Hollow Knight: Silksong (PPSA12544) on Linux

[awake-devel]

Bring-up notes for Hollow Knight: Silksong (PPSA12544) on Linux, kept separate from the Windows thread (#726). Everything below can be reproduced from a ready-made branch.

This work was done and written up by Claude Code (AI), and posted by me after review.

Status (2026-10-08)

• ✅ Loads, reaches the title screen, save select, the screen-size calibration screen, and New Game plays the intro cinematic.
• ❌ At the end of the intro cinematic, the game ends with Terminating due to uncaught exception of type Il2CppExceptionWrapper. This is being investigated; I will post updates here.
• ⚠️ 8–23 fps, and some flickering with short "rollbacks" during the cinematic. The likely suspect is the storage-image write-back workaround (guest memory changed while GPU results were pending; keeping the CPU's blocks, agc-driver: storage image write-back drops GPU results for CPU-stamped 64 KiB edge blocks (boxes in Silksong's menu text) #865), but that is not confirmed.

Test machine: Arch Linux, i7-8750H, GTX 1050 Ti Max-Q 4 GB, NVIDIA 580.178.04, Wayland (Hyprland), GCC.

Ready-made branch

awake-devel/AnyPS5@test/silksong-linux (fbb383e). It is main 296d68a plus the open PRs below and one local workaround. It is a test stack, not a PR.

git clone --recursive -b test/silksong-linux https://github.com/awake-devel/AnyPS5
cd AnyPS5
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF
cmake --build build && cmake --build build --target libs

What is in it, in the order each one became necessary

Stop on Linux	What fixes it
Crash with no output while modules load: the bundled sce_module/libc.prx's malloc returns NULL inside Il2cppUserAssemblies.prx	Regression from #1156, reported in #1500. The branch carries a local workaround (a6b6439, undoes the Linux half of #1156). This is the only commit that is not an upstream PR.
_umtx_op missing	#1482
Two PSN imports missing	#1436
Runtime-loaded modules (Media/Plugins)	#846
sceKernelLoadStartModule with start arguments not implemented	#1102, merged with #846 using #ifdef _WIN32 in sceKernelLoadStartModule: Windows keeps #846's call, Linux uses #1102's block
Unity GC stops threads through signals	#1431
Il2CppUserAssemblies.prx requested, but the file is Il2cppUserAssemblies.prx (ENOENT, then a jump to address 0)	#1079 (case-insensitive paths)
CppWebApi::Common::initialize() failed (0x80553504): the bundled libc refuses a heap above the PS5 address window	#861
First flip: display buffer ... reads mixed DCC keys ... not implemented	#788
Calibration screen: skipped draws when depth-surface memory is reused as a resized colour texture	#1495, then #454 (which brings #425 and #374). The conflict in DepthSurfaceTexture is resolved by keeping #454's function and applying #1495's rule only on its single-slice sampled path.

Running it

1. You need your own dump of the game: the executable and modules (eboot.bin, sce_module/, Media/Modules, Media/Plugins) as relinker input, and the full game folder as app0/.
2. Relink, making sure app0/ is the folder the relinker writes the *.guest.prx files into:

   build/core/relinker/relinker --module-dir Media/Plugins <dump>/eboot.bin <runtime>/app.elf
   mkdir -p <runtime>/libs && cp build/core/libs/libs/*.prx <runtime>/libs/
   cd <runtime> && ./app.elf

3. No environment variables are needed on a current main. fix(agc): keep dma-buf import off the NVIDIA proprietary driver #1329 fixed the NVIDIA device creation (AGC: Vulkan device creation fails on the NVIDIA proprietary driver since #943 (dma-buf import probe) #1289).

Notes

• The fix(relinker): suppress duplicate system library lifecycle #1156 workaround brings back Linux: when a title bundles its own libc.prx, both libc instances run the application-heap handshake, so the title initialises its main pool twice #693 (both libc copies initialise the heap). Nothing suspicious so far, up to the cinematic.
• On Ctrl+C, the game's handler calls sceKernelGetModuleInfoFromAddr with an uninitialised size, which throws (ModuleInfo.cpp). This only happens at shutdown.
• Only Linux was tested. The #ifdef _WIN32 side of the feat(relinker): add --module-dir for modules loaded at run time #846/fix(libkernel): pass module args and argp to runtime-loaded modules on Linux #1102 merge was not built on Windows.

If you try it on another distribution or GPU, please post how far it gets and the last lines of the log.

[awake-devel]

Update: gameplay reached on Linux. The game now gets past the intro cinematic into gameplay. It runs slowly (7–19 fps on a GTX 1050 Ti) and the cinematic still flickers; both are being looked at.

The branch test/silksong-linux now has two more local commits (b7acea0, b3953f0). Together they fix the crash at the end of the cinematic (Terminating due to uncaught exception of type Il2CppExceptionWrapper).

• What crashed: the managed exception was System.Reflection.AmbiguousMatchException ("Ambiguous match found."). Unity normally catches it itself. Here it could not be caught, because on Linux no C++ exception thrown in guest code can be caught at all:
  • 1. The relinker dropped the unwind header. It removed PT_GNU_EH_FRAME from the relinked executable, because SegmentFilter.cpp drops the whole OS-specific range 0x61000000–0x6fffffff, and that header falls inside it. Fixed in b7acea0.
  • 2. The unwinder was told there are no tables. The bundled libc's unwinder finds tables through sceKernelGetModuleInfoForUnwind, which on Linux always answered "no tables" (fix(libkernel): report guest unwind tables in sceKernelGetModuleInfoForUnwind #904 covers Windows only). Fixed in b3953f0: it now reports the unwind tables of the executable and of the *.guest.prx modules, using the header from b7acea0.
• Status: neither has an open PR yet. Both are generic and I plan to propose them once they have tests.

Also: the white box sometimes drawn over the "O" of OPTIONS on the title screen comes and goes between runs, and is older than these changes.

This update was investigated and written by Claude Code (AI), and posted by me after review.

[awake-devel]

Investigated and written by Claude Code (AI), and posted by me after review.

Update: the Linux fixes are now PRs

The two changes that got Silksong past the end of the intro cinematic, plus the flicker fix, are now separate PRs on current main (cba8d86). Each one passed the full test suite and the validation-layer run, with no regressions, and has a regression test:

PR	What it fixes
#1586	The relinker dropped PT_GNU_EH_FRAME from the relinked executable, so nothing could find its unwind tables. It is now kept, and the relinker falls back to the old behaviour with a warning when the program-header table is full.
#1600	On Linux, sceKernelGetModuleInfoForUnwind always reported "no tables" (#904 fixed only Windows). The title's own libc uses it to unwind, so no C++ exception could be caught, which caused the Il2CppExceptionWrapper terminate after the cinematic.
#1601 (draft, work in progress)	Cinematic flicker: with 2 output buffers, AvPlayer reused a frame's buffer while the GPU could still be reading it. It now keeps the 5 newest handed-out frames. The count comes from experiment, so it needs review. In testing, the jumps went from 63 to 3, and the title-screen version text and the cinematic's skip prompt reappeared.

With #1586 and #1600 (on the test stack), the cinematic plays through and the game reaches gameplay: the tutorial area loads and the game saves.

Still open:

• Slow: about 7–19 fps. A benchmark baseline is being set up to measure the open performance PRs.
• Sound well ahead of the picture during the cinematic: possibly up to a minute by the end. The picture doesn't look slowed down, so it may be an offset. It was already there before fix(avplayer): keep handed-out video frames until later frames are handed out #1601, and is being investigated.
• Shutdown: closing the game throws sceKernelGetModuleInfoFromAddr: unsupported st_size (the title passes an uninitialised size). It only happens at shutdown.
• Linux: since #1156, a bundled sce_module/libc.prx is never initialised, but titles still call its malloc, which returns NULL #1500: the bundled libc regression from fix(relinker): suppress duplicate system library lifecycle #1156. The test branch still carries the local workaround.

[fisherjoey]

Linux NVIDIA here too (RTX 5070 Ti, 595.91.07, Nobara 44), running Demon's Souls (PPSA01341). Before you set the baseline, these are the Linux costs that skewed our numbers. All are quick to check.

Window placement. When our window sat off-screen (on an output that was switched off), KWin on NVIDIA throttled swapchain acquire in FIFO, and the title's WAIT_FLIP_DONE waited about 129 ms per flip. The movie went from 5.5 to 19-27 fps when the window moved to a visible output. I haven't tried Hyprland. A window on a hidden workspace may be throttled the same way, so keep it visible during baseline runs.

PROCMAP_QUERY per guest-range check. Since #157, each page-access check is one PROCMAP_QUERY ioctl, and nothing caches the answer. In our tutorial level that was 2.1M ioctls per 20 s, nearly all on the queue-0 worker. A per-page cache brought it down to 2.7k. #1330 targets the same cost. To see whether Silksong pays it:

sudo bpftrace -e 'tracepoint:syscalls:sys_enter_ioctl /pid == $1/ { @[args->cmd] = count(); }
  interval:s:10 { print(@); clear(@); }' $(pgrep -f app.elf)

Title threads starving the queue worker. DS's 12 job workers spin on all 12 hardware threads. Over 25 s of movie, the queue-0 worker spent 7.6 s preempted and 3.8 s waiting in the run queue. Raising only the driver's queue threads (nice -10 through rtkit) took the movie from 25 to 33-35 fps. top -H during a slow stretch shows whether Unity's workers do the same. I have that as an opt-in patch on a branch if you want to try it.

GPU or CPU. With APS5_PROFILE_DRAW=1, the [present] line gives GPU busy and idle time per present. If busy is close to the frame time, the 1050 Ti is the limit and the CPU-side PRs won't show.

Method. Single runs misled us more than once. After a recompiler change, the first run builds the shader cache (we saw 0.04-0.2 fps in that run), so we throw it away, interleave A and B (ABAB, 2-4 runs each) and compare medians over the steady windows. I have a small script that digests an APS5_PROFILE_DRAW log per phase and diffs two runs. It isn't published, but I can put it in a gist. #1286 proposes documenting those lines.

AI-assisted: yes (Claude Code; the runs are on my machine).

[awake-devel]

I'll work on this later, I'm first making sure my fixes work and can be merged with the main branch of the project with no conflict

[awake-devel]

Investigated and written by Claude Code (AI), and posted by me after review.

Update: sound sync, the load crash, and first performance numbers

Thanks to @hikazey for testing #1586 on Windows: the PE output is byte-identical with and without the PR, and the full suite passes there.

Sound ahead of the picture (issue #1635)

• Cause: Silksong turns off AvPlayer's A/V sync (sceAvPlayerSetAvSyncMode, mode none) and plays the soundtrack through its own audio. It only takes video from AvPlayer. In that mode AvPlayer hands out the next frame on every call and never skips one. Whenever the game asks for frames slower than the film's frame rate, the picture falls further behind the sound. Here it drifted 32–52 s behind by the end of the cinematic.
• Local fix: at normal speed, the video also follows AvPlayer's own real-time clock and skips the frames that are already late, even when sync is off. With it, 79.5 s of film played in 80.0 s, and the picture stayed about 0.5 s behind from start to end instead of drifting. A PR referencing AvPlayer: in no-sync mode the video plays at the game's frame rate, so cinematics fall far behind their audio #1635 will follow once the full checks pass.
• What is left: the picture is now choppy, because in that run the game only asked for a frame every 167 ms (see the performance part below).

Crash while loading the first area: unsupported st_size
This is not only a shutdown problem. The bundled libc's __cxa_finalize calls sceKernelGetModuleInfoFromAddr with an uninitialised st_size. It does so when the runtime-loaded CommonDialog.prx is unloaded while the first area loads, about 5 minutes in. The console's own libc never sets the field, so the fix ignores the input st_size. The fix has passed the full suite and its mutation check, and its PR will follow after an in-game check.

White boxes over menu letters
These are #865: a font atlas that isn't 64 KiB aligned gets its first block overwritten by stale CPU memory. #1599 is being tested on Linux with this title, over several runs, since the bug comes and goes.

Performance, first measurements (Hyprland, GTX 1050 Ti)

• Window visibility does the opposite of what was reported for KWin. On a hidden workspace, the title screen runs at its 60 fps cap and the cinematic at about 21 fps. With the window visible and focused, the game asked for a frame every 167 ms (about 6 fps, exactly 10 refreshes at 60 Hz). This is the biggest slowdown seen so far and has not been explained yet.
• Title-screen benchmark: the GPU is busy about 4.9 ms of each 16.7 ms frame, and the queue-0 worker waits for a CPU less than 1% of the time, so neither is the limit there.
• Open performance PRs: being measured one at a time against a fixed baseline, using CPU cycles per frame because the title screen is capped at 60 fps. No open PR for the kernel-wait or shader-recompiler items showed a gain on this title. perf(memory): restrict registry work to affected ranges #1324 and perf(agc): cache page permissions within registered allocations #1330 are being benchmarked now, with a PROCMAP_QUERY call counter.