[LeandroLP23]

What

Fixes #10833 the way agreed there (option 1): replacement stays atomic, and the tests skip only what an SMB share cannot do.

On Windows WriteFileAtomically replaces a cache entry with a POSIX-semantics rename (SetFileInformationByHandle class 22, FileRenameInfoEx), so readers and mappings of the old entry keep working. Through an SMB share that rename fails with ERROR_INVALID_PARAMETER, and the std::filesystem::rename fallback is refused while the destination is open or mapped. The issue's NAS and \\localhost\C$ (Windows' own SMB server, over NTFS) behave the same; the error codes are in #10833 (comment).

With the first skip in place, the tests reached three more places that need the same operations:

• verifyStore and verifyPersist replace an entry that ShaderDiskCache itself has mapped with MapViewOfFile after loading it. A mapping blocks the plain rename just like an open reader.
• The cleanup removeCacheDirectory deletes with FileDispositionInfoEx and FILE_DISPOSITION_FLAG_POSIX_SEMANTICS, which SMB also rejects with ERROR_INVALID_PARAMETER, and the entries the cache still maps can't be deleted there.
• shader_cache_file_read built the \\?\ alias of a UNC path as \\?\\\server\share\…; the extended form of a UNC path is \\?\UNC\server\share\….

Changes:

• ShaderCacheDirectory.hpp: the POSIX rename moves unchanged into PosixReplaceByHandle, which WriteFileAtomically calls as before, so production behaviour does not change. PosixReplaceUnsupported(directory) probes that rename on two scratch files in a directory and reports only ERROR_INVALID_PARAMETER.
• agc_shader_disk_cache: where the probe reports it, verifyAtomicReplacement skips the retained-reader replacement, verifyStore skips the truncated- and damaged-entry checks, and verifyPersist skips replacing the mapped persisted entry, each printing why. Any other replacement failure still fails. The cleanup falls back to a plain delete only when the POSIX delete returns ERROR_INVALID_PARAMETER, and then leaves the directory behind with a note instead of failing.
• shader_cache_file_read: the UNC alias fix, and the retained-reader replacement is skipped under the same probe. A plain replacement without a reader is still checked there.
• docs/dev/debt/shader.md: a "Shader cache directories on SMB shares" entry records that such a cache directory cannot update an open or mapped entry, including one this process's cache has mapped, and why replacement stays atomic.

Tested

Rebased onto main 9656d366be (head 8259ecb26d); on the RTX 5060 Ti the whole table below and the validation run were repeated on the rebased head and on 9656d366be, with the same results. The first runs, on both machines, were on base 7b415520b6.

Same commands on two machines, TEMP/TMP set to a local NTFS directory or to \\localhost\C$\aps5-smbtmp, each test 3 times, on this head and on its base main 7b415520b6:

• Windows 11 10.0.26300, AMD Ryzen AI 7 350 / Radeon 860M, AMD 24.30.58.08 (32.0.13058.8001), WinLibs GCC 15.2.0 r7, Release.
• Windows 11 10.0.26300, Intel i5-10600K / NVIDIA GeForce RTX 5060 Ti, NVIDIA 621.01, same toolchain, Release.

Test	TEMP	this head (both machines)	main (both machines)
agc_shader_disk_cache	NTFS	3/3	3/3
shader_cache_file_read	NTFS	3/3	3/3
agc_shader_disk_cache	SMB	3/3, prints the skips	0/3, cannot replace cache entry with a retained reader
shader_cache_file_read	SMB	3/3, prints the skips	0/3

• Before the verifyStore/verifyPersist/cleanup changes, a local build that ran every verify* case under try/catch listed the failing SMB cases: only verifyStore (cannot truncate the entry) and verifyPersist (cannot replace the persisted entry), then the cleanup's POSIX delete. The other 31 cases pass on the share unchanged.
• Validation layer: ctest -V -R "shader_disk_cache|shader_cache|pipeline_cache|pipeline_journal" with ANYPS5_REQUIRE_VULKAN=1, ANYPS5_NO_SHADER_CACHE=1 and VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation (Vulkan SDK 1.4.363.0): 27/27 on this head and on main, on both GPUs, with no VUID- lines and no was not found but was requested by env var VK_INSTANCE_LAYERS. With VK_LOADER_DEBUG=layer, four of these create a device and show Insert instance layer "VK_LAYER_KHRONOS_validation": agc_pipeline_cache_header, agc_pipeline_cache, agc_pipeline_cache_file and agc_driver_pipeline_journal; the disk cache tests themselves are CPU-only.
• A real NAS was not available; \\localhost\C$ reproduces the issue's failure and error codes.
• check_conventions.py --base origin/main passes. pr_overlap.py finds no conflicting comparison; PRs that already conflict with main (fix(shader): preserve raw fragment exports for integer attachments #6235, fix(agc): keep texture images inside their extent's mip chain #6280, fix(recompiler): defer the TYPE check of a guarded read-only static V# #6412, feat(recompiler): run waterfall loops over per-lane T#s through image tables #6488, perf(recompiler): keep runtime-specialized shader modules in the shader disk cache #6924, perf(recompiler): leave the descriptor-range comparison out of direct buffer loads when the device zeroes past the range #7484, fix(shader): bound image-table snapshots by table-key ranges #7684, perf(shader): let spanned waves run at any host subgroup size where it can be narrower than 32 #8004, feat(agc): emulate mesh-fallback barycentrics through tessellation #9387, feat(agc): blend into 10_11_11 unorm color targets in the pixel shader #9748, fix(agc): retire a multisampled color target a write-only storage image overwrites #10049, fix(agc): read the compressed FMASK of a resident two-sample color target #10233) share ShaderDiskCache.cpp or the shader debt file and were not compared.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: Fixes Windows SMB: shader disk-cache retained-reader replacement fails; local NTFS passes #10833
• Unblocks: none
• GPUs tested: AMD Radeon 860M (integrated), AMD 24.30.58.08, Windows 11 10.0.26300; NVIDIA GeForce RTX 5060 Ti, NVIDIA 621.01, Windows 11 10.0.26300
• AI-assisted: yes (Claude Code)

[MarcZX10]

The tested base in the description is 7b415520b6, while current main is 035ac96edd (9 commits ahead). Although the intervening commits do not change the shader-cache implementation paths, CONTRIBUTING requires PRs to stay up to date. Please sync to current main and rerun conventions before final review.

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 5021	docs/dev/debt/shader.md
PR 5371	docs/dev/debt/shader.md
PR 5609	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 6235	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 6280	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 6412	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 6488	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 6858	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 6924	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 6959	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 6970	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 7395	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 7410	docs/dev/debt/shader.md
PR 7684	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 7694	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 7922	docs/dev/debt/shader.md
PR 8004	docs/dev/debt/shader.md
PR 8939	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 9020	docs/dev/debt/shader.md
PR 9193	docs/dev/debt/shader.md
PR 9335	docs/dev/debt/shader.md
PR 9387	docs/dev/debt/shader.md
PR 9457	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 9633	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 9748	docs/dev/debt/shader.md
PR 9749	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 9790	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 9852	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 9889	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 9897	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 9969	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 10000	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 10049	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 10233	docs/dev/debt/shader.md
PR 10489	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 10556	docs/dev/debt/shader.md
PR 10803	docs/dev/debt/shader.md
PR 10846	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp docs/dev/debt/shader.md
PR 11483	docs/dev/debt/shader.md
PR 11551	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp
PR 11557	docs/dev/debt/shader.md

[LeandroLP23]

@MarcZX10 rebased onto current main 9656d366be (head 8259ecb26d, same single commit; between the old and new base only StencilReplay.cpp changed under core/shader/recompiler). check_conventions.py --base origin/main passes. I also reran the full check on the rebased head and on 9656d366be on the RTX 5060 Ti: agc_shader_disk_cache and shader_cache_file_read pass 3/3 on NTFS and on \\localhost\C$ with this head, while main fails 0/3 on the share; the 27 shader/pipeline cache tests pass with the validation layer on both, with no VUID- lines. The description is updated.

AI-assisted: yes (Claude Code)

[MarcZX10]

I reviewed the Windows SMB probing/fallback and its test changes on exact head 8259ecb26dac1890717b823dd42f9558eb4fa9a4. I found no implementation defect in the bounded FILE_RENAME_INFO_EX buffer construction or the fallback behavior. The author reports repeated Windows runs on two GPUs and local NTFS plus \\localhost\C$, including the cache test matrix and validation-layer comparison. I did not independently run those Windows tests. Current conventions, git diff --check, and the merge-tree preview pass; hosted Linux/Windows product jobs are path-skipped.

Before merge, please sync this branch to current main (b0f1183ca12b). Its tested base is 9656d366be; the intervening commits do not touch this PR's implementation, tests, or debt entry, and the preview merges cleanly. This is the remaining branch-freshness requirement from CONTRIBUTING.md.

[LeandroLP23]

@MarcZX10 synced: rebased onto current main da89f18ea7 (head b95b22cf16, the same single commit). None of the PR's files (ShaderCacheDirectory.hpp, CacheFileReadTests.cpp, ShaderDiskCache.cpp, docs/dev/debt/shader.md) changed between the tested base 9656d366be and da89f18ea7, and check_conventions.py --base origin/main passes.

AI-assisted: yes (Claude Code)

[MarcZX10]

Re-reviewed exact head b95b22cf1622, now based on live main da89f18ea7c1. The rebase preserves the reviewed one-commit implementation and reported Windows SMB/NTFS regression coverage; no source or test files changed between the tested base and this main. The current-main merge preview, conventions and whitespace checks pass. Linux and Windows relinker plus ASan checks pass; macOS relinker jobs are still running/queued, and runtime product tests are skipped by path policy. I did not independently run the Windows/GPU tests. The earlier branch-freshness request is resolved. Approved.