AudioOut2 turns controller haptic PCM into motor rumble only when no
controller sound card plays it, which is the case for a DualSense over
Bluetooth. PadInput then dropped that rumble for every DualSense, so
titles that vibrate through haptic audio never rumbled one over
Bluetooth. A DualSense now takes it like any other controller; with a
sound card no haptic rumble is submitted, so nothing plays twice.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Every change of the pad output state sent a rumble, light bar and trigger
effect update to SDL, and the trigger effect was sent again even when it
had not changed. Titles that animate the light bar or drive rumble from
haptic audio change that state every frame, so at 120 Hz about 110
reports a second were queued while SDL's HIDAPI rumble thread sends at
most one per 10 ms. Different report kinds do not merge in its queue, so
the backlog grew: rumble arrived ever later, and disconnecting the
controller blocked the presenter until the queue drained.
Updates now go out at most every 30 ms with the latest state, and the
trigger effect is sent only when it changes or a controller connects.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A per-game relinker-patches.toml (or --patches) replaces verified bytes in
the input executable and bundled modules before conversion. Entries are
checked against the expected bytes and against every in-place writer,
the ELF header range and relocation targets. An <output-stem>.patches.toml
manifest records the applied entries, and --verify-patches checks an input
or a converted output. With no patch file the output is unchanged.
agc_driver_host_import_lifetime takes the settle alias path only on a
device whose graphics family has a single queue, such as AMD's Windows
driver; elsewhere the settles share the graphics family and make no
alias. agc_driver_host_import_lifetime_other_family runs the same tests
with the settles forced onto another queue family, as
agc_import_settle_deferred_other_family does, so a device with spare
graphics queues covers that path too. It is registered on Windows and
skips where the device has no other queue family.
forgetSettle handed back an import's settle batch only while the batch
was still running. A batch the GPU had finished but that had not been
reaped yet still held its aliases, so retireImport released the import
at once, freeing its memory before the alias bound to it was destroyed.
Hand back any busy batch. The import then stays with its batch until
the batch is reaped or an unmap releases it, and both destroy the
aliases first. A finished batch now keeps its imports' Vulkan objects
until it is reaped (#7601).
On Windows a host import under 1 MiB is settled by a batched 4-byte
read. When the settle queue is in another queue family than the
import's, as on AMD's Windows driver with its single graphics queue,
the batch reads through an alias buffer bound to the import's memory.
Reaping the batch destroys those aliases before it drops the imports
retired while the batch held them.
An unmap that needs such an import (releaseSettleRetired) waited for
the batch and then released the import, destroying its buffer and
freeing its memory while the alias was still bound to that memory; the
alias went at the next reap. Destroy the batch's aliases first, as
reapSettleBatches and drainSettles do.
sceSslConnect verifies the server's chain under MBEDTLS_SSL_VERIFY_OPTIONAL and returns the first failure the connection's verify options select, with the SSL_ERROR_* codes KytyPS5 lists. The initial options are all five checks, so a connection whose title never sets them fails closed: with no CA loaded it returns SSL_ERROR_UNKNOWN_CA instead of throwing. The result is final for the connection: a failure sends close_notify, later sceSslConnect calls return the same code, and sceSslRead and sceSslWrite throw. A verify callback runs once after the handshake, outside the session lock, with the connection id, that code, handles to a copy of the peer chain that the context's certificate readers accept, the chain length and the user argument; a negative return fails the connect with SSL_ERROR_APPLICATION_VERIFICATION_CALLBACK_CHECK. The ALPN list set with sceSslSetAlpn is split on commas and offered, sceSslGetAlpnSelected returns the protocol the server chose, and the minimum version floors the handshake at TLS 1.2. A socket that would block returns SSL_ERROR_WANT_POLLIN or WANT_POLLOUT from connect, read and write and the next sceSslConnect resumes the handshake; the BIO maps a reset to MBEDTLS_ERR_NET_CONN_RESET, which returns SSL_ERROR_SOCKET_CLOSED, and tells a blocking socket's expired receive timeout, which returns SSL_ERROR_READ_TIMEOUT, apart from a non-blocking would-block. sceSslClose sends close_notify once and marks the connection closed. The debt entry records the default options and the situation-to-code choices as unverified and drops the title IDs.
Repacked copies and depth stencil copies allocated device-local memory
directly and stopped the title on a refusal; they now take the same
allocation path as the other images. Device buffers choose their host
fallback type the same way, the host placement line carries the cumulative
MiB, and the debt file records that a placed image stays in host memory
for its lifetime.
Textures, storage textures, block, depth and depth plane copies and
attachments now share one allocation path: the pool trims its idle tiers,
releases the texture caches, retries device-local memory and places the
image in host memory only when that retry is refused too. Block and depth
plane copies no longer fall back on their first refusal.
Render, depth and multisampled targets and attachment proxies that the
driver refuses device memory now go through the buffer pool's recovery:
the pool releases its idle allocations and the attachment is retried in
device-local memory. Only a second refusal places it in host memory, and
that placement is logged with the bytes reclaimed.
Depth targets, depth plane copies, multisampled and render targets, block and
depth copies and attachment proxies allocated device-local memory without the
fallback, so a full device-local heap still stopped the title at them.
The budget check in DeviceMemoryType trusts VK_EXT_memory_budget, which lags
behind the driver: with the device-local heap nearly full the budget can still
report room, the allocation fails with VK_ERROR_OUT_OF_DEVICE_MEMORY, and the
title stops. Device buffers (after the pool's reclaim and retry), textures and
storage textures now retry in a host memory type allowed by their requirements.
Evgeniy Kazarin