http: Skip req and beresp trailers #4125
Conversation
|
All failing platforms appear to overflow the workspace used in the h2 test case, except the sanitizers job that fails on something else. I will have a closer look later. |
|
As it turned out, the failing tests were exposing a race revolving around the request body status. Transitioning to I took care of that and added a couple patches next to the ones picked from #3798 at the beginning of this patch series. I'm no longer able to stress test cases and cause failures and I believe the race is gone. I have not yet studied the remaining CI failure, triggering a panic from the workspace emulator. I suspect either a logic error in the workspace pipelining (expanding from req keep-alive only to req and beresp trailers) or an error in how it translates in the workspace emulator. I also started an effort to improve feature parity between HTTP/1 and h2 |
|
As I suspected, the emulated |
|
bugwash: The following commits can be merged: 23ec1fc...9875258 The workspace refactoring commits can be submitted independently. |
|
I'm fine with you commiting those first six commits while we wait for @nigoroll |
|
As per bugwash:
|
Until now, we read the (CR)?LF at the end of a chunk as part of the next chunk header (see: /* Skip leading whitespace */). For a follow up commit, we are going to want to know if the next chunk header is available for read, so we now consume the chunk end as part of the chunk itself. This also fixes a corner case: We previously accepted chunks with a missing end-of-chunk (see fix of r01729.vtc). Ref: https://datatracker.ietf.org/doc/html/rfc7230#section-4.1
... which we are going to need in a follow up commit. No functional changes, diff best viewed with -b
While working on request body caching improvements, it was noticed that, for chunked encoding, we did not know the request body size before we attempted to read at least one additional byte because the chunked fetch did not return VFP_END with the last chunk, but rather only as a zero-length additional chunk. This commit changes chunked encoding processing to return VFP_END opportunistically if the next chunk header is available. Implementation: Unless there is pipeline readahead data available, the test for available data implies a poll(2) call. Relative to the existing cost of the current implementation, this cost is not considered relevant. To improve efficiency, we should consider a generic readahead for chunked encoding, probably by making the file descriptor non blocking and adding a readahead buffer.
It seems this test now shows no data loss more frequently, which I hope should be fine for the purpose of the test?
Now that WS_Pipeline() can be used on backend side too, we may have situations where we run out of workspace when copying pipelined data. We should gracefully handle these failures by failing the task instead of panicking. For existing fail-safe call sites, we check the result.
This will be the method responsible for reading and parsing http trailers. This commit only prepares the structure, the method is implemented in follow up commits.
This will allow us to check that we have received a complete end of chunked body, possibly including HTTP traillers. Co-authored-by: Dridi Boukelmoune <[email protected]>
It will be used to convey that the body was fully fetched and trailers are expected or were already encountered.
For now, we simply do nothing about them.
It cannot be reused for req headers, but it will apply to beresp and req trailers.
At this point trailer fields are just checked for correctness and discarded. Co-authored-by: Walid Boudebouda <[email protected]>
It is no longer the responsibility of the VFP to read the final (CR)?LF in chunked bodies, because trailers are not part of the request or response body. For backend responses trailers are handled separately. The only reason why this does not break keep-alive for the client is that (CR)?LF sequences are legitimately ignored between requests, so an empty trailer list just turns into a no-op (CR)?LF for the next client request. This should also be the case for backend responses, but having extra unread empty lines breaks connection pooling instead. We can probably as part of recycling connections clear remaining (CR)?LF sequences and close the connection with SC_JUNK if something remains.
This is not allowed for HEADERS frames containing trailers.
This is currently the only state where HEADERS frames are processed, but only a subset of the operations will be shared by trailers. Better diff with the --ignore-all-space option.
It will not necessarily be a new stream once trailers are involved.
It is now possible to receive a HEADERS frame on an OPEN stream. The HPACK block is processed but trailers are skipped. Therefore a workspace overflow is not a failure. Trailers are processed from the h2 session, and the request might be processed concurrently in a different worker, so we can't use the request workspace, even as a temporary buffer. Instead of relying on scratch space on the stack, we can use the h2 session's req workspace. This will make things easier once we collect trailers instead of merely skipping them.
|
Force-pushed to rebase against trunk after merging #4130. |
This is the first step of the following plan:
experimental::pass_trailersflagThe first step, this pull request, should allow applications producing trailers to work with Varnish when the trailers are not strictly required.
The second step should allow protocols like gRPC to pass through Varnish, with the caveats that body filters could break the meaning of trailers (for example checksums), hence the
experimentalflag.It will require internal changes, and from the look of it we may not be able to directly use the existing data structures and we may need to a dedicated
OA_TRAILERSobject attribute.The last step will likely require changes to the VCL state machines in addition to new symbols.
Most of the work on trailers was done by @walid-git under my supervision.
This pull request includes work from other pull requests too:
BS_CACHEDinto a request flagBS_TRAILERSbody status for coordinationgettrls()callback for directors (missing docs update)beresptrailersreqtrailersThis is a lot of commits, but they should all have a reasonable size for reviewers. I already reviewed Walid's work, and already addressed my own review items while Walid is away. I added Walid or myself as a co-author when I made significant changes (at the scale of individual commits).