sceNpCalloutStartOnCtx64 cleared and filled the entry before looking for
it in the queue, so restarting a queued callout cut off every callout
queued after it and changed its deadline while the callout thread could
read it. The entry is now written only after the duplicate check under
the context mutex. sceNpArchInit and sceNpArchTerm now hold a lock, so a
second caller no longer returns before the global heap exists. Both
differences from the firmware are recorded in the debt document.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013XFS3FUfBtqCVfwmBrRQ6q
sceNetCtlGetResult threw for every event after its initialization and
address checks, and Worms Armageddon: Anniversary Edition (PPSA20052)
and Monster Hunter Wilds (PPSA24191) stop on it at boot; MHW's first
call asks about event 1, a disconnection.
It now answers as sceNetCtlGetResultV6 does: for event types 1 to 3,
disconnected to IP obtained (the numbers sceNetCtlCheckCallback
delivers), it writes 0 to the result and returns 0. Other event types
throw, naming the type. The initialization and null checks come first,
as before. The error code a console reports per event is unconfirmed,
which the library's debt page now records.
The guest NetCtl test covers each supported event, three unsupported
ones with the new message and the result left untouched, and a null
result answered before the event is looked at.
kern_fpathconf answers _PC_ASYNC_IO with _POSIX_ASYNCHRONOUS_IO (200112) after the descriptor lookup, and a pipe or socket (no vnode) with PIPE_BUF for _PC_PIPE_BUF and EINVAL for every other name. Drop the unused File.hpp include and the duplicate KernelErrors.hpp include.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
A finished or joined thread returns ESRCH for any priority, and scePthreadSetprio sets errno for a refused priority like pthread_setprio.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
BunnyFl measured these functions against firmware 13.42's libkernel and
libkernel_sys in the PR thread (behaviour only):
- pthread_rwlockattr_settype_np stores only types 1 and 2. Any other
type, 0 included, and a null attribute or object return EINVAL and
store nothing; type 0 was stored before.
- scePthreadMutexattrSetkind returns 0x8001FFFF, sceKernelError(-1), for
a null attribute or object instead of throwing. Kinds 1 to 4 are set as
before; other kinds still throw, since the firmware stores them
unchecked and a mutex's behaviour with them is not measured.
- scePthreadMutexattrGetkind takes the attribute itself and returns
sceKernelError of its kind, or 0x8001FFFF for a null one.
CONVENTIONS.md still listed stable resident runs of at least 256 KiB as
the pages the write watch can leave read-only between collections on
Windows, and pointed at the debt/libc.md entry this PR removes. Resident
runs are no longer write-protected. What stays read-only between
collections on Windows are the pages of shared views (shared memory,
shared file mappings, and direct memory mapped at more than one guest
address), which Collect arms; the Windows direct-memory backing entry in
debt/libc.md describes the direct-memory case.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The UTF-16 and UTF-32 string forms and the UTF-8 to Shift-JIS and EUC-JP
calls now share the measured string loop and UTF-8 decoder, and the
Japanese tests expect JIS X 0212 to be neither decoded nor encoded.
sceCesMbcsUcsContextInit rejects a null profile.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
orezraey