[davidfowl]

Reduce per-request allocations in StaticFileMiddleware

• You've read the Contributor Guide and Code of Conduct.
• You've included unit or integration tests for your change, where applicable.
• You've included inline docs for your change, where applicable.
• There's an open issue for the PR that you are making.

Remove avoidable allocations from the UseStaticFiles request path.

Description

Every request that reaches StaticFileMiddleware calls FileExtensionContentTypeProvider.TryGetContentType. That call allocated an extension substring for the lookup. Each served or revalidated file also paid for a few small allocations in StaticFileContext. This PR removes them. All changes are internal.

• FileExtensionContentTypeProvider
  • When Mappings is exactly a Dictionary<string, string> whose comparer supports span lookups, the constructor caches an AlternateLookup<ReadOnlySpan<char>>. This covers the default instance (OrdinalIgnoreCase) and user dictionaries with Ordinal or OrdinalIgnoreCase.
  • The extension is then looked up as a span, so no substring is allocated. Every other mapping type, including derived dictionaries, still uses IDictionary.TryGetValue(string).
  • The alternate lookup reads the same dictionary instance, so later changes to Mappings are still seen. The extension rules (IsNullOrWhiteSpace, then last . to the end) are unchanged.
• StaticFileContext.LookupFileInfo
  • The ETag string is built with a single string.Create(InvariantCulture, $"\"{hash:x}\""), replacing '"' + Convert.ToString(hash, 16) + '"' (two strings).
  • The output is the same lowercase two's-complement hex. I checked this against Convert.ToString for 1M random values plus edge cases, and a new test covers negative hashes.
• StaticFileContext.ApplyResponseHeadersAsync
  • Writes HeaderUtilities.FormatDate(lastModified) and etag.ToString() straight into Response.Headers, so it no longer allocates a typed ResponseHeaders wrapper.
  • These are exactly the values the typed setters wrote. The typed wrapper is still created lazily for Content-Range.
• GetMaxPreconditionState: no longer takes a params PreconditionState[], which allocated an array per request.

Headers, status codes and log output are unchanged.

Benchmark

I used a local BenchmarkDotNet harness (not included). It runs StaticFileMiddleware with a PhysicalFileProvider on a DefaultHttpContext. Timings are on 2 cores, 3 runs each.

FileExtensionContentTypeProvider.TryGetContentType (default provider)

Path	Before	After
/css/site.css	25.6–25.9 ns, 32 B	18.3–19.0 ns, 0 B (−27%)
/lib/bootstrap/dist/js/bootstrap.bundle.min.js	26.2–26.7 ns, 32 B	16.2–16.3 ns, 0 B (−38%)
/unknown.foo	20.2–21.9 ns, 32 B	13.0–13.2 ns, 0 B (−36%)
/api/values (no extension)	3.69–3.72 ns, 0 B	3.68–3.78 ns, 0 B (unchanged)

StaticFileMiddleware.Invoke: time is dominated by the physical file stat, so these show only the allocation drop.

Request	Before	After
HEAD (200)	2765–2825 ns, 2344 B	2800–2875 ns, 2200 B (−144 B)
GET If-None-Match (304)	3013–3143 ns, 2752 B	3033–3063 ns, 2608 B (−144 B)
GET If-Modified-Since (304)	2960–2995 ns, 2560 B	2966–3047 ns, 2416 B (−144 B)
GET file not found	1502–1567 ns, 1592 B	1530–1533 ns, 1560 B (−32 B)
GET full body	~38 µs, ~20.6 KB	~39 µs, ~20.5 KB (I/O noise)

That is four fewer allocations per served or revalidated file, and one fewer for every request that has an extension.

Tests

• Microsoft.AspNetCore.StaticFiles.Tests: 315 passed, 51 skipped.
• Microsoft.AspNetCore.StaticFiles.FunctionalTests: 20 passed, 5 skipped.
• Microsoft.AspNetCore.StaticAssets.Tests: 64/64.
• New tests:
  • ApplyResponseHeaders_WritesExpectedETagAndLastModified checks the exact ETag against the old Convert.ToString(hash, 16) formula, including negative hashes, and checks Last-Modified, Accept-Ranges and Content-Type.
  • Content-type provider tests cover a case-sensitive custom mapping, changes to Mappings after construction, a trailing-dot key, a non-Dictionary mapping (SortedDictionary), and a derived dictionary that re-implements TryGetValue.

Benchmark (3 interleaved rounds)

New StaticFileBenchmark in src/Middleware/StaticFiles/perf/Microbenchmarks, added in this PR. It replaces the local harness described above. StaticFileMiddleware serves /css/site.css (1 KB) from an in-memory IFileProvider, so the numbers measure the changed CPU and allocation path without disk I/O.

• TryGetContentType_Default (span lookup): 20.1–21.1 ns → 12.0–12.1 ns (about −40%), 32 B → 0 B.
• StaticFileMiddleware.Invoke (Head, Get, GetIfNoneMatch → 304): −144 B per request in every case. Time is within noise: the PR was 2–6% faster in rounds 1–2 and 2–6% slower in round 3.
• TryGetContentType_NonDictionaryMappings (a SortedDictionary mapping on the unchanged string fallback path): still 32 B. It is about 1 ns slower (12.7–13.2 ns → 13.6–14.9 ns) because of the added _useSpanLookup check. This affects only non-Dictionary custom mappings.

Round	Method	main Mean	PR Mean	main Allocated	PR Allocated
1	TryGetContentType_Default	20.1 ns	12.0 ns	32 B	0 B
1	TryGetContentType_NonDictionaryMappings	13.2 ns	13.6 ns	32 B	32 B
1	Head	688.3 ns	665.7 ns	1,912 B	1,768 B
1	Get	914.7 ns	883.7 ns	2,000 B	1,856 B
1	GetIfNoneMatch	1.042 us	984.1 ns	2,288 B	2,144 B
2	TryGetContentType_Default	20.1 ns	12.0 ns	32 B	0 B
2	TryGetContentType_NonDictionaryMappings	12.7 ns	14.9 ns	32 B	32 B
2	Head	676.0 ns	656.0 ns	1,912 B	1,768 B
2	Get	911.5 ns	889.4 ns	2,000 B	1,856 B
2	GetIfNoneMatch	1.003 us	949.9 ns	2,288 B	2,144 B
3	TryGetContentType_Default	21.1 ns	12.1 ns	32 B	0 B
3	TryGetContentType_NonDictionaryMappings	12.7 ns	13.8 ns	32 B	32 B
3	Head	659.6 ns	700.8 ns	1,912 B	1,768 B
3	Get	871.4 ns	897.2 ns	2,000 B	1,856 B
3	GetIfNoneMatch	984.2 ns	1.003 us	2,288 B	2,144 B

Machine: AMD EPYC 9V74, 2 cores, Ubuntu 24.04 container; .NET SDK 12.0.100-alpha.1.26521.103, runtime 12.0.0 (12.0.26.52911); BenchmarkDotNet 0.13.0 --config perflab (InProcess, MemoryDiagnoser), WarmupCount=3, IterationCount=15. Order per round: main, then PR. "main" is the PR branch with the product source files reverted to the merge base (4f556f7), so both builds run the same benchmark source. Means and bytes come from the BDN JSON exporter.

🤖 Posted by ASP.NET Core Perf 2 (AI Agent)

[Copilot]

🟢 Approval recommended

No active review issues remain for the static-file allocation optimizations.

No open issues 🎉

What changed in this PR

Reduces allocations in the static-file request path while preserving content-type lookup behavior and response headers.

Changes:

• Uses span-based extension lookups for compatible dictionaries, retaining the string-based fallback for other mappings.
• Removes intermediate ETag strings, typed response-header wrappers, and precondition arrays.
• Adds regression tests for custom mappings, mapping mutations, and exact response-header values.

File	Description
src/​Middleware/​StaticFiles/​test/​UnitTests/​StaticFileContextTest.cs	Checks exact ETag and response headers, including negative hash values.
src/​Middleware/​StaticFiles/​test/​UnitTests/​DefaultContentTypeProviderTests.cs	Covers mapping compatibility, mutations, and extension edge cases.
src/​Middleware/​StaticFiles/​src/​StaticFileContext.cs	Reduces allocations in ETag formatting, header writing, and precondition evaluation.
src/​Middleware/​StaticFiles/​src/​FileExtensionContentTypeProvider.cs	Adds allocation-free span lookups for compatible dictionaries.

_{Review effort: Balanced}

[davidfowl]

Retriggering CI with an empty commit. The failures are unrelated to this change (it only touches StaticFiles content-type lookup and header writes):

• Build: Windows x64/x86/arm64: eng\Npm.Workspace.nodeproj(168), npm run build exited with code -2147483645 (0x80000003, a crash in the node process). This is an infra/JS build failure, not a compile or test failure.
• Platform tests: Windows Server x64: Templates.Blazor.Tests failed. The same job failed the same way on the unrelated Reduce allocations in output formatter content negotiation #70166 (MVC output formatters). Across the 35 most recent PRs this job failed on 2 of them (~6%), both with Templates.Blazor.Tests.
• Blazor E2E (Mono server): RoutingTest.NavigationLock_CanBlockExternalNavigation timed out waiting for the browser URL (/subdir/ vs /myexternalpath3). This is a navigation-timing test; the Blazor E2E jobs failed on 2 of the ~25 recent PRs that ran them (Avoid resolving the default hub activator from DI for every hub invocation #70144 and Reduce per-request allocations in W3CLoggingMiddleware #70176, both unrelated).

🤖 Posted by ASP.NET Core Perf 2 (AI Agent)

[davidfowl]

Pushed empty commit a1d008f to retrigger CI. The only failures were in aspnetcore-components-e2e (Blazor E2E, Linux, Mono server):

• ServerRoutingTest.NavigationLock_CanBlockExternalNavigation (expected //myexternalpath3, got /subdir/): the same failure hit unrelated Reduce per-request allocations in W3CLoggingMiddleware #70176 (HttpLogging) and Avoid per-request allocations in ForwardedHeadersMiddleware #70177 (HttpOverrides) in the same window, ~3/25 recent runs.
• ServerVirtualizationTest.QuickGrid_AnchorMode_End_PrependAtTop_ViewportStaysStable(useItemsProvider: True) (scroll position assertion): a known flaky test, previously quarantined in Quarantine Microsoft.AspNetCore.Components.E2ETest.Tests.VirtualizationTest.QuickGrid_AnchorMode_End_PrependAtTop_ViewportStaysS [Content truncated due to length] #68730.

Neither test depends on content-type mapping or ETag formatting. The static file responses those tests receive keep byte-identical headers, which the StaticFiles tests cover.

🤖 Posted by ASP.NET Core Perf 2 (AI Agent)

[davidfowl]

Pushed empty commit 38d1685 to retrigger CI. Everything except aspnetcore-components-e2e passed. These e2e failures, in the Mono client/WebAssembly and server legs:

• VirtualizationTest.VariableHeight_HtmlTable_DoesNotProduceNestedTrElements
• RedirectionTest.NavigationException_InAsyncContext_DoesNotBecomeUnobservedTaskException
• BlazorWebJsInitializersTest.InitializersRunsModernCallbacksByDefaultWhenPresent (2 cases)

are the same set failing right now on unrelated PRs: #69752 (only touches the RequestDelegateGenerator) fails exactly these four, and #70068 fails the first two. That's 3 of the ~12 recent PRs I sampled. Static file content types are unchanged by this PR. FileExtensionContentTypeProvider returns the same mapping for every extension (covered by the existing provider tests plus the new ones), and .js, .mjs, .wasm and .dll resolve identically.

🤖 Posted by ASP.NET Core Perf 2 (AI Agent)

[PureWeen]

/review

[davidfowl]

On 55c4a17, aspnetcore-ci passed. The only failure is in the aspnetcore-components-e2e Mono server leg (build 1667861), and it isn't caused by this PR.

Failing test: Microsoft.AspNetCore.Components.E2ETest.ServerExecutionTests.ServerRoutingTest.NavigationLock_CanBlockExternalNavigation. Build Analysis gives it a 31.89% failure rate.
BrowserAssertFailedException: Expected: http://127.0.0.1:41537//myexternalpath3, Actual: http://127.0.0.1:41537/subdir/

Why it's unrelated

• The same test fails on many unrelated PRs: 23 of the last 120 PRs I checked, including Avoid per-line MemoryStream allocation in BufferedReadStream.ReadLine #70194, Add review-pull-request reviewer regression Vally suite #70238, Unquarantine TempData cookie persistence test #70261 and [blazor] Extract current dotnet.d.ts from runtime pack for repository build #70344.
• The Mono server leg is also red on main at 1a03092, where a different Blazor test failed.
• This PR only changes StaticFiles content-type lookup and header writes, plus a new microbenchmark project. A Blazor Server NavigationLock external-navigation test doesn't depend on either.

I didn't retrigger CI because main E2E is currently flaky. A maintainer can rerun the failed leg if needed.

🤖 Posted by ASP.NET Core Perf 2 (AI Agent)