Max Karolinskiy
da886f9678
ObliviousHttpResponse headers field cannot be skipped because it's not a nullable field and the upstream change below now validates fields. The header value is not important since we don't check it in the test. Chromium change: https://chromium.googlesource.com/chromium/src.git/+/d8255385dd669 commit d8255385dd669a6580e605b8d9b04a86ed91b5c9 Author: Daniel Cheng <dcheng@chromium.org> Date: Tue Apr 14 19:18:52 2026 -0700 Reland "Remove use deprecated [Native] types in network_param.mojom" This is a reland of commit c565345fc495c824f8a1c54f11461dbfaf235551 The only additional fix is to mark HTTPResponseHeaders as nullable in NetworkService::ParseHeaders(). Original change's description: > Remove use deprecated [Native] types in network_param.mojom > > This ends up being large change since it is easiest to convert an entire > tree of `[Native]` types at once, as it's a awkward to mix and match > `[Native]` and non-`[Native]` types without introducing temporary > `[Native]` types to bridge the transition. The opaque nature of > `[Native]` types also means that there were types that Mojo didn't even > know about at all, further increasing the size of the change. > > To avoid making network_param.mojom a dumping ground of arbitrary types, > define the new types in distinct .mojom files, using a simple heuristic: > if the C++ types are defined in a single header, the Mojo types should > be defined in a single .mojom file as well. Several types previously > defined in network_param.mojom are also moved out to match this > heuristic. > > The typemap definitions are also reorganized and split into smaller > blocks; in theory, this should allow more granular dependencies, though > this CL does not take full advantage of this since all the trait > definitions are built as one source set still. Additional cleanups: > - normalize how .cc trait sources are defined in the GN rules–the > typemap definitions no longer use `trait_sources` and depend on the > `source_set` that includes all the trait sources > - normalize where trait declarations and definitions live: the traits > for X509Certificate are now in x509_certificate_mojom_traits.{cc,h} > rather than being split across x509_certificate_mojom_traits.h and > network_param_mojom_traits.cc. > > Many typemaps also need to become shared typemaps now. This is necessary > because `URLLoaderCompletionStatus` transitively depends on `SSLInfo`, > which was previously a `[Native]` type. Unfortunately, this results in a > large compile size increase which is unavoidable: `[Native]` types are > entirely opaque to Mojo and as such, very little generated code was > needed to support them. Teaching Mojo about these types results in a lot > more generated code. > > `HttpResponseHeaders` now has an internal, shared serialization helper, > since serializing to a pickle first and then serializing to Mojo is an > additional wasted copy. This helper is used by the original `Persist()` > method, as well as the new `SerializeForMojoIpc()` method. It would be > possible to do even better here, i.e. by teaching Mojo how to serialize > a segmented string, but that is left for a future followup if additional > efficiency is needed. > > Since `HttpResponseHeaders` and `X509Certificate` are no longer opaque > types, nullability checks are now enforced by Mojo deserialization: this > revealed several bugs (in `CertVerifierService`, `NetworkContext`, and > `URLResponseHead`) where fields/parameters are not marked as nullable > despite comments indicating it could be null. > > Finally, though it's not strictly necessary, make `HttpResponseHeaders` > a shared typemap as well. The raw form of `HttpResponseHeaders` is hard > to hold correctly, and it's safer to use the `net::HttpResponseHeaders` > to handle it. > > Bug: 393179188