion. Isn't the buffer the thing we use for breaking the header into pieces? Does your proposed setting limit the size of each piece or is it a different buffer?

Each piece - the reason I keep coming back to this idea is because my understanding was that the problem is allocating a really large buffer, and this would be an easy check. But as discussed above, let me try to pursue validating the total headers before the write operation.

[ladeak]

I am thinking (and moving - wip) the limit logic to HttpProtocol after the headers collection is marked readonly. I believe it is going to be a common place for all protocol versions.

[amcasey]

I was actually thinking of a single check for each protocol, but I'm fine with merging the checks if that's an option.

[amcasey]

Each piece - the reason I keep coming back to this idea is because my understanding was that the problem is allocating a really large buffer, and this would be an easy check. But as discussed above, let me try to pursue validating the total headers before the write operation.

Yeah, I think I was just mixing myself up. Preventing the actual problem seems like a viable way forward. We just need to make sure we're able to give the user an intelligible setting.

[ladeak]

I was actually thinking of a single check for each protocol, but I'm fine with merging the checks if that's an option.

In HttpProtocol I would need to iterate the headers an additional time, so I don't think it would be viable option eventually.

What if we capped the header size before hpack?

I am still looking at this idea. If I want to avoid iterating all headers an additional time, I find it very suitable to calculate the header length in Http2FrameWriter where the headers are iterated, and in HPackHeaderWriter's EncodeHeadersCore method where the encoder available at hand to calculate a "more-or-less" length before applying the HPACK compression.

I will prepare a prototype of this solution. The confusing part is that EncodeHeadersCore already has an out int length parameter, and I feel the urge to introduce a second out int rawLength or so parameter.

[amcasey]

Thanks! I agree that an additional iteration would be undesirable.

I'm going to be out of town next week, but @mgravell should be around to answer your questions (except Monday, which is a holiday for him).

[ladeak]

@amcasey (and @mgravell ) I moved the size validation into HPackHeaderWriter. It is using the payload length when available from the underlying HPack encoder, or it 'calculates' a header size disregarding static and dynamic tables. If the encoder used those tables, it would have been very likely that header still fits within the buffer anyway.

This PR is still not covering HTTP/1.1 and HTTP/3, but I remove the Draft label, so then we can discuss the other HTTP version and the new limit.
For HTTP/3, I believe the same frame size issue is present, so that might worth having a look in a separate PR maybe?
For HTTP/1.1 I am slightly concerned about the default value of the new limit being too aggressive.

Another consideration on the way the headers length is aggregated: because it is HTTP version specific, and because in the current HTTP/2 the implementation piggy backs onto HPack's calculation for perf reasons, it means the shared KestrelServerLimits will result different behavior on the different version of HTTP, given different compression/encoding/static-dynamic tables used.

So, if this approach is kept, that we have the limit calculated for each protocol (and to me it seems reasonable), maybe separate settings would make more sense? (unless it gets too granular, or these differences are acceptable)