[harisibrar1]

What

pselect (merged in #9273) never returned EINTR. With the canonical pattern of blocking a signal, raising it and waiting with a mask that unblocks it:

sigprocmask(SIG_BLOCK, {SIGUSR1}); raise(SIGUSR1);
pselect(0, NULL, NULL, NULL, &{0, 300 ms}, &empty);

FreeBSD 11.4 runs the handler and returns -1/EINTR at once. Here the handler ran inside the mask swap (_sigprocmask delivers a pending signal that SIG_SETMASK unblocks), and select then waited the full 300 ms and returned 0. With a NULL timeout it never returned. This was raised in the review of #9273.

In FreeBSD's kern_pselect the mask is installed and kern_select sleeps with PCATCH. Before it sleeps, kern_select checks nd, the descriptors and the timeout, then scans once: a ready descriptor wins, a zero timeout returns 0, and only then does a pending signal end the sleep with EINTR (sys/kern/sys_generic.c, release/11.4.0). On EINTR the descriptor sets are not copied out.

The change:

• Signals.cpp: main's ChangeBlockedMask (from fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499) gains an optional bool* caught, and DeliverUnblockedPending returns whether a guest handler ran. SetSignalMaskCatching_nid_no_patch uses it with SIG_SETMASK, releasing holds like the sigprocmask call it replaces. _sigprocmask itself behaves as before.
• SetSignalMaskCatching_nid_no_patch is declared once, in prx/libkernel/include/Api.hpp, which both Signals.cpp and Pselect.cpp include.
• Pselect.cpp: when the swap ran a handler, pselect scans the descriptors once with a zero timeout instead of waiting. A ready descriptor, a zero timeout or an invalid argument (nfds, timeout) keep select's answer. Otherwise the call returns -1/EINTR and leaves the descriptor sets as they were.
• The pselect debt entry is rewritten. The old text described a narrow race; the remaining gap is that a signal raised during the wait (another thread's pthread_kill, or raise from a handler) still doesn't interrupt it, because select has no EINTR path. It also notes that the mask is the process-wide one.

Tested

Windows 11, WinLibs GCC 15.2.0 r7 (UCRT), Release. Full build of 7784604 (current main 47f9cee merged in): 0 failed steps.

• guest_kernel_sys_pselect passes 3 out of 3 on 7784604. The new cases cover: the pending-signal idiom returns EINTR in under 150 ms with the handler run once, the descriptor set unchanged and the mask restored; a NULL timeout returns EINTR instead of blocking; a ready descriptor still wins (returns 1); a zero timeout still returns 0; an invalid timeout still returns EINVAL; and a mask that keeps the signal blocked still waits the full timeout, with the handler running only when the signal is later unblocked.
• Negative control: with main's Pselect.cpp and Signals.cpp, the test fails at the first EINTR check (line 129).
• The related signal, timer and select tests pass (11 of 11, 3 runs), including guest_kernel_sys, guest_signals, guest_sigtimedwait, guest_getitimer, guest_itimer and guest_itimer_pending_default.
• tools/check_conventions.py --base origin/main passes.
• Not tested: Linux and macOS.

Overlap

• fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 (@EagleBearerYS) landed first as agreed; this now builds on its ChangeBlockedMask instead of keeping a second helper.
• docs(libkernel): record select limits as measured on a PS5 #10310 edits docs/dev/debt/libkernel.md near the select entries; a rebase there is a text merge only.

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: fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 (proposed order: fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 first, then this PR rebased onto its ChangeBlockedMask); docs(libkernel): record select limits as measured on a PS5 #10310 in either order, as a text merge in docs/dev/debt/libkernel.md (asked both authors to confirm below)
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes

🤖 Generated with Claude Code

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
PR 11273	core/libs/prx/libkernel/System/src/Signals.cpp L543, L578

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 3986	core/libs/prx/libkernel/System/src/Signals.cpp
PR 4630	docs/dev/debt/libkernel.md
PR 5021	docs/dev/debt/libkernel.md
PR 6149	docs/dev/debt/libkernel.md
PR 6740	docs/dev/debt/libkernel.md
PR 6830	docs/dev/debt/libkernel.md
PR 6936	docs/dev/debt/libkernel.md
PR 7088	docs/dev/debt/libkernel.md
PR 7105	docs/dev/debt/libkernel.md
PR 7121	docs/dev/debt/libkernel.md
PR 7149	docs/dev/debt/libkernel.md
PR 7192	docs/dev/debt/libkernel.md
PR 7201	core/libs/prx/libkernel/System/src/Signals.cpp
PR 7447	docs/dev/debt/libkernel.md
PR 7509	docs/dev/debt/libkernel.md
PR 7804	docs/dev/debt/libkernel.md
PR 8010	core/libs/prx/libkernel/include/Api.hpp docs/dev/debt/libkernel.md
PR 8605	docs/dev/debt/libkernel.md
PR 9156	docs/dev/debt/libkernel.md
PR 9199	docs/dev/debt/libkernel.md
PR 9242	docs/dev/debt/libkernel.md
PR 9284	docs/dev/debt/libkernel.md
PR 9293	docs/dev/debt/libkernel.md
PR 9297	docs/dev/debt/libkernel.md
PR 9302	docs/dev/debt/libkernel.md
PR 9389	core/libs/prx/libkernel/System/src/Signals.cpp docs/dev/debt/libkernel.md
PR 9509	docs/dev/debt/libkernel.md
PR 9557	docs/dev/debt/libkernel.md
PR 9566	docs/dev/debt/libkernel.md
PR 9575	docs/dev/debt/libkernel.md
PR 9669	docs/dev/debt/libkernel.md
PR 9693	core/libs/prx/libkernel/include/Api.hpp docs/dev/debt/libkernel.md
PR 10061	docs/dev/debt/libkernel.md
PR 10319	docs/dev/debt/libkernel.md
PR 10442	docs/dev/debt/libkernel.md
PR 10621	docs/dev/debt/libkernel.md
PR 10666	docs/dev/debt/libkernel.md
PR 10898	docs/dev/debt/libkernel.md
PR 11002	core/libs/prx/libkernel/System/src/Signals.cpp docs/dev/debt/libkernel.md
PR 11047	docs/dev/debt/libkernel.md

[MarcZX10]

The pselect change is narrowly scoped and the negative control plus ready-descriptor, timeout, mask-restore, and signal-blocking cases are useful. The author reports the related Windows tests passing on an earlier base with an identical patch-id; current hosted PR build jobs are skipped, and I did not rerun the test suite locally.

Please resolve two merge-order declarations before approval. The body explicitly says #9499 remains open, changes the same Signals.cpp helper, and that #9499 should land first; however the checklist says Depends on: none. The body also says open #10310 changes the pselect debt text and needs a text merge, but no order is declared for that overlap. Please coordinate with those authors and list the agreed order(s) in the checklist, following CONTRIBUTING. The current head is mergeable with main, so I found no separate rebase blocker or code defect in this pass.

[harisibrar1]

@MarcZX10 Thanks. I've updated the checklist with the order. @EagleBearerYS, @orezraey, could you confirm it works for you?

• fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 first, then this PR. Both move _sigprocmask's body into a helper in Signals.cpp. Once fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 lands, I'll rebase this onto its ChangeBlockedMask and add the "a guest handler ran" result there, instead of keeping a second helper. If you'd rather this one went first, I'm fine with that too; then fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 would only need to keep the caught out-parameter when it rebases.
• docs(libkernel): record select limits as measured on a PS5 #10310 in either order. It only touches the select text in docs/dev/debt/libkernel.md next to the pselect entry this PR rewrites, so whichever lands second keeps both sentences.

Reviewed with AI assistance (Claude Code).

[EagleBearerYS]

@harisibrar1 that works for us: #9499 first, then #10501 on its ChangeBlockedMask. Thanks for taking the rebase.

[Dom3005]

Linux run (Ubuntu 26.04 container, GCC 15, glibc 2.43, CMake + Ninja, Release, -DBUILD_TESTING=ON, ctest --repeat until-fail:3)

• c7cd45c0d merged onto main 39a36bbec: guest_kernel_sys_pselect, guest_kernel_sys, guest_signals, guest_sigtimedwait, guest_getitimer, guest_itimer and guest_itimer_pending_default pass, 7 of 7, 3 runs each.
• Control, main 39a36bbec with only this PR's GuestKernelSysPselect.cpp: guest_kernel_sys_pselect fails with check failed at line 129, the same first EINTR check as on Windows.
• main d53b52be7 on its own: the same seven tests pass.

This covers the Linux part of "Not tested: Linux and macOS".

[MarcZX10]

Thanks, the #9499 author has now confirmed the proposed order, and the Linux follow-up supplies the missing platform run. The remaining coordination item is #10310: the checklist says its debt-text overlap can land in either order, but I do not see confirmation from its author yet. Please record that agreement before merge; the prior review request remains open until then.

[orezraey]

@harisibrar1 Either order works for #10310. It only rewrites the select entry in docs/dev/debt/libkernel.md, so whichever of the two lands second will keep both the select sentence and this PR's pselect entry; if #10310 is second I will refresh it after this one merges.

[MarcZX10]

Reviewed current head c7cd45c. The previously missing merge-order agreements are now explicit: #9499 confirmed it lands first, and #10310 confirmed either order for the adjacent debt text; the checklist records both. The implementation and regression coverage are unchanged from my prior review. The author reports the Windows focused signal suite and the Linux follow-up reports seven related tests passing three consecutive runs, including a negative control where main fails the new EINTR case. I did not reproduce those runs locally. Hosted structural checks pass; product build jobs are skipped by eligibility. No remaining blocker found.

[harisibrar1]

Merged main e7e3e9c into the branch (head b8b4214) to resolve the conflict from the libkernel header refactor (#10706):

• Pselect.cpp now includes prx/libkernel/include/Api.hpp and prx/libc/include/Api.hpp, like main's version. It keeps only a local declaration of the internal SetSignalMaskCatching_nid_no_patch, typed with GuestSignalSet to match its definition in Signals.cpp.
• GuestKernelSysPselect.cpp drops its local extern "C" block in favour of Api.hpp, and uses main's GuestHandler.

The behaviour is unchanged. Rerun on Windows 11, WinLibs GCC 15.2.0 r7, full Release build: no failed steps. guest_kernel_sys_pselect, the signal and timer tests (guest_signals, guest_sigtimedwait, guest_getitimer, guest_itimer, guest_itimer_pending_default) and the select tests pass, 11/11 in 3 runs. With main's Pselect.cpp and Signals.cpp, guest_kernel_sys_pselect still fails. check_conventions.py --base origin/main passes. #9499 is still open, so the agreed order stands.

Reviewed with AI assistance (Claude Code); the checks above were run locally on my machine.

[oneandonlydean]

@harisibrar1 Thanks for this. One change before it can go in: the new APS5_VABI SetSignalMaskCatching_nid_no_patch bridge is only declared locally in Pselect.cpp. Please move that declaration into the shared libkernel header that the umbrella reaches, and include it from both translation units, so the definition and its caller can't drift apart. That's the repo's one-declaration-per-export rule. Please keep the agreed order with #9499 (it lands first), and keep the pending-signal, descriptor-priority and mask-restoration tests.

[harisibrar1]

@oneandonlydean done in 7784604b0b, which also merges current main (47f9cee):

• SetSignalMaskCatching_nid_no_patch is now declared once, in prx/libkernel/include/Api.hpp next to _sigprocmask. The libkernel.hpp umbrella includes that header, and so do both Signals.cpp and Pselect.cpp. The local declaration in Pselect.cpp is gone.
• Main had replaced the mask helper with ChangeBlockedMask (handler holds). Instead of keeping a second copy, the bridge now calls main's helper: it gained an optional bool* caught, and DeliverUnblockedPending returns whether a handler ran. The bridge releases holds like the sigprocmask(SIG_SETMASK) call it replaces.
• The pending-signal, descriptor-priority and mask-restoration tests are unchanged, and fix(libkernel): block a signal while its handler runs and honour SA_RESETHAND #9499 stays first.

Windows 11, GCC 15.2.0: full build clean, and the 11 signal/pselect/select tests pass in 3 runs. With main's Pselect.cpp and Signals.cpp, guest_kernel_sys_pselect fails (2 runs). check_conventions passes. Could you take another look?

Reviewed with AI assistance (Claude Code); the checks above were run locally on my machine.

[harisibrar1]

Overlap note: #11273 (@Gruzzberry, EFAULT for unusable signal pointers) also adds a trailing optional parameter to ChangeBlockedMask (GuestSignalSet* guestPrevious), so the two conflict on its signature line and on the DeliverUnblockedPending line. Whichever lands second keeps both parameters; if #11273 lands first I'll merge main here and do that. Details on #11273.

Reviewed with AI assistance (Claude Code); the checks above were run locally on my machine.